unsloth/studio/backend/core/inference/llama_cpp.py

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

"""llama-server inference backend for GGUF models.

Manages a llama-server subprocess and proxies chat completions through its
OpenAI-compatible /v1/chat/completions endpoint.
"""

import atexit
import contextlib
import json
import os
import re
import struct
from loggers import get_logger
import shutil
import signal
import socket
import subprocess
import sys
import threading
import time
from pathlib import Path
from typing import Callable, Collection, Generator, Iterable, List, Mapping, Optional, Union

import httpx

from core.inference.llama_server_args import (
    _effective_tensor_parallel,
    _tensor_parallel_matches_loaded,
    extra_args_disable_mmproj,
    parse_cache_override,
    parse_cache_override_per_axis,
    parse_ctx_override,
    parse_split_mode_override,
    resolve_requested_ctx,
    strip_shadowing_flags,
    strip_split_mode_only,
)

# Share strip / signal constants with the multi-format parser so BUFFERING also
# catches Llama-3 / Mistral / Gemma 4 (legacy helper only knew <tool_call> / <function=).
from core.inference.tool_call_parser import (
    _GEMMA_BARE_TC_PREFIX_RE,
    _GEMMA_BARE_TC_RE,
    _TOOL_ALL_PATS as _PARSER_TOOL_ALL_PATS,
    _TOOL_CLOSED_PATS as _PARSER_TOOL_CLOSED_PATS,
    _balanced_brace_end,
    _strip_function_xml_calls,
    _strip_gemma_wrapperless_calls,
    _strip_glm_calls,
    _strip_mistral_closed_calls,
    TOOL_XML_SIGNALS as _SHARED_TOOL_XML_SIGNALS,
    RAG_MAX_SEARCHES_PER_TURN,
    RAG_SEARCH_CAP_NUDGE,
    parse_tool_calls_from_text as _shared_parse_tool_calls_from_text,
    strip_leading_bare_json_call,
    strip_llama3_leading_sentinels,
    strip_tool_markup as _shared_strip_tool_markup,
)

# The healer owns the bracket-tag + rehearsal strip helpers and their name-gated
# pattern lists, so the GGUF streaming strip stays aligned with the parser.
from core.tool_healing import (
    _REHEARSAL_TAIL_STRIP_RE,
    _strip_bracket_tag_calls,
    apply_tool_strip_patterns,
    strip_outside_think,
)
from utils.native_path_leases import child_env_without_native_path_secret
from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
from utils.subprocess_compat import (
    windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
)
from utils.process_lifetime import child_popen_kwargs as _child_popen_kwargs
from core.inference.tool_call_parser import (
    MAX_ACT_REPROMPTS as _MAX_REPROMPTS,
    REPROMPT_MAX_CHARS as _REPROMPT_MAX_CHARS,
    is_short_intent_without_action as _is_short_intent_without_action,
    reprompt_to_act_message as _reprompt_to_act_message,
)
from core.inference.tool_loop_controller import (
    ToolLoopController,
    tool_event_provenance,
)
from state.tool_approvals import (
    TOOL_REJECTED_MESSAGE,
    abort_tool_decision,
    begin_tool_decision,
    new_approval_id,
    wait_tool_decision,
)

logger = get_logger(__name__)


class LlamaServerNotFoundError(RuntimeError):
    """GGUF model needs the llama.cpp runtime but no llama-server is installed.
    Subclasses RuntimeError so existing handlers still catch it."""


# Shared so the from_identifier preflight and the load-time raise stay in sync.
LLAMA_SERVER_NOT_FOUND_DETAIL = (
    "This is a GGUF model, but the llama.cpp runtime (llama-server) is not "
    "installed. Run `unsloth studio setup` to download the prebuilt runtime, "
    "then try again. (Advanced: set LLAMA_SERVER_PATH to an existing binary.)"
)


# llama-server can serve HTTP 200 while running a model entirely on CPU when a
# GPU backend fails to init (#5807 / #5106 / #5830). Classify the startup log so
# Studio can warn. Priority: explicit "offloaded N/M layers to GPU" counts
# (authoritative), then GPU "model buffer size" lines (host-pinned _Host
# excluded), then the "device_info:" device table (disconfirm only).
_GPU_OFFLOAD_MARKERS = (
    "CUDA",
    "ROCm",
    "ROCM",
    "HIP",
    "Metal",
    "Vulkan",
    "OpenCL",
    "SYCL",
    "MUSA",
    "CANN",
)
_OFFLOADED_LAYERS_RE = re.compile(
    r"offloaded\s+(\d+)\s*/\s*(\d+)\s+layers?\s+to\s+gpu", re.IGNORECASE
)
_DEVICE_ROW_RE = re.compile(
    r"-\s*(CUDA|ROCm|ROCM|HIP|Metal|Vulkan|SYCL|OpenCL|MUSA|CANN|CPU)\w*\s*:",
    re.IGNORECASE,
)
_GPU_DEVICE_PREFIXES = (
    "cuda",
    "rocm",
    "hip",
    "metal",
    "vulkan",
    "sycl",
    "opencl",
    "musa",
    "cann",
)


def classify_gpu_offload_lines(lines: "list[str]") -> Optional[bool]:
    """True if the model landed on a GPU, False if it stayed on CPU despite GPU
    intent, None when the log has no usable signal."""
    # Counted offload is authoritative, keyed on the model with the most layers.
    # A separate MTP/draft model logs its own (much smaller) "offloaded N/M"
    # line, so decide on the largest-M line: a drafter that fits on GPU must not
    # mask a main model running on CPU. N>0 on that model is True, 0 is False.
    max_total = -1
    offloaded_at_max = 0
    for line in lines:
        match = _OFFLOADED_LAYERS_RE.search(line)
        if not match:
            continue
        offloaded, total = int(match.group(1)), int(match.group(2))
        if total > max_total or (total == max_total and offloaded > offloaded_at_max):
            max_total, offloaded_at_max = total, offloaded
    if max_total >= 0:
        return offloaded_at_max > 0

    # GPU marker on a *model* buffer; _Host buffers are CPU-pinned, not offload.
    # Buffer lines are authoritative: present but none on a GPU means CPU-only,
    # so do not let the device table below override that.
    saw_model_buffer = False
    for line in lines:
        if "model buffer size" not in line:
            continue
        saw_model_buffer = True
        if "_Host" not in line and any(m in line for m in _GPU_OFFLOAD_MARKERS):
            return True
    if saw_model_buffer:
        return False

    # device_info: lists *available* devices (printed whenever a GPU backend is
    # visible), not where the model loaded, so it can only disconfirm: an
    # all-CPU table means no usable GPU. A visible GPU device is not proof the
    # model used it, so it does not return True. Rows after the header only.
    after_header = False
    saw_device_row = False
    saw_gpu_device = False
    for line in lines:
        if "device_info:" in line:
            after_header = True
            continue
        if not after_header:
            continue
        match = _DEVICE_ROW_RE.search(line)
        if not match:
            continue
        saw_device_row = True
        if match.group(1).lower().startswith(_GPU_DEVICE_PREFIXES):
            saw_gpu_device = True
    if saw_device_row and not saw_gpu_device:
        return False
    return None


def _wsl_system_rocm_lib_dirs() -> "list[str]":
    """System ROCm lib dir(s) to load before a prebuilt's bundled HIP, on WSL.

    The bundled bare-metal HIP can't drive WSL's /dev/dxg and segfaults on the
    first GPU call; the system ROCm libs (libamdhip64 + librocdxg) can, while
    the bundle still supplies libggml-hip / librocblas (gfx1151 kernels).
    Mirrors install_llama_prebuilt._wsl_system_rocm_lib_dirs so a prebuilt that
    passed install validation runs the same at serve time. No-op off a ROCDXG
    WSL host (needs /dev/dxg, "microsoft" /proc/version, librocdxg in /opt/rocm).
    """
    try:
        if not os.path.exists("/dev/dxg"):
            return []
        with open("/proc/version", encoding = "utf-8", errors = "replace") as fh:
            if "microsoft" not in fh.read().lower():
                return []
    except OSError:
        return []
    out: "list[str]" = []
    for d in ("/opt/rocm/lib", "/opt/rocm/lib64"):
        if os.path.exists(os.path.join(d, "librocdxg.so")) or os.path.exists(
            os.path.join(d, "librocdxg.so.1")
        ):
            out.append(d)
    return out


# Plan-without-action re-prompt state (intent signal, caps, message) now lives
# in tool_call_parser, imported above under its old aliases.

# Default max_tokens to the effective context when known. The floor is high
# enough for reasoning-heavy GGUFs and max_tokens-omitting API clients.
_DEFAULT_MAX_TOKENS_FLOOR = 32768
_DEFAULT_FIRST_TOKEN_TIMEOUT_S = 1200.0  # 20 min

# Only large streamed tool payloads get an early provisional card; render_html
# is exempt because it needs immediate artifact feedback.
_PROVISIONAL_ARGS_MIN_CHARS = 256
_DEFAULT_STREAM_STALL_TIMEOUT_S = 120.0  # 2 min
# Cap tool calls from a single TEXTUAL-fallback turn (mirrors the safetensors
# loop). Structured delta.tool_calls are grammar-bounded by llama-server; text
# parsed from content is not, so one runaway turn could fan out unbounded.
_MAX_TOOL_CALLS_PER_TURN = 8
_FORCED_REPEAT_PLAN_SIGNAL = re.compile(
    r"\b(?:i\s+will|i'll|let\s+me|going\s+to|need\s+to|call|use|run|search|fetch|render)\b",
    re.I,
)
_FINAL_ANSWER_SIGNAL = re.compile(
    r"\b(?:final\s+answer|answer\s*:|here\s+is|here's|in\s+summary|result\s*:)\b",
    re.I,
)


def _gguf_active_tool_names(active_tools: list[dict]) -> list[str]:
    names = [
        (tool.get("function") or {}).get("name")
        for tool in (active_tools or [])
        if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
    ]
    return [name for name in names if name]


# Rehearsal NAME chars (word + hyphen, matching the parser); the lookbehind excludes the
# Mistral [CALL_ID]...[ARGS] shape.
_GGUF_REHEARSAL_ARGS_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]")


def _gguf_rehearsal_signal_pos(text: str, active_tools: list[dict]) -> int:
    """Index of the first ``NAME[ARGS]`` whose NAME is an active tool, else -1. A
    bare/inactive-name ``foo[ARGS]`` in prose is not a call; mirrors the safetensors
    ``_earliest_tool_signal`` name-gating (no unrestricted GGUF mode)."""
    active = set(_gguf_active_tool_names(active_tools))
    if not active:
        return -1
    for m in _GGUF_REHEARSAL_ARGS_RE.finditer(text):
        if m.group(1) in active:
            return m.start()
    return -1


def _gguf_has_genuine_tool_signal(text: str, signals, active_tools: list[dict]) -> bool:
    """True when ``text`` holds a genuine tool-call boundary for one of ``signals``.

    Unambiguous markers (``<tool_call>``, ``[TOOL_CALLS]``, ``<function=``) count on a
    plain substring hit; an ``[ARGS]`` hit is genuine only when an active tool name
    precedes it, so inactive-name prose is neither drained nor parsed."""
    for sig in signals:
        if sig == "[ARGS]":
            if _gguf_rehearsal_signal_pos(text, active_tools) >= 0:
                return True
            continue
        if sig in text:
            return True
    return False


def _is_rehearsal_prefix(stripped: str, active_tools: list[dict]) -> bool:
    """True if ``stripped`` is a (possibly partial) prefix of ``NAME[ARGS]`` for an
    active tool -- the bare tool name arriving in its own chunk before ``[ARGS]{...}``.
    Mirrors the safetensors loop so the split rehearsal call is not streamed."""
    if not stripped or any(ch.isspace() for ch in stripped):
        return False
    for name in _gguf_active_tool_names(active_tools):
        if stripped == name or f"{name}[ARGS]".startswith(stripped):
            return True
    return False


def _held_rehearsal_tail_len(text: str, active_tools: list[dict]) -> int:
    """Length of a trailing bare tool-name token that may be a split rehearsal call
    (``...web_search`` with ``[ARGS]{...}`` still to arrive), so STREAMING can hold it
    instead of leaking the name. Returns 0 for ordinary prose. Mirrors safetensors."""
    i = len(text)
    while i > 0 and not text[i - 1].isspace():
        i -= 1
    tail = text[i:]
    return len(tail) if tail and _is_rehearsal_prefix(tail, active_tools) else 0


def _should_suppress_forced_no_tool_output(text: str) -> bool:
    """Suppress only repeated forced-turn planning text, not final answers."""
    stripped = text.strip()
    if not stripped or len(stripped) >= _REPROMPT_MAX_CHARS:
        return False
    if _FINAL_ANSWER_SIGNAL.search(stripped):
        return False
    return _FORCED_REPEAT_PLAN_SIGNAL.search(stripped) is not None


# ── Pre-compiled patterns for GGUF shard detection ───────────
_SHARD_FULL_RE = re.compile(r"^(.*)-(\d{5})-of-(\d{5})\.gguf$", re.IGNORECASE)
_SHARD_RE = re.compile(r"^(.*)-\d{5}-of-\d{5}\.gguf$", re.IGNORECASE)


# ── Sliding-window-pattern resolver ───────────────────────────
# Resolves the per-layer SWA mask when a GGUF reports a sliding window but
# no `sliding_window_pattern` field. Tier order in `_resolve_swa_pattern`:
# GGUF metadata, on-disk cache, bootstrap dict below, transformers
# introspection, HF Hub config.json, legacy 1/4 fallback. Period N means
# layer i is SWA iff `(i + 1) % N != 0`, matching transformers. Skipped on
# purpose: phi3 (no key/val length in GGUF, window >= ctx anyway), qwen2
# family (converter strips sliding_window when use_sliding_window=False),
# mistral v0.1/v0.2 (all-SWA can't be a period).
_BOOTSTRAP_SWA_DEFAULTS: dict[str, int] = {
    "gemma2": 2,  # Gemma2Config.sliding_window_pattern
    "gemma3": 6,  # Gemma3TextConfig.sliding_window_pattern
    "gemma3n": 5,  # text_config.layer_types: SWA*4 + FULL
    "gpt_oss": 2,  # text_config.layer_types: alternating
    "cohere2": 4,  # Cohere2Config.sliding_window_pattern
}

# Process-wide cache backed by JSON on disk. Values are int period or
# list[bool] mask. Lazy-loaded.
_SWA_CACHE: Optional[dict] = None
_SWA_CACHE_LOCK = threading.Lock()


def _probe_dns_dead(host: str = "huggingface.co", timeout: float = 2.0) -> bool:
    """Quick DNS check on a daemon thread, so concurrent sockets aren't
    affected by socket.setdefaulttimeout."""
    result: list[Optional[bool]] = [None]

    def _probe() -> None:
        try:
            socket.gethostbyname(host)
            result[0] = False
        except Exception:
            result[0] = True

    t = threading.Thread(target = _probe, daemon = True)
    t.start()
    t.join(timeout)
    # Thread still running -> resolver wedged -> dead.
    return True if result[0] is None else result[0]


def _hf_env_offline() -> bool:
    """True when an HF offline env var is set to any truthy value (1/true/yes/on).

    Mirrors utils.models.model_config._env_offline so a user-set HF_HUB_OFFLINE=true
    (not just "1") still routes through the local-cache reuse path below.
    """
    try:
        from utils.models.model_config import _env_offline
        return _env_offline()
    except Exception:
        return os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in {"1", "true", "yes", "on"}


@contextlib.contextmanager
def _hf_offline_if_dns_dead():
    """Set HF_HUB_OFFLINE for this block only when DNS to huggingface.co fails;
    restores env on exit so a transient hiccup can't quarantine the process.
    No-op if the user already set it."""
    if "HF_HUB_OFFLINE" in os.environ:
        yield False
        return
    if not _probe_dns_dead():
        yield False
        return

    transformers_was_set = "TRANSFORMERS_OFFLINE" in os.environ
    os.environ["HF_HUB_OFFLINE"] = "1"
    if not transformers_was_set:
        os.environ["TRANSFORMERS_OFFLINE"] = "1"
    logger.warning("huggingface.co unreachable; using local HF cache for this load.")
    try:
        yield True
    finally:
        os.environ.pop("HF_HUB_OFFLINE", None)
        if not transformers_was_set:
            os.environ.pop("TRANSFORMERS_OFFLINE", None)


def _swa_cache_path() -> Path:
    home = os.environ.get("UNSLOTH_STUDIO_HOME") or os.environ.get("STUDIO_HOME")
    base = Path(home) if home else Path.home() / ".unsloth" / "studio"
    return base / "swa_cache.json"


def _load_swa_cache() -> dict:
    global _SWA_CACHE
    with _SWA_CACHE_LOCK:
        if _SWA_CACHE is not None:
            return _SWA_CACHE
        try:
            with open(_swa_cache_path()) as f:
                _SWA_CACHE = json.load(f)
                if not isinstance(_SWA_CACHE, dict):
                    _SWA_CACHE = {}
        except (FileNotFoundError, json.JSONDecodeError, OSError):
            _SWA_CACHE = {}
        return _SWA_CACHE


def _save_swa_cache(cache: dict) -> None:
    try:
        path = _swa_cache_path()
        path.parent.mkdir(parents = True, exist_ok = True)
        tmp = path.with_suffix(".json.tmp")
        with open(tmp, "w") as f:
            json.dump(cache, f, indent = 2, sort_keys = True)
        tmp.replace(path)
    except OSError:
        pass


def _period_from_layer_types(layer_types: list) -> Optional[int]:
    """Smallest period N where `(i+1) % N != 0` matches the SWA mask, else None."""
    if not layer_types:
        return None
    is_swa = ["full" not in str(t).lower() for t in layer_types]
    n = len(is_swa)
    for N in range(1, n + 1):
        if all(((i + 1) % N != 0) == is_swa[i] for i in range(n)):
            return N
    return None


def _swa_entry_from_layer_types(lt) -> Optional[object]:
    """Period int, or per-layer bool mask, from a transformers ``layer_types`` list."""
    if isinstance(lt, list) and lt:
        return _period_from_layer_types(lt) or ["full" not in str(t).lower() for t in lt]
    return None


def _fetch_swa_entry_from_hf(repo_id: str) -> Optional[object]:
    try:
        from huggingface_hub import hf_hub_download
        cfg_path = hf_hub_download(repo_id, "config.json", repo_type = "model")
        with open(cfg_path) as f:
            cfg = json.load(f)
    except Exception:
        return None

    src = cfg.get("text_config") if isinstance(cfg.get("text_config"), dict) else cfg
    period = src.get("sliding_window_pattern")
    if isinstance(period, int) and period > 0:
        return period
    return _swa_entry_from_layer_types(src.get("layer_types"))


def _arch_aliases(arch: str) -> tuple:
    # GGUF emits `falcon-h1`; HF model_type is `falcon_h1`. Normalise both ways.
    seen = []
    for a in (arch, arch.replace("-", "_"), arch.replace("_", "-")):
        if a and a not in seen:
            seen.append(a)
    return tuple(seen)


def _swa_entry_from_config_obj(cfg) -> Optional[object]:
    src = getattr(cfg, "text_config", None) or cfg
    period = getattr(src, "sliding_window_pattern", None)
    if isinstance(period, int) and period > 0:
        return period
    return _swa_entry_from_layer_types(getattr(src, "layer_types", None))


_SWA_PATTERN_SOURCE_RE = re.compile(r"sliding_window_pattern\s*(?::\s*[\w\[\], ]*)?\s*=\s*(\d+)")


def _resolve_swa_entry_from_transformers(arch: str) -> Optional[object]:
    """Default-instantiate the matching Config; on failure, regex-parse its
    source for `sliding_window_pattern = N`."""
    try:
        from transformers.models.auto.configuration_auto import (
            CONFIG_MAPPING,
            CONFIG_MAPPING_NAMES,
        )
    except Exception:
        return None

    cfg_class = None
    for alias in _arch_aliases(arch):
        if alias in CONFIG_MAPPING_NAMES:
            try:
                cfg_class = CONFIG_MAPPING[alias]
                break
            except Exception:
                cfg_class = None
    if cfg_class is None:
        return None

    try:
        if (entry := _swa_entry_from_config_obj(cfg_class())) is not None:
            return entry
    except Exception:
        pass

    import inspect

    candidates = [cfg_class]
    text_cfg_class = getattr(cfg_class, "sub_configs", {}).get("text_config")
    if text_cfg_class is not None:
        candidates.append(text_cfg_class)
    for cls in candidates:
        try:
            src = inspect.getsource(cls)
        except (OSError, TypeError):
            continue
        if m := _SWA_PATTERN_SOURCE_RE.search(src):
            period = int(m.group(1))
            if period > 0:
                return period
    return None


def _resolve_swa_pattern(
    arch: Optional[str],
    n_layers: Optional[int],
    source_repo_candidates: tuple = (),
    *,
    allow_network: Optional[bool] = None,
) -> Optional[list]:
    if not arch or not n_layers:
        return None
    if allow_network is None:
        allow_network = os.environ.get("UNSLOTH_STUDIO_OFFLINE", "0") not in (
            "1",
            "true",
            "True",
            "yes",
        )

    cache = _load_swa_cache()

    def _entry_to_mask(entry):
        if isinstance(entry, int) and entry > 0:
            return [(i + 1) % entry != 0 for i in range(n_layers)]
        if isinstance(entry, list) and entry:
            return [bool(entry[i % len(entry)]) for i in range(n_layers)]
        return None

    def _persist(entry):
        with _SWA_CACHE_LOCK:
            cache[arch] = entry
        _save_swa_cache(cache)

    if (entry := cache.get(arch)) is not None:
        if (mask := _entry_to_mask(entry)) is not None:
            return mask

    if (entry := _BOOTSTRAP_SWA_DEFAULTS.get(arch)) is not None:
        return _entry_to_mask(entry)

    entry = _resolve_swa_entry_from_transformers(arch)
    if entry is not None:
        _persist(entry)
        return _entry_to_mask(entry)

    # Tier 3: live HF fetch (result persistently cached)
    if allow_network:
        for repo_id in source_repo_candidates:
            if not repo_id:
                continue
            entry = _fetch_swa_entry_from_hf(repo_id)
            if entry is not None:
                _persist(entry)
                return _entry_to_mask(entry)

    return None


def _hf_repo_from_url(url: Optional[str]) -> Optional[str]:
    """Strip `https://huggingface.co/owner/name(/...)` -> `owner/name`."""
    if not url or "huggingface.co/" not in url:
        return None
    tail = url.split("huggingface.co/", 1)[1].rstrip("/")
    parts = tail.split("/")
    if len(parts) < 2:
        return None
    return f"{parts[0]}/{parts[1]}"


# Lazy import to avoid pulling transformers in at module level.
def _extract_model_size_b(model_id: str):
    from utils.models import extract_model_size_b
    return extract_model_size_b(model_id)


_TOOL_TEMPLATE_MARKERS = (
    "{%- if tools %}",
    "{%- if tools -%}",
    "{% if tools %}",
    "{% if tools -%}",
    '"role" == "tool"',
    "'role' == 'tool'",
    'message.role == "tool"',
    "message.role == 'tool'",
    # DeepSeek: no top-level ``{% if tools %}`` block; it gates emission on
    # ``message['role'] == 'tool'`` plus ``message['tool_calls'] is defined``.
    "message['role'] == 'tool'",
    'message["role"] == "tool"',
    "message['tool_calls']",
    'message["tool_calls"]',
    "tool_calls is defined",
)


# Canonical reasoning_effort levels, weakest -> strongest. Used to read the
# discrete set a template branches on (e.g. GLM-5.2 uses 'high' | 'max') so we
# only ever offer levels the template actually understands.
_REASONING_EFFORT_SCALE = ("minimal", "low", "medium", "high", "max")


def _extract_reasoning_effort_levels(chat_template: str) -> list:
    """Return the reasoning_effort levels a template references, in canonical
    (weakest -> strongest) order.

    Looks for the quoted literals (e.g. ``'high'`` / ``"max"``) the template
    compares ``reasoning_effort`` against, so we surface exactly the levels it
    branches on and nothing else.
    """
    return [
        level
        for level in _REASONING_EFFORT_SCALE
        if f"'{level}'" in chat_template or f'"{level}"' in chat_template
    ]


def detect_reasoning_flags(
    chat_template: Optional[str],
    model_identifier: Optional[str] = None,
    *,
    log_source: Optional[str] = None,
) -> dict:
    """Classify a chat template's reasoning and tool-calling capabilities.

    Returns the same six keys as the GGUF sniffer: ``supports_reasoning``,
    ``reasoning_style`` (``"enable_thinking"`` | ``"reasoning_effort"`` |
    ``"enable_thinking_effort"``), ``reasoning_always_on``,
    ``reasoning_effort_levels``, ``supports_preserve_thinking``,
    ``supports_tools``. A falsy ``chat_template`` yields the all-default dict.
    Used by both the llama-server backend at load time and the
    safetensors/transformers paths in ``routes/inference`` so they agree on
    what the frontend sees.
    """
    flags = {
        "supports_reasoning": False,
        "reasoning_style": "enable_thinking",
        "reasoning_always_on": False,
        "reasoning_effort_levels": [],
        "supports_preserve_thinking": False,
        "supports_tools": False,
    }
    if not chat_template:
        return flags
    tpl = chat_template
    prefix = f"{log_source}: " if log_source else ""

    effort_levels = (
        _extract_reasoning_effort_levels(tpl)
        if ("reasoning_effort" in tpl and "enable_thinking" in tpl)
        else []
    )
    if effort_levels:
        # DeepSeek-V4's encoder accepts reasoning_effort {'high', 'max'} but its
        # template only branches on 'max', so the literal scan misses 'high'. Add it
        # (matched on whole repo-name segments, so 'deepseek-v40' won't false-match)
        # to expose the full none/high/max ladder instead of none/max.
        segments = re.split(r"[-_.]", (model_identifier or "").lower().split("/")[-1])
        is_dsv4 = "deepseek4" in segments or any(
            a == "deepseek" and b == "v4" for a, b in zip(segments, segments[1:])
        )
        if is_dsv4 and "high" not in effort_levels:
            effort_levels = sorted(set(effort_levels) | {"high"}, key = _REASONING_EFFORT_SCALE.index)
        # GLM-5.2-style: an enable_thinking on/off gate PLUS a reasoning_effort
        # level among a discrete set (e.g. 'high' | 'max'). Distinct from
        # gpt-oss (reasoning_effort only, no on/off gate) and Qwen
        # (enable_thinking only). Disabling is enable_thinking=false; the levels
        # are the quoted effort literals the template actually branches on.
        flags["supports_reasoning"] = True
        flags["reasoning_style"] = "enable_thinking_effort"
        flags["reasoning_effort_levels"] = effort_levels
        logger.info(
            f"{prefix}model supports reasoning "
            f"(enable_thinking + reasoning_effort: {effort_levels})"
        )
    elif "enable_thinking" in tpl:
        flags["supports_reasoning"] = True
        flags["reasoning_style"] = "enable_thinking"
        logger.info(f"{prefix}model supports reasoning (enable_thinking)")
    elif "reasoning_effort" in tpl:
        # gpt-oss / Harmony use reasoning_effort
        # ("low" | "medium" | "high"), not a boolean.
        flags["supports_reasoning"] = True
        flags["reasoning_style"] = "reasoning_effort"
        logger.info(f"{prefix}model supports reasoning (reasoning_effort)")
    elif "thinking" in tpl:
        # DeepSeek uses 'thinking', not 'enable_thinking'
        normalized_id = (model_identifier or "").lower()
        if "deepseek" in normalized_id:
            flags["supports_reasoning"] = True
            logger.info(f"{prefix}model supports reasoning (DeepSeek thinking)")

    # Hardcoded <think> tags or reasoning_content in the template mean
    # thinking is always on (no toggle).
    if not flags["supports_reasoning"]:
        if ("<think>" in tpl and "</think>" in tpl) or "reasoning_content" in tpl:
            flags["supports_reasoning"] = True
            flags["reasoning_always_on"] = True
            logger.info(f"{prefix}model always reasons (<think> tags in template)")

    # preserve_thinking: independent kwarg on some Qwen templates that
    # keeps historical <think> blocks in prior assistant turns.
    if "preserve_thinking" in tpl:
        flags["supports_preserve_thinking"] = True
        logger.info(f"{prefix}model supports preserve_thinking")

    if any(marker in tpl for marker in _TOOL_TEMPLATE_MARKERS):
        flags["supports_tools"] = True
        logger.info(f"{prefix}model supports tool calling")

    return flags


# Gemma 4 ships MTP as a separate drafter (no "-mtp" in the name). Gemma 3n
# ships no drafter, so it is excluded -- it takes the normal non-MTP path.
_GEMMA_MTP_FAMILY_RE = re.compile(r"gemma[-_]?4[-_]", re.IGNORECASE)


def _is_gemma_mtp_family(name: Optional[str]) -> bool:
    """Match Gemma 4 by name."""
    return bool(name) and bool(_GEMMA_MTP_FAMILY_RE.search(name))


def _is_gemma_mtp_name(model_identifier: Optional[str], gguf_path: Optional[str] = None) -> bool:
    """Match Gemma 4 by id or GGUF filename."""
    return _is_gemma_mtp_family(model_identifier) or _is_gemma_mtp_family(
        Path(gguf_path).name if gguf_path else None
    )


def _is_mtp_model_name(model_identifier: Optional[str], gguf_path: Optional[str] = None) -> bool:
    """Name-based MTP detector. Fallback for the metadata signal."""
    for cand in (model_identifier, Path(gguf_path).name if gguf_path else None):
        if cand and "-mtp" in cand.lower():
            return True
        # Recognise Gemma 4 too, so a failed drafter download surfaces a
        # fallback reason instead of silently defaulting.
        if cand and _is_gemma_mtp_family(cand):
            return True
    return False


def _is_companion_gguf_path(path: str) -> bool:
    """True for a non-main GGUF: vision mmproj or a separate MTP drafter
    (repo-root ``mtp-*.gguf`` or the ``MTP/`` subdir copies, Gemma 4).

    Mirrors hub.utils.gguf so variant resolution never picks a companion as
    the main model -- e.g. a Gemma ``Q8_0`` request must not resolve to the
    ``MTP/...-Q8_0-MTP.gguf`` drafter, which sorts ahead of the real weight.
    """
    p = path.lower()
    if not p.endswith(".gguf"):
        return False
    if "mmproj" in p:
        return True
    name = p.rsplit("/", 1)[-1]
    return name.startswith("mtp-") or "/mtp/" in f"/{p}"


_BIG_ENDIAN_GGUF_FILENAME_RE = re.compile(r"(^|[-_])be(?:[._-]|$)", re.IGNORECASE)
_GGUF_KNOWN_QUANT_RE = re.compile(
    r"(UD-)?"
    r"(MXFP[0-9]+(?:_[A-Z0-9]+)*"
    r"|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?"
    r"|TQ[0-9]+_[0-9]+"
    r"|Q[0-9]+_K_[A-Z]+"
    r"|Q[0-9]+_[0-9]+"
    r"|Q[0-9]+_K"
    r"|BF16|F16|F32)",
    re.IGNORECASE,
)


def _is_big_endian_gguf_path(path: str, variant_key: str = "") -> bool:
    normalized = path.replace("\\", "/")
    name = normalized.rsplit("/", 1)[-1]
    stem = name.rsplit(".", 1)[0].lower()
    variant_key = variant_key.strip().lower()
    variant_index = stem.find(variant_key) if variant_key else -1
    parent = normalized.rsplit("/", 1)[0].lower() if "/" in normalized else ""
    variant_in_parent_only = (
        bool(parent)
        and variant_index < 0
        and (
            (variant_key and variant_key in parent)
            or (not variant_key and _GGUF_KNOWN_QUANT_RE.search(parent) is not None)
        )
    )
    for match in _BIG_ENDIAN_GGUF_FILENAME_RE.finditer(stem):
        if variant_index >= 0 and variant_index < match.start():
            return True
        tail = stem[match.end() :].lstrip("._-")
        if not tail or _GGUF_KNOWN_QUANT_RE.search(tail) is None:
            return not variant_in_parent_only
    return False


def _gguf_snapshot_files(snapshot: Path) -> list[str]:
    return [
        p.relative_to(snapshot).as_posix()
        for p in snapshot.rglob("*")
        if p.is_file() and p.name.lower().endswith(".gguf")
    ]


def _cached_hf_snapshot_file(
    repo_id: str,
    filename: str,
    *,
    expected_size: Optional[int] = None,
) -> Optional[str]:
    """Return a cached snapshot file even when HF's current-ref probe misses it."""
    if not filename:
        return None
    parts = [part for part in filename.replace("\\", "/").split("/") if part]
    if not parts or any(part in (".", "..") for part in parts):
        return None
    try:
        from utils.models.model_config import _iter_hf_cache_snapshots
        for snap in _iter_hf_cache_snapshots(repo_id):
            candidate = snap.joinpath(*parts)
            if not candidate.is_file():
                continue
            if expected_size:
                try:
                    if candidate.stat().st_size < expected_size:
                        continue
                except OSError:
                    continue
            return str(candidate)
    except Exception as e:
        logger.debug("Snapshot cache lookup failed for %s/%s: %s", repo_id, filename, e)
    return None


def _snapshot_has_all_shards(
    main_path: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
) -> bool:
    """True when every shard sits beside ``main_path`` in the same cache snapshot.

    llama.cpp loads a split GGUF by resolving its siblings from the main shard's
    directory, so a cached main shard is only safe to reuse when the rest of the
    set is co-located; otherwise the caller must fetch the whole set together.
    """
    root = Path(main_path)
    for _ in [part for part in main_filename.replace("\\", "/").split("/") if part]:
        root = root.parent
    for shard in shards:
        parts = [part for part in shard.replace("\\", "/").split("/") if part]
        if not parts or any(part in (".", "..") for part in parts):
            return False
        sibling = root.joinpath(*parts)
        try:
            if not sibling.is_file():
                return False
            expected = expected_sizes.get(shard)
            if expected and sibling.stat().st_size < expected:
                return False
        except OSError:
            return False
    return True


def _resolve_repo_id_casing(hf_repo: str) -> str:
    """Map a requested repo id to its cached canonical casing, or return it unchanged.

    A case-variant request (for example a lowercased id) resolves to the
    canonical-cased cache directory so the main GGUF and its companions
    (mmproj / MTP drafter) all read the same cache entry. Returns ``hf_repo``
    unchanged when resolution is unavailable or errors.
    """
    try:
        from utils.paths import resolve_cached_repo_id_case
        return resolve_cached_repo_id_case(hf_repo)
    except Exception:
        return hf_repo


def _cached_colocated_split_main(
    repo_id: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
) -> Optional[str]:
    """Main-shard path from a cache snapshot that also holds every sibling shard.

    A newer snapshot may hold only the first shard while an older snapshot has the
    complete split set. ``_cached_hf_snapshot_file`` would return that newer partial
    main and the co-location check would then force a refetch, so scan snapshots for
    one where the whole set is present and return that main path instead. None when
    no snapshot holds the full set.
    """
    main_parts = [part for part in main_filename.replace("\\", "/").split("/") if part]
    if not main_parts or any(part in (".", "..") for part in main_parts):
        return None
    try:
        from utils.models.model_config import _iter_hf_cache_snapshots
        for snap in _iter_hf_cache_snapshots(repo_id):
            main_path = snap.joinpath(*main_parts)
            if not main_path.is_file():
                continue
            expected_main = expected_sizes.get(main_filename)
            try:
                if expected_main and main_path.stat().st_size < expected_main:
                    continue
            except OSError:
                continue
            if _snapshot_has_all_shards(str(main_path), main_filename, shards, expected_sizes):
                return str(main_path)
    except Exception as e:
        logger.debug("Co-located split snapshot lookup failed for %s: %s", repo_id, e)
    return None


def _gguf_extra_shards(files: Iterable[str], first_shard: str) -> list[str]:
    m = _SHARD_FULL_RE.match(first_shard)
    if not m:
        return []
    prefix = m.group(1)
    total = m.group(3)
    sibling_pat = re.compile(
        r"^" + re.escape(prefix) + r"-\d{5}-of-" + re.escape(total) + r"\.gguf$",
        re.IGNORECASE,
    )
    return sorted(f for f in files if f != first_shard and sibling_pat.match(f))


def _gguf_files_for_variant(files: Iterable[str], variant: str) -> list[str]:
    """Return main GGUF files matching a requested variant.

    Prefer exact quant-label matches over loose substring matches so a request
    for ``stories260K`` does not resolve to ``stories260K-be.gguf``.
    """
    variant_key = variant.strip().lower()
    main_files = [
        f
        for f in files
        if f.lower().endswith(".gguf")
        and not _is_companion_gguf_path(f)
        and not _is_big_endian_gguf_path(f, variant_key)
    ]
    if not variant_key:
        return sorted(main_files)

    try:
        from utils.models.model_config import _extract_quant_label
    except Exception:
        _extract_quant_label = None

    if _extract_quant_label is not None:
        try:
            exact = sorted(f for f in main_files if _extract_quant_label(f).lower() == variant_key)
            if exact:
                return exact
        except Exception as e:
            logger.warning("Failed to extract GGUF quant labels: %s", e)

    boundary = re.compile(r"(?<![a-zA-Z0-9])" + re.escape(variant_key) + r"(?![a-zA-Z0-9])")
    return sorted(f for f in main_files if boundary.search(f.lower()))


# Below this many B params, draft-mtp regresses vs spec-off (bench in
# _build_speculative_flags); auto mode drops MTP under it.
_MTP_MIN_SIZE_B = 3.0

# Cap total GPU occupancy at this fraction of the card. The fit reserves an
# absolute (1 - frac) * total per GPU when total VRAM is known, else a fraction
# of free (see _fit_context_to_vram), plus a byte-accurate MTP draft reserve.
# 3%: the context-linear compute buffer is now modelled (_compute_buffer_ctx_bytes),
# so this cushion no longer covers it - only fragmentation, the per-device CUDA
# context on a multi-GPU split, and MoE routing, which measure ~2-3% (Qwen3.5-397B on
# 3 GPUs under-predicts by 2.7%). Below 3% one fragmentation spike overflows to CPU.
_CTX_FIT_VRAM_FRACTION = 0.97

# Apple unified memory is shared with the OS, so tighter than VRAM. Matches the
# 0.85 MLX uses in mlx_inference.py (_configure_memory_limits); not kept in sync.
_APPLE_UNIFIED_MEMORY_FRACTION = 0.85

# Flat MTP reserve, used only when GGUF dims are too sparse for the byte-accurate
# reserve (_estimate_mtp_overhead_bytes). Applied to both the fit budget and pin.
_MTP_VRAM_RESERVE_FRAC = 0.05


def _kv_bytes_per_elem(cache_type: Optional[str]) -> float:
    """Bytes per KV-cache element for a llama.cpp cache type (f16 default)."""
    return {
        "f32": 4.0,
        "f16": 2.0,
        "bf16": 2.0,
        "q8_0": 34 / 32,
        "q5_1": 0.75,
        "q5_0": 0.6875,
        "q4_1": 0.625,
        "q4_0": 0.5625,
        "iq4_nl": 0.5625,
    }.get((cache_type or "f16").strip().lower(), 2.0)


def _env_main_cache_type_for_budget(env: Optional[Mapping[str, str]] = None) -> Optional[str]:
    """Heavier of the inherited LLAMA_ARG_CACHE_TYPE_K/_V env types when it
    exceeds the f16 default, else None. Studio emits --cache-type only for the
    param/extras path, so a heavier env (f32) would otherwise reach the child
    unbudgeted; quantized env types stay over-reserved by f16 (-> None)."""
    e = os.environ if env is None else env
    f16_bpe = _kv_bytes_per_elem("f16")
    heaviest: Optional[str] = None
    heaviest_bpe = f16_bpe
    for var in ("LLAMA_ARG_CACHE_TYPE_K", "LLAMA_ARG_CACHE_TYPE_V"):
        raw = (e.get(var) or "").strip().lower()
        if not raw:
            continue
        bpe = _kv_bytes_per_elem(raw)
        if bpe > heaviest_bpe:
            heaviest, heaviest_bpe = raw, bpe
    return heaviest


def _extra_args_main_cache_type_for_budget(extra_args: Optional[Iterable[str]]) -> Optional[str]:
    """Heavier (max bytes/elem) of the explicit --cache-type-k/-v extras, or None.

    Extras are appended last and win per axis, so an asymmetric K=f32,V=f16 must be
    budgeted by its heavier axis. resolve_cache_type_kv returns only the last-wins
    single type, which under-reserves the heavier axis when the lighter one is last."""
    k, v = parse_cache_override_per_axis(extra_args)
    candidates = [c for c in (k, v) if c]
    if not candidates:
        return None
    return max(candidates, key = _kv_bytes_per_elem)


def _auto_mode_drops_mtp(
    req_mode: Optional[str],
    size_b: Optional[float],
    *,
    has_separate_drafter: bool = False,
) -> bool:
    """Auto mode drops MTP below _MTP_MIN_SIZE_B for an embedded draft head
    (its per-token cost regresses there); a separate drafter (Gemma) is a tiny
    standalone model that still speeds up below 3B, so it never drops. Forced
    mtp / mtp+ngram engage regardless of size."""
    if has_separate_drafter:
        return False
    return req_mode == "auto" and size_b is not None and size_b < _MTP_MIN_SIZE_B


def _mla_mtp_auto_enabled() -> bool:
    """Whether Auto may pick embedded MTP for an MLA model (GLM-5.2/DeepSeek/Kimi).

    Off by default: llama.cpp's MLA/DSA MTP path keeps a duplicated full target-KV
    context and recomputes the sparse-attention indexer every draft step, so it runs
    ~2x slower than no speculation (GLM-5.2 bench: 27 vs 45 tok/s, flat across draft
    depth and 96-100% acceptance) -- the opposite of the vLLM/SGLang speedup on the
    same model. Set UNSLOTH_MLA_MTP_ENABLED=1 to let Auto promote MLA MTP again once
    that path is optimized upstream. Forced mtp / mtp+ngram ignore this gate."""
    return os.environ.get("UNSLOTH_MLA_MTP_ENABLED", "0").strip().lower() in (
        "1",
        "true",
        "yes",
        "on",
    )


def _extra_args_set_spec_type(extra_args: Optional[Iterable[str]]) -> bool:
    """User passed --spec-type / --spec-default? llama-server takes one
    --spec-type (comma-separated to chain), so suppress auto-emit."""
    return _extra_args_set_any_flag(extra_args, {"--spec-type", "--spec-default"})


_GPU_OFFLOAD_OVERRIDE_FLAGS = frozenset({"-ngl", "--gpu-layers", "--n-gpu-layers", "-fit", "--fit"})
_THREAD_OVERRIDE_FLAGS = frozenset({"-t", "--threads"})


def _extra_arg_flag_name(token: str) -> Optional[str]:
    if not token.startswith("-") or token in {"-", "--"}:
        return None
    if len(token) >= 2 and (token[1].isdigit() or token[1] == "."):
        return None
    return token.split("=", 1)[0]


def _extra_args_set_any_flag(extra_args: Optional[Iterable[str]], flags: Collection[str]) -> bool:
    if not extra_args:
        return False
    for raw in extra_args:
        flag = _extra_arg_flag_name(str(raw))
        if flag in flags:
            return True
    return False


def _effective_spec_type(
    extra_args: Optional[Iterable[str]], env: Optional[Mapping[str, str]] = None
) -> Optional[str]:
    """The --spec-type llama-server will use: the last CLI --spec-type (or
    --spec-default, which resolves non-MTP), else LLAMA_ARG_SPEC_TYPE. A CLI flag
    overrides the env (matching llama.cpp), so a stale MTP env can't make the
    budget reserve a drafter the launch won't load. None if neither sets it."""
    args = [str(a) for a in extra_args] if extra_args else []
    cli_present = False
    cli_value: Optional[str] = None
    for i, raw in enumerate(args):
        flag, eq, inline = raw.partition("=")
        if flag == "--spec-default":
            cli_present = True
            cli_value = "default"
            continue
        if flag != "--spec-type":
            continue
        cli_present = True
        cli_value = inline if eq else (args[i + 1] if i + 1 < len(args) else "")
    if cli_present:
        return cli_value
    return (os.environ if env is None else env).get("LLAMA_ARG_SPEC_TYPE")


def _extra_args_requests_mtp(
    extra_args: Optional[Iterable[str]], env: Optional[Mapping[str, str]] = None
) -> bool:
    """True if the effective --spec-type selects MTP (mtp/draft-mtp), so the
    budget must reserve for it."""
    value = _effective_spec_type(extra_args, env)
    if not value:
        return False
    return any(p.strip().lower() in ("mtp", "draft-mtp") for p in value.split(","))


def _extra_args_requests_separate_draft(
    extra_args: Optional[Iterable[str]], env: Optional[Mapping[str, str]] = None
) -> bool:
    """True if the effective --spec-type selects a non-MTP model draft mode
    (draft-simple/draft-eagle3), which loads a separate draft model the budget
    must reserve (draft-mtp -> _extra_args_requests_mtp; ngram-* load no model)."""
    value = _effective_spec_type(extra_args, env)
    if not value:
        return False
    return any(p.strip().lower() in ("draft-simple", "draft-eagle3") for p in value.split(","))


def _extra_args_spec_draft_n_max(extra_args: Optional[Iterable[str]]) -> Optional[int]:
    """Draft depth from extras (``--spec-draft-n-max`` or legacy ``--draft-max``), else None."""
    if not extra_args:
        return None
    args = [str(a) for a in extra_args]
    found: Optional[int] = None
    for i, raw in enumerate(args):
        flag, eq, inline = raw.partition("=")
        if flag not in ("--spec-draft-n-max", "--draft-max"):
            continue
        value = inline if eq else (args[i + 1] if i + 1 < len(args) else "")
        try:
            found = int(value)
        except (TypeError, ValueError):
            continue
    return found


def _extra_args_mtp_draft_path(
    extra_args: Optional[Iterable[str]], env: Optional[Mapping[str, str]] = None
) -> Optional[str]:
    """Separate drafter path from extras (local --model-draft/-md or HF
    --spec-draft-hf/-hfd/...), else the LLAMA_ARG_SPEC_DRAFT_MODEL/_HF_REPO env,
    else None. An HF repo isn't a local file, so the budget can't size it (falls
    back to the flat reserve), but recognizing it avoids sizing the wrong one."""
    flags = {
        "--model-draft",
        "--spec-draft-model",
        "-md",
        "--spec-draft-hf",
        "-hfd",
        "-hfrd",
        "--hf-repo-draft",
    }
    args = [str(a) for a in extra_args] if extra_args else []
    found: Optional[str] = None
    for i, raw in enumerate(args):
        flag, eq, inline = raw.partition("=")
        if flag not in flags:
            continue
        value = inline if eq else (args[i + 1] if i + 1 < len(args) else "")
        if value and not value.startswith("-"):
            found = value
    if found is not None:
        return found
    e = os.environ if env is None else env
    return e.get("LLAMA_ARG_SPEC_DRAFT_MODEL") or e.get("LLAMA_ARG_SPEC_DRAFT_HF_REPO") or None


def _extra_args_draft_cache_types(
    extra_args: Optional[Iterable[str]], env: Optional[Mapping[str, str]] = None
) -> tuple[Optional[str], Optional[str]]:
    """Draft KV cache types (k_type, v_type), each from extras else the
    LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_K/_V env, else None (f16). K and V are
    independent: a one-sided override must not apply to both."""
    args = [str(a) for a in extra_args] if extra_args else []
    k_flags = {"--cache-type-k-draft", "--spec-draft-type-k", "-ctkd"}
    v_flags = {"--cache-type-v-draft", "--spec-draft-type-v", "-ctvd"}
    k_type: Optional[str] = None
    v_type: Optional[str] = None
    for i, raw in enumerate(args):
        flag, eq, inline = raw.partition("=")
        if flag not in k_flags and flag not in v_flags:
            continue
        value = inline if eq else (args[i + 1] if i + 1 < len(args) else "")
        if not value or value.startswith("-"):
            continue
        if flag in k_flags:
            k_type = value
        else:
            v_type = value
    e = os.environ if env is None else env
    if k_type is None:
        k_type = e.get("LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_K") or None
    if v_type is None:
        v_type = e.get("LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_V") or None
    return k_type, v_type


def _extra_args_draft_offloaded_to_cpu(
    extra_args: Optional[Iterable[str]], env: Optional[Mapping[str, str]] = None
) -> bool:
    """True if the SEPARATE draft model is on CPU (so the budget must not charge
    its weights+KV): --spec-draft-ngl 0, or --spec-draft-device naming only
    cpu/none, else the LLAMA_ARG_N_GPU_LAYERS_DRAFT env the child honors (the
    device flag has no env). An embedded MTP head follows the main -ngl, so these
    draft-only flags don't move it. Last-wins, so only each flag's final value counts."""
    ngl_flags = {"--spec-draft-ngl", "-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}
    dev_flags = {"--spec-draft-device", "-devd", "--device-draft"}
    args = [str(a) for a in extra_args] if extra_args else []
    last_ngl: Optional[str] = None
    last_dev: Optional[str] = None
    for i, raw in enumerate(args):
        flag, eq, inline = raw.partition("=")
        value = inline if eq else (args[i + 1] if i + 1 < len(args) else "")
        if flag in ngl_flags:
            last_ngl = value
        elif flag in dev_flags:
            last_dev = value
    if last_ngl is None:
        last_ngl = (os.environ if env is None else env).get("LLAMA_ARG_N_GPU_LAYERS_DRAFT")
    if last_ngl is not None:
        try:
            if int(last_ngl) == 0:
                return True
        except (TypeError, ValueError):
            pass
    if last_dev is not None:
        devs = [d.strip().lower() for d in last_dev.split(",") if d.strip()]
        if devs and all(d in ("cpu", "none") for d in devs):
            return True
    return False


def _extra_args_n_ubatch(
    extra_args: Optional[Iterable[str]], env: Optional[Mapping[str, str]] = None
) -> Optional[int]:
    """Physical micro-batch from extras (--ubatch-size/-ub) else the LLAMA_ARG_UBATCH
    env, else None. It sizes the compute-graph buffer, so an override must reach
    the VRAM reserve."""
    args = [str(a) for a in extra_args] if extra_args else []
    found: Optional[int] = None
    for i, raw in enumerate(args):
        flag, eq, inline = raw.partition("=")
        if flag not in ("--ubatch-size", "-ub"):
            continue
        value = inline if eq else (args[i + 1] if i + 1 < len(args) else "")
        try:
            found = int(value)
        except (TypeError, ValueError):
            continue
    if found is not None:
        return found
    raw = (os.environ if env is None else env).get("LLAMA_ARG_UBATCH")
    if raw:
        try:
            return int(raw)
        except (TypeError, ValueError):
            pass
    return None


def _build_ngram_mod_flags(
    caps: Optional[dict],
    n_match: int = 24,
    n_min: int = 48,
    n_max: int = 64,
) -> list[str]:
    """Emit the right ngram-mod knob flags for the running llama-server.

    Post-rename builds expose ``--spec-ngram-mod-n-{match,min,max}``;
    pre-rename builds expose legacy ``--spec-ngram-size-n`` /
    ``--draft-min`` / ``--draft-max``. ``caps`` comes from
    ``probe_server_capabilities``; ``ngram_mod_flavor`` says which set is
    real (vs a removal-stub). Returns ``[]`` when neither is available so
    the caller can drop ngram-mod entirely.
    """
    flavor = caps.get("ngram_mod_flavor") if caps else None
    if flavor == "new":
        return [
            "--spec-ngram-mod-n-match",
            str(n_match),
            "--spec-ngram-mod-n-min",
            str(n_min),
            "--spec-ngram-mod-n-max",
            str(n_max),
        ]
    if flavor == "legacy":
        # Pre-rename llama.cpp: same knobs lived under --spec-ngram-size-n
        # (lookup length) and generic --draft-min / --draft-max (N range).
        return [
            "--spec-ngram-size-n",
            str(n_match),
            "--draft-min",
            str(n_min),
            "--draft-max",
            str(n_max),
        ]
    return []


# Canonical Speculative Decoding modes exposed by the Studio chat UI.
# Dropdown renders five (auto, mtp, ngram, mtp+ngram, off); the load API
# also accepts legacy values the original Switch and external callers emit
# (default, draft-mtp, ngram-mod, ngram-simple).
_CANONICAL_SPEC_MODES = {"auto", "mtp", "ngram", "mtp+ngram", "off", "ngram-simple"}
_LEGACY_SPEC_MODE_MAP = {
    "default": "auto",
    "draft-mtp": "mtp",
    "ngram-mod": "ngram",
}


def _canonicalize_spec_mode(value):
    """Map any accepted ``speculative_type`` input onto a canonical mode.

    Returns ``auto``, ``mtp``, ``ngram``, ``mtp+ngram``, ``off``,
    ``ngram-simple``, or ``None`` (callers treat ``None`` as ``auto``).
    Unknown strings collapse to ``auto`` so a stale UI value or typo falls
    back to the safe platform-aware path.
    """
    if value is None:
        return None
    if not isinstance(value, str):
        return None
    stripped = value.strip().lower()
    if not stripped:
        return None
    if stripped in _CANONICAL_SPEC_MODES:
        return stripped
    if stripped in _LEGACY_SPEC_MODE_MAP:
        return _LEGACY_SPEC_MODE_MAP[stripped]
    # Old persisted state emits llama.cpp comma-chains e.g.
    # "ngram-mod,draft-mtp"; collapse the most common one explicitly.
    pieces = [p.strip() for p in stripped.split(",") if p.strip()]
    has_mtp = any(p in ("mtp", "draft-mtp") for p in pieces)
    has_ngram = any(p in ("ngram", "ngram-mod") for p in pieces)
    if has_mtp and has_ngram:
        return "mtp+ngram"
    if has_mtp:
        return "mtp"
    if has_ngram:
        return "ngram"
    return "auto"


def _backfill_usage_from_timings(usage, timings):
    """Synthesize ``usage`` from llama-server's ``timings`` when the
    OpenAI-style usage block is missing or reports zero tokens.

    The Studio chat UI computes generation t/s from
    ``meta.usage.completion_tokens / totalStreamTime``. llama-server always
    populates ``timings.predicted_n`` (true decoded count) and
    ``timings.prompt_n``, but the final SSE chunk's ``usage`` can be absent
    or zero on some server builds / streaming configs, making the UI fall
    back to wall-clock t/s and dilute speculative-decoding speedups.
    """
    if not timings:
        return usage
    if usage and usage.get("completion_tokens"):
        return usage
    predicted_n = timings.get("predicted_n")
    prompt_n = timings.get("prompt_n")
    if predicted_n is None and prompt_n is None:
        return usage
    out = dict(usage or {})
    if not out.get("completion_tokens") and predicted_n is not None:
        out["completion_tokens"] = predicted_n
    if not out.get("prompt_tokens") and prompt_n is not None:
        out["prompt_tokens"] = prompt_n
    out["total_tokens"] = int(out.get("prompt_tokens") or 0) + int(
        out.get("completion_tokens") or 0
    )
    return out


def _is_external_link(path: Path) -> bool:
    """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink
    or a Windows directory junction / reparse point. Such a link resolves into
    the user's own llama.cpp checkout, which Studio does not own."""
    try:
        if os.path.islink(path):
            return True
    except OSError:
        return False
    if os.name == "nt":
        try:
            import stat
            attrs = os.lstat(path).st_file_attributes  # type: ignore[attr-defined]
            return bool(attrs & stat.FILE_ATTRIBUTE_REPARSE_POINT)
        except (OSError, AttributeError):
            return False
    return False


class LlamaCppBackend:
    """Manages a llama-server subprocess for GGUF model inference.

    Lifecycle:
        1. load_model(): start llama-server with the GGUF file
        2. generate_chat_completion(): proxy to /v1/chat/completions, stream back
        3. unload_model(): terminate the subprocess
    """

    def __init__(self):
        self._process: Optional[subprocess.Popen] = None
        self._port: Optional[int] = None
        self._model_identifier: Optional[str] = None
        self._gguf_path: Optional[str] = None
        self._hf_repo: Optional[str] = None
        # Separate MTP drafter launched with the current model; reload-dedup
        # key so a drafter that appears next to the weights forces a reload.
        self._mtp_draft_path: Optional[str] = None
        # Why MTP was disabled on the last load that asked for it (auto on an
        # MTP model, or forced mtp / mtp+ngram), else None. Drives the "update
        # llama.cpp" hint in the UI. "binary_no_mtp" / "binary_outdated" ->
        # a newer prebuilt would help; "runtime_error" -> it may not.
        self._spec_fallback_reason: Optional[str] = None
        self._hf_variant: Optional[str] = None
        self._is_vision: bool = False
        # Block-diffusion model (e.g. DiffusionGemma): served by the diffusion
        # runner, not llama-server. Set from the GGUF architecture at load.
        self._architecture: Optional[str] = None
        self._is_diffusion: bool = False
        self._diffusion_visual_bin: Optional[str] = None
        self._healthy = False
        self._load_rss_hwm = (None, 0)  # (pid, peak VmRSS) for load_progress
        self._stats_logger = None  # vLLM-style engine-stats poller, set on load
        # Set by _classify_gpu_offload after _wait_for_health.
        self._gpu_offload_active: Optional[bool] = None
        self._context_length: Optional[int] = None
        self._effective_context_length: Optional[int] = None
        self._max_context_length: Optional[int] = None
        self._chat_template: Optional[str] = None
        self._chat_template_override: Optional[str] = None
        self._supports_reasoning: bool = False
        self._reasoning_always_on: bool = False
        self._reasoning_style: str = "enable_thinking"
        self._reasoning_effort_levels: list = []
        self._supports_preserve_thinking: bool = False
        self._supports_tools: bool = False
        self._cache_type_kv: Optional[str] = None
        # Whether --split-mode tensor was applied on the active load.
        self._tensor_parallel: bool = False
        # Layer load kept multi-GPU only to honor a downgraded tensor request, so a
        # later explicit tensor-off reloads instead of deduping to it (#6659).
        self._layer_preserves_tensor_intent: bool = False
        self._reasoning_default: bool = True
        self._speculative_type: Optional[str] = None
        # Canonical UI-facing mode the user requested
        # (auto/mtp/ngram/mtp+ngram/off/ngram-simple). Round-tripped through the
        # status API so the dropdown reflects the picked mode, not the resolved
        # flag set (auto on a 27B MTP GGUF resolves to draft-mtp but reads "Auto").
        self._requested_spec_mode: Optional[str] = None
        # User --spec-draft-n-max override (None = platform default).
        self._spec_draft_n_max: Optional[int] = None
        # KV-cache estimation fields (populated by _read_gguf_metadata)
        self._n_layers: Optional[int] = None
        self._n_kv_heads: Optional[int] = None
        self._n_kv_heads_by_layer: Optional[list[int]] = None
        self._n_heads: Optional[int] = None
        self._embedding_length: Optional[int] = None
        # For the compute-graph buffer estimate; vocab from the tokens array len.
        self._feed_forward_length: Optional[int] = None
        self._vocab_size: Optional[int] = None
        # Architecture-aware KV fields for 5-path estimation
        self._kv_key_length: Optional[int] = None
        self._kv_value_length: Optional[int] = None
        self._sliding_window: Optional[int] = None
        self._sliding_window_pattern: Optional[list[bool]] = None
        self._full_attention_interval: Optional[int] = None
        self._kv_lora_rank: Optional[int] = None
        self._key_length_mla: Optional[int] = None
        self._kv_key_length_swa: Optional[int] = None
        self._kv_value_length_swa: Optional[int] = None
        self._ssm_inner_size: Optional[int] = None
        self._ssm_state_size: Optional[int] = None
        # Last N layers reuse earlier layers' KV and don't allocate their own
        # cache (Gemma 3n / Gemma 4: <arch>.attention.shared_kv_layers).
        self._shared_kv_layers: Optional[int] = None
        # MTP head count (llama.cpp #22673); >0 enables --spec-type draft-mtp.
        self._nextn_predict_layers: Optional[int] = None
        self._lock = threading.Lock()
        # Wraps load_model() end-to-end so concurrent loads serialise and never
        # coexist as two llama-server processes (#5401). RLock so MTP-crash
        # recovery can re-acquire it for its nested load_model.
        self._serial_load_lock = threading.RLock()
        # Serialises mid-session respawns so many generations hitting a killed
        # server trigger at most one reload (see _respawn_if_dead).
        self._respawn_lock = threading.Lock()
        # Set by the in-app updater while it swaps prebuilt binaries; load_model()
        # rejects fast so no server starts from a half-swapped binary.
        self._llama_update_in_progress = False
        # Last extra_args / requested n_ctx, preserved across unload so the chat
        # UI's /unload+/load Apply path can inherit them (#5401).
        # ``_extra_args_source`` records the (model_identifier, hf_variant) the
        # stored args came from so the route can refuse cross-model inheritance.
        self._extra_args: Optional[List[str]] = None
        self._extra_args_source: Optional[tuple[str, Optional[str]]] = None
        self._requested_n_ctx: int = 0
        # Raw kwargs of the last healthy load, for the MTP-crash reload. Memory-only
        # (carries hf_token, never logged); single-flight via the lock below.
        self._last_load_kwargs: Optional[dict] = None
        self._mtp_runtime_fallback_lock = threading.Lock()
        self._mtp_runtime_fallback_in_progress = False
        # Background watchdog so an MTP+tensor crash recovers even when no request
        # observes it (direct proxy endpoints, or nothing in flight).
        self._mtp_watchdog_thread: Optional[threading.Thread] = None
        self._mtp_watchdog_stop = threading.Event()
        # True when the launch actually runs MTP+tensor (Studio- or user/env-driven);
        # gates the probe, watchdog, and recovery so pass-through MTP is covered.
        self._mtp_runtime_fallback_active = False
        self._stdout_lines: list[str] = []
        self._stdout_thread: Optional[threading.Thread] = None
        # llama-server tee log (see _drain_stdout / _kill_process).
        self._llama_log_fh = None
        self._llama_log_path: Optional[Path] = None
        self._cancel_event = threading.Event()
        self._api_key: Optional[str] = None
        # True once a probe has completed; cleared on transient failure.
        self._is_audio: bool = False
        self._audio_type: Optional[str] = None
        self._audio_probed: bool = False
        # Audio INPUT capability (distinct from _is_audio, which is TTS output).
        self._has_audio_input: bool = False
        self._mmproj_has_audio: bool = False  # clip.has_audio_encoder, set at load
        # Monotonic timestamp set in _kill_process; read by load_model
        # to decide whether to wait for the VRAM reclaim to finish.
        self._last_kill_monotonic: float = 0.0

        _reaped = self._kill_orphaned_servers()
        if _reaped:
            # Reaped VRAM frees lazily; arm the settle wait so the first load
            # waits before ranking GPUs by free memory.
            self._last_kill_monotonic = time.monotonic()
        atexit.register(self._cleanup)

    # ── Properties ────────────────────────────────────────────────

    @property
    def is_loaded(self) -> bool:
        return self._process is not None and self._healthy

    @property
    def is_active(self) -> bool:
        """True if a llama-server process exists (loading or loaded)."""
        return self._process is not None

    @property
    def base_url(self) -> str:
        return f"http://127.0.0.1:{self._port}"

    @property
    def _auth_headers(self) -> "Optional[dict[str, str]]":
        """Bearer header matching the --api-key direct-stream mode uses, else
        None (so unauthenticated llama-server calls don't get a spurious 401)."""
        return {"Authorization": f"Bearer {self._api_key}"} if self._api_key else None

    @property
    def model_identifier(self) -> Optional[str]:
        return self._model_identifier

    @property
    def is_vision(self) -> bool:
        return self._is_vision

    @property
    def is_diffusion(self) -> bool:
        """True when the loaded GGUF is a block-diffusion model (DiffusionGemma)."""
        return self._is_diffusion

    @property
    def hf_variant(self) -> Optional[str]:
        return self._hf_variant

    @property
    def gguf_path(self) -> Optional[str]:
        return self._gguf_path

    @property
    def hf_repo(self) -> Optional[str]:
        """HF repo of the loaded model, or None for local/native file loads."""
        return self._hf_repo

    @property
    def mtp_draft_path(self) -> Optional[str]:
        return self._mtp_draft_path

    @property
    def spec_fallback_reason(self) -> Optional[str]:
        """Why MTP was disabled on the last MTP-requesting load, else None."""
        return self._spec_fallback_reason

    @property
    def extra_args(self) -> Optional[List[str]]:
        """Extra llama-server flags from the last load (a copy). None =
        never set, [] = explicitly cleared. Used by the route for
        inheritance."""
        return list(self._extra_args) if self._extra_args is not None else None

    @property
    def requested_n_ctx(self) -> int:
        """n_ctx the last load was invoked with (not the effective cap).
        0 means Auto. Used by the route to detect Auto-vs-explicit flips."""
        return self._requested_n_ctx

    @property
    def extra_args_source(self) -> Optional[tuple[str, Optional[str]]]:
        """(model_identifier, hf_variant) the stored extra_args came from.
        ``None`` if no extras have ever been recorded. Used by the route
        to refuse cross-model inheritance (#5401)."""
        return self._extra_args_source

    @property
    def context_length(self) -> Optional[int]:
        """Return the effective context length the server is running at."""
        return self._effective_context_length or self._context_length

    @property
    def max_context_length(self) -> Optional[int]:
        """Return the largest context that fits on this hardware at load time.

        The UI's "safe zone" warning threshold: the ``_fit_context_to_vram``
        binary-search cap for the best GPU subset, or the 4096 fallback if the
        weights exceed 90% of every subset. The slider ceiling is
        ``native_context_length``; dragging above this triggers the warning.
        """
        return self._max_context_length or self._context_length

    @property
    def native_context_length(self) -> Optional[int]:
        """Return the model's native context length from GGUF metadata."""
        return self._context_length

    @staticmethod
    def _read_rss_bytes(pid: int) -> Optional[int]:
        """Resident set size of ``pid`` in bytes, from /proc/<pid>/status (Linux).
        0 when the status has no VmRSS line (zombie / kernel thread); None where
        /proc is unavailable (macOS/Windows) or the value is unreadable."""
        try:
            with open(f"/proc/{pid}/status", "r", encoding = "utf-8") as f:
                for line in f:
                    if line.startswith("VmRSS:"):
                        # IndexError guards a "VmRSS:" line with no value column.
                        return int(line.split()[1]) * 1024  # kB -> bytes
        except (FileNotFoundError, PermissionError, ValueError, IndexError, OSError):
            return None
        return 0  # readable but no VmRSS line

    def load_progress(self) -> Optional[dict]:
        """Return live model-load progress, or None if not loading.

        During warm-up llama-server mmaps weight shards into page cache before
        pushing layers to VRAM, a window where status only reports ``loading``
        and the UI spinner looks stuck for minutes on large MoEs. Samples
        ``/proc/<pid>/status VmRSS`` against the sum of GGUF shard sizes for a
        real progress bar. Returns ``None`` when no load is in flight.

        Shape::

            {
                "phase": "mmap" | "ready",
                "bytes_loaded": int,   # VmRSS of the llama-server
                "bytes_total":  int,   # sum of shard file sizes
                "fraction": float,     # bytes_loaded / bytes_total, 0..1
            }

        Linux-only; returns ``None`` where ``/proc/<pid>/status`` is unavailable.
        """
        proc = self._process
        if proc is None:
            return None
        pid = proc.pid
        if pid is None:
            return None

        # Sum shard sizes (primary + any extras alongside).
        bytes_total = 0
        gguf_path = self._gguf_path
        if gguf_path:
            primary = Path(gguf_path)
            try:
                if primary.is_file():
                    bytes_total += primary.stat().st_size
            except OSError:
                pass
            # Extra shards share the primary's prefix before the shard index.
            try:
                parent = primary.parent
                stem = primary.name
                m = _SHARD_RE.match(stem)
                prefix = m.group(1) if m else None
                if prefix and parent.is_dir():
                    prefix_lower = prefix.lower()
                    for sibling in parent.iterdir():
                        if (
                            sibling.is_file()
                            and sibling.name.lower().startswith(prefix_lower)
                            and sibling.name != stem
                            and sibling.suffix.lower() == ".gguf"
                        ):
                            try:
                                bytes_total += sibling.stat().st_size
                            except OSError:
                                pass
            except OSError:
                pass

        # VmRSS of the llama-server; None where /proc is unavailable.
        bytes_loaded = LlamaCppBackend._read_rss_bytes(pid)
        if bytes_loaded is None:
            return None

        # RSS climbs as weights page in, then drops once -ngl offloads them to
        # VRAM and the mmap pages are freed. Hold a per-process high-water mark
        # so the bar never regresses to ~8% mid-load (#5740).
        hwm_pid, hwm = getattr(self, "_load_rss_hwm", (None, 0))
        hwm = bytes_loaded if hwm_pid != pid else max(hwm, bytes_loaded)
        self._load_rss_hwm = (pid, hwm)
        bytes_loaded = hwm

        phase = "ready" if self._healthy else "mmap"
        fraction = 0.0
        if bytes_total > 0:
            fraction = min(1.0, bytes_loaded / bytes_total)
        # Once llama-server is healthy the load is complete by definition. With
        # layers offloaded to VRAM (-ngl) the process releases the mmap'd weight
        # pages, so VmRSS sinks back well below the shard total; the raw RSS
        # fraction would then report a partial (~8%) load indefinitely and freeze
        # a fraction-driven progress bar even though the model is ready (#5740).
        if self._healthy:
            if bytes_total > 0:
                bytes_loaded = bytes_total
            fraction = 1.0
        return {
            "phase": phase,
            "bytes_loaded": bytes_loaded,
            "bytes_total": bytes_total,
            "fraction": round(fraction, 4),
        }

    @property
    def chat_template(self) -> Optional[str]:
        return self._chat_template

    @property
    def chat_template_override(self) -> Optional[str]:
        return self._chat_template_override

    @property
    def supports_reasoning(self) -> bool:
        return self._supports_reasoning

    @property
    def reasoning_always_on(self) -> bool:
        return self._reasoning_always_on

    @property
    def reasoning_style(self) -> str:
        return self._reasoning_style

    @property
    def reasoning_effort_levels(self) -> list:
        """Discrete reasoning_effort levels the template offers (e.g. GLM-5.2's
        ['high', 'max']). Empty unless reasoning_style == 'enable_thinking_effort'."""
        return self._reasoning_effort_levels

    @property
    def supports_preserve_thinking(self) -> bool:
        return self._supports_preserve_thinking

    @property
    def reasoning_default(self) -> bool:
        return self._reasoning_default

    def _reasoning_kwargs(self, enable_thinking: bool) -> dict:
        if self._reasoning_style == "enable_thinking_effort":
            # GLM-5.2-style: enable_thinking is the on/off gate; when on, leave
            # the template's default effort (max) in place.
            return {"enable_thinking": enable_thinking}
        if self._reasoning_style == "reasoning_effort":
            return {"reasoning_effort": "high" if enable_thinking else "low"}
        return {"enable_thinking": enable_thinking}

    def _request_reasoning_kwargs(
        self,
        enable_thinking: Optional[bool],
        reasoning_effort: Optional[str] = None,
        preserve_thinking: Optional[bool] = None,
    ) -> Optional[dict]:
        """Build chat_template_kwargs from per-request reasoning fields.

        Merges the active model's reasoning style (``enable_thinking`` or
        ``reasoning_effort``) plus the independent ``preserve_thinking``
        kwarg when the template supports it.
        """
        kwargs: dict = {}
        # Always-on reasoning models hardcode <think> tags and don't consume
        # enable_thinking / reasoning_effort -- skip.
        if self._supports_reasoning and not self._reasoning_always_on:
            if self._reasoning_style == "enable_thinking_effort":
                # GLM-5.2-style: enable_thinking gates thinking on/off, and the
                # reasoning_effort level (e.g. 'high' | 'max') is only meaningful
                # while thinking is on. Disabling is enable_thinking=false; a raw
                # API caller can also disable via the OpenAI-style
                # reasoning_effort="none" sentinel. We never coerce off into a
                # 'low' effort the way gpt-oss does (those models genuinely
                # cannot disable).
                thinking_off = enable_thinking is False or reasoning_effort == "none"
                # A named effort level implies thinking on, so emit enable_thinking
                # even if the caller sent only reasoning_effort (else the template
                # defaults it off and the requested level never renders).
                effort_on = reasoning_effort in self._reasoning_effort_levels
                if enable_thinking is not None or reasoning_effort == "none" or effort_on:
                    kwargs["enable_thinking"] = not thinking_off
                if not thinking_off and effort_on:
                    kwargs["reasoning_effort"] = reasoning_effort
            elif self._reasoning_style == "reasoning_effort":
                if reasoning_effort in ("none", "low", "medium", "high"):
                    kwargs["reasoning_effort"] = reasoning_effort
                elif reasoning_effort == "minimal":
                    kwargs["reasoning_effort"] = "low"
                elif enable_thinking is not None:
                    kwargs["reasoning_effort"] = "high" if enable_thinking else "low"
            else:
                if enable_thinking is not None:
                    kwargs["enable_thinking"] = enable_thinking
        if self._supports_preserve_thinking and preserve_thinking is not None:
            kwargs["preserve_thinking"] = preserve_thinking
        return kwargs or None

    @property
    def supports_tools(self) -> bool:
        # DiffusionGemma serves via the visual runner, whose live per-step canvas
        # frames are dropped by the agentic tool loop; never route it through tools.
        if self._is_diffusion:
            return False
        return self._supports_tools

    @property
    def supports_tool_passthrough(self) -> bool:
        # supports_tools is forced off for DiffusionGemma (its agentic loop drops the
        # per-step canvas frames), but client passthrough skips that loop, so it uses
        # the real _supports_tools.
        return self._supports_tools

    @property
    def cache_type_kv(self) -> Optional[str]:
        return self._cache_type_kv

    @property
    def tensor_parallel(self) -> bool:
        """Whether --split-mode tensor is active on the loaded server."""
        return self._tensor_parallel

    @property
    def layer_preserves_tensor_intent(self) -> bool:
        """True when a downgraded tensor request kept this layer load multi-GPU."""
        return self._layer_preserves_tensor_intent

    @property
    def speculative_type(self) -> Optional[str]:
        return self._speculative_type

    @property
    def requested_spec_mode(self) -> Optional[str]:
        """Canonical UI-facing mode the user requested (see field doc)."""
        return self._requested_spec_mode

    @property
    def spec_draft_n_max(self) -> Optional[int]:
        """User --spec-draft-n-max override active on the load, or None when
        the platform default (6 GPU / 3 CPU) is in effect."""
        return self._spec_draft_n_max

    # ── Binary discovery ──────────────────────────────────────────

    @staticmethod
    def _resolved_studio_root_and_is_legacy() -> "tuple[Optional[Path], bool]":
        """Resolve the Studio install root and classify it as the legacy
        ~/.unsloth/studio root vs. a custom (env/venv-inferred) root.

        Returns (resolved_root, is_legacy). On any import/resolution failure the
        root is treated as legacy and resolved_root is None -- callers must read
        resolved_root only when is_legacy is False. Shared by
        _find_llama_server_binary (discovery) and _kill_orphaned_servers
        (cleanup) so the two never disagree on which root is legacy.
        """
        try:
            from utils.paths.storage_roots import studio_root as _sr  # noqa: WPS433

            resolved = _sr()
            legacy_studio = Path.home() / ".unsloth" / "studio"
            try:
                is_legacy = resolved.resolve() == legacy_studio.resolve()
            except (OSError, ValueError):
                is_legacy = resolved == legacy_studio
            return (None if is_legacy else resolved), is_legacy
        except (ImportError, OSError, ValueError):
            return None, True

    @staticmethod
    def _find_llama_server_binary(*, include_denied: bool = False) -> Optional[str]:
        """
        Locate the llama-server binary.

        Search order:
        1.  LLAMA_SERVER_PATH environment variable (direct path to binary)
        1b. UNSLOTH_LLAMA_CPP_PATH env var (custom llama.cpp install dir)
        2.  ~/.unsloth/llama.cpp/llama-server        (make build, root dir)
        3.  ~/.unsloth/llama.cpp/build/bin/llama-server  (cmake build, Linux)
        4.  ~/.unsloth/llama.cpp/build/bin/Release/llama-server.exe  (cmake build, Windows)
        5.  ./llama.cpp/llama-server                 (legacy: make build, root dir)
        6.  ./llama.cpp/build/bin/llama-server        (legacy: cmake in-tree build)
        7.  llama-server on PATH                     (system install)
        8.  ./bin/llama-server                       (legacy: extracted binary)
        """
        binary_name = "llama-server.exe" if sys.platform == "win32" else "llama-server"

        def _file_status(p: Path) -> str:
            # "file", "absent", or "denied" (exists but stays access-denied
            # across a short retry: Windows AV/ACL or an install replace in
            # flight). is_file() raises PermissionError (WinError 5) instead of
            # returning False for the locked case, so never treat it as missing.
            for _ in range(5):
                try:
                    return "file" if p.is_file() else "absent"
                except PermissionError:
                    time.sleep(0.2)
                except OSError:
                    return "absent"
            return "denied"

        def _is_file(p: Path) -> bool:
            return _file_status(p) == "file"

        def _layout_candidates(d: Path) -> list:
            # build layouts probed under a llama.cpp dir, highest priority first
            cands = [d / binary_name, d / "build" / "bin" / binary_name]
            if sys.platform == "win32":
                cands.append(d / "build" / "bin" / "Release" / binary_name)
            return cands

        def _unavailable(p: object) -> None:
            # a pinned or managed binary that exists but is access-denied: report
            # it instead of silently downgrading to a lower-priority llama-server
            logger.warning(
                f"llama-server at {p} exists but is access-denied (antivirus or "
                "an in-flight install); not falling back to another binary, "
                "retry once it is released"
            )
            return None

        def _scan_pinned(paths: list):
            # first existing candidate wins -> (path, None); a present-but-denied
            # one -> (None, denied_path) so the caller reports it rather than
            # skipping to a lower-priority location. include_denied returns the
            # locked path instead: diffusion asset lookup only needs its dir.
            for p in paths:
                st = _file_status(p)
                if st == "file":
                    return str(p), None
                if st == "denied":
                    return (str(p), None) if include_denied else (None, p)
            return None, None

        # 1. Env var: direct path to binary
        env_path = os.environ.get("LLAMA_SERVER_PATH")
        if env_path:
            hit, locked = _scan_pinned([Path(env_path)])
            if locked is not None:
                return _unavailable(locked)
            if hit:
                return hit

        # 1b. UNSLOTH_LLAMA_CPP_PATH: custom llama.cpp install dir
        custom_llama_cpp = os.environ.get("UNSLOTH_LLAMA_CPP_PATH")
        if custom_llama_cpp:
            hit, locked = _scan_pinned(_layout_candidates(Path(custom_llama_cpp)))
            if locked is not None:
                return _unavailable(locked)
            if hit:
                return hit

        # 2-4. Match installer layout: env-mode -> $STUDIO_HOME/llama.cpp;
        # default/HOME-redirect -> ~/.unsloth/llama.cpp (sibling of studio).
        legacy_llama = Path.home() / ".unsloth" / "llama.cpp"
        _resolved_sr, _is_legacy = LlamaCppBackend._resolved_studio_root_and_is_legacy()
        if _is_legacy:
            search_roots = [legacy_llama]
        else:
            # _kill_orphaned_servers excludes the legacy root in custom mode;
            # discovery must match so we never spawn a server we then refuse to
            # clean up. UNSLOTH_LLAMA_CPP_PATH (handled earlier) is the explicit
            # way to share a build across roots.
            search_roots = [_resolved_sr / "llama.cpp"]
        for unsloth_home in search_roots:
            hit, locked = _scan_pinned(_layout_candidates(unsloth_home))
            if locked is not None:
                return _unavailable(locked)
            if hit:
                return hit

        # 5-6. Legacy: in-tree build (older setup.sh / setup.ps1). A fallback,
        # so a denied candidate here just continues (no no-fallback halt).
        project_root = Path(__file__).resolve().parents[4]
        for p in _layout_candidates(project_root / "llama.cpp"):
            if _is_file(p):
                return str(p)

        # 7. System PATH
        system_path = shutil.which("llama-server")
        if system_path:
            return system_path

        # 8. Legacy: extracted to bin/
        bin_path = project_root / "bin" / binary_name
        if _is_file(bin_path):
            return str(bin_path)

        return None

    # ── llama-server capability probe ─────────────────────────────

    # Cached on (path, mtime); `unsloth studio update` bumps mtime.
    _capability_cache: dict[tuple[str, int], dict[str, object]] = {}

    @classmethod
    def probe_server_capabilities(cls, binary: Optional[str] = None) -> dict[str, object]:
        """Parse `llama-server --help` for feature flags. Returns
        {found, mtp_token, supports_mtp, ngram_mod_flavor,
        supports_ngram_mod, spec_draft_n_max_flag, cache flag support}.

        ``ngram_mod_flavor``: ``"new"`` when the post-rename
        ``--spec-ngram-mod-n-match / -n-min / -n-max`` are real args;
        ``"legacy"`` when only the pre-rename
        ``--spec-ngram-size-n / --draft-min / --draft-max`` are real (the
        rename ships stub removal entries for legacy names, told apart by
        the "argument has been removed" description); ``None`` if neither
        set is usable.

        ``spec_draft_n_max_flag``: the flag the binary accepts --
        ``--spec-draft-n-max`` post-rename, ``--draft-max`` on legacy.
        ``None`` means n_max cannot be set.
        """
        bin_path = binary or cls._find_llama_server_binary()
        if not bin_path or not Path(bin_path).is_file():
            return {
                "found": False,
                "mtp_token": None,
                "supports_mtp": False,
                "ngram_mod_flavor": None,
                "supports_ngram_mod": False,
                "spec_draft_n_max_flag": None,
                "supports_kv_unified": False,
                "supports_fit_ctx": False,
                "supports_cache_ram": False,
                "supports_ctx_checkpoints": False,
                "supports_no_cache_prompt": False,
                "supports_metrics": False,
            }
        try:
            mtime = int(Path(bin_path).stat().st_mtime)
        except OSError:
            mtime = 0
        cache_key = (bin_path, mtime)
        cached = cls._capability_cache.get(cache_key)
        if cached is not None:
            return cached

        mtp_token: Optional[str] = None
        ngram_mod_flavor: Optional[str] = None
        spec_draft_n_max_flag: Optional[str] = None
        supports_kv_unified = False
        supports_fit_ctx = False
        supports_cache_ram = False
        supports_ctx_checkpoints = False
        supports_no_cache_prompt = False
        supports_metrics = False
        try:
            probe_env = cls._llama_server_env_for_binary(bin_path)
            result = subprocess.run(
                [bin_path, "--help"],
                capture_output = True,
                text = True,
                errors = "replace",
                timeout = 10,
                check = False,
                env = probe_env,
            )
            help_text = (result.stdout or "") + "\n" + (result.stderr or "")
            # Split into per-flag blocks (each --flag line + its indented
            # continuation), so the "argument has been removed" description
            # sits with its flag.
            blocks: dict[str, str] = {}
            current_flags: list[str] = []
            current_desc: list[str] = []
            for line in help_text.splitlines():
                stripped = line.strip()
                if stripped.startswith("-") and not line.startswith(" "):
                    # New flag line; flush previous.
                    if current_flags:
                        desc = " ".join(current_desc)
                        for f in current_flags:
                            blocks[f] = desc
                    current_flags = []
                    current_desc = [stripped]
                    # Extract long-form flag tokens from the DECLARATION
                    # prefix only (comma-separated aliases). Stop at the
                    # first non-flag token so flag references inside
                    # descriptions are ignored.
                    for tok in re.split(r"[,\s]+", stripped):
                        if tok.startswith("--") and re.match(r"--[A-Za-z][A-Za-z0-9_-]*$", tok):
                            current_flags.append(tok)
                        elif tok.startswith("-") and len(tok) > 1:
                            # short alias like -fa; keep scanning aliases.
                            continue
                        else:
                            # First non-flag token marks end of decl.
                            break
                else:
                    current_desc.append(stripped)
            if current_flags:
                desc = " ".join(current_desc)
                for f in current_flags:
                    blocks[f] = desc

            def _is_real(flag: str) -> bool:
                """True if the flag exists AND is not a removal stub."""
                desc = blocks.get(flag)
                if desc is None:
                    return False
                return "argument has been removed" not in desc

            # MTP token from the --spec-type line.
            spec_line = ""
            for line in help_text.splitlines():
                if "--spec-type" in line:
                    spec_line = line
                    break
            # PR #22673 used draft-mtp; later renamed to mtp.
            if "draft-mtp" in spec_line:
                mtp_token = "draft-mtp"
            elif re.search(r"[|,\[]mtp[|,\]]", spec_line):
                mtp_token = "mtp"

            # ngram-mod flag flavor. Post-rename builds advertise both new
            # args (real) and legacy ones (stubs); pre-rename builds only
            # have legacy ones as real.
            new_ngram_real = (
                _is_real("--spec-ngram-mod-n-match")
                and _is_real("--spec-ngram-mod-n-min")
                and _is_real("--spec-ngram-mod-n-max")
            )
            legacy_ngram_real = (
                _is_real("--spec-ngram-size-n")
                and _is_real("--draft-max")
                and _is_real("--draft-min")
            )
            if new_ngram_real:
                ngram_mod_flavor = "new"
            elif legacy_ngram_real:
                ngram_mod_flavor = "legacy"

            # n_max flag: prefer post-rename, fall back to legacy.
            if _is_real("--spec-draft-n-max"):
                spec_draft_n_max_flag = "--spec-draft-n-max"
            elif _is_real("--draft-max"):
                spec_draft_n_max_flag = "--draft-max"

            supports_kv_unified = _is_real("--kv-unified")
            supports_fit_ctx = _is_real("--fit-ctx")
            supports_cache_ram = _is_real("--cache-ram")
            supports_ctx_checkpoints = _is_real("--ctx-checkpoints")
            supports_no_cache_prompt = _is_real("--no-cache-prompt")
            supports_metrics = _is_real("--metrics")
        except (OSError, subprocess.SubprocessError) as exc:
            logger.debug(f"llama-server --help probe failed: {exc}")

        info = {
            "found": True,
            "mtp_token": mtp_token,
            "supports_mtp": mtp_token is not None,
            "ngram_mod_flavor": ngram_mod_flavor,
            "supports_ngram_mod": ngram_mod_flavor is not None,
            "spec_draft_n_max_flag": spec_draft_n_max_flag,
            "supports_kv_unified": supports_kv_unified,
            "supports_fit_ctx": supports_fit_ctx,
            "supports_cache_ram": supports_cache_ram,
            "supports_ctx_checkpoints": supports_ctx_checkpoints,
            "supports_no_cache_prompt": supports_no_cache_prompt,
            "supports_metrics": supports_metrics,
        }
        cls._capability_cache[cache_key] = info
        return info

    # ── GPU allocation ────────────────────────────────────────────

    @staticmethod
    def _get_gguf_size_bytes(model_path: str) -> int:
        """Total GGUF size in bytes, including split shards."""
        main = Path(model_path)
        total = main.stat().st_size

        # Check for split shards (e.g. model-00001-of-00003.gguf)
        m = _SHARD_FULL_RE.match(main.name)
        if m:
            prefix, _, num_total = m.group(1), m.group(2), m.group(3)
            sibling_pat = re.compile(
                r"^" + re.escape(prefix) + r"-\d{5}-of-" + re.escape(num_total) + r"\.gguf$",
                re.IGNORECASE,
            )
            for sibling in main.parent.iterdir():
                if sibling != main and sibling_pat.match(sibling.name):
                    total += sibling.stat().st_size

        return total

    @staticmethod
    def _resolve_visible_physical_ids() -> Optional[list[int]]:
        """Physical GPU ids behind the active visibility mask (HIP/ROCR/CUDA on
        ROCm, CUDA otherwise). None when no mask is set; empty list for an empty
        mask. Shared by the APU / datacenter / free-memory probes so they agree
        on the ordinal->physical mapping."""
        try:
            import torch
            is_rocm = getattr(torch.version, "hip", None) is not None
        except Exception:
            is_rocm = False
        if is_rocm:
            hip_v = os.environ.get("HIP_VISIBLE_DEVICES")
            rocr_v = os.environ.get("ROCR_VISIBLE_DEVICES")
            cvd = (
                hip_v
                if hip_v is not None
                else rocr_v
                if rocr_v is not None
                else os.environ.get("CUDA_VISIBLE_DEVICES")
            )
        else:
            cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
        if cvd is None:
            return None
        try:
            return [int(x.strip()) for x in cvd.split(",") if x.strip()]
        except ValueError:
            return None

    @staticmethod
    def _amd_apu_wants_unified_memory(gpu_indices = None) -> bool:
        """True only for AMD unified-memory APUs (gfx1150/gfx1151), where
        GGML_CUDA_ENABLE_UNIFIED_MEMORY lets llama.cpp use shared system RAM (it
        hurts discrete GPUs). gpu_indices (PHYSICAL ids) scopes the check to the
        selected GPUs, so a dGPU on a mixed host is not treated as unified-memory;
        None means every visible GPU."""
        try:
            import torch

            if getattr(torch.version, "hip", None) is None:
                return False
            if not (hasattr(torch, "cuda") and torch.cuda.is_available()):
                return False
            # Map visible ordinal -> physical id via the active ROCm mask (HIP,
            # then ROCR, then CUDA), mirroring _get_gpu_memory's ROCm branch.
            physical_ids = LlamaCppBackend._resolve_visible_physical_ids()
            arch_by_id: dict[int, str] = {}
            for ordinal in range(torch.cuda.device_count()):
                try:
                    _arch = (
                        getattr(torch.cuda.get_device_properties(ordinal), "gcnArchName", "") or ""
                    )
                except Exception:
                    continue
                pid = (
                    physical_ids[ordinal]
                    if physical_ids is not None and ordinal < len(physical_ids)
                    else ordinal
                )
                arch_by_id[pid] = _arch.split(":")[0].strip().lower()
            for _i in list(gpu_indices) if gpu_indices is not None else list(arch_by_id):
                if arch_by_id.get(_i) in {"gfx1150", "gfx1151"}:
                    return True
        except Exception:
            return False
        return False

    # Datacenter / professional NVIDIA parts that benefit from the llama.cpp
    # FP32-accum / P2P tunings. Whole-word (\b) so short markers don't match
    # workstation parts as substrings: "a100" must not fire on "RTX A1000".
    _DATACENTER_GPU_RE = re.compile(
        r"\b(?:a100|a30|h100|h200|h800|gh200|b200|b100|b300|gb200|gb300|"
        r"l40s?|l4|rtx pro 6000|rtx 6000 ada)\b"
    )

    @staticmethod
    def _is_datacenter_gpu(gpu_indices = None) -> bool:
        """True iff every selected NVIDIA GPU is a datacenter/professional part.
        NVIDIA-only, fails open to False (consumer GeForce, ROCm, CPU and errors
        are left untouched); a mixed DC+consumer selection counts as non-DC.

        gpu_indices are PHYSICAL ids (see _get_gpu_free_memory), but
        get_device_properties wants mask-relative ordinals, so we rebuild the
        ordinal->physical map from CUDA_VISIBLE_DEVICES and key names by physical
        id. Otherwise a masked host (CUDA_VISIBLE_DEVICES=4,5,6,7, selection [4,5])
        would drop the tuning or probe the wrong GPU."""
        try:
            import torch

            if getattr(torch.version, "hip", None) is not None:
                return False  # ROCm reuses torch.cuda.*; not a CUDA part
            if not (hasattr(torch, "cuda") and torch.cuda.is_available()):
                return False
            count = torch.cuda.device_count()

            # Mirror _get_gpu_free_memory: map visible ordinal -> physical id via
            # CUDA_VISIBLE_DEVICES; unset/unparsable leaves physical id == ordinal.
            physical_ids = LlamaCppBackend._resolve_visible_physical_ids()

            pattern = LlamaCppBackend._DATACENTER_GPU_RE
            names_by_id: dict[int, str] = {}
            for ordinal in range(count):
                try:
                    name = (torch.cuda.get_device_properties(ordinal).name or "").lower()
                except Exception:
                    continue
                pid = (
                    physical_ids[ordinal]
                    if physical_ids is not None and ordinal < len(physical_ids)
                    else ordinal
                )
                names_by_id[pid] = name

            indices = list(gpu_indices) if gpu_indices else list(names_by_id)
            saw = False
            for _i in indices:
                name = names_by_id.get(_i)
                if name is None:
                    continue  # not visible -> skip (fail conservative)
                saw = True
                if not pattern.search(name):
                    return False
            return saw
        except Exception:
            return False

    @staticmethod
    def _effective_gpu_count(gpu_indices = None) -> int:
        """GPUs llama-server will use: len(selection), else the visible CUDA
        device count (None = every visible GPU). 0 on error so multi-GPU tuning
        stays off when the count is unknown."""
        if gpu_indices is not None:
            return len(gpu_indices)
        try:
            import torch
            if hasattr(torch, "cuda") and torch.cuda.is_available():
                return torch.cuda.device_count()
        except Exception:
            return 0
        return 0

    @staticmethod
    def _apply_datacenter_env(env: dict, gpu_indices = None) -> bool:
        """Inject DC llama.cpp tuning into env in place via setdefault (user
        values win); return whether the box qualified. Opt out with
        UNSLOTH_DISABLE_DC_TUNING=1; only datacenter NVIDIA parts qualify
        (consumer/ROCm/CPU/error are a no-op). Sets GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F
        for any qualifying GPU (FP32 accum: ~0% cost on B200, real cost on GeForce),
        plus GGML_CUDA_P2P + CUDA_SCALE_LAUNCH_QUEUES=4x for multi-GPU (+33-51% pp
        tensor-split, +8-16% pipeline split on B200)."""
        if os.environ.get("UNSLOTH_DISABLE_DC_TUNING") == "1":
            return False
        if not LlamaCppBackend._is_datacenter_gpu(gpu_indices):
            return False
        env.setdefault("GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F", "1")
        if LlamaCppBackend._effective_gpu_count(gpu_indices) > 1:
            env.setdefault("GGML_CUDA_P2P", "1")
            env.setdefault("CUDA_SCALE_LAUNCH_QUEUES", "4x")
        return True

    @staticmethod
    def _get_gpu_free_memory() -> list[tuple[int, int]]:
        """Query free memory per GPU. Returns ``(gpu_index, free_mib)`` sorted by
        index; empty if no supported GPU is reachable. Thin wrapper over
        ``_get_gpu_memory`` for callers that only need free VRAM."""
        return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory()]

    @staticmethod
    def _apple_metal_memory_budget_bytes() -> int:
        """Unified-memory budget for GGUF context fitting on Apple Silicon.

        No GPU is enumerated on Metal, so the context would default to native and
        over-commit unified memory ("Compute error." at decode, #5118/#6529). Use a
        fraction of MLX's Metal working-set, else total RAM; 0 off Apple Silicon or
        when unresolvable, so callers skip the cap.
        """
        from utils.hardware import is_apple_silicon

        if not is_apple_silicon():
            return 0
        rec_bytes = 0
        try:
            import mlx.core as mx
            if mx.metal.is_available():
                rec_bytes = int(mx.device_info().get("max_recommended_working_set_size") or 0)
        except Exception:
            rec_bytes = 0
        if rec_bytes <= 0:
            try:
                import psutil
                rec_bytes = int(psutil.virtual_memory().total)
            except Exception:
                return 0
        return int(rec_bytes * _APPLE_UNIFIED_MEMORY_FRACTION)

    @staticmethod
    def _get_gpu_memory() -> list[tuple[int, int, int]]:
        """Query free AND total memory per GPU.

        Order:
          1. ``nvidia-smi`` (NVIDIA CUDA hosts) -- respects
             ``CUDA_VISIBLE_DEVICES``.
          2. ``torch.cuda.mem_get_info`` -- universal fallback that works
             on AMD ROCm too (HIP runtime reuses the ``torch.cuda.*``
             namespace). Covers the AMD case for issue #5106 (nvidia-smi
             probe returned [] on AMD) and NVIDIA hosts missing
             ``nvidia-smi`` from PATH.

        Returns (gpu_index, free_mib, total_mib) sorted by index; empty if no
        supported GPU is reachable. ``total`` lets the fit reserve absolute headroom.
        """
        # ── NVIDIA via nvidia-smi ────────────────────────────────────
        try:
            result = subprocess.run(
                [
                    "nvidia-smi",
                    "--query-gpu=index,memory.free,memory.total",
                    "--format=csv,noheader,nounits",
                ],
                capture_output = True,
                text = True,
                timeout = 10,
                env = child_env_without_native_path_secret(),
                **_windows_hidden_subprocess_kwargs(),
            )
            if result.returncode == 0:
                allowed: Optional[set[int]] = None
                cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
                if cvd is not None:
                    try:
                        # `if x.strip()` filters trailing-comma masks ("0,1,").
                        # Empty mask (CVD="") yields an empty set -> all GPUs
                        # filtered out, per codebase convention.
                        allowed = set(int(x.strip()) for x in cvd.split(",") if x.strip())
                    except ValueError:
                        pass
                gpus: list[tuple[int, int, int]] = []
                for line in result.stdout.strip().splitlines():
                    parts = [p.strip() for p in line.split(",")]
                    if len(parts) < 2:
                        continue
                    # Index and free required; skip a bad line rather than abandon
                    # the probe to the torch fallback.
                    try:
                        idx = int(parts[0])
                        free_mib = int(parts[1])
                    except ValueError:
                        continue
                    # Total parsed separately: a two-column line or a non-integer
                    # total ("N/A" on MIG/vGPU) keeps the GPU at total 0 (fit uses
                    # the free*frac fallback) instead of dropping it.
                    total_mib = 0
                    if len(parts) >= 3 and parts[2]:
                        try:
                            total_mib = int(parts[2])
                        except ValueError:
                            total_mib = 0
                    if allowed is not None and idx not in allowed:
                        continue
                    gpus.append((idx, free_mib, total_mib))
                # Match the docstring's sort-by-id guarantee (driver order isn't).
                gpus.sort(key = lambda g: g[0])
                if gpus:
                    return gpus
        except Exception as e:
            logger.debug(f"nvidia-smi probe failed: {e}")

        # ── Torch fallback (covers AMD ROCm and missing nvidia-smi) ──
        try:
            import torch

            if not hasattr(torch, "cuda") or not torch.cuda.is_available():
                return []
            if not hasattr(torch.cuda, "mem_get_info"):
                return []
            # torch.cuda enumerates GPUs RELATIVE to the visibility mask. We
            # feed these IDs back into the subprocess as CVD, so visible ordinals
            # must be translated to physical indices first; otherwise CVD=2,3
            # gets rewritten to 0,1 and targets the wrong GPUs.
            # Match utils/hardware/hardware.py::_get_parent_visible_gpu_spec:
            # treat an empty mask (HIP_VISIBLE_DEVICES="") as "no GPUs" rather
            # than falling through. ``or`` would coerce "" to the wrong source.
            # Empty mask (CVD="") yields an empty list -> no GPUs, consistent
            # with the nvidia-smi path.
            physical_ids = LlamaCppBackend._resolve_visible_physical_ids()
            gpus = []
            for ordinal in range(torch.cuda.device_count()):
                free_bytes, total_bytes = torch.cuda.mem_get_info(ordinal)
                idx = (
                    physical_ids[ordinal]
                    if physical_ids is not None and ordinal < len(physical_ids)
                    else ordinal
                )
                gpus.append((idx, free_bytes // (1024 * 1024), total_bytes // (1024 * 1024)))
            # Match the nvidia-smi path's docstring guarantee of sorted-by-id.
            return sorted(gpus, key = lambda g: g[0])
        except Exception as e:
            logger.debug(f"torch GPU probe failed: {e}")
            return []

    @staticmethod
    def _available_system_memory_mib() -> Optional[int]:
        """Available system RAM in MiB (psutil, then /proc/meminfo), or None if
        neither is readable. On a unified-memory APU this, not the ROCm-reported
        VRAM, is the real ceiling: the weights load into shared system RAM."""
        try:
            import psutil
            return int(psutil.virtual_memory().available // (1024 * 1024))
        except Exception:
            pass
        try:
            with open("/proc/meminfo") as f:
                for line in f:
                    if line.startswith("MemAvailable:"):
                        return int(line.split()[1]) // 1024  # kB -> MiB
        except Exception:
            pass
        return None

    @staticmethod
    def _apu_ram_shortfall_message(
        model_size_bytes: int,
        avail_mib: Optional[int],
        headroom_mib: int = 2048,
    ) -> Optional[str]:
        """On a unified-memory APU, return a user-facing refusal when the weights
        cannot fit in available system RAM (else None). Weights only: KV/context
        auto-reduce, so counting them too would refuse loads that would succeed.
        None avail (unknown RAM) never refuses."""
        if avail_mib is None:
            return None
        need_mib = model_size_bytes / (1024 * 1024)
        if need_mib <= avail_mib - headroom_mib:
            return None
        return (
            f"This model needs about {need_mib / 1024:.0f} GB but only about "
            f"{avail_mib / 1024:.0f} GB of memory is available. On a unified-memory "
            "APU the weights load into system RAM, so a larger model is stopped by "
            "the OS mid-load. Use a smaller or more quantized GGUF, or free memory "
            "(on WSL, raise the memory limit in .wslconfig)."
        )

    # Skip the wait when the last kill is older than this; the driver has
    # already reclaimed the prior process's allocations.
    _VRAM_SETTLE_WINDOW_S: float = 15.0

    @staticmethod
    def _wait_for_vram_settle(
        max_wait: float = 2.0,
        interval: float = 0.25,
        tolerance_mib: int = 256,
        since_kill: float = 0.0,
    ) -> None:
        """Poll ``_get_gpu_free_memory`` until free VRAM stabilises.

        The driver reclaims a dead process's allocations asynchronously, so
        sampling free memory in the kill-to-spawn window reads artificially low
        and pushes GPU selection toward needless CPU offload (the Apply-reload
        OOM bare-shell launches never see).

        Short-circuits on cold start, stale kill (older than
        ``_VRAM_SETTLE_WINDOW_S``), CPU-only hosts, probe exceptions, and GPU-set
        changes. ``max_wait`` bounds wall-clock time so a wedged ``nvidia-smi``
        can't extend the reload.
        """
        now = time.monotonic()
        if since_kill <= 0.0:
            return
        if now - since_kill > LlamaCppBackend._VRAM_SETTLE_WINDOW_S:
            return
        deadline = now + max_wait

        def _probe_or_none():
            if time.monotonic() >= deadline:
                return None
            try:
                return LlamaCppBackend._get_gpu_free_memory()
            except Exception:
                return None

        prev = _probe_or_none()
        if prev is None or not prev:
            return
        while time.monotonic() < deadline:
            remaining = deadline - time.monotonic()
            if remaining <= 0:
                return
            # Clip the nap so a near-zero ``max_wait`` is respected.
            time.sleep(min(interval, remaining))
            curr = _probe_or_none()
            if curr is None or not curr or len(curr) != len(prev):
                return
            prev_map = dict(prev)
            stable = True
            for idx, free in curr:
                if idx not in prev_map:
                    stable = False
                    break
                prev_free = prev_map[idx]
                # Adaptive: 2% of the larger sample dominates the 256 MiB floor.
                per_gpu_tol = max(tolerance_mib, int(max(free, prev_free) * 0.02))
                if abs(free - prev_free) >= per_gpu_tol:
                    stable = False
                    break
            if stable:
                return
            prev = curr

    # Free-VRAM fraction at which Studio pins the GPU directly instead of
    # deferring to ``--fit on``. 3% headroom: the compute buffer is now modelled in
    # the fit, so this only guards fragmentation + multi-GPU per-device CUDA context
    # (~2-3%); kept >= 3% as a floor (0.90 dropped 91-94% fits to CPU offload, #5106).
    _GPU_PIN_VRAM_FRACTION = 0.97

    # Fallback per-device tensor-mode compute buffer (MiB), used only when GGUF
    # dims are unavailable so _estimate_compute_buffer_bytes (the primary, derived
    # path) returns 0.
    _TENSOR_PARALLEL_BUFFER_RESERVE_MIB = 5120

    # Fixed per-device overhead on every GPU of a LAYER split (CUDA context +
    # scratch), beyond the conserved slot-scaling buffer. ~0.9 GB/device measured
    # (Qwen3.6-27B, b9625), independent of --parallel; reserved per extra GPU so a
    # tight layer split can't advertise a context that OOMs at load.
    _PIPELINE_PER_DEVICE_OVERHEAD_MIB = 1024

    # KV cache types llama.cpp accepts in tensor mode. A quantized KV cache
    # aborts a --split-mode tensor load, so it's dropped for the tensor attempt.
    _TENSOR_PARALLEL_KV_TYPES = frozenset({"f16", "bf16", "f32"})

    # (binary, mtime, model) that aborted on --split-mode tensor this process (#6415
    # geometry limit, e.g. MQA n_head_kv=1). Model-keyed so one model's abort doesn't
    # skip tensor for others; tensor is tried by default, recorded only on a real abort.
    _tensor_split_abort_keys: set[tuple[str, int, str]] = set()

    @classmethod
    def _tensor_split_cache_key(
        cls, binary: Optional[str], model: Optional[str]
    ) -> Optional[tuple[str, int, str]]:
        """(path, mtime_ns, model) key; ns mtime re-probes a same-second binary swap."""
        if not binary or not model:
            return None
        try:
            mtime = Path(binary).stat().st_mtime_ns
        except OSError:
            mtime = 0
        return (binary, mtime, model)

    @classmethod
    def _tensor_split_aborts(cls, binary: Optional[str], model: Optional[str]) -> bool:
        """True if (binary, model) aborted on --split-mode tensor this session."""
        key = cls._tensor_split_cache_key(binary, model)
        return key is not None and key in cls._tensor_split_abort_keys

    @classmethod
    def _record_tensor_split_abort(cls, binary: Optional[str], model: Optional[str]) -> None:
        """Remember a (binary, model) that aborts on --split-mode tensor."""
        key = cls._tensor_split_cache_key(binary, model)
        if key is not None:
            cls._tensor_split_abort_keys.add(key)

    @staticmethod
    def _windows_pip_nvidia_dll_dirs(prefix: str) -> list[str]:
        """Return DLL dirs from pip-installed CUDA wheels under
        ``<prefix>/Lib/site-packages/`` so llama-server.exe can load
        ``cudart64_X.dll`` / ``cublas64_X.dll`` without a system CUDA toolkit.
        Mirrors the Linux ``nvidia/cu*/lib`` LD_LIBRARY_PATH block, covering the
        Windows wheel layouts seen in the wild:
          * ``nvidia/<pkg>/bin`` -- legacy modular wheels.
          * ``nvidia/<pkg>/bin/x86_64`` and ``.../bin/x64`` -- CUDA 13 layout
            for unsuffixed packages (#5106).
          * ``nvidia/<pkg>/Library/bin`` (and arch subdirs) -- conda repacks.
          * ``torch/lib`` -- PyTorch's CUDA-bundled wheel can ship
            ``cudart64_*.dll`` here; mirrors install_llama_prebuilt.py.

        Walks with ``Path.iterdir`` not ``glob.glob`` so it's safe against
        Windows paths containing ``[`` or ``]`` (valid in usernames)."""
        site_packages = Path(prefix) / "Lib" / "site-packages"
        out: list[str] = []
        seen: set[str] = set()

        def _add(path: Path) -> None:
            if not path.is_dir():
                return
            key = os.path.normcase(os.path.abspath(str(path)))
            if key in seen:
                return
            seen.add(key)
            out.append(str(path))

        nvidia_root = site_packages / "nvidia"
        if nvidia_root.is_dir():
            for pkg_dir in nvidia_root.iterdir():
                if not pkg_dir.is_dir():
                    continue
                # Arch-specific subdirs first so the explicit cudart64_X.dll
                # location wins over an empty sibling ``bin``.
                for sub in (
                    pkg_dir / "bin" / "x86_64",
                    pkg_dir / "bin" / "x64",
                    pkg_dir / "bin",
                    pkg_dir / "Library" / "bin" / "x86_64",
                    pkg_dir / "Library" / "bin" / "x64",
                    pkg_dir / "Library" / "bin",
                ):
                    _add(sub)
        _add(site_packages / "torch" / "lib")
        return out

    @staticmethod
    def _build_windows_path_dirs(binary_dir: str, prefix: str, cuda_path: str) -> list[str]:
        """Ordered PATH entries prepended so llama-server.exe resolves cudart /
        cublas DLLs: binary_dir, pip nvidia wheels, CUDA_PATH/bin, .../bin/x64.
        Extracted so test_windows_gpu_detection_mock tests the real logic. #5106."""
        path_dirs = [binary_dir]
        path_dirs.extend(LlamaCppBackend._windows_pip_nvidia_dll_dirs(prefix))
        if cuda_path:
            cuda_bin = os.path.join(cuda_path, "bin")
            if os.path.isdir(cuda_bin):
                path_dirs.append(cuda_bin)
            cuda_bin_x64 = os.path.join(cuda_path, "bin", "x64")
            if os.path.isdir(cuda_bin_x64):
                path_dirs.append(cuda_bin_x64)
        return path_dirs

    @staticmethod
    def _llama_server_env_for_binary(binary: str) -> dict[str, str]:
        """Build a subprocess env that lets llama-server resolve native libs."""
        env = child_env_without_native_path_secret()
        binary_dir = str(Path(binary).parent)

        if sys.platform == "win32":
            # Ordering: see _build_windows_path_dirs. #5106.
            path_dirs = LlamaCppBackend._build_windows_path_dirs(
                binary_dir,
                sys.prefix,
                os.environ.get("CUDA_PATH", ""),
            )
            existing_path = env.get("PATH", "")
            env["PATH"] = ";".join(path_dirs) + ";" + existing_path

            # ROCm: the prebuilt bundles rocblas.dll but NOT the Tensile
            # kernel files (rocblas/library/*.dat + *.hsaco); the DLL searches
            # <binary_dir>/rocblas/library/ which doesn't exist.
            _hip_path = os.environ.get("HIP_PATH", os.environ.get("ROCM_PATH", ""))
            if _hip_path:
                _rocblas_lib = os.path.join(_hip_path, "bin", "rocblas", "library")
                if os.path.isdir(_rocblas_lib):
                    env.setdefault("ROCBLAS_TENSILE_LIBPATH", _rocblas_lib)
        else:
            # Linux: LD_LIBRARY_PATH for shared libs next to the binary plus
            # CUDA runtime libs (libcudart, libcublas, etc.)
            import platform

            lib_dirs = []
            # WSL: system HIP before the bundle's (which segfaults on /dev/dxg).
            lib_dirs.extend(_wsl_system_rocm_lib_dirs())
            if lib_dirs:
                env.setdefault("HSA_ENABLE_DXG_DETECTION", "1")
            lib_dirs.append(binary_dir)
            _arch = platform.machine()  # x86_64, aarch64, etc.

            # Pip-installed nvidia CUDA runtime libs. The prebuilt binary links
            # libcudart.so.13 / libcublas.so.13 which live here, not in
            # /usr/local/cuda.
            import glob as _glob

            for _nv_pattern in [
                os.path.join(sys.prefix, "lib", "python*", "site-packages", "nvidia", _sub, "lib")
                for _sub in ("cu*", "cudnn", "nvjitlink")
            ]:
                for _nv_dir in _glob.glob(_nv_pattern):
                    if os.path.isdir(_nv_dir):
                        lib_dirs.append(_nv_dir)

            for cuda_lib in [
                "/usr/local/cuda/lib64",
                f"/usr/local/cuda/targets/{_arch}-linux/lib",
                # Fallback CUDA compat paths (e.g. binary built with CUDA 12
                # where default /usr/local/cuda is CUDA 13+).
                "/usr/local/cuda-12/lib64",
                "/usr/local/cuda-12.8/lib64",
                f"/usr/local/cuda-12/targets/{_arch}-linux/lib",
                f"/usr/local/cuda-12.8/targets/{_arch}-linux/lib",
            ]:
                if os.path.isdir(cuda_lib):
                    lib_dirs.append(cuda_lib)
            existing_ld = env.get("LD_LIBRARY_PATH", "")
            new_ld = ":".join(lib_dirs)
            env["LD_LIBRARY_PATH"] = f"{new_ld}:{existing_ld}" if existing_ld else new_ld

        return env

    @staticmethod
    def _select_gpus(
        model_size_bytes: int,
        gpus: list[tuple[int, int]],
        usable_fraction: Optional[float] = None,
        total_by_idx: Optional[dict[int, int]] = None,
        per_device_overhead_bytes: int = 0,
        min_gpus: int = 1,
    ) -> tuple[Optional[list[int]], bool]:
        """Pick GPU(s) for a model from estimated VRAM and free memory.

        ``min_gpus`` (default 1, capped at ``len(gpus)``) keeps a downgraded
        tensor/multi-GPU request spread instead of collapsing to one card.

        ``model_size_bytes`` should include weights and estimated KV cache.
        ``usable_fraction`` (default ``_GPU_PIN_VRAM_FRACTION``) provides
        headroom for compute buffers, CUDA context, and other runtime
        overhead; callers lower it when MTP reserves VRAM for a draft model.
        ``total_by_idx`` (index -> total MiB) makes the headroom an ABSOLUTE
        ``(1 - fraction) * total`` per GPU instead of a fraction of free.
        ``per_device_overhead_bytes`` is the fixed layer-split cost per GPU beyond
        the first; a k-GPU pin must hold ``model + (k-1) * overhead`` or it can OOM
        a device after -ngl -1 (no --fit fallback). Single-GPU adds none.

        Returns (gpu_indices, use_fit):
          - ([1], False)       fits on 1 GPU at the headroom threshold
          - ([1, 2], False)    needs 2 GPUs
          - (None, True)       too large, let --fit handle it
        """
        if not gpus:
            return None, True

        min_gpus = max(1, min(min_gpus, len(gpus)))
        model_size_mib = model_size_bytes / (1024 * 1024)
        if usable_fraction is None:
            usable_fraction = LlamaCppBackend._GPU_PIN_VRAM_FRACTION
        overhead_mib = per_device_overhead_bytes / (1024 * 1024)

        # Per-GPU usable budget: free - (1-frac)*total when total is known, else
        # the legacy free*frac (also covers a total-0 two-column probe).
        def _usable(idx: int, free_mib: int) -> float:
            t = total_by_idx.get(idx, 0) if total_by_idx else 0
            if t > 0:
                return max(0.0, free_mib - (1.0 - usable_fraction) * t)
            return free_mib * usable_fraction

        # Rank by usable budget (free - reserve), not raw free: a more-used large
        # card can have less usable room than a less-used small one.
        ranked = sorted(gpus, key = lambda g: _usable(g[0], g[1]), reverse = True)

        # Cap a downgraded multi-GPU request to the usable count so it doesn't pull
        # in a near-full card to hit min_gpus. No-op for the default min_gpus == 1.
        usable_count = sum(1 for idx, free_mib in ranked if _usable(idx, free_mib) > overhead_mib)
        min_gpus = max(1, min(min_gpus, usable_count or 1))

        # Try 1 GPU at the usable-VRAM threshold (only when one device is allowed).
        if min_gpus <= 1 and _usable(ranked[0][0], ranked[0][1]) >= model_size_mib:
            return [ranked[0][0]], False

        # Try N GPUs (most-free first); each past the first adds per-device overhead.
        # Require at least min_gpus devices before accepting a fit.
        cumulative = 0.0
        selected = []
        for idx, free_mib in ranked:
            selected.append(idx)
            cumulative += _usable(idx, free_mib)
            if (
                len(selected) >= min_gpus
                and cumulative >= model_size_mib + (len(selected) - 1) * overhead_mib
            ):
                return sorted(selected), False

        # Too large even for all GPUs; let --fit handle it
        logger.debug(
            "Model does not fit in available GPU memory, falling back to --fit",
            model_size_mib = round(model_size_mib, 2),
            ranked_gpus = ranked,
        )
        return None, True

    # ── KV cache VRAM estimation ─────────────────────────────────────

    def _can_estimate_kv(self) -> bool:
        """True if we have enough GGUF metadata to estimate KV cache size."""
        if self._n_layers is None:
            return False
        # MLA: kv_lora_rank suffices (K-only cache).
        if self._kv_lora_rank is not None:
            return True
        # New-style: need explicit key AND value dimensions.
        if self._kv_key_length is not None and self._kv_value_length is not None:
            return True
        # Legacy: need embedding_length + a head count (scalar or per-layer).
        return self._embedding_length is not None and (
            self._n_kv_heads is not None
            or self._n_heads is not None
            or self._n_kv_heads_by_layer is not None
        )

    def _kv_heads_for_layer(self, layer_idx: int, fallback: int) -> int:
        if self._n_kv_heads_by_layer is not None and layer_idx < len(self._n_kv_heads_by_layer):
            return self._n_kv_heads_by_layer[layer_idx]
        return fallback

    def _legacy_head_dim(self) -> int:
        """Head-dim fallback for GGUFs without explicit key/value dims. Reached
        only via the legacy branch of _can_estimate_kv(), so _embedding_length
        is non-None here."""
        return self._embedding_length // self._n_heads if self._n_heads else 128  # type: ignore[operator]

    def _estimate_kv_cache_bytes(
        self,
        n_ctx: int,
        cache_type_kv: Optional[str] = None,
        *,
        swa_full: bool = False,
        n_parallel: int = 1,
        kv_unified: bool = True,
        ctx_checkpoints: int = 0,
    ) -> int:
        """Estimate KV cache VRAM for a given context length.

        5-path architecture-aware estimation:
          1. MLA      -- compressed KV latent + RoPE, K-only (no separate V)
          2. Hybrid   -- only attention layers need KV (Mamba layers don't)
          3. SWA      -- sliding-window layers cache min(ctx, window) tokens
          4. GQA      -- standard full KV with explicit key/value dimensions
          5. Legacy   -- fallback using embed // n_heads

        Server-flag knobs (mirror llama-server's CLI):
          swa_full        -- --swa-full: SWA layers cache full n_ctx (path 3->4).
          n_parallel      -- --parallel slots: non-SWA constant, SWA scale linearly.
          kv_unified      -- --kv-unified: memory no-op (API forward-compat).
          ctx_checkpoints -- --ctx-checkpoints: N SWA snapshots per slot.

        Returns 0 if metadata is insufficient.
        """
        if not self._can_estimate_kv() or n_ctx <= 0:
            return 0

        n_layers = self._n_layers  # type: ignore[assignment]
        # Gemma 3n / Gemma 4 reuse earlier KV in the last ``shared_kv_layers``
        # blocks (no cache). Floor at 1 so a bad GGUF can't zero out KV.
        shared = self._shared_kv_layers or 0
        n_layers_kv = max(1, n_layers - shared)
        n_kv = self._n_kv_heads or self._n_heads or 1  # type: ignore[assignment]

        # Bytes per element depends on KV cache quantization
        bpe = _kv_bytes_per_elem(cache_type_kv)

        slots = max(1, n_parallel)

        # Path 1: MLA (DeepSeek-V2/V3, GLM-4.7, GLM-5, Kimi-K2.5)
        # One compressed KV latent per token/layer (shared across heads); V is
        # reconstructed from it, no separate V cache. key_length = kv_lora_rank
        # + rope_dim. MLA GGUFs set head_count_kv=1; default to 1 if absent to
        # avoid falling back to n_heads (e.g. 128 for DeepSeek-V3) which 128x's.
        if self._kv_lora_rank is not None:
            n_kv_mla = self._n_kv_heads or 1
            rope_dim = self._key_length_mla or 64
            key_len = self._kv_key_length or (self._kv_lora_rank + rope_dim)
            return int(n_layers_kv * n_ctx * n_kv_mla * key_len * bpe)

        key_len = self._kv_key_length
        val_len = self._kv_value_length

        # Path 2: Hybrid Mamba/Attention (Qwen3.5-27B, Qwen3.5-35B-A3B)
        # Only 1 in N layers is attention; the rest are Mamba (no KV cache).
        if self._ssm_inner_size is not None and self._full_attention_interval is not None:
            fai = self._full_attention_interval
            n_attn = -(-n_layers // fai) if fai > 0 else n_layers  # ceiling division
            if key_len is not None and val_len is not None:
                return int(n_attn * n_ctx * n_kv * (key_len + val_len) * bpe)
            head_dim = self._legacy_head_dim()
            return int(n_attn * n_ctx * n_kv * 2 * head_dim * bpe)

        # Path 3: Sliding window (Gemma 2/3/3n/4, gpt-oss, Cohere2 ...). Pattern
        # from the resolver; if absent, falls through to the legacy 1/4-global
        # heuristic. --parallel N accounting (verified against llama-server):
        # non-SWA cells = n_ctx split across slots (CONSTANT); SWA per-slot cells
        # = 2*sliding_window (capped at n_ctx/per_slot_ctx) -> LINEAR in slots.
        # --swa-full forces full n_ctx for SWA; --ctx-checkpoints N adds snapshots.
        if (
            self._sliding_window is not None
            and self._sliding_window > 0
            and key_len is not None
            and val_len is not None
        ):
            swa = self._sliding_window
            per_slot_ctx = max(1, n_ctx // slots)
            # --swa-full caches full per_slot_ctx (constant n_ctx total); else SWA
            # caches 2*sliding_window per slot, clamped at per-slot ctx.
            swa_cells_per_slot = per_slot_ctx if swa_full else min(n_ctx, 2 * swa, per_slot_ctx)
            key_len_swa = self._kv_key_length_swa or key_len
            val_len_swa = self._kv_value_length_swa or val_len
            if self._sliding_window_pattern is not None:
                global_bytes = 0.0  # constant across slots
                swa_bytes_per_slot = 0.0  # multiplied by slots
                checkpoint_extra_per_slot = 0.0
                # Only layers that allocate their own KV; trailing shared layers
                # reuse earlier caches.
                for layer_idx in range(n_layers_kv):
                    layer_n_kv = self._kv_heads_for_layer(layer_idx, n_kv)
                    is_swa = (
                        layer_idx < len(self._sliding_window_pattern)
                        and self._sliding_window_pattern[layer_idx]
                    )
                    if is_swa:
                        swa_bytes_per_slot += (
                            swa_cells_per_slot * layer_n_kv * (key_len_swa + val_len_swa) * bpe
                        )
                        if ctx_checkpoints > 0 and not swa_full:
                            checkpoint_extra_per_slot += (
                                ctx_checkpoints
                                * swa
                                * layer_n_kv
                                * (key_len_swa + val_len_swa)
                                * bpe
                            )
                    else:
                        global_bytes += n_ctx * layer_n_kv * (key_len + val_len) * bpe
                return int(global_bytes + slots * (swa_bytes_per_slot + checkpoint_extra_per_slot))
            n_global = max(1, n_layers_kv // 4)
            n_swa = n_layers_kv - n_global
            kv_per_token = n_kv * (key_len + val_len) * bpe
            kv_per_token_swa = n_kv * (key_len_swa + val_len_swa) * bpe
            global_bytes = n_global * n_ctx * kv_per_token
            swa_bytes_per_slot = n_swa * swa_cells_per_slot * kv_per_token_swa
            checkpoint_extra_per_slot = (
                ctx_checkpoints * n_swa * swa * kv_per_token_swa
                if ctx_checkpoints > 0 and not swa_full
                else 0.0
            )
            return int(global_bytes + slots * (swa_bytes_per_slot + checkpoint_extra_per_slot))

        # Path 4: Standard GQA with explicit key/value dimensions
        if key_len is not None and val_len is not None:
            return int(n_layers_kv * n_ctx * n_kv * (key_len + val_len) * bpe)

        # Path 5: Legacy fallback (old GGUFs without explicit dimensions)
        head_dim = self._legacy_head_dim()
        return int(2 * n_kv * head_dim * n_layers_kv * n_ctx * bpe)

    def _draft_backend_for(self, drafter_path: str) -> Optional["LlamaCppBackend"]:
        """Lightweight backend with a drafter GGUF's metadata, to size its own KV
        via _estimate_kv_cache_bytes. Cached per path; None if unreadable."""
        cache = getattr(self, "_draft_backend_cache", None)
        if cache is not None and cache[0] == drafter_path:
            return cache[1]
        db: Optional[LlamaCppBackend] = None
        try:
            db = LlamaCppBackend.__new__(LlamaCppBackend)
            for attr in (
                "_context_length",
                "_n_layers",
                "_n_kv_heads",
                "_n_heads",
                "_embedding_length",
                "_kv_key_length",
                "_kv_value_length",
                "_kv_lora_rank",
                "_sliding_window",
                "_sliding_window_pattern",
                "_ssm_inner_size",
                "_full_attention_interval",
                "_key_length_mla",
                "_n_kv_heads_by_layer",
                "_kv_key_length_swa",
                "_kv_value_length_swa",
                "_shared_kv_layers",
                "_nextn_predict_layers",
            ):
                setattr(db, attr, None)
            db._model_identifier = "mtp-draft"
            db._read_gguf_metadata(drafter_path)
        except Exception as e:  # unreadable drafter -> caller falls back
            logger.debug(f"Could not read drafter GGUF for MTP budget: {e}")
            db = None
        self._draft_backend_cache = (drafter_path, db)
        return db

    def _mtp_draft_kv_bytes(
        self,
        n_ctx: int,
        *,
        drafter_path: Optional[str] = None,
        draft_cache_type_k: Optional[str] = None,
        draft_cache_type_v: Optional[str] = None,
        n_parallel: int = 1,
    ) -> Optional[int]:
        """Draft KV cache bytes at n_ctx, sized from GGUF dims (K and V types are
        independent). Separate drafter (Gemma): its own KV via _estimate_kv_cache_bytes
        at the heavier type. Embedded head (Qwen): nextn_predict_layers attention
        layers from the main dims. None when dims are missing (flat fallback)."""
        if n_ctx <= 0:
            return None
        bpe_k = _kv_bytes_per_elem(draft_cache_type_k)
        bpe_v = _kv_bytes_per_elem(draft_cache_type_v)
        if drafter_path:
            db = self._draft_backend_for(drafter_path)
            if db is None or not db._can_estimate_kv():
                return None
            heavier = draft_cache_type_k if bpe_k >= bpe_v else draft_cache_type_v
            # The drafter is served under the same --parallel slot count as the
            # main model, so price its KV per slot too: a sliding-window drafter
            # (Gemma) grows KV with slots and would otherwise be under-reserved.
            kv = db._estimate_kv_cache_bytes(n_ctx, heavier, n_parallel = n_parallel)
            return kv or None
        nextn = self._nextn_predict_layers or 0
        n_kv = self._n_kv_heads or self._n_heads
        k_len = self._kv_key_length
        v_len = self._kv_value_length
        if not (nextn and n_kv and k_len and v_len):
            return None
        # The embedded MTP head is one draft layer, so a quantized draft KV can't
        # amortize its overhead and fits *less* context than f16 (llama.cpp#24102).
        # Floor it at f16: a quantized override is priced as f16, f32 keeps its 4
        # bytes. The separate-drafter branch is multi-layer, so it keeps its type.
        f16_bpe = _kv_bytes_per_elem("f16")
        bpe_k = max(bpe_k, f16_bpe)
        bpe_v = max(bpe_v, f16_bpe)
        return int(nextn * n_kv * (k_len * bpe_k + v_len * bpe_v) * n_ctx)

    def _estimate_mtp_overhead_bytes(
        self,
        n_ctx: int,
        *,
        spec_draft_n_max: int = 0,
        draft_cache_type_k: Optional[str] = None,
        draft_cache_type_v: Optional[str] = None,
        drafter_path: Optional[str] = None,
        draft_weights_bytes: int = 0,
        n_parallel: int = 1,
        mtp_keeps_target_ctx: bool = True,
    ) -> Optional[int]:
        """MTP draft reserve at ``n_ctx`` = draft KV (grows with ctx) + separate-
        drafter weights + (MTP + MLA only) a duplicated target KV context. The
        verify buffer rides in the ctx-fit headroom (no tuned constant). None when
        the draft KV can't be sized (caller keeps the flat fallback).
        ``draft_weights_bytes`` is the drafter file size (0 for embedded).
        ``mtp_keeps_target_ctx`` is True for MTP draft modes (which keep the
        duplicated target context) and False for separate-drafter spec modes
        (draft-simple/draft-eagle3), which do not."""
        draft_kv = self._mtp_draft_kv_bytes(
            n_ctx,
            drafter_path = drafter_path,
            draft_cache_type_k = draft_cache_type_k,
            draft_cache_type_v = draft_cache_type_v,
            n_parallel = n_parallel,
        )
        weights = max(0, draft_weights_bytes)
        # MLA models (GLM-5.x, DeepSeek, Kimi-K2) under MTP keep a *second* full copy
        # of the target model's KV context for draft verification -- llama.cpp's
        # `ctx_tgt=yes` -- allocated at f16 regardless of the main cache type. It is
        # ~the main KV again and dwarfs the embedded draft head (GLM-5.2 @ 1M ctx:
        # a ~2 GiB head next to a ~89 GiB target copy), so omitting it lets auto-fit
        # pick a context that fits on paper but OOMs cublasCreate at the first
        # decode. Gated on both MLA (kv_lora_rank present) and the engaged mode
        # actually being MTP: non-MLA MTP (Qwen/Gemma) keeps no such copy, and the
        # separate-drafter spec modes (draft-simple/draft-eagle3) load a small
        # distinct drafter with its own KV -- already counted in draft_kv/weights --
        # rather than duplicating the target, so they must not be charged for it.
        target_ctx_copy = 0
        if mtp_keeps_target_ctx and self._kv_lora_rank is not None:
            target_ctx_copy = self._estimate_kv_cache_bytes(n_ctx, "f16", n_parallel = n_parallel)
        if draft_kv is None:
            # KV unsized (exotic/remote drafter): still reserve known weights + any
            # MLA target copy so a large config can't launch over budget (the small
            # unsized draft KV rides in the cushion). Nothing known -> None, so the
            # caller keeps the flat fallback.
            total = weights + target_ctx_copy
            return total if total > 0 else None
        return draft_kv + weights + target_ctx_copy

    _DEFAULT_N_UBATCH = 512  # llama.cpp --ubatch default; Studio does not override it
    _COMPUTE_BUFFER_SAFETY = 1.15  # upper-bound margin on the compute-buffer estimate
    # Soft VRAM the modeled terms omit; charged to the fit budget on tight tiers (#6682).
    _CUDA_CONTEXT_RESERVE_BYTES = 320 * 1024 * 1024  # CUDA ctx + cuBLAS workspace (~330 MiB)
    _MMPROJ_VRAM_SAFETY = 1.4  # mmproj worst-case buffer vs file size (runtime ~1.3x)
    _MTP_DRAFT_COMPUTE_BYTES = 224 * 1024 * 1024  # MTP draft decode graph beyond its KV
    # The flash-attn KQ mask + attention scratch grow ~linearly with context; the flat
    # _estimate_compute_buffer_bytes term only covers ctx -> 0. The per-token rate
    # depends on the KV cache type: a QUANTIZED cache (q8_0/q5/q4/iq4) needs a
    # context-sized dequant scratch that scales with n_embd, measured at 0.74-2.02 x
    # n_embd across Qwen3.5/3.6 (2B/4B/9B/27B) and Gemma-4 (12B/31B) at q8_0; an
    # f16/bf16/f32 cache skips the dequant and pays only the KQ mask, a flat n_ubatch*2
    # bytes per context token regardless of n_embd (measured 1024 B/tok on Qwen-9B and
    # Gemma-31B alike). So Qwen3.5-4B at 256k is 1.30 GiB at q8_0 vs 0.31 GiB at f16.
    # 2.25 covers the worst quantized case (Qwen3.5-4B, ~2.0x) plus the under-modeled
    # flat base; the mask safety covers the f16 base gap. Without this term, tight tiers
    # at extreme context over-pin and spill to CPU (the 3% cushion is only ~0.25 GiB on
    # an 8 GB card, far below the ~1-2.4 GiB quantized buffer at 256k): e.g. Qwen3.5-4B
    # Q4 at 256k needs ~8.5 GiB on a real 8 GB card (weights 2.4 + KV 4.3 + compute 1.3
    # + CUDA ctx) -> CPU spill; with this reserve the auto context caps to ~210k, fits.
    _CTX_COMPUTE_BYTES_PER_EMBD = 2.25  # quantized KV, regular attention (dequant scratch)
    _CTX_COMPUTE_BYTES_PER_EMBD_MLA = 1.25  # quantized KV, MLA (compressed attn: measured 0.94x)
    _CTX_COMPUTE_F16_MASK_SAFETY = 1.5  # f16/bf16/f32 KV: KQ mask only (n_ubatch*2 B/tok)
    # DeepSeek-V4 (deepseek4): its lightning indexer + sparse attention reserve a large
    # context-scaling compute buffer the rates above miss (present even with an f16
    # cache). Measured on UD-Q4_K_XL (ub=512): ~2 GiB at 16k -> ~65.5 GiB at 1M. Without
    # it auto-fit commits the full 1M train context, OOMs the reserve, and spills to CPU.
    _DSV4_CTX_COMPUTE_FLAT_BYTES = 2 * 1024**3  # ctx-independent indexer scratch
    _DSV4_CTX_COMPUTE_BYTES_PER_TOK = 72000  # per token at ub=512 (~72 GiB at 1M)

    def _estimate_compute_buffer_bytes(
        self,
        *,
        n_ubatch: Optional[int] = None,
        n_parallel: int = 1,
        per_device_tensor: bool = False,
    ) -> int:
        """Per-device compute-graph buffer (bytes) from GGUF dims: a vocab-width
        output buffer + activation scratch. Context-independent; scales with
        ``--parallel`` (serving slots). Tensor mode materializes it on every device.
        A slight upper bound over measured allocations; 0 when dims are missing."""
        n_vocab = self._vocab_size or 0
        n_embd = self._embedding_length or 0
        if n_vocab <= 0 or n_embd <= 0:
            return 0
        ub = max(1, int(n_ubatch if n_ubatch else self._DEFAULT_N_UBATCH))
        par = max(1, int(n_parallel))
        out_buffer = n_vocab * ub * 4  # f32 output/logits buffer
        act_scratch = 4 * n_embd * ub * 4  # a few resident hidden-width buffers
        if per_device_tensor:
            # Output + comm/staging materialized on every device, every slot.
            compute = 2 * act_scratch + out_buffer * par
        else:
            # Each extra concurrent slot adds one output buffer (chat decode sizes
            # ~one logit row per slot; would under-count embeddings/--logits-all,
            # not run here). Matches measured {1:36,2:492,4:1388,8:3220} MiB.
            compute = act_scratch + out_buffer * max(0, par - 1)
        return int(compute * self._COMPUTE_BUFFER_SAFETY)

    def _compute_buffer_ctx_bytes(
        self,
        n_ctx: int,
        n_ubatch: Optional[int] = None,
        cache_type_kv: Optional[str] = None,
    ) -> int:
        """Context-linear growth of the per-device compute buffer (bytes), charged
        on top of the flat ``_estimate_compute_buffer_bytes``. The flash-attn KQ
        mask + attention scratch scale ~linearly with context and with the micro-
        batch; the flat term only covers ctx -> 0. A quantized KV cache adds a
        context-sized dequant scratch that scales with n_embd; f16/bf16/f32 pays only
        the KQ mask, a flat n_ubatch*2 bytes per context token. ``cache_type_kv`` None
        -> f16 (llama.cpp's default; an env-set quantized cache is budgeted as f16 on
        the KV side, whose over-reservation absorbs the dequant scratch). Returns 0
        when dims are missing or ``n_ctx`` <= 0."""
        n_embd = self._embedding_length or 0
        if n_embd <= 0 or n_ctx <= 0:
            return 0
        ub = max(1, int(n_ubatch if n_ubatch else self._DEFAULT_N_UBATCH))
        if getattr(self, "_architecture", None) == "deepseek4":
            # DSV4 indexer/CSA buffer (see constants): flat + linear, ub-scaled. Fires
            # for any KV type -- the indexer scratch is present even with an f16 cache.
            ub_scale = ub / self._DEFAULT_N_UBATCH
            return int(
                self._DSV4_CTX_COMPUTE_FLAT_BYTES
                + self._DSV4_CTX_COMPUTE_BYTES_PER_TOK * n_ctx * ub_scale
            )
        if _kv_bytes_per_elem(cache_type_kv) < 2.0:
            # Quantized cache: the dequant scratch dominates and scales with n_embd.
            # MLA (compressed KV) needs far less of it: measured 0.94 x n_embd on
            # GLM-5.2 and Kimi-K2.7 vs up to 2.02x on regular attention.
            ub_scale = ub / self._DEFAULT_N_UBATCH
            rate = (
                self._CTX_COMPUTE_BYTES_PER_EMBD_MLA
                if self._key_length_mla
                else self._CTX_COMPUTE_BYTES_PER_EMBD
            )
            per_tok = rate * n_embd * ub_scale
        else:
            # f16/bf16/f32: only the KQ mask ([n_kv, n_ubatch] f16), n_embd-independent.
            per_tok = ub * 2 * self._CTX_COMPUTE_F16_MASK_SAFETY
        return int(per_tok * n_ctx)

    def _slots_that_fit_on_gpu(
        self,
        n_parallel: int,
        effective_ctx: int,
        gpus: list[tuple[int, int]],
        total_by_idx: Optional[dict[int, int]],
        base_footprint_bytes: int,
        cache_type_kv: Optional[str],
        pin_fraction: float,
        per_device_overhead_bytes: int,
        min_gpus: int,
        n_ubatch: Optional[int] = None,
    ) -> tuple[Optional[list[int]], bool, int]:
        """Largest serving-slot count in [1, n_parallel) whose fully-on-GPU footprint fits,
        so Studio keeps the model on GPU (-ngl -1) instead of --fit on, which offloads layers
        to host and collapses decode ~3x (oobabooga #6718). ``base_footprint_bytes`` is the
        slot-independent footprint (weights + soft overhead + MTP + context-linear compute,
        minus the folded compute buffer); each candidate re-adds the slot-sized compute buffer
        and KV, then re-selects GPUs like the explicit-context path. Returns (gpu_indices,
        use_fit=False, slots) for the largest fitting count, else (None, True, n_parallel).
        Only ever reduces; deterministic and unit-testable with synthetic VRAM maps."""
        for slots in range(n_parallel - 1, 0, -1):
            cb = self._estimate_compute_buffer_bytes(
                n_ubatch = n_ubatch, n_parallel = slots, per_device_tensor = False
            )
            if cb <= 0:
                cb = self._TENSOR_PARALLEL_BUFFER_RESERVE_MIB * 1024 * 1024
            total = (
                base_footprint_bytes
                + cb
                + self._estimate_kv_cache_bytes(effective_ctx, cache_type_kv, n_parallel = slots)
            )
            gpu_indices, use_fit = self._select_gpus(
                total,
                gpus,
                usable_fraction = pin_fraction,
                total_by_idx = total_by_idx,
                per_device_overhead_bytes = per_device_overhead_bytes,
                min_gpus = min_gpus,
            )
            if not use_fit:
                return gpu_indices, False, slots
        return None, True, n_parallel

    def _fit_context_to_vram(
        self,
        requested_ctx: int,
        available_mib: int,
        model_size_bytes: int,
        cache_type_kv: Optional[str] = None,
        min_ctx: int = 4096,
        *,
        swa_full: bool = False,
        n_parallel: int = 1,
        kv_unified: bool = True,
        ctx_checkpoints: int = 0,
        kv_on_gpu: bool = True,
        mtp_engaged: bool = False,
        mtp_overhead_fn: Optional[Callable[[int], int]] = None,
        compute_ctx_bytes_fn: Optional[Callable[[int], int]] = None,
        budget_frac: Optional[float] = None,
        total_mib: Optional[int] = None,
    ) -> int:
        """Return the largest context length that fits in GPU VRAM.

        Budget caps occupancy at ``_CTX_FIT_VRAM_FRACTION`` of the card: an
        absolute ``free - (1 - frac) * total`` when ``total_mib`` is given, else
        ``free * frac``. Weights alone over budget returns ``requested_ctx``.

        ``kv_on_gpu`` mirrors ``--kv-offload`` (default on); when False the KV
        cache lives in CPU RAM and the requested context is honored verbatim.
        Other keyword args mirror ``_estimate_kv_cache_bytes``.

        ``mtp_engaged`` reserves extra VRAM for the MTP draft model's KV cache +
        compute buffers, else tight tiers (e.g. 32 GB) spill to a slower path.
        """
        if not self._can_estimate_kv():
            logger.debug(
                "Skipping context fit because KV cache metadata is unavailable",
                requested_ctx = requested_ctx,
                available_mib = available_mib,
            )
            return requested_ctx

        # KV lives off-GPU: no VRAM accounting needed for the cache itself.
        if not kv_on_gpu:
            return requested_ctx

        kv_kwargs = dict(
            swa_full = swa_full,
            n_parallel = n_parallel,
            kv_unified = kv_unified,
            ctx_checkpoints = ctx_checkpoints,
        )

        # byte-accurate mtp_overhead_fn supersedes the flat fraction (the fallback
        # when dims can't size the draft KV); callers may override budget_frac.
        if budget_frac is None:
            flat_mtp = mtp_engaged and mtp_overhead_fn is None
            budget_frac = _CTX_FIT_VRAM_FRACTION - (_MTP_VRAM_RESERVE_FRAC if flat_mtp else 0.0)
        # Absolute reserve off total when known, else fraction-of-free; clamp >=0.
        if total_mib is not None and total_mib > 0:
            budget_mib = max(0.0, available_mib - (1.0 - budget_frac) * total_mib)
        else:
            budget_mib = available_mib * budget_frac
        budget_bytes = budget_mib * 1024 * 1024
        model_footprint = model_size_bytes

        def _mtp_at(ctx: int) -> int:
            return mtp_overhead_fn(ctx) if mtp_overhead_fn is not None else 0

        def _cc_at(ctx: int) -> int:
            # Context-linear compute-buffer growth (flash-attn KQ mask + scratch);
            # the flat term in model_footprint only covers ctx -> 0.
            return compute_ctx_bytes_fn(ctx) if compute_ctx_bytes_fn is not None else 0

        # Already fits?
        kv = self._estimate_kv_cache_bytes(requested_ctx, cache_type_kv, **kv_kwargs)
        if model_footprint + kv + _mtp_at(requested_ctx) + _cc_at(requested_ctx) <= budget_bytes:
            return requested_ctx

        # Weights + compute buffer alone exceed budget -- reducing ctx can't help.
        if model_footprint >= budget_bytes:
            logger.debug(
                "Model footprint exceeds GPU budget before KV cache",
                requested_ctx = requested_ctx,
                available_mib = available_mib,
                model_size_gb = round(model_footprint / (1024**3), 2),
            )
            return requested_ctx

        # Binary search for max context that fits (KV + MTP draft reserve at that ctx)
        remaining = budget_bytes - model_footprint
        effective_min = min(min_ctx, requested_ctx)
        lo, hi = effective_min, requested_ctx
        best = effective_min
        while lo <= hi:
            mid = (lo + hi) // 2
            kv = self._estimate_kv_cache_bytes(mid, cache_type_kv, **kv_kwargs)
            if kv + _mtp_at(mid) + _cc_at(mid) <= remaining:
                best = mid
                lo = mid + 1
            else:
                hi = mid - 1

        # Round down to nearest 256 for alignment, never above requested_ctx
        best = (best // 256) * 256
        best = max(effective_min, best)
        best = min(best, requested_ctx)
        return best

    # ── Variant fallback ────────────────────────────────────────────

    @staticmethod
    def _find_smallest_fitting_variant(
        hf_repo: str,
        free_bytes: int,
        hf_token: Optional[str] = None,
    ) -> Optional[tuple[str, int]]:
        """Find the smallest GGUF variant (including all shards) that fits.

        Groups split shards by variant prefix and sums their sizes (e.g.
        UD-Q4_K_XL with 9 shards of 50 GB each = 450 GB total).

        Returns (first_shard_filename, total_size_bytes) or None.
        """
        try:
            from huggingface_hub import get_paths_info, list_repo_files

            files = list_repo_files(hf_repo, token = hf_token)
            gguf_files = [
                f
                for f in files
                if f.lower().endswith(".gguf")
                and not _is_companion_gguf_path(f)
                and not _is_big_endian_gguf_path(f)
            ]
            if not gguf_files:
                return None

            # Sizes for all GGUF files
            path_infos = list(get_paths_info(hf_repo, gguf_files, token = hf_token))
            size_map = {p.path: (p.size or 0) for p in path_infos}

            # Group by variant: shards share a prefix before -NNNNN-of-NNNNN
            variants: dict[str, list[str]] = {}
            for f in gguf_files:
                m = _SHARD_RE.match(f)
                key = m.group(1) if m else f
                variants.setdefault(key, []).append(f)

            # Sum shard sizes per variant, track the first shard (for download)
            variant_sizes: list[tuple[str, int, list[str]]] = []
            for key, shard_files in variants.items():
                total = sum(size_map.get(f, 0) for f in shard_files)
                first = sorted(shard_files)[0]
                variant_sizes.append((first, total, shard_files))

            # Smallest that fits
            variant_sizes.sort(key = lambda x: x[1])
            for first_file, total_size, _ in variant_sizes:
                if total_size > 0 and total_size <= free_bytes:
                    return first_file, total_size

            return None
        except Exception:
            return None

    # ── Port allocation ───────────────────────────────────────────

    @staticmethod
    def _find_free_port() -> int:
        """Find an available TCP port."""
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
            s.bind(("127.0.0.1", 0))
            return s.getsockname()[1]

    # ── Stdout drain (prevents pipe deadlock on Windows) ─────────

    def _drain_stdout(self):
        """Read subprocess stdout lines in a background thread.

        Prevents a pipe-buffer deadlock on Windows (~4 KB buffer): without
        draining, llama-server blocks on writes and never becomes healthy.
        Each line is also teed to ``self._llama_log_fh`` when set, so a
        post-mortem has the full output even if the crash predates the
        drain-thread join in ``_wait_for_health``.
        """
        try:
            for line in self._process.stdout:
                line = line.rstrip()
                if line:
                    self._stdout_lines.append(line)
                    logger.debug(f"[llama-server] {line}")
                    fh = getattr(self, "_llama_log_fh", None)
                    if fh is not None:
                        try:
                            fh.write(line + "\n")
                            fh.flush()
                        except (ValueError, OSError):
                            # Log file closed under us; tee silently.
                            pass
        except Exception:
            # Never let the drain thread die: a full stdout pipe can deadlock
            # llama-server (Windows). Pipe-closed on exit is the common case.
            logger.debug("llama-server stdout drain stopped", exc_info = True)

    # GGUF KV type sizes for fast skipping
    _GGUF_TYPE_SIZE = {
        0: 1,
        1: 1,
        2: 2,
        3: 2,
        4: 4,
        5: 4,
        6: 4,
        7: 1,
        10: 8,
        11: 8,
        12: 8,
    }

    @staticmethod
    def _gguf_skip_value(f, vtype: int) -> None:
        """Skip a GGUF KV value without reading it."""
        sz = LlamaCppBackend._GGUF_TYPE_SIZE.get(vtype)
        if sz is not None:
            f.seek(sz, 1)
        elif vtype == 8:  # STRING
            slen = struct.unpack("<Q", f.read(8))[0]
            f.seek(slen, 1)
        elif vtype == 9:  # ARRAY
            atype = struct.unpack("<I", f.read(4))[0]
            alen = struct.unpack("<Q", f.read(8))[0]
            elem_sz = LlamaCppBackend._GGUF_TYPE_SIZE.get(atype)
            if elem_sz is not None:
                f.seek(elem_sz * alen, 1)
            elif atype == 8:
                for _ in range(alen):
                    slen = struct.unpack("<Q", f.read(8))[0]
                    f.seek(slen, 1)
            else:
                for _ in range(alen):
                    LlamaCppBackend._gguf_skip_value(f, atype)

    @staticmethod
    def _gguf_read_array_value(f, atype: int, alen: int) -> Optional[list]:
        if atype == 4:  # UINT32
            return [struct.unpack("<I", f.read(4))[0] for _ in range(alen)]
        if atype == 5:  # INT32
            return [struct.unpack("<i", f.read(4))[0] for _ in range(alen)]
        if atype == 7:  # BOOL
            return [struct.unpack("<?", f.read(1))[0] for _ in range(alen)]

        for _ in range(alen):
            LlamaCppBackend._gguf_skip_value(f, atype)
        return None

    def _read_gguf_metadata(self, gguf_path: str) -> None:
        """Read context_length, architecture params, and chat_template from a GGUF header.

        Parses only the KV pairs we need (~30ms even for multi-GB files).
        For split GGUFs, metadata is always in shard 1.
        """
        # Reset metadata so stale flags (e.g. _supports_reasoning) don't
        # carry over when switching models.
        self._context_length = None
        self._chat_template = None
        self._supports_reasoning = False
        self._reasoning_always_on = False
        self._reasoning_style = "enable_thinking"
        self._reasoning_effort_levels = []
        self._reasoning_default = True
        self._supports_preserve_thinking = False
        self._supports_tools = False
        self._n_layers = None
        self._n_kv_heads = None
        self._n_kv_heads_by_layer = None
        self._n_heads = None
        self._embedding_length = None
        self._feed_forward_length = None
        self._vocab_size = None
        self._kv_key_length = None
        self._kv_value_length = None
        self._sliding_window = None
        self._sliding_window_pattern = None
        self._full_attention_interval = None
        self._kv_lora_rank = None
        self._key_length_mla = None
        self._kv_key_length_swa = None
        self._kv_value_length_swa = None
        self._ssm_inner_size = None
        self._ssm_state_size = None
        self._shared_kv_layers = None
        self._nextn_predict_layers = None
        self._architecture = None
        self._is_diffusion = False

        try:
            canvas_seen = False
            WANTED = {
                "general.architecture",
                "tokenizer.chat_template",
                # Vocab size = tokens array length (no vocab_size key in many GGUFs).
                "tokenizer.ggml.tokens",
                # Block-diffusion marker (DiffusionGemma); routes to the diffusion runner.
                "diffusion.canvas_length",
                # Source-repo hints for the SWA resolver's HF fallback.
                "general.source.huggingface.repository",
                "general.source.url",
                "general.source.repo_url",
                "general.base_model.0.repo_url",
                "general.base_model.0.organization",
                "general.base_model.0.name",
                "general.basename",
                "general.organization",
                "general.size_label",
                "general.finetune",
            }
            # Arch-specific keys added dynamically once we know the arch.
            arch_keys: dict[str, str] = {}  # gguf_key -> attribute name
            arch = None
            sliding_window_pattern_period: Optional[int] = None
            general: dict[str, str] = {}

            with open(gguf_path, "rb") as f:
                magic = struct.unpack("<I", f.read(4))[0]
                if magic != 0x46554747:  # b"GGUF" as little-endian u32
                    return
                _version = struct.unpack("<I", f.read(4))[0]
                _tensor_count, kv_count = struct.unpack("<QQ", f.read(16))

                for _ in range(kv_count):
                    # Tolerate truncated input (e.g. a partial header from an
                    # HTTP byte-range fetch): bail out so the resolver
                    # fallback runs on whatever we parsed.
                    try:
                        key_len_bytes = f.read(8)
                        if len(key_len_bytes) < 8:
                            break
                        key_len = struct.unpack("<Q", key_len_bytes)[0]
                        key_bytes = f.read(key_len)
                        if len(key_bytes) < key_len:
                            break
                        key = key_bytes.decode("utf-8")
                        vtype_bytes = f.read(4)
                        if len(vtype_bytes) < 4:
                            break
                        vtype = struct.unpack("<I", vtype_bytes)[0]
                    except (struct.error, UnicodeDecodeError):
                        break

                    try:
                        if key in WANTED or key in arch_keys:
                            if vtype == 8:  # STRING
                                slen = struct.unpack("<Q", f.read(8))[0]
                                val_s = f.read(slen).decode("utf-8")
                                if key.startswith("general.") and key != "general.architecture":
                                    general[key] = val_s
                                if key == "general.architecture":
                                    arch = val_s
                                    self._architecture = val_s
                                    arch_keys = {
                                        f"{arch}.context_length": "context_length",
                                        f"{arch}.block_count": "n_layers",
                                        f"{arch}.attention.head_count_kv": "n_kv_heads",
                                        f"{arch}.attention.head_count": "n_heads",
                                        f"{arch}.embedding_length": "embedding_length",
                                        f"{arch}.feed_forward_length": "feed_forward_length",
                                        f"{arch}.attention.key_length": "kv_key_length",
                                        f"{arch}.attention.value_length": "kv_value_length",
                                        f"{arch}.attention.sliding_window": "sliding_window",
                                        f"{arch}.attention.sliding_window_pattern": "sliding_window_pattern",
                                        f"{arch}.full_attention_interval": "full_attention_interval",
                                        f"{arch}.attention.kv_lora_rank": "kv_lora_rank",
                                        f"{arch}.attention.key_length_mla": "key_length_mla",
                                        f"{arch}.attention.key_length_swa": "kv_key_length_swa",
                                        f"{arch}.attention.value_length_swa": "kv_value_length_swa",
                                        f"{arch}.attention.shared_kv_layers": "shared_kv_layers",
                                        f"{arch}.ssm.inner_size": "ssm_inner_size",
                                        f"{arch}.ssm.state_size": "ssm_state_size",
                                        f"{arch}.nextn_predict_layers": "nextn_predict_layers",
                                    }
                                elif key == "tokenizer.chat_template":
                                    self._chat_template = val_s
                            elif vtype in (4, 10):  # UINT32 or UINT64
                                val_i = (
                                    struct.unpack("<I", f.read(4))[0]
                                    if vtype == 4
                                    else struct.unpack("<Q", f.read(8))[0]
                                )
                                if key == "diffusion.canvas_length":
                                    canvas_seen = True
                                attr = arch_keys.get(key)
                                if attr:
                                    if attr == "sliding_window_pattern":
                                        sliding_window_pattern_period = val_i
                                    else:
                                        setattr(self, f"_{attr}", val_i)
                            elif vtype == 9:  # ARRAY
                                atype = struct.unpack("<I", f.read(4))[0]
                                alen = struct.unpack("<Q", f.read(8))[0]
                                # Vocab size = token count; keep the length, not the strings.
                                if key == "tokenizer.ggml.tokens":
                                    self._vocab_size = int(alen)
                                val_a = self._gguf_read_array_value(f, atype, alen)
                                attr = arch_keys.get(key)
                                if attr == "n_kv_heads" and val_a is not None:
                                    self._n_kv_heads_by_layer = [int(x) for x in val_a]
                                    if self._n_kv_heads is None and self._n_kv_heads_by_layer:
                                        self._n_kv_heads = max(self._n_kv_heads_by_layer)
                                elif attr == "sliding_window_pattern" and val_a is not None:
                                    self._sliding_window_pattern = [bool(x) for x in val_a]
                                    sliding_window_pattern_period = None
                            else:
                                self._gguf_skip_value(f, vtype)
                        else:
                            self._gguf_skip_value(f, vtype)
                    except (struct.error, UnicodeDecodeError):
                        # Truncated input (e.g. HTTP byte-range header
                        # fetch); break so the resolver fallback runs on
                        # what we have.
                        break

            # Decide diffusion routing before the SWA resolver below: it can raise on an arch transformers
            # does not know, which would otherwise drop a DiffusionGemma model to plain llama-server.
            self._is_diffusion = bool(
                (arch and arch.lower().startswith("diffusion")) or canvas_seen
            )
            if self._is_diffusion:
                logger.info(
                    f"GGUF metadata: diffusion model detected (architecture={arch}); "
                    "will serve via the diffusion runner"
                )

            # Expand a scalar period straight from the GGUF first.
            if (
                self._sliding_window_pattern is None
                and sliding_window_pattern_period
                and self._n_layers
            ):
                self._sliding_window_pattern = [
                    (i + 1) % sliding_window_pattern_period != 0 for i in range(self._n_layers)
                ]

            # Otherwise hand off to the resolver (cache / bootstrap / transformers / HF). Diffusion models
            # skip it: they do not use Studio's SWA pattern and the resolver can raise for them.
            if (
                self._sliding_window_pattern is None
                and self._sliding_window
                and self._n_layers
                and not self._is_diffusion
            ):
                hf_repo_candidates = (
                    general.get("general.source.huggingface.repository"),
                    _hf_repo_from_url(general.get("general.source.url")),
                    _hf_repo_from_url(general.get("general.source.repo_url")),
                    _hf_repo_from_url(general.get("general.base_model.0.repo_url")),
                    (
                        f"{general['general.base_model.0.organization']}/"
                        f"{general['general.base_model.0.name']}".replace(" ", "-")
                        if general.get("general.base_model.0.organization")
                        and general.get("general.base_model.0.name")
                        else None
                    ),
                    (
                        f"{general['general.organization']}/{general['general.basename']}".replace(
                            " ", "-"
                        )
                        if general.get("general.organization") and general.get("general.basename")
                        else None
                    ),
                )
                self._sliding_window_pattern = _resolve_swa_pattern(
                    arch,
                    self._n_layers,
                    hf_repo_candidates,
                )

            if self._context_length:
                logger.info(f"GGUF metadata: context_length={self._context_length}")
            if self._chat_template:
                logger.info(f"GGUF metadata: chat_template={len(self._chat_template)} chars")
                # Detect thinking/reasoning support from chat template.
                flags = detect_reasoning_flags(
                    self._chat_template,
                    self._model_identifier,
                    log_source = "GGUF metadata",
                )
                self._supports_reasoning = flags["supports_reasoning"]
                self._reasoning_style = flags["reasoning_style"]
                self._reasoning_effort_levels = flags.get("reasoning_effort_levels", [])
                self._reasoning_always_on = flags["reasoning_always_on"]
                self._supports_preserve_thinking = flags["supports_preserve_thinking"]
                self._supports_tools = flags["supports_tools"]
        except Exception as e:
            logger.warning(f"Failed to read GGUF metadata: {e}")

    # ── Diffusion runner (DiffusionGemma) ──

    def _find_diffusion_assets(self) -> Optional[tuple[list, str, Optional[str]]]:
        """Resolve how to launch the DiffusionGemma runner: (shim argv prefix,
        visual-server binary, optional extra PYTHONPATH dir for the file override).

        Shim: UNSLOTH_DG_SHIM (a .py file) first, else the installed
        unsloth_zoo.diffusion_studio.shim. Binary: DG_VISUAL_BIN first, else
        alongside llama-server. Returns None if neither can be found.
        """
        import importlib.util

        # Visual-server binary: env override, else next to llama-server or in the
        # install's build/bin (where the prebuilt/installer puts it). .exe on Windows.
        visual_bin = os.environ.get("DG_VISUAL_BIN")
        if not visual_bin:
            name = "llama-diffusion-gemma-visual-server" + (".exe" if os.name == "nt" else "")
            # include_denied: a transiently locked llama-server still pins the
            # install dir so the adjacent visual-server can be found
            base = self._find_llama_server_binary(include_denied = True)
            if base:
                base_dir = Path(base).parent
                for cand in (
                    base_dir / name,
                    base_dir / "build" / "bin" / name,
                    base_dir / "build" / "bin" / "Release" / name,
                ):
                    if cand.is_file():
                        visual_bin = str(cand)
                        break
        if not (visual_bin and Path(visual_bin).is_file()):
            return None

        # Shim: a file override (its dir goes on PYTHONPATH), else the zoo package via -m.
        shim_file = os.environ.get("UNSLOTH_DG_SHIM")
        if shim_file and Path(shim_file).is_file():
            return ([sys.executable, shim_file], visual_bin, str(Path(shim_file).parent))

        # Find the installed shim without importing the heavy unsloth_zoo package
        # (find_spec on the top-level package does not run its __init__).
        try:
            spec = importlib.util.find_spec("unsloth_zoo")
        except Exception:
            spec = None
        if spec is not None and spec.submodule_search_locations:
            pkg_dir = Path(list(spec.submodule_search_locations)[0])
            if (pkg_dir / "diffusion_studio" / "shim.py").is_file():
                return (
                    [sys.executable, "-m", "unsloth_zoo.diffusion_studio.shim"],
                    visual_bin,
                    None,
                )

        return None

    def _start_diffusion_server(
        self,
        *,
        model_path: str,
        gguf_path: Optional[str],
        hf_repo: Optional[str],
        hf_variant: Optional[str],
        model_identifier: str,
        n_ctx: int,
        extra_args: Optional[List[str]],
    ) -> bool:
        """Launch the OpenAI-compat diffusion shim (which drives the on-device
        visual decoder) and wait for health. Presents the same /v1 + /health
        interface as llama-server, so the rest of Studio is unchanged.
        """
        assets = self._find_diffusion_assets()
        if assets is None:
            raise RuntimeError(
                "DiffusionGemma runner not found. Install unsloth_zoo (which ships "
                "unsloth_zoo.diffusion_studio.shim) or set UNSLOTH_DG_SHIM to a shim "
                "file, and provide the visual-server binary via DG_VISUAL_BIN or next "
                "to llama-server in the install tree."
            )
        shim_cmd, visual_bin, extra_pythonpath = assets
        self._diffusion_visual_bin = visual_bin

        self._kill_process()
        self._port = self._find_free_port()
        # Auto-size (0): the visual server probes the largest context that fits this GPU's VRAM
        # (capped at the training context). An explicit in-range n_ctx overrides it.
        maxtok = n_ctx if (n_ctx and 0 < n_ctx <= 65536) else 0
        gpu = os.environ.get("DG_GPU", "0")

        cmd = list(shim_cmd) + [
            "--gguf",
            model_path,
            "--host",
            "127.0.0.1",
            "--port",
            str(self._port),
            "--gpu",
            gpu,
            "--maxtok",
            str(maxtok),
        ]

        env = child_env_without_native_path_secret()
        # `python -m unsloth_zoo.diffusion_studio.shim` imports unsloth_zoo, which
        # refuses to load unless UNSLOTH_IS_PRESENT is set (normally by `import
        # unsloth`). The shim never imports unsloth, so set it here as unsloth does.
        env["UNSLOTH_IS_PRESENT"] = "1"
        env["DG_VISUAL_BIN"] = visual_bin
        env["DG_GPU"] = gpu
        # The file-override shim imports its sibling visual_engine; put its dir on PYTHONPATH.
        # (The zoo-package shim is an installed module and needs no PYTHONPATH change.)
        if extra_pythonpath:
            existing = env.get("PYTHONPATH")
            env["PYTHONPATH"] = (
                (extra_pythonpath + os.pathsep + existing) if existing else extra_pythonpath
            )

        logger.info(f"Starting DiffusionGemma runner: {' '.join(cmd)}")
        self._stdout_lines = []
        self._llama_log_fh = None
        self._llama_log_path = None
        try:
            log_dir = _swa_cache_path().parent / "logs" / "diffusion-server"
            log_dir.mkdir(parents = True, exist_ok = True)
            self._llama_log_path = log_dir / f"diffusion-{int(time.time())}-port-{self._port}.log"
            self._llama_log_fh = open(self._llama_log_path, "w", encoding = "utf-8", buffering = 1)
            logger.info(f"diffusion runner stdout/stderr -> {self._llama_log_path}")
        except OSError as e:
            logger.debug(f"Could not open diffusion runner log file: {e}")

        # The shim (and its visual server) die with this backend process, so a
        # Studio crash/restart never orphans a GPU process.
        self._process = subprocess.Popen(
            cmd,
            stdout = subprocess.PIPE,
            stderr = subprocess.STDOUT,
            text = True,
            env = env,
            **_windows_hidden_subprocess_kwargs(),
            **_child_popen_kwargs(),
        )
        self._stdout_thread = threading.Thread(
            target = self._drain_stdout, daemon = True, name = "diffusion-stdout"
        )
        self._stdout_thread.start()

        # Publish state before the health wait (mirrors the llama-server path).
        self._gguf_path = model_path
        self._hf_repo = hf_repo
        self._is_vision = False
        self._is_audio = False  # clear any prior TTS/audio model's routing flag
        self._model_identifier = model_identifier
        self._cache_type_kv = None
        self._gpu_offload_active = True
        if hf_variant:
            self._hf_variant = hf_variant
        elif gguf_path:
            try:
                from utils.models.model_config import _extract_quant_label
                self._hf_variant = _extract_quant_label(gguf_path)
            except Exception:
                self._hf_variant = None
        else:
            self._hf_variant = None
        # Provisional until the server reports the budget it resolved (auto-size picks it from VRAM).
        self._effective_context_length = maxtok or self._context_length
        self._max_context_length = self._context_length or maxtok or None

        healthy = self._wait_for_health(timeout = 600.0)
        if healthy:
            self._healthy = True
            self._gpu_offload_active = True
            if extra_args is not None:
                self._extra_args = list(extra_args)
                self._extra_args_source = (model_identifier, hf_variant)
            # The visual server logs "MAXTOK=<N>" with the context budget it actually resolved
            # (auto-sized to VRAM). Read it back so the UI context bar shows the real budget.
            chosen = maxtok
            try:
                for _ln in reversed(self._stdout_lines):
                    _m = re.search(r"MAXTOK=(\d+)", _ln)
                    if _m:
                        chosen = int(_m.group(1))
                        break
            except Exception:
                pass
            if chosen and chosen > 0:
                self._effective_context_length = chosen
                self._max_context_length = chosen
            self._requested_n_ctx = int(n_ctx)
        else:
            self._healthy = False
            logger.error("DiffusionGemma runner failed to become healthy")
        return healthy

    # ── HF download (no lock held) ───────────────────────────────

    def _download_gguf(
        self,
        *,
        hf_repo: str,
        hf_variant: Optional[str] = None,
        hf_token: Optional[str] = None,
        force: bool = False,
        allow_smaller_fallback: bool = True,
        cancel_event: Optional[threading.Event] = None,
    ) -> str:
        """Download GGUF file(s) from HuggingFace. Returns local path.

        Runs WITHOUT self._lock so unload_model() can set _cancel_event at
        any time; checks it between each shard download.

        ``force`` re-fetches even when a (possibly stale) blob is cached.
        ``allow_smaller_fallback=False`` raises on low disk instead of silently
        switching to a smaller quant. ``cancel_event`` overrides
        ``self._cancel_event`` so an update can use a private event without
        touching the shared one; defaults to the shared event.
        """
        cancel_event = cancel_event if cancel_event is not None else self._cancel_event
        try:
            import huggingface_hub  # noqa: F401 -- presence check only
        except ImportError:
            raise RuntimeError(
                "huggingface_hub is required for HF model loading. "
                "Install it with: pip install huggingface_hub"
            )

        resolved_hf_repo = _resolve_repo_id_casing(hf_repo)
        if resolved_hf_repo != hf_repo:
            logger.info(
                "Using cached repo_id casing '%s' for requested '%s'",
                resolved_hf_repo,
                hf_repo,
            )
            hf_repo = resolved_hf_repo

        # Resolve the filename from the variant
        gguf_filename = None
        gguf_extra_shards: list[str] = []
        if hf_variant:
            try:
                from huggingface_hub import list_repo_files

                files = list_repo_files(hf_repo, token = hf_token)
                gguf_files = _gguf_files_for_variant(files, hf_variant)
                if gguf_files:
                    gguf_filename = gguf_files[0]
                    gguf_extra_shards = _gguf_extra_shards(gguf_files, gguf_filename)
            except Exception as e:
                logger.warning(f"Could not list repo files: {e}")

            # Offline: resolve variant -> filename from the local HF cache.
            # The heuristic below assumes filenames echo the repo name, which
            # breaks for e.g. Qwen3.6-27B-MTP-GGUF (no "MTP" in file). Match
            # against the rel path (not just basename) so subdir layouts like
            # ``BF16/foo.gguf`` are findable.
            if not gguf_filename:
                try:
                    from utils.models.model_config import _iter_hf_cache_snapshots
                    for snap in _iter_hf_cache_snapshots(hf_repo):
                        cached_files = _gguf_snapshot_files(snap)
                        matches = _gguf_files_for_variant(cached_files, hf_variant)
                        if not matches:
                            continue
                        gguf_filename = matches[0]
                        gguf_extra_shards = _gguf_extra_shards(matches, gguf_filename)
                        logger.info(
                            "Resolved variant %s -> %s from local HF cache",
                            hf_variant,
                            gguf_filename,
                        )
                        break
                except Exception as e:
                    logger.debug(f"Offline cache lookup for variant failed: {e}")

            if not gguf_filename:
                repo_name = hf_repo.split("/")[-1].replace("-GGUF", "")
                gguf_filename = f"{repo_name}-{hf_variant}.gguf"

        # Check disk space; fall back to a smaller variant if needed
        all_gguf_files = [gguf_filename] + gguf_extra_shards
        expected_sizes: dict[str, int] = {}
        try:
            from huggingface_hub import get_paths_info, try_to_load_from_cache

            path_infos = list(get_paths_info(hf_repo, all_gguf_files, token = hf_token))
            expected_sizes = {p.path: p.size for p in path_infos if p.size}
            total_bytes = sum((p.size or 0) for p in path_infos)

            # Subtract bytes already in the HF cache so we only preflight
            # against what we must download. Without this, re-loading a
            # cached large model (e.g. MiniMax-M2.7-GGUF at 131 GB) fails
            # cold whenever free disk is below the full weight footprint,
            # even though nothing needs downloading.
            already_cached_bytes = 0
            # Cross-snapshot / case-variant cache reuse is offline-only (see the download
            # path below); online, hf_hub_download fetches the current revision and
            # resumes partials, so an old snapshot must not be counted as cached here or
            # the preflight would under-count the download and skip the disk fallback.
            offline = _hf_env_offline()
            # A split GGUF whose shards are not co-located in a single snapshot is
            # refetched as a whole set later, so it must not be counted as cached here.
            split_needs_refetch = False
            if offline and not force and gguf_extra_shards:
                # Scan all snapshots for one that holds the whole set co-located, so a
                # newer snapshot with only the first shard does not mask an older
                # complete one and needlessly trip the disk fallback.
                if (
                    _cached_colocated_split_main(
                        hf_repo, gguf_filename, gguf_extra_shards, expected_sizes
                    )
                    is None
                ):
                    split_needs_refetch = True
            if not force and not split_needs_refetch:
                for p in path_infos:
                    if not p.size:
                        continue
                    try:
                        cached_path = try_to_load_from_cache(hf_repo, p.path)
                    except Exception:
                        cached_path = None
                    if (
                        not (isinstance(cached_path, str) and os.path.exists(cached_path))
                        and offline
                    ):
                        cached_path = _cached_hf_snapshot_file(
                            hf_repo,
                            p.path,
                            expected_size = p.size,
                        )
                    if isinstance(cached_path, str) and os.path.exists(cached_path):
                        try:
                            on_disk = os.path.getsize(cached_path)
                        except OSError:
                            on_disk = 0
                        # Satisfied only when the full blob is present.
                        if on_disk >= p.size:
                            already_cached_bytes += p.size

            total_download_bytes = max(0, total_bytes - already_cached_bytes)

            if total_download_bytes > 0:
                cache_dir = os.environ.get(
                    "HF_HUB_CACHE",
                    str(Path.home() / ".cache" / "huggingface" / "hub"),
                )
                Path(cache_dir).mkdir(parents = True, exist_ok = True)
                free_bytes = shutil.disk_usage(cache_dir).free

                total_gb = total_download_bytes / (1024**3)
                free_gb = free_bytes / (1024**3)
                cached_gb = already_cached_bytes / (1024**3)

                logger.info(
                    f"GGUF download: {total_gb:.1f} GB needed "
                    f"({cached_gb:.1f} GB already cached), "
                    f"{free_gb:.1f} GB free on disk"
                )

                if total_download_bytes > free_bytes:
                    if not allow_smaller_fallback:
                        # Update path: never silently switch to a smaller quant;
                        # surface the disk shortfall for the requested variant.
                        raise RuntimeError(
                            f"Not enough disk space to download {gguf_filename}. "
                            f"Only {free_gb:.1f} GB free in {cache_dir}"
                        )
                    smaller = self._find_smallest_fitting_variant(
                        hf_repo,
                        free_bytes,
                        hf_token,
                    )
                    if smaller:
                        fallback_file, fallback_size = smaller
                        logger.info(
                            f"Selected variant too large ({total_gb:.1f} GB), "
                            f"falling back to {fallback_file} ({fallback_size / (1024**3):.1f} GB)"
                        )
                        gguf_filename = fallback_file
                        _m = _SHARD_RE.match(gguf_filename)
                        _prefix = _m.group(1) if _m else None
                        if _prefix:
                            prefix_lower = _prefix.lower()
                            gguf_extra_shards = sorted(
                                f
                                for f in all_gguf_files
                                if f.lower().startswith(prefix_lower)
                                and f != gguf_filename
                                and not _is_companion_gguf_path(f)
                            )
                        else:
                            gguf_extra_shards = []
                        # Record the fallback's size so the later cache-reuse probe can
                        # size-verify it; only for a single-file fallback, since
                        # _find_smallest_fitting_variant returns the whole-variant size
                        # and using that as the first shard's expected size would reject
                        # a valid cached first shard of a split fallback.
                        if not gguf_extra_shards:
                            expected_sizes[fallback_file] = fallback_size
                    else:
                        raise RuntimeError(
                            f"Not enough disk space to download any variant. "
                            f"Only {free_gb:.1f} GB free in {cache_dir}"
                        )
        except RuntimeError:
            raise
        except Exception as e:
            logger.warning(f"Could not check disk space: {e}")

        gguf_label = f"{hf_repo}/{gguf_filename}" + (
            f" (+{len(gguf_extra_shards)} shards)" if gguf_extra_shards else ""
        )
        logger.info(f"Resolving GGUF: {gguf_label}")
        try:
            if cancel_event.is_set():
                raise RuntimeError("Cancelled")
            dl_start = time.monotonic()
            # Xet primary, HTTP fallback on stall; per-file so finished shards stay cached.
            local_path = None
            # Reuse a cached copy from another snapshot / case-variant repo dir only when
            # offline. Online, fall through to hf_hub_download so its revision/etag check
            # fetches the current file (and resumes a partial) instead of serving a stale
            # same-name blob from an older revision.
            if not force and _hf_env_offline():
                if gguf_extra_shards:
                    # A split GGUF must load every shard from one snapshot; reuse only a
                    # snapshot that holds the whole set co-located, scanning past a newer
                    # snapshot that has just the first shard while an older one is complete.
                    local_path = _cached_colocated_split_main(
                        hf_repo, gguf_filename, gguf_extra_shards, expected_sizes
                    )
                else:
                    local_path = _cached_hf_snapshot_file(
                        hf_repo,
                        gguf_filename,
                        expected_size = expected_sizes.get(gguf_filename),
                    )
            if local_path is None:
                local_path = hf_hub_download_with_xet_fallback(
                    hf_repo,
                    gguf_filename,
                    hf_token,
                    cancel_event = cancel_event,
                    on_status = lambda m: logger.info(m),
                    force_download = force,
                )
                for shard in gguf_extra_shards:
                    if cancel_event.is_set():
                        raise RuntimeError("Cancelled")
                    logger.info(f"Resolving GGUF shard: {shard}")
                    hf_hub_download_with_xet_fallback(
                        hf_repo,
                        shard,
                        hf_token,
                        cancel_event = cancel_event,
                        force_download = force,
                    )
        except Exception as e:
            if isinstance(e, RuntimeError) and "Cancelled" in str(e):
                raise
            raise RuntimeError(
                f"Failed to download GGUF file '{gguf_filename}' from {hf_repo}: {e}"
            )

        dl_elapsed = time.monotonic() - dl_start
        if dl_elapsed < 2.0:
            logger.info(f"GGUF resolved from cache: {local_path}")
        else:
            logger.info(f"GGUF downloaded in {dl_elapsed:.1f}s: {local_path}")
        return local_path

    def _download_companion_gguf(
        self,
        *,
        hf_repo: str,
        hf_token: Optional[str],
        pick: Callable[[list[str]], Optional[str]],
        label: str,
        cancel_event: Optional[threading.Event] = None,
    ) -> Optional[str]:
        """Resolve and fetch a companion GGUF (mmproj / MTP drafter) by name.

        Tries the live repo file list, then the local HF cache snapshots
        (offline, same fallback as _download_gguf), then hf_hub_download.
        Runs WITHOUT self._lock (like _download_gguf); honors _cancel_event so
        an /unload between the main download and here skips the fetch.
        ``cancel_event`` overrides ``self._cancel_event`` (defaults to it).
        """
        cancel_event = cancel_event if cancel_event is not None else self._cancel_event
        if cancel_event.is_set():
            return None

        target: Optional[str] = None
        from huggingface_hub import list_repo_files

        # Retry a transient listing blip; permanent repo/auth errors and offline
        # mode are not retried (offline raises at once -> fall through to cache).
        for attempt in range(3):
            if cancel_event.is_set():
                return None
            try:
                target = pick(list_repo_files(hf_repo, token = hf_token))
                break
            except Exception as e:
                if type(e).__name__ in (
                    "RepositoryNotFoundError",
                    "GatedRepoError",
                    "RevisionNotFoundError",
                    "EntryNotFoundError",
                    "OfflineModeIsEnabled",
                ):
                    logger.debug(f"Could not list repo files for {label}: {e}")
                    break
                logger.debug(
                    f"Could not list repo files for {label} (attempt {attempt + 1}/3): {e}"
                )
                if attempt < 2:
                    cancel_event.wait(2**attempt)

        if target is None:
            try:
                from utils.models.model_config import _iter_hf_cache_snapshots
                for snap in _iter_hf_cache_snapshots(hf_repo):
                    rel_files = _gguf_snapshot_files(snap)
                    target = pick(rel_files)
                    if target is not None:
                        logger.info("Resolved %s %s from local HF cache", label, target)
                        break
            except Exception as e:
                logger.debug(f"Offline cache lookup for {label} failed: {e}")

        if target is None or cancel_event.is_set():
            return None

        # Offline, resolve the companion straight from the cache snapshot that
        # holds it. resolve_cached_repo_id_case can return a partial lower-case
        # spelling when any dir exists under the requested casing, so calling
        # hf_hub_download with hf_repo would miss the canonical file and silently
        # drop the companion. _cached_hf_snapshot_file scans every case variant.
        if _hf_env_offline():
            cached = _cached_hf_snapshot_file(hf_repo, target)
            if cached:
                logger.info("Resolved %s from local HF cache: %s", label, cached)
                return cached

        try:
            logger.info(f"Downloading {label}: {hf_repo}/{target}")
            # Same policy; companions are best-effort (caller below swallows failures to None).
            return hf_hub_download_with_xet_fallback(
                hf_repo,
                target,
                hf_token,
                cancel_event = cancel_event,
            )
        except Exception as e:
            logger.warning(f"Could not download {label}: {e}")
            return None

    def _download_mmproj(
        self,
        *,
        hf_repo: str,
        hf_token: Optional[str] = None,
        cancel_event: Optional[threading.Event] = None,
    ) -> Optional[str]:
        """Download the mmproj (vision projection) file from a GGUF repo.

        Prefers mmproj-F16.gguf, else any mmproj*.gguf. Returns the local
        path, or None if none exists. ``cancel_event`` overrides
        ``self._cancel_event`` (defaults to it).
        """

        def _pick_mmproj(candidates: list[str]) -> Optional[str]:
            mmproj_files = sorted(
                f
                for f in candidates
                if f.lower().endswith(".gguf") and "mmproj" in Path(f).name.lower()
            )
            if not mmproj_files:
                return None
            for f in mmproj_files:
                if f.lower().endswith("-f16.gguf"):
                    return f
            return mmproj_files[0]

        return self._download_companion_gguf(
            hf_repo = hf_repo,
            hf_token = hf_token,
            pick = _pick_mmproj,
            label = "mmproj",
            cancel_event = cancel_event,
        )

    def _download_mtp(
        self,
        *,
        hf_repo: str,
        hf_token: Optional[str] = None,
    ) -> Optional[str]:
        """Download the separate MTP drafter (speculative head) from a GGUF repo.

        Targets the repo-root ``mtp-*.gguf`` companion -- the Q8_0 drafter
        unsloth mirrors there for llama.cpp ``-hf`` auto-discovery (smallest,
        recommended for speculation). Repos that bake the MTP head into the
        main GGUF (e.g. Qwen) ship no such sibling and this returns None. The
        higher-precision copies under ``MTP/`` are for explicit selection and
        are intentionally skipped. Returns the local path, or None.
        """

        def _pick_mtp(candidates: list[str]) -> Optional[str]:
            mtp_files = sorted(
                f
                for f in candidates
                if f.lower().endswith(".gguf") and Path(f).name.lower().startswith("mtp-")
            )
            return mtp_files[0] if mtp_files else None

        return self._download_companion_gguf(
            hf_repo = hf_repo,
            hf_token = hf_token,
            pick = _pick_mtp,
            label = "MTP drafter",
        )

    def _resolve_launch_mmproj_path(
        self, *, model_path: str, mmproj_path: Optional[str]
    ) -> Optional[str]:
        """Return mmproj_path iff it exists on disk AND matches the model family.

        None if mmproj_path is None, missing, or family-mismatched.
        """
        if not mmproj_path:
            return None

        mmproj = Path(mmproj_path)
        if not mmproj.is_file():
            logger.warning(f"mmproj file not found: {mmproj_path}")
            return None

        from utils.models.model_config import mmproj_matches_model_family

        if not mmproj_matches_model_family(model_path, str(mmproj)):
            logger.warning(
                f"mmproj does not match model family: model={Path(model_path).name} "
                f"mmproj={mmproj.name}"
            )
            return None

        return str(mmproj)

    def _mmproj_vram_bytes(self, launch_mmproj_path: Optional[str]) -> int:
        """Return resolved mmproj VRAM bytes, or 0 when absent/unreadable."""
        if not launch_mmproj_path:
            return 0
        try:
            return self._get_gguf_size_bytes(launch_mmproj_path)
        except OSError as e:
            logger.debug(f"Could not size mmproj {launch_mmproj_path}: {e}")
            return 0

    def _resolve_launch_mtp_path(self, *, mtp_draft_path: Optional[str]) -> Optional[str]:
        """Return mtp_draft_path iff it exists on disk, else None.

        No family check needed: the drafter is only ever auto-resolved from
        the same repo as the main GGUF (see _download_mtp).
        """
        if not mtp_draft_path:
            return None
        if not Path(mtp_draft_path).is_file():
            logger.warning(f"MTP drafter file not found: {mtp_draft_path}")
            return None
        return str(mtp_draft_path)

    # ── Lifecycle ─────────────────────────────────────────────────

    # GGUF ``general.architecture`` values for diffusion / image models.
    # llama.cpp has no such architectures, so loading one as a chat model dies
    # with "unknown model architecture: '<arch>'". These match the patched
    # stable-diffusion.cpp / ComfyUI-GGUF enums. Unsloth publishes FLUX and
    # Qwen-Image GGUFs under
    # https://huggingface.co/collections/unsloth/unsloth-diffusion-ggufs.
    # Matched exactly (not a substring) so a chat arch containing "wan"/"sd1"
    # (e.g. "taiwan") isn't misrouted to Images.
    _DIFFUSION_ARCHES = frozenset(
        (
            "qwen_image",
            "flux",
            "sd1",
            "sdxl",
            "sd3",
            "aura",
            "hidream",
            "cosmos",
            "ltxv",
            "hyvid",
            "wan",
            "lumina2",
        )
    )

    @staticmethod
    def _classify_llama_start_failure(
        output: str,
        gguf_path: Optional[str],
        model_identifier: Optional[str],
        returncode: Optional[int] = None,
    ) -> str:
        """Explain *why* llama-server failed to start, from its output.

        Several distinct failures otherwise collapse into the same opaque
        "invalid GGUF or out of memory" message. Worst case: a diffusion GGUF
        loaded as a chat model -- valid file, plenty of memory, but llama.cpp
        has no such architecture, so the user is told to free memory that was
        never the problem (#5842). Pick the most specific message we can.
        """
        lowered = (output or "").lower()

        # Tensor parallelism (--split-mode tensor) is arch-gated in llama.cpp;
        # unsupported architectures abort the load with this marker. Point the
        # user at the toggle instead of a generic invalid-GGUF/OOM message.
        if "split_mode_tensor not implemented" in lowered:
            return (
                "Tensor parallelism is not supported for this model's "
                "architecture. Turn off Tensor Parallelism in the model "
                "settings and reload."
            )

        # Detect Ollama source up front so the arch branch can keep the
        # Ollama hint instead of the generic "unsupported arch" message.
        gguf = gguf_path or ""
        is_ollama = (
            ".studio_links" in gguf
            or os.sep + "ollama_links" + os.sep in gguf
            or os.sep + ".cache" + os.sep + "ollama" + os.sep in gguf
            or (model_identifier or "").startswith("ollama/")
        )

        # "unknown model architecture: '<arch>'": diffusion -> Images page,
        # Ollama -> Ollama hint, else a precise "unsupported" message. Exact
        # match so chat archs aren't misrouted.
        arch_match = re.search(r"unknown model architecture:\s*'([^']+)'", lowered)
        if arch_match:
            arch = arch_match.group(1)
            if arch in LlamaCppBackend._DIFFUSION_ARCHES:
                return (
                    f"'{arch}' is a diffusion (image-generation) GGUF, which "
                    "llama-server cannot run as a chat/completion model. Use "
                    "Studio's Images page to generate with local diffusion "
                    "GGUFs such as FLUX and Qwen-Image."
                )
            if is_ollama:
                return (
                    "Some Ollama models do not work with llama.cpp. Try a "
                    "different model, or use this model directly through "
                    "Ollama instead."
                )
            return (
                f"llama.cpp does not support this GGUF's model architecture "
                f"('{arch}'). The file is valid, but this model type cannot "
                "be run with llama-server."
            )

        # Other Ollama compat failures that don't name an arch. Only when
        # the output shows a GGUF compat issue, not OOM / missing binaries.
        if is_ollama:
            gguf_compat_hints = (
                "key not found",
                "unknown model architecture",
                "failed to load model",
            )
            if any(h in lowered for h in gguf_compat_hints):
                return (
                    "Some Ollama models do not work with llama.cpp. Try a "
                    "different model, or use this model directly through "
                    "Ollama instead."
                )

        # SIGKILL with no diagnostic output is the OOM killer (e.g. a model too
        # large for the WSL VM's RAM cap); name it actionably.
        if returncode == -9:
            return (
                "llama-server was stopped by the operating system (signal 9), "
                "most likely out of memory. Try a smaller or more quantized "
                "GGUF, lower the context length, or free memory (on WSL, raise "
                "the memory limit in .wslconfig)."
            )
        # SIGTERM is also how an unload/cancel or a supervisor stops the server,
        # so report it neutrally rather than blaming memory.
        if returncode == -15:
            return (
                "llama-server was terminated (signal 15) before it became "
                "healthy. If you cancelled or unloaded the model this is "
                "expected; otherwise check the llama-server log for the cause."
            )

        # A live server that never answered 200 on /health is not a bad GGUF:
        # the load is too large for VRAM/context, or a local proxy/VPN grabbed
        # the loopback probe (#5740).
        if "health check timed out" in lowered:
            return (
                "llama-server started but never became healthy on its local "
                "/health endpoint. Try a smaller context length or a more "
                "quantized GGUF, and if you use a VPN or HTTP proxy make sure "
                "localhost bypasses it (NO_PROXY=127.0.0.1,localhost)."
            )

        # Fallback: genuinely unknown failure (OOM, missing binary ...).
        return (
            "llama-server failed to start. "
            "Check that the GGUF file is valid and you have enough memory."
        )

    def _plan_tensor_parallel(
        self,
        gpus: list[tuple[int, int]],
        model_size: int,
        target_ctx: int,
        cache_type_kv: Optional[str] = None,
        n_parallel: int = 1,
        mtp_engaged: bool = False,
        mtp_overhead_fn: Optional[Callable[[int], int]] = None,
        mtp_flat_reserve_bytes: int = 0,
        max_target_ctx: Optional[int] = None,
        total_by_idx: Optional[dict[int, int]] = None,
        n_ubatch: Optional[int] = None,
        soft_overhead_bytes: int = 0,
    ) -> tuple[int, int, list[int], Optional[list[int]]]:
        """Plan a ``--split-mode tensor`` load. Pure: no model or GPU needed.

        ``gpus`` is a list of ``(gpu_index, free_mib)``; ``model_size`` is the
        weight size in bytes; ``target_ctx`` is the context to fit (the explicit
        request, or the model's native length for auto). ``max_target_ctx`` is
        the native/hardware ceiling used only for the UI bound (defaults to
        ``target_ctx``). Returns
        ``(effective_ctx, max_available_ctx, gpu_indices, tensor_split)``.

        Policy (assumes >= 2 GPUs; the caller drops the toggle below that):
        - Cap context to the KV that fits the pooled VRAM after the weights, one
          per-device flat compute-graph buffer (``_estimate_compute_buffer_bytes``,
          deterministic from dims; flat fallback when dims are unavailable), and the
          per-device context-linear compute growth (``_compute_buffer_ctx_bytes``,
          replicated on every device in tensor mode, so summed over the split).
          llama.cpp's ``--fit`` is a no-op in tensor mode, so this is the only
          cap, honored even for an explicit ``-c``. It is more accurate than the
          0.80 whole-pool heuristic, which over-reserves and leaves VRAM unused.
        - ``tensor_split`` is None (llama.cpp's even default, safe for every arch
          incl. Gemma 3n which GGML_ASSERTs on a weighted split) when an even
          share fits the smallest GPU; otherwise it is weighted by usable budget
          so the roomier GPU absorbs more weight and the smallest keeps room for KV.
        ``total_by_idx`` enables the total-based occupancy cap; ``n_ubatch`` sizes
        the compute buffer. ``soft_overhead_bytes`` is the CUDA-context / mmproj /
        MTP-draft-graph reserve the layer path folds into ``model_size_fit``;
        charged against the pooled budget so tensor mode reserves the same overhead.
        """

        # Per-GPU usable budget: free - (1-frac)*total, else (unknown total, e.g. a
        # two-column probe) the legacy free*frac. Mirrors _select_gpus and
        # _gpu_usable so the 5% cushion is kept on every path, not dropped here.
        def _usable(idx: int, free_mib: int) -> float:
            t = total_by_idx.get(idx, 0) if total_by_idx else 0
            if t > 0:
                return max(0.0, free_mib - (1.0 - _CTX_FIT_VRAM_FRACTION) * t)
            return max(0.0, free_mib * _CTX_FIT_VRAM_FRACTION)

        # Drop GPUs whose usable budget can't hold the per-device compute-graph
        # buffer; they'd OOM in tensor mode. Admitting on raw free would let a
        # partly-used big card in with no budget left. Defense-in-depth (load_model
        # gates too). Derived per-device reserve; flat fallback.
        _reserve_bytes = self._estimate_compute_buffer_bytes(
            n_ubatch = n_ubatch, n_parallel = n_parallel, per_device_tensor = True
        )
        reserve_mib = (
            _reserve_bytes // (1024 * 1024)
            if _reserve_bytes > 0
            else self._TENSOR_PARALLEL_BUFFER_RESERVE_MIB
        )
        usable_gpus = [g for g in gpus if _usable(g[0], g[1]) >= reserve_mib]
        gpu_indices = sorted(idx for idx, _ in usable_gpus)
        if len(gpu_indices) < 2:
            # Tensor parallelism is meaningless on <2 GPUs (the caller drops the
            # toggle before this); be defensive and never emit a split here.
            return (
                target_ctx if target_ctx > 0 else 4096,
                target_ctx if target_ctx > 0 else 4096,
                gpu_indices,
                None,
            )
        free_by_idx = {idx: free for idx, free in usable_gpus}
        usable_by_idx = {idx: _usable(idx, free_by_idx[idx]) for idx in gpu_indices}
        pool_mib = sum(usable_by_idx.values())
        # MTP reserve: byte-accurate per-ctx inside _fit_ctx (mtp_overhead_fn) plus
        # a flat cushion that the byte fn can't size -- 2 GiB when dims are wholly
        # unavailable (no fn), or mtp_flat_reserve_bytes when the fn is weights-only
        # because the draft KV couldn't be sized (_mtp_kv_unsized). Without this the
        # binary search spends the unsized-KV cushion on main context and OOMs.
        flat_mtp_bytes = max(0, mtp_flat_reserve_bytes)
        if mtp_engaged and mtp_overhead_fn is None:
            flat_mtp_bytes = max(flat_mtp_bytes, 2 * 1024**3)
        # soft_overhead_bytes is the CUDA-context / mmproj / MTP-draft-graph reserve
        # the layer path folds into model_size_fit. Tensor mode has no --fit valve, so
        # an unreserved overshoot OOMs at startup rather than offloading; charge it here
        # too. Once (pooled), mirroring the layer path -- the per-device CUDA context is
        # a known slight under-charge, left for real multi-GPU data.
        kv_budget_b = (
            (pool_mib - len(gpu_indices) * reserve_mib) * 1024 * 1024
            - model_size
            - flat_mtp_bytes
            - max(0, soft_overhead_bytes)
        )

        def _mtp_at(ctx: int) -> int:
            return mtp_overhead_fn(ctx) if mtp_overhead_fn is not None else 0

        # Context-linear compute buffer, summed over the split. Tensor mode
        # replicates the compute graph on EVERY device (measured: the per-device
        # buffer grows a flat n_ubatch*2 bytes/token, ~1024 B/tok on Qwen3.5-9B at
        # f16, independent of n_embd), so the growth is n_dev x the per-device
        # term. cache_type_kv here is always non-quantized (tensor forces f16), so
        # _compute_buffer_ctx_bytes returns the light KQ-mask term, not the heavy
        # quantized dequant scratch. The flat reserve_mib above only covers ctx->0;
        # without this the fit over-pins and OOMs at high context on a tight pool
        # (0.5-4 GiB unreserved at 262k-1M across 2-4 GPUs), the tensor-mode analog
        # of the layer-split compute bug.
        n_dev = len(gpu_indices)

        def _cc_ctx(ctx: int) -> int:
            return n_dev * self._compute_buffer_ctx_bytes(ctx, n_ubatch, cache_type_kv)

        def _fit_ctx(ctx: int) -> int:
            # Largest context whose KV (+ MTP draft reserve + context-linear
            # compute) fits the pooled budget. Floors small, but never raises an
            # explicit ctx above asked.
            if self._can_estimate_kv() and ctx > 0:
                ctx_floor = min(2048, ctx)
                if kv_budget_b <= 0:
                    # Weights + buffers exceed the pool -> floor; the load then
                    # falls back to layer split.
                    return ctx_floor
                if mtp_overhead_fn is not None:
                    # kv(ctx)+mtp(ctx)+compute(ctx) is not single-linear, so binary search.
                    def _consumer(c: int) -> int:
                        return (
                            self._estimate_kv_cache_bytes(c, cache_type_kv, n_parallel = n_parallel)
                            + _mtp_at(c)
                            + _cc_ctx(c)
                        )

                    if _consumer(ctx) <= kv_budget_b:
                        return ctx
                    lo, hi, best = ctx_floor, ctx, ctx_floor
                    while lo <= hi:
                        mid = (lo + hi) // 2
                        if _consumer(mid) <= kv_budget_b:
                            best = mid
                            lo = mid + 1
                        else:
                            hi = mid - 1
                    return best
                kv_at = self._estimate_kv_cache_bytes(ctx, cache_type_kv, n_parallel = n_parallel)
                total_at = kv_at + _cc_ctx(ctx)  # both ~linear through the origin
                if total_at <= kv_budget_b:
                    return ctx
                return max(ctx_floor, int(ctx * kv_budget_b / total_at))
            # KV size unknown -> can't prove a safe cap; floor.
            return min(4096, ctx) if ctx > 0 else 4096

        # max_available_ctx is the hardware ceiling for the UI bound, sized from
        # the native context independent of an explicit small -c (which only
        # caps effective_ctx).
        max_ctx_target = max_target_ctx if (max_target_ctx and max_target_ctx > 0) else target_ctx
        max_available_ctx = _fit_ctx(max_ctx_target)
        effective_ctx = min(_fit_ctx(target_ctx), max_available_ctx)

        min_usable_mib = min(usable_by_idx.values())
        kv_bytes = (
            self._estimate_kv_cache_bytes(effective_ctx, cache_type_kv, n_parallel = n_parallel)
            if (self._can_estimate_kv() and effective_ctx > 0)
            else 0
        )
        # The MTP reserve also has to fit the even split (mirror the pooled budget):
        # byte-accurate per-ctx (0 when no fn) plus the same flat cushion as above.
        mtp_bytes = (_mtp_at(effective_ctx) if effective_ctx > 0 else 0) + flat_mtp_bytes
        # Context-linear compute is replicated per device; charge the whole split so
        # the weighted ratio reflects it (mirrors kv_budget_b's per-device reserve).
        cc_bytes = _cc_ctx(effective_ctx) if effective_ctx > 0 else 0
        even_share_mib = (
            (model_size + kv_bytes + mtp_bytes + cc_bytes) / len(gpu_indices) / (1024 * 1024)
        )
        tensor_split: Optional[list[int]] = None
        if even_share_mib > (min_usable_mib - reserve_mib):
            # Each device also holds its replicated share of the context-linear
            # compute (cc_bytes/n_dev) on top of the flat reserve. The even-share
            # gate above charges cc_bytes; the split weights must subtract it too, or
            # the smaller card is weighted above its real usable budget and OOMs (the
            # per-device analog of the layer path's per-GPU overhead in _select_gpus).
            cc_per_dev_mib = (cc_bytes // len(gpu_indices)) // (1024 * 1024) if cc_bytes else 0
            adj = [
                max(0, int(usable_by_idx[i] - reserve_mib - cc_per_dev_mib)) for i in gpu_indices
            ]
            if sum(adj) > 0:
                tensor_split = adj
        return effective_ctx, max_available_ctx, gpu_indices, tensor_split

    @staticmethod
    def _is_projector_incompatibility(output: str) -> bool:
        """True when llama-server aborted because it cannot load the model's
        vision/audio projector (mmproj), typically an installed llama.cpp
        that predates the projector format. Conservative: only matches
        projector-format errors so unrelated failures (OOM, bad GGUF, port
        bind, ...) keep their own handling, and a bare 'clip'/'mmproj'
        mention in a normal startup log does not match.
        """
        text = (output or "").lower()
        if any(
            m in text
            for m in (
                "unknown projector type",
                "unsupported projector",
                "unsupported mmproj",
            )
        ):
            return True
        # Builds that phrase it via clip.cpp without the exact words above.
        return (
            "clip" in text
            and "projector" in text
            and ("unknown" in text or "unsupported" in text or "not supported" in text)
        )

    @staticmethod
    def _output_has_nonprojector_diagnostic(output: str) -> bool:
        """True when the output already names a concrete non-projector cause (out
        of memory, an unsupported architecture, a tensor-parallel limit). A hard
        crash carrying such a marker must surface that error, not be silently
        retried text-only as if the vision projector were at fault; a bare crash
        with no marker still gets the text-only retry.
        """
        text = (output or "").lower()
        return any(
            m in text
            for m in (
                "out of memory",
                "failed to allocate",
                "unknown model architecture",
                "split_mode_tensor not implemented",
            )
        )

    @staticmethod
    def _is_tensor_split_assert(output: str) -> bool:
        """True only for the #6415 split-axis warmup assert (GGML_BACKEND_SPLIT_AXIS_*),
        not any ggml assert/abort, so an unrelated invariant isn't cached. stderr is
        merged into output."""
        text = (output or "").lower()
        if "ggml_assert" not in text and "ggml_abort" not in text:
            return False
        # the split-axis enum token, unique to this assert (not the source file).
        return "split_axis" in text

    @staticmethod
    def _is_signal_crash(returncode: Optional[int]) -> bool:
        """True only on a hard fault (SIGSEGV/SIGABRT/SIGILL/SIGFPE/SIGBUS or a
        Windows 0xC0000000+ status), not SIGKILL/SIGTERM/SIGINT (OOM killer /
        unload) nor a clean exit or still-running (None) process.
        """
        if returncode is None:
            return False
        if returncode >= 0xC0000000:  # Windows access violation / illegal instruction
            return True
        return -returncode in (4, 6, 7, 8, 11)  # SIGILL SIGABRT SIGBUS SIGFPE SIGSEGV

    @staticmethod
    def _is_abort_exit(returncode: Optional[int]) -> bool:
        """Windows CRT abort() exit code (3) from GGML_ASSERT on MSVC -- not a POSIX
        signal or 0xC0000000+ NTSTATUS."""
        return returncode == 3

    @classmethod
    def _should_record_tensor_split_abort(cls, returncode: Optional[int], output: str) -> bool:
        """The #6415 split-axis abort: the marker plus a hard crash (POSIX signal or
        Windows abort exit). Marker required so a generic crash isn't cached."""
        return cls._is_tensor_split_assert(output) and (
            cls._is_signal_crash(returncode) or cls._is_abort_exit(returncode)
        )

    @staticmethod
    def _with_flash_attn_off(cmd: list[str]) -> Optional[list[str]]:
        """Return cmd with flash attention forced off, or None when its effective
        (last-wins) value is already off/absent so there is nothing to retry. FA
        kernels hard-crash at startup on some ROCm builds; disabling FA keeps
        vision and MTP, the least destructive rung. A bare --flash-attn/-fa reads
        as on, so it counts toward the effective value and is neutralised too;
        every form is flipped in place (length preserved for downstream slices)."""
        out = list(cmd)

        def explicit(i):
            nxt = out[i + 1] if i + 1 < len(out) else None
            return nxt if nxt in ("on", "auto", "off") else None

        effective = None
        for i, tok in enumerate(out):
            if tok.startswith(("--flash-attn=", "-fa=")):
                effective = tok.partition("=")[2]
            elif tok in ("--flash-attn", "-fa"):
                effective = explicit(i) or "on"
        if effective not in ("on", "auto"):
            return None
        for i, tok in enumerate(out):
            if tok.startswith(("--flash-attn=", "-fa=")):
                flag, _, value = tok.partition("=")
                if value in ("on", "auto"):
                    out[i] = f"{flag}=off"
            elif tok in ("--flash-attn", "-fa"):
                if explicit(i) in ("on", "auto"):
                    out[i + 1] = "off"
                elif explicit(i) is None:  # bare flag (reads as on) -> explicit off
                    out[i] = f"{tok}=off"
        return out

    @staticmethod
    def _strip_mmproj_args(cmd: list[str]) -> list[str]:
        """Return cmd without the '--mmproj <path>' pair (text-only retry).
        Every other flag is preserved; a no-op when --mmproj is absent.
        """
        out: list[str] = []
        skip_value = False
        for tok in cmd:
            if skip_value:
                skip_value = False
                continue
            if tok == "--mmproj":
                skip_value = True
                continue
            out.append(tok)
        return out

    @staticmethod
    def _redacted_cmd_for_log(cmd: "list[str]") -> "list[str]":
        """Copy of cmd with the value after --api-key replaced by <redacted>."""
        out = list(cmd)
        if "--api-key" in out:
            ki = out.index("--api-key") + 1
            if ki < len(out):
                out[ki] = "<redacted>"
        return out

    def _start_llama_process(self, cmd: list[str], env: dict) -> None:
        """Spawn llama-server from cmd and start draining its output.

        Caller holds self._lock. Resets the stdout buffer, opens a fresh
        per-attempt tee log, launches the process, and starts the drain
        thread. Used for the initial start and the text-only mmproj retry.
        """
        # Defensive kill: if a concurrent load slipped past Phase 1
        # (because its `self._process` was None at the time) and already
        # stored a Popen handle here, drop that orphan before we overwrite
        # the reference. See issue #5161.
        self._kill_process()

        self._stdout_lines = []
        # Tee llama-server output to a dedicated log file so a post-mortem
        # in CI (or after a remote-debug session) has the full subprocess
        # trail even when the parent only stored the last 50 lines.
        self._llama_log_fh = None
        try:
            log_dir = _swa_cache_path().parent / "logs" / "llama-server"
            log_dir.mkdir(parents = True, exist_ok = True)
            self._llama_log_path = log_dir / f"llama-{int(time.time())}-port-{self._port}.log"
            self._llama_log_fh = open(
                self._llama_log_path,
                "w",
                encoding = "utf-8",
                buffering = 1,
            )
            logger.info(f"llama-server stdout/stderr -> {self._llama_log_path}")
        except OSError as e:
            # Best-effort; never block the load on logging.
            logger.debug(f"Could not open llama-server log file: {e}")
            self._llama_log_path = None

        # Log the argv per attempt (the text-only mmproj retry re-enters here
        # with --mmproj stripped), redacting the API key.
        logger.info(f"Starting llama-server: {' '.join(self._redacted_cmd_for_log(cmd))}")

        self._process = subprocess.Popen(
            cmd,
            stdout = subprocess.PIPE,
            stderr = subprocess.STDOUT,
            text = True,
            env = env,
            **_windows_hidden_subprocess_kwargs(),
            **_child_popen_kwargs(),
        )
        # Cross-session backstop: record the PID so a later startup can reap this
        # server if parent-death cleanup did not run (macOS / best-effort failure).
        self._record_server_pid(self._process.pid)

        # Start background thread to drain stdout and prevent pipe deadlock
        self._stdout_thread = threading.Thread(
            target = self._drain_stdout, daemon = True, name = "llama-stdout"
        )
        self._stdout_thread.start()

    def load_model(
        self,
        *,
        # Local mode: pass a path to a .gguf file
        gguf_path: Optional[str] = None,
        # Vision projection (mmproj) for local vision models
        mmproj_path: Optional[str] = None,
        # Separate MTP drafter for local Gemma loads (HF loads auto-resolve it)
        mtp_draft_path: Optional[str] = None,
        # HF mode: let llama-server download via -hf "repo:quant"
        hf_repo: Optional[str] = None,
        hf_variant: Optional[str] = None,
        hf_token: Optional[str] = None,
        # Common
        model_identifier: str,
        is_vision: bool = False,
        n_ctx: int = 4096,
        chat_template_override: Optional[str] = None,
        cache_type_kv: Optional[str] = None,
        speculative_type: Optional[str] = None,
        spec_draft_n_max: Optional[int] = None,
        tensor_parallel: bool = False,
        n_threads: Optional[int] = None,
        n_gpu_layers: Optional[int] = None,  # caller compat, unused
        n_parallel: int = 1,
        extra_args: Optional[List[str]] = None,
        # Route-level tensor->layer fallback retry: keep the layer split multi-GPU.
        preserve_multi_gpu_on_layer: bool = False,
    ) -> bool:
        """Start llama-server with a GGUF model.

        Two modes:
        - Local: ``gguf_path="/path/to/model.gguf"`` → uses ``-m``
        - HF:    ``hf_repo="...-GGUF", hf_variant="Q4_K_M"`` → uses ``-hf``

        Returns True if the server started and the health check passed.
        """
        # Raw load inputs so the runtime MTP-crash reload can replay this model
        # without MTP. Committed to _last_load_kwargs only on a healthy load.
        _pending_load_kwargs = {
            "gguf_path": gguf_path,
            "mmproj_path": mmproj_path,
            "mtp_draft_path": mtp_draft_path,
            "hf_repo": hf_repo,
            "hf_variant": hf_variant,
            "hf_token": hf_token,
            "model_identifier": model_identifier,
            "is_vision": is_vision,
            "n_ctx": n_ctx,
            "chat_template_override": chat_template_override,
            "cache_type_kv": cache_type_kv,
            "speculative_type": speculative_type,
            "spec_draft_n_max": spec_draft_n_max,
            "tensor_parallel": tensor_parallel,
            "n_threads": n_threads,
            "n_gpu_layers": n_gpu_layers,
            "n_parallel": n_parallel,
            "extra_args": list(extra_args) if extra_args is not None else None,
            # Replayed by _respawn_if_dead so a downgraded model stays multi-GPU.
            "preserve_multi_gpu_on_layer": preserve_multi_gpu_on_layer,
        }
        # Serialise the whole load so concurrent /load calls never leave two
        # llama-server processes alive (#5401 / #5161). Doesn't block /unload.
        with self._serial_load_lock:
            # In-app update swapping binaries: refuse fast (set under this lock,
            # so any in-flight load has drained) instead of using a half-swapped one.
            if getattr(self, "_llama_update_in_progress", False):
                raise RuntimeError("llama.cpp is updating; try again in a moment.")
            # Duplicate /load that raced past the route check: do nothing if the
            # live server already satisfies this request.
            if self._already_in_target_state(
                gguf_path = gguf_path,
                mtp_draft_path = mtp_draft_path,
                model_identifier = model_identifier,
                hf_variant = hf_variant,
                n_ctx = n_ctx,
                cache_type_kv = cache_type_kv,
                speculative_type = speculative_type,
                spec_draft_n_max = spec_draft_n_max,
                tensor_parallel = tensor_parallel,
                chat_template_override = chat_template_override,
                extra_args = extra_args,
                is_vision = is_vision,
                preserve_multi_gpu_on_layer = preserve_multi_gpu_on_layer,
            ):
                logger.info(
                    f"load_model: backend already in target state for "
                    f"'{model_identifier}', skipping reload"
                )
                # Retry probe only if a prior attempt didn't finish.
                if not self._audio_probed:
                    try:
                        detected = self._detect_audio_type_strict()
                        self._audio_probed = True
                    except Exception as exc:
                        logger.debug("Fast-path audio probe failed: %s", exc)
                        detected = None
                    if not self._apply_detected_audio(detected):
                        return False
                if not self._healthy:
                    return False
                return True

            self._cancel_event.clear()

            # ── Phase 1: kill old process (under lock, fast) ──────────
            with self._lock:
                self._kill_process()

            # Resolve llama-server now but defer a not-found error: a block-diffusion
            # GGUF uses the diffusion runner, and its arch is only known after the header.
            binary = self._find_llama_server_binary()

            # ── Phase 2: download (NO lock held, so cancel can proceed) ──
            # mtp_draft_path arrives set for local Gemma loads (detected
            # sibling); for -hf loads it's None here and resolved just below.
            # Scope HF_HUB_OFFLINE to the download block only when DNS is
            # dead; cleanup runs even on exception so a transient hiccup
            # can't quarantine future loads.
            if hf_repo:
                # Resolve the requested repo id to its cached canonical casing once,
                # up front, so the main GGUF and its companions (mmproj / MTP drafter)
                # all resolve from the same cache entry. Otherwise a case-variant
                # request resolves the main file from the canonical cache dir while the
                # companions keep the requested casing and miss the cached files.
                _resolved_repo = _resolve_repo_id_casing(hf_repo)
                if _resolved_repo != hf_repo:
                    logger.info(
                        "Using cached repo_id casing '%s' for requested '%s'",
                        _resolved_repo,
                        hf_repo,
                    )
                    hf_repo = _resolved_repo
                with _hf_offline_if_dns_dead():
                    model_path = self._download_gguf(
                        hf_repo = hf_repo,
                        hf_variant = hf_variant,
                        hf_token = hf_token,
                    )
                    # Auto-download mmproj for vision models unless opted out.
                    if is_vision and not mmproj_path and not extra_args_disable_mmproj(extra_args):
                        mmproj_path = self._download_mmproj(
                            hf_repo = hf_repo,
                            hf_token = hf_token,
                        )
                    # Auto-download the separate MTP drafter (e.g. Gemma) when
                    # the requested spec mode can use it. Repos with the head
                    # baked into the main GGUF (Qwen) have no mtp- sibling and
                    # this no-ops, so the size gate stays out of it: a separate
                    # drafter speeds up even sub-3B (Gemma E2B), and the resolver
                    # below decides the final emission. Skipped only when the
                    # user disabled MTP or drives --spec-type manually.
                    _spec_canon = _canonicalize_spec_mode(speculative_type) or "auto"
                    if (
                        not mtp_draft_path
                        and _spec_canon in ("auto", "mtp", "mtp+ngram")
                        and not _extra_args_set_spec_type(extra_args)
                    ):
                        mtp_draft_path = self._download_mtp(
                            hf_repo = hf_repo,
                            hf_token = hf_token,
                        )
            elif gguf_path:
                if not Path(gguf_path).is_file():
                    raise FileNotFoundError(f"GGUF file not found: {gguf_path}")
                model_path = gguf_path
            else:
                raise ValueError("Either gguf_path or hf_repo must be provided")

            # Set identifier early so _read_gguf_metadata can use it (DeepSeek).
            self._model_identifier = model_identifier

            # Read GGUF metadata (context_length, chat_template); header-only.
            self._read_gguf_metadata(model_path)

            if self._cancel_event.is_set():
                logger.info("Load cancelled after download phase")
                return False

            # Block-diffusion GGUFs (DiffusionGemma) cannot run on llama-server;
            # serve them with the diffusion runner (same OpenAI-compat interface).
            if self._is_diffusion:
                # Not a tensor/layer GGUF: clear any preserved-fallback flag from a
                # prior load (this path skips the command builder that clears it).
                self._layer_preserves_tensor_intent = False
                with self._lock:
                    if self._cancel_event.is_set():
                        logger.info("Load cancelled before diffusion server start")
                        return False
                    return self._start_diffusion_server(
                        model_path = model_path,
                        gguf_path = gguf_path,
                        hf_repo = hf_repo,
                        hf_variant = hf_variant,
                        model_identifier = model_identifier,
                        n_ctx = n_ctx,
                        extra_args = extra_args,
                    )

            if not binary:
                # distinguish a transiently locked binary (antivirus / in-flight
                # install) from a missing one so the user retries, not reinstalls
                locked = self._find_llama_server_binary(include_denied = True)
                if locked:
                    raise RuntimeError(
                        f"llama-server at {locked} is temporarily unavailable "
                        "(access-denied; antivirus or an in-flight install). "
                        "Retry the load once it is released."
                    )
                # Reached only after the diffusion early-return above, so this is a
                # genuine llama-server-backed GGUF with no runtime. Raise the typed
                # error so /load returns the actionable 400 (not a generic 500), the
                # same message remote validation already shows.
                raise LlamaServerNotFoundError(LLAMA_SERVER_NOT_FOUND_DETAIL)

            # Outside ``self._lock`` so /unload, /cancel, /status aren't
            # blocked. ``unload_model`` also records the kill, so the
            # frontend /unload+/load Apply path engages the wait here even
            # without an in-process kill.
            self._wait_for_vram_settle(since_kill = self._last_kill_monotonic)

            # ── Phase 3: start llama-server (under lock) ──────────────
            with self._lock:
                # Re-check cancel inside lock
                if self._cancel_event.is_set():
                    logger.info("Load cancelled before server start")
                    return False

                self._port = self._find_free_port()

                # Select GPU(s) from model size + estimated KV cache. Seed
                # safe defaults before probing so the except path has valid
                # state to publish.
                ctx_override = parse_ctx_override(extra_args)
                requested_ctx = resolve_requested_ctx(extra_args, n_ctx)
                cache_override = parse_cache_override(extra_args)
                # Budget the heavier of asymmetric --cache-type-k/-v extras (they
                # win per axis at launch, appended last); resolve_cache_type_kv only
                # returns the last-wins type, which under-reserves the heavier axis.
                # The user's extras still set the real (possibly asymmetric) child
                # cache, so this only affects the reserve, not the emitted command.
                _extras_cache = _extra_args_main_cache_type_for_budget(extra_args)
                cache_type_kv = _extras_cache if _extras_cache is not None else cache_type_kv
                _cache_type_from_env = False
                if cache_type_kv is None:
                    # Param/extras set nothing, so the child inherits
                    # LLAMA_ARG_CACHE_TYPE_K/_V. Adopt a heavier env type (f32) for
                    # the reserve only; the launch does NOT re-emit it (that would
                    # rewrite an asymmetric K=f32,V=f16 env into symmetric flags),
                    # so _cache_type_from_env keeps it out of the emitted flags.
                    cache_type_kv = _env_main_cache_type_for_budget()
                    _cache_type_from_env = cache_type_kv is not None
                # A user --split-mode in extras last-wins-overrides the toggle, and
                # an inherited tensor LLAMA_ARG_SPLIT_MODE flips it on (the child
                # would run tensor unbudgeted otherwise). The duplicate-load matchers
                # use the same helper so a healthy env-driven tensor server matches.
                split_mode_override = parse_split_mode_override(extra_args)
                tensor_parallel = _effective_tensor_parallel(extra_args, tensor_parallel)
                # Tensor mode aborts on a quantized KV cache, so drop it for the
                # tensor attempt (and strip any inherited/explicit --cache-type
                # that would re-impose it when appended last). Layer split does
                # support it, so remember the dropped type and the original extras
                # to restore (verbatim, incl. an asymmetric K/V) if we later fall
                # back to layer split below.
                _tensor_dropped_cache_type_kv: Optional[str] = None
                _tensor_dropped_extra_args: Optional[list] = None
                # Tensor mode rejects any quantized axis. cache_type_kv is the
                # heavier-by-bytes budget type, which can mask a quantized axis (an
                # f16 budget hides a paired q4_0), so also test each explicit
                # --cache-type-k/-v extra, not just the budget type.
                _ck_extra, _cv_extra = parse_cache_override_per_axis(extra_args)
                _cache_non_tensor_safe = any(
                    c and c.strip().lower() not in self._TENSOR_PARALLEL_KV_TYPES
                    for c in (cache_type_kv, _ck_extra, _cv_extra)
                )
                if tensor_parallel and _cache_non_tensor_safe:
                    logger.info(
                        "Tensor parallelism requires a non-quantized KV cache; "
                        "ignoring cache type %s for the tensor attempt.",
                        cache_type_kv,
                    )
                    _tensor_dropped_cache_type_kv = cache_type_kv
                    cache_type_kv = None
                    if extra_args:
                        # Keep the originals so a layer downgrade restores the real
                        # (possibly asymmetric) --cache-type-k/-v the layer path
                        # supports, not just the scalar heavier type.
                        _tensor_dropped_extra_args = list(extra_args)
                        extra_args = strip_shadowing_flags(
                            extra_args,
                            strip_context = False,
                            strip_cache = True,
                            strip_spec = False,
                            strip_template = False,
                            strip_split_mode = False,
                        )
                    # The launch keeps an inherited tensor-safe env cache type (the
                    # env cleanup only pops quantized ones), so re-adopt a heavier
                    # env type (f32) for the budget here too -- mirrors the initial
                    # adoption, which was skipped because the param/extras set the
                    # (now-dropped) quantized type. Else the child allocates f32 KV
                    # against an f16 budget.
                    _env_tensor_cache = _env_main_cache_type_for_budget()
                    if _env_tensor_cache is not None:
                        cache_type_kv = _env_tensor_cache
                        _cache_type_from_env = True
                if ctx_override is not None and ctx_override > 0:
                    logger.info(f"User --ctx-size {ctx_override} honored; skipping auto-reduce")
                if cache_override is not None:
                    _ck, _cv = parse_cache_override_per_axis(extra_args)
                    logger.info(
                        f"User --cache-type-k/-v (k={_ck}, v={_cv}) honored; "
                        "KV estimate budgets the heavier axis"
                    )
                if split_mode_override is not None:
                    logger.info(
                        f"User --split-mode {split_mode_override} honored; "
                        "reconciled into tensor_parallel state"
                    )
                effective_ctx = requested_ctx if requested_ctx > 0 else (self._context_length or 0)
                max_available_ctx = self._context_length or effective_ctx
                gpus: list[tuple[int, int]] = []
                # Keep fit-budget and launch-flag mmproj resolution in sync.
                launch_mmproj_path = None
                if not extra_args_disable_mmproj(extra_args):
                    launch_mmproj_path = self._resolve_launch_mmproj_path(
                        model_path = model_path,
                        mmproj_path = mmproj_path,
                    )
                # Need both a resolved mmproj AND the config vision flag; a stray
                # mmproj passing the family-name heuristic must not flip a non-VLM
                # GGUF into vision mode.
                effective_is_vision = bool(launch_mmproj_path) and bool(is_vision)
                if is_vision and not effective_is_vision:
                    logger.warning(
                        "Vision-capable GGUF loaded without a usable mmproj; "
                        "image input will be disabled for this session"
                    )
                model_size = None  # set in the fit try; used by the APU RAM guard
                # Layer-fallback min GPUs; raised below on a tensor downgrade. Bound
                # before the try so the --fit-on except path still has it (no UnboundLocal).
                _layer_min_gpus = 1
                try:
                    gguf_size = self._get_gguf_size_bytes(model_path)
                    # Include GPU-loaded mmproj in the fit budget (#5825).
                    mmproj_size = (
                        self._mmproj_vram_bytes(launch_mmproj_path) if effective_is_vision else 0
                    )
                    model_size = gguf_size + mmproj_size
                    # 2-tuple gpus for existing logic + a total map for the absolute
                    # per-GPU headroom (correct when the GPU is already partly used).
                    _gpu_mem = self._get_gpu_memory()
                    gpus = [(idx, free) for idx, free, _t in _gpu_mem]
                    total_by_idx = {idx: total for idx, _f, total in _gpu_mem}

                    def _gpu_usable(g, frac = _CTX_FIT_VRAM_FRACTION):
                        # Per-GPU usable budget for ranking: free - (1-frac)*total.
                        # Callers pass the ACTIVE fraction so the ranking matches the
                        # budget the fit then tests (else mixed totals mis-order).
                        idx, free = g
                        t = total_by_idx.get(idx, 0)
                        if t > 0:
                            return free - (1.0 - frac) * t
                        return free * frac

                    def _pool_budget_mib(subset, frac):
                        # Sum each GPU's own usable budget. Pooling free and total
                        # separately would let an unknown-total GPU (MIG/vGPU/N/A)
                        # add full free with no cushion among known-total GPUs.
                        return sum(max(0.0, _gpu_usable(g, frac)) for g in subset)

                    # Resolve effective context: 0 means let llama-server use
                    # the model's native length. Only expand to a known native
                    # length if metadata exists; else keep 0 as a sentinel.
                    if requested_ctx > 0:
                        effective_ctx = requested_ctx
                    elif self._context_length is not None:
                        effective_ctx = self._context_length
                    else:
                        effective_ctx = 0
                    original_ctx = effective_ctx
                    # Default UI ceiling to the native context length;
                    # GPU/VRAM-fit logic below may shrink it on limited HW.
                    max_available_ctx = self._context_length or effective_ctx

                    # Will MTP engage? If so, auto-fit reserves draft-model VRAM.
                    # Mirrors _build_speculative_flags: forced mtp/mtp+ngram always
                    # engage; auto only on an MTP model >= 3B; ngram/off never. A
                    # separate drafter (Gemma) counts as an MTP model.
                    _mtp_canonical = _canonicalize_spec_mode(speculative_type)
                    _mtp_effective = _mtp_canonical or "auto"
                    _mtp_size_for_fit = _extract_model_size_b(model_identifier)
                    # Sub-3B drops MTP only for an embedded head; a separate
                    # drafter (Gemma) engages and needs its VRAM reserved.
                    _mtp_sub_3b_for_fit = (
                        _mtp_size_for_fit is not None
                        and _mtp_size_for_fit < _MTP_MIN_SIZE_B
                        and not bool(mtp_draft_path)
                    )
                    # LLAMA_ARG_SPEC_TYPE only reaches the child when neither extras
                    # nor Studio emit a spec flag (mode "off", no user --spec-type),
                    # since _build_speculative_flags emits one for every other mode.
                    # Consult the env for the reserve only then, else a stale MTP env
                    # would over-reserve.
                    _spec_env: Mapping[str, str] = (
                        os.environ
                        if (not _extra_args_set_spec_type(extra_args) and _mtp_canonical == "off")
                        else {}
                    )
                    # Extras can run MTP even when Studio suppresses its own emission.
                    _user_mtp_via_extras = _extra_args_requests_mtp(extra_args, env = _spec_env)
                    # A non-MTP model-based draft mode (draft-simple/draft-eagle3) in
                    # extras also loads a separate draft model that needs reserving;
                    # engage only when extras actually name a drafter for it.
                    _user_draft_via_extras = _extra_args_requests_separate_draft(
                        extra_args, env = _spec_env
                    ) and bool(_extra_args_mtp_draft_path(extra_args))
                    # Mirror _build_speculative_flags: reserve only for MTP the launch
                    # resolver will actually emit (needs a head/drafter and a binary
                    # that supports --spec-type mtp).
                    _mtp_model_for_fit = bool(
                        self._nextn_predict_layers
                        or _is_mtp_model_name(model_identifier, model_path)
                        or bool(mtp_draft_path)
                    ) and not (
                        # Drafterless Gemma falls back to ngram-mod; reserve no
                        # drafter VRAM for it (mirrors the launch resolver).
                        _is_gemma_mtp_name(model_identifier, model_path)
                        and not mtp_draft_path
                        and not self._nextn_predict_layers
                    )
                    _mtp_binary_ok = True
                    _mtp_probe_raised = False
                    if not _user_mtp_via_extras:
                        try:
                            _mtp_binary_ok = bool(
                                (self.probe_server_capabilities(binary) or {}).get("mtp_token")
                            )
                        except Exception:
                            _mtp_binary_ok = False
                            _mtp_probe_raised = True
                    _auto_studio_mtp = (
                        not _extra_args_set_spec_type(extra_args)
                        and _mtp_model_for_fit
                        and (
                            _mtp_effective in ("mtp", "mtp+ngram")
                            or (_mtp_effective == "auto" and not _mtp_sub_3b_for_fit)
                        )
                        and (
                            _mtp_binary_ok
                            # Reserve on a raised (uncached) probe too: it re-probes in
                            # _build_speculative_flags and may still engage MTP (embedded
                            # head or separate drafter -- _mtp_model_for_fit covers both).
                            or _mtp_probe_raised
                        )
                    )
                    _mtp_will_engage = bool(
                        _user_mtp_via_extras or _user_draft_via_extras or _auto_studio_mtp
                    )
                    # The duplicated full target-KV copy (ctx_tgt) is an MTP-only
                    # cost: the MTP head runs a second context over the target
                    # model's own KV geometry. The separate-drafter spec modes
                    # (draft-simple/draft-eagle3, reached via _user_draft_via_extras)
                    # load a small distinct drafter with its own KV and keep no such
                    # copy, so only charge it when the engaged mode is truly MTP.
                    _engaged_is_mtp = bool(_user_mtp_via_extras or _auto_studio_mtp)

                    # Effective draft depth: extras win (last-wins at launch), else
                    # the field, else the platform default (2 GPU / 3 CPU).
                    _extra_n_max = _extra_args_spec_draft_n_max(extra_args)
                    _mtp_eff_n_max = _extra_n_max if _extra_n_max is not None else spec_draft_n_max
                    if _mtp_eff_n_max is None:
                        _mtp_eff_n_max = 2 if gpus else 3
                    # Separate-drafter weights live on GPU (an embedded head is
                    # already in model_size). Size the drafter the launch loads, by
                    # precedence: extras --model-draft (last-wins), else Studio's
                    # emitted mtp_draft_path, else the env drafter. Sizing the wrong
                    # one would under-reserve and OOM.
                    _cli_draft_for_budget = _extra_args_mtp_draft_path(extra_args, env = {})
                    _studio_draft_for_budget = (
                        mtp_draft_path
                        if (
                            _mtp_will_engage
                            and mtp_draft_path
                            and not _extra_args_set_spec_type(extra_args)
                        )
                        else None
                    )
                    _env_draft_for_budget = _extra_args_mtp_draft_path([], env = os.environ)
                    _mtp_draft_for_budget = (
                        _cli_draft_for_budget or _studio_draft_for_budget or _env_draft_for_budget
                    )
                    # Drafter offloaded to CPU keeps its weights+KV off the GPU, so
                    # drop it from the budget (an embedded head stays in the model).
                    # Consult the env too: the child honors LLAMA_ARG_N_GPU_LAYERS_DRAFT.
                    _draft_on_cpu = _extra_args_draft_offloaded_to_cpu(extra_args, env = os.environ)
                    if _draft_on_cpu:
                        _mtp_draft_for_budget = None
                    _mtp_draft_weights = 0
                    if _mtp_draft_for_budget:
                        try:
                            _mtp_draft_weights = self._get_gguf_size_bytes(_mtp_draft_for_budget)
                        except Exception:
                            _mtp_draft_weights = 0
                    # Draft K/V types (f16 by default; independent extras overrides).
                    _mtp_draft_ck, _mtp_draft_cv = _extra_args_draft_cache_types(extra_args)

                    # Byte-accurate reserve when dims allow, else None -> flat fallback.
                    mtp_overhead_fn: Optional[Callable[[int], int]] = None
                    # True when the byte reserve is the drafter weights ONLY because
                    # its KV couldn't be sized; the flat fraction must then stay on
                    # as the cushion for that unsized draft KV (it is not covered by
                    # the weights-only mtp_overhead_fn).
                    _mtp_kv_unsized = False
                    if _mtp_will_engage:
                        _probe_ctx = self._context_length or (
                            effective_ctx if effective_ctx > 0 else 4096
                        )
                        _draft_kv_probe = self._mtp_draft_kv_bytes(
                            _probe_ctx,
                            drafter_path = _mtp_draft_for_budget,
                            draft_cache_type_k = _mtp_draft_ck,
                            draft_cache_type_v = _mtp_draft_cv,
                            n_parallel = n_parallel,
                        )
                        if (
                            self._estimate_mtp_overhead_bytes(
                                _probe_ctx,
                                spec_draft_n_max = _mtp_eff_n_max,
                                draft_cache_type_k = _mtp_draft_ck,
                                draft_cache_type_v = _mtp_draft_cv,
                                drafter_path = _mtp_draft_for_budget,
                                draft_weights_bytes = _mtp_draft_weights,
                                n_parallel = n_parallel,
                                mtp_keeps_target_ctx = _engaged_is_mtp,
                            )
                            is not None
                        ):
                            # Reserve is weights-only when the draft KV is unsizable.
                            _mtp_kv_unsized = _draft_kv_probe is None

                            # Closure binding this load's draft params; ctx varies.
                            def mtp_overhead_fn(
                                ctx: int,
                                _n: int = _mtp_eff_n_max,
                                _ck: Optional[str] = _mtp_draft_ck,
                                _cv: Optional[str] = _mtp_draft_cv,
                                _dp: Optional[str] = _mtp_draft_for_budget,
                                _w: int = _mtp_draft_weights,
                                _np: int = n_parallel,
                                _mtp: bool = _engaged_is_mtp,
                            ) -> int:
                                v = self._estimate_mtp_overhead_bytes(
                                    ctx,
                                    spec_draft_n_max = _n,
                                    draft_cache_type_k = _ck,
                                    draft_cache_type_v = _cv,
                                    drafter_path = _dp,
                                    draft_weights_bytes = _w,
                                    n_parallel = _np,
                                    mtp_keeps_target_ctx = _mtp,
                                )
                                return v if v is not None else 0

                    def _mtp_bytes(ctx: int) -> int:
                        return mtp_overhead_fn(ctx) if mtp_overhead_fn is not None else 0

                    # Effective micro-batch (a user --ubatch override scales the
                    # compute buffer); None -> the 512 default in the estimate.
                    _effective_ubatch = _extra_args_n_ubatch(extra_args)

                    def _cc_bytes(ctx: int, n_gpus: int = 1) -> int:
                        # Context-linear compute-buffer growth (flash-attn KQ mask +
                        # attention scratch); the flat _compute_buffer_pipeline folded
                        # into model_size_fit only covers ctx -> 0. Charged per
                        # candidate context so the fit can't over-pin and spill. The
                        # rate depends on the KV cache type (quantized adds a dequant
                        # scratch), so pass it through. In a layer split this buffer is
                        # replicated on EVERY device (measured ~equal per GPU), so scale
                        # by the device count; a large model at high context otherwise
                        # under-reserves ~(n-1)x it (e.g. Qwen3.5-397B on 3 GPUs).
                        return max(1, n_gpus) * self._compute_buffer_ctx_bytes(
                            ctx, _effective_ubatch, cache_type_kv
                        )

                    # Layer-split compute buffer (one lump; tensor mode reserves it
                    # per device in _plan_tensor_parallel). Context-independent, so
                    # fold it into the model footprint for the branches below. Falls
                    # back to the flat reserve when dims are missing (returns 0), a
                    # safe upper bound since the tensor buffer >= the layer one.
                    _compute_buffer_pipeline = self._estimate_compute_buffer_bytes(
                        n_ubatch = _effective_ubatch,
                        n_parallel = n_parallel,
                        per_device_tensor = False,
                    )
                    if _compute_buffer_pipeline <= 0:
                        _compute_buffer_pipeline = (
                            self._TENSOR_PARALLEL_BUFFER_RESERVE_MIB * 1024 * 1024
                        )

                    # Layer split adds a fixed per-device overhead on every GPU. The
                    # folded buffer covers one device; reserve the extra devices'
                    # share so a k-GPU split can't pin a context that OOMs a device
                    # (k=1 adds nothing).
                    _pipeline_overhead_bytes = self._PIPELINE_PER_DEVICE_OVERHEAD_MIB * 1024 * 1024

                    # Auto-cap context to fit VRAM and select GPUs. Explicit n_ctx:
                    # honor it, cap only if it fits no combination. Auto (native):
                    # prefer fewer GPUs with reduced context (multi-GPU is slower).
                    gpu_indices, use_fit = None, True
                    # Per-GPU weight proportions for tensor mode (None = even).
                    tp_tensor_split: Optional[list[int]] = None
                    explicit_ctx = requested_ctx > 0
                    # Flat MTP reserve fraction: used only as the fallback when the
                    # byte-accurate mtp_overhead_fn can't size the draft KV (dims
                    # unavailable, or _mtp_kv_unsized = weights-only). A separate
                    # drafter on CPU uses no GPU (no reserve); an embedded head is on
                    # GPU regardless of draft-offload flags (keep its reserve).
                    _flat_mtp_engages = _mtp_will_engage and (
                        mtp_overhead_fn is None or _mtp_kv_unsized
                    )
                    _draft_cpu_no_embedded = _draft_on_cpu and not self._nextn_predict_layers
                    # MTP reserves GPU VRAM unless its only drafter is a separate
                    # CPU-offloaded one (an embedded head stays on GPU). The tensor
                    # path reserves like the layer path; gate both on this.
                    _mtp_reserves_gpu = _mtp_will_engage and not _draft_cpu_no_embedded
                    _flat_mtp_reserve = (
                        _MTP_VRAM_RESERVE_FRAC
                        if (_flat_mtp_engages and not _draft_cpu_no_embedded)
                        else 0.0
                    )
                    _pin_fraction = self._GPU_PIN_VRAM_FRACTION - _flat_mtp_reserve

                    # Charge the soft overhead _CTX_FIT_VRAM_FRACTION under-covers on tight
                    # tiers, gated so plain dense loads (#5106) only pay the CUDA-ctx base.
                    # CUDA/cuBLAS context is discrete-GPU only (not Metal); the mmproj and
                    # MTP draft-graph buffers exist on every backend.
                    _soft_overhead = self._CUDA_CONTEXT_RESERVE_BYTES if gpus else 0
                    if effective_is_vision and mmproj_size > 0:
                        _soft_overhead += int(mmproj_size * (self._MMPROJ_VRAM_SAFETY - 1.0))
                    if _mtp_reserves_gpu:
                        _soft_overhead += self._MTP_DRAFT_COMPUTE_BYTES
                    model_size_fit = model_size + _compute_buffer_pipeline + _soft_overhead

                    def _subset_model_size(n_gpus: int) -> int:
                        return model_size_fit + max(0, n_gpus - 1) * _pipeline_overhead_bytes

                    # Unified-memory budget (0 off Apple Silicon) for the no-GPU Metal cap below.
                    _apple_budget_mib = self._apple_metal_memory_budget_bytes() // (1024 * 1024)

                    def _restore_after_tensor_downgrade():
                        # Restore the quantized KV + extras tensor dropped (layer
                        # split supports them), minus --split-mode.
                        nonlocal cache_type_kv, _cache_type_from_env, extra_args
                        if _tensor_dropped_cache_type_kv is not None:
                            cache_type_kv = _tensor_dropped_cache_type_kv
                            _cache_type_from_env = False
                        extra_args = strip_split_mode_only(
                            _tensor_dropped_extra_args
                            if _tensor_dropped_extra_args is not None
                            else extra_args
                        )

                    # The route fallback retry is tensor-off; keep it multi-GPU.
                    if preserve_multi_gpu_on_layer:
                        _layer_min_gpus = max(_layer_min_gpus, len(gpus))

                    if tensor_parallel and self._tensor_split_aborts(binary, model_identifier):
                        # Aborted on tensor for this model this session (#6415); skip
                        # tensor upfront, layer split serves it.
                        logger.info(
                            "Tensor parallelism skipped: this llama.cpp build aborted "
                            "on --split-mode tensor for this model earlier this "
                            "session; using layer split across %d GPU(s).",
                            len(gpus),
                        )
                        tensor_parallel = False
                        # Keep the multi-GPU request (gated on it, not the cache).
                        _layer_min_gpus = max(_layer_min_gpus, len(gpus))
                        _restore_after_tensor_downgrade()

                    # Tensor mode replicates a compute buffer on every GPU, so drop
                    # GPUs below that reserve from the set up front (gpu_indices
                    # becomes the CUDA_VISIBLE_DEVICES mask, fully excluding them).
                    tp_gpus = gpus
                    if tensor_parallel:
                        # Deterministic per-device compute buffer (replicated on
                        # every device in tensor mode); flat fallback when dims
                        # are unavailable. _plan_tensor_parallel uses the same.
                        _tp_reserve_bytes = self._estimate_compute_buffer_bytes(
                            n_ubatch = _effective_ubatch,
                            n_parallel = n_parallel,
                            per_device_tensor = True,
                        )
                        reserve_mib = (
                            _tp_reserve_bytes // (1024 * 1024)
                            if _tp_reserve_bytes > 0
                            else self._TENSOR_PARALLEL_BUFFER_RESERVE_MIB
                        )
                        # Admit by usable budget (free - (1-frac)*total), not raw
                        # free: a partly-used big card can clear the reserve on raw
                        # free yet have no budget left.
                        tp_gpus = [g for g in gpus if _gpu_usable(g) >= reserve_mib]

                    if tensor_parallel and len(tp_gpus) < 2:
                        # Tensor parallelism needs >= 2 usable GPUs. On a single
                        # GPU --split-mode tensor is a no-op; with 0 GPUs (CPU-only
                        # or probe failed) it must not reach llama-server; and a
                        # GPU below the buffer reserve can't participate. Drop the
                        # flag and fall through to normal layer/CPU allocation.
                        logger.info(
                            "Tensor parallelism requested but only %d of %d GPU(s) "
                            "have enough free VRAM for the compute buffer; "
                            "ignoring (needs >= 2).",
                            len(tp_gpus),
                            len(gpus),
                        )
                        tensor_parallel = False
                        # GPUs below tensor's compute-buffer reserve can still do layer
                        # split, so keep multi-GPU (mirrors the budget/geometry drops);
                        # _select_gpus caps unusable cards.
                        if len(gpus) >= 2:
                            _layer_min_gpus = max(_layer_min_gpus, len(gpus))
                        # Layer split supports a quantized KV the tensor attempt
                        # dropped; restore the original cache type + extras (minus
                        # --split-mode) so the layer launch re-emits them.
                        _restore_after_tensor_downgrade()

                    if tensor_parallel and tp_gpus:
                        # Pooled usable budget (after each device's compute buffer)
                        # must hold the non-shrinkable footprint: weights + the MTP
                        # reserve. The planner can shrink ctx/KV, not these.
                        _tp_weight_budget_mib = (
                            sum(_gpu_usable(g) for g in tp_gpus) - len(tp_gpus) * reserve_mib
                        )
                        _tp_flat_mtp = 2 * 1024**3  # flat reserve when dims unavailable
                        if not _mtp_reserves_gpu:
                            # No MTP, or its only drafter is CPU-offloaded (no GPU).
                            _tp_mtp_floor = 0
                        elif mtp_overhead_fn is not None and not _mtp_kv_unsized:
                            _tp_mtp_floor = _mtp_bytes(
                                min(2048, effective_ctx) if effective_ctx > 0 else 2048
                            )
                        else:
                            # Dims unavailable / weights-only: tensor mode has no
                            # --fit valve, so keep the flat reserve as the unsized-KV
                            # cushion, never below the known byte reserve.
                            _tp_mtp_floor = max(
                                _tp_flat_mtp,
                                _mtp_bytes(min(2048, effective_ctx) if effective_ctx > 0 else 2048),
                            )
                        _tp_required_mib = (model_size + _tp_mtp_floor + _soft_overhead) / (
                            1024 * 1024
                        )
                        if _tp_weight_budget_mib <= _tp_required_mib:
                            logger.info(
                                "Tensor parallelism requested but the pooled VRAM "
                                "budget cannot hold the weights, MTP reserve, and "
                                "per-device compute buffers; falling back to layer split."
                            )
                            tensor_parallel = False
                            # Weights needed >1 card, so keep multi-GPU across the
                            # usable tensor GPUs.
                            if len(tp_gpus) >= 2:
                                _layer_min_gpus = max(_layer_min_gpus, len(tp_gpus))
                            # Restore the dropped quantized KV + cache extras (minus
                            # --split-mode); layer split supports them.
                            _restore_after_tensor_downgrade()

                    if tensor_parallel and tp_gpus:
                        # Tensor-parallel allocation; see _plan_tensor_parallel.
                        target_ctx = (
                            effective_ctx
                            if explicit_ctx
                            else (self._context_length or effective_ctx)
                        )
                        # When the draft KV couldn't be sized (weights-only reserve),
                        # the planner's mtp_overhead_fn is non-None but covers only
                        # weights, so pass the flat cushion for the unsized KV (else
                        # the binary search spends it on context).
                        _tp_unsized_mtp_reserve = (
                            2 * 1024**3 if (_mtp_reserves_gpu and _mtp_kv_unsized) else 0
                        )
                        (
                            effective_ctx,
                            max_available_ctx,
                            gpu_indices,
                            tp_tensor_split,
                        ) = self._plan_tensor_parallel(
                            tp_gpus,
                            model_size,
                            target_ctx,
                            cache_type_kv = cache_type_kv,
                            n_parallel = n_parallel,
                            mtp_engaged = _mtp_reserves_gpu,
                            mtp_overhead_fn = mtp_overhead_fn,
                            mtp_flat_reserve_bytes = _tp_unsized_mtp_reserve,
                            # Report the UI ceiling from native ctx, not the
                            # explicit small request.
                            max_target_ctx = self._context_length or target_ctx,
                            total_by_idx = total_by_idx,
                            n_ubatch = _effective_ubatch,
                            soft_overhead_bytes = _soft_overhead,
                        )
                        use_fit = False
                    elif gpus and self._can_estimate_kv() and effective_ctx > 0:
                        # Compute the largest hardware-aware cap from the model's
                        # native context across all usable GPU subsets (for UI
                        # bounds), independent of the currently requested context.
                        native_ctx_for_cap = self._context_length or effective_ctx
                        if native_ctx_for_cap > 0:
                            ranked_for_cap = sorted(
                                gpus,
                                key = lambda g: _gpu_usable(
                                    g, _CTX_FIT_VRAM_FRACTION - _flat_mtp_reserve
                                ),
                                reverse = True,
                            )
                            best_cap = 0
                            _cap_fraction = _CTX_FIT_VRAM_FRACTION - _flat_mtp_reserve
                            for n_gpus in range(1, len(ranked_for_cap) + 1):
                                subset = ranked_for_cap[:n_gpus]
                                # Per-GPU-consistent pool budget (fixes mixed
                                # known/unknown totals); pass it as an absolute
                                # budget so the fit and the check below agree.
                                pool_budget = _pool_budget_mib(subset, _cap_fraction)
                                _ms = _subset_model_size(n_gpus)
                                # Compute buffer is replicated per device in a layer
                                # split, so scale the context term by the subset size.
                                _cc_sub = lambda c, n = n_gpus: _cc_bytes(c, n)
                                capped = self._fit_context_to_vram(
                                    native_ctx_for_cap,
                                    pool_budget,
                                    _ms,
                                    cache_type_kv,
                                    n_parallel = n_parallel,
                                    mtp_engaged = _mtp_reserves_gpu,
                                    mtp_overhead_fn = mtp_overhead_fn,
                                    compute_ctx_bytes_fn = _cc_sub,
                                    budget_frac = 1.0,
                                    total_mib = None,
                                )
                                kv = self._estimate_kv_cache_bytes(
                                    capped, cache_type_kv, n_parallel = n_parallel
                                )
                                footprint_mib = (
                                    _ms + kv + _mtp_bytes(capped) + _cc_sub(capped)
                                ) / (1024 * 1024)
                                if footprint_mib <= pool_budget:
                                    best_cap = max(best_cap, capped)
                            if best_cap > 0:
                                max_available_ctx = best_cap
                            else:
                                # Weights exceed 90% of every GPU subset, so no
                                # context fits. Anchor the UI "safe zone" at 4096
                                # so the slider warns above the fallback.
                                max_available_ctx = min(4096, native_ctx_for_cap)

                        if explicit_ctx:
                            # Honor the requested context verbatim. If it fits,
                            # pin GPUs and skip --fit; else ship -c <ctx> --fit
                            # on and let llama-server flex -ngl (CPU offload).
                            requested_total = (
                                model_size_fit
                                + self._estimate_kv_cache_bytes(
                                    effective_ctx, cache_type_kv, n_parallel = n_parallel
                                )
                                + _mtp_bytes(effective_ctx)
                                + _cc_bytes(effective_ctx)
                            )
                            # The compute buffer is replicated on every device in a
                            # layer split; fold it into the per-device reserve so a
                            # multi-GPU pin sizes each card for its own copy.
                            gpu_indices, use_fit = self._select_gpus(
                                requested_total,
                                gpus,
                                usable_fraction = _pin_fraction,
                                total_by_idx = total_by_idx,
                                per_device_overhead_bytes = _pipeline_overhead_bytes
                                + _cc_bytes(effective_ctx),
                                min_gpus = _layer_min_gpus,
                            )
                            # No silent shrink: effective_ctx stays == requested_ctx.
                        else:
                            # Auto context: prefer fewer GPUs, cap to fit. Same
                            # headroom threshold as _select_gpus (#5106). Rank by the
                            # active pin fraction so the order matches the fit budget.
                            pin_fraction = _pin_fraction
                            ranked = sorted(
                                gpus, key = lambda g: _gpu_usable(g, pin_fraction), reverse = True
                            )
                            # Skips _select_gpus, so apply its cap: count only cards
                            # whose usable VRAM clears the per-device layer overhead.
                            _pipeline_overhead_mib = _pipeline_overhead_bytes / (1024 * 1024)
                            _auto_min_gpus = max(
                                1,
                                min(
                                    _layer_min_gpus,
                                    sum(
                                        1
                                        for g in ranked
                                        if _gpu_usable(g, pin_fraction) > _pipeline_overhead_mib
                                    )
                                    or 1,
                                ),
                            )
                            for n_gpus in range(_auto_min_gpus, len(ranked) + 1):
                                subset = ranked[:n_gpus]
                                pool_budget = _pool_budget_mib(subset, pin_fraction)
                                _ms = _subset_model_size(n_gpus)
                                # Compute buffer is replicated per device in a layer
                                # split, so scale the context term by the subset size.
                                _cc_sub = lambda c, n = n_gpus: _cc_bytes(c, n)
                                capped = self._fit_context_to_vram(
                                    effective_ctx,
                                    pool_budget,
                                    _ms,
                                    cache_type_kv,
                                    n_parallel = n_parallel,
                                    mtp_engaged = _mtp_reserves_gpu,
                                    mtp_overhead_fn = mtp_overhead_fn,
                                    compute_ctx_bytes_fn = _cc_sub,
                                    budget_frac = 1.0,
                                    total_mib = None,
                                )
                                kv = self._estimate_kv_cache_bytes(
                                    capped, cache_type_kv, n_parallel = n_parallel
                                )
                                footprint_mib = (
                                    _ms + kv + _mtp_bytes(capped) + _cc_sub(capped)
                                ) / (1024 * 1024)
                                if footprint_mib <= pool_budget:
                                    effective_ctx = capped
                                    gpu_indices = sorted(idx for idx, _ in subset)
                                    use_fit = False
                                    break
                            else:
                                # Native ctx doesn't fit. Drop to 4096 and
                                # re-check before --fit on: a model overflowing
                                # at 131k may pin fine with a 4096 KV (#5106).
                                effective_ctx = min(4096, effective_ctx)
                                if effective_ctx > 0:
                                    for n_gpus in range(_auto_min_gpus, len(ranked) + 1):
                                        subset = ranked[:n_gpus]
                                        kv = self._estimate_kv_cache_bytes(
                                            effective_ctx,
                                            cache_type_kv,
                                            n_parallel = n_parallel,
                                        )
                                        footprint_mib = (
                                            _subset_model_size(n_gpus)
                                            + kv
                                            + _mtp_bytes(effective_ctx)
                                            + _cc_bytes(effective_ctx, n_gpus)
                                        ) / (1024 * 1024)
                                        if footprint_mib <= _pool_budget_mib(subset, pin_fraction):
                                            gpu_indices = sorted(idx for idx, _ in subset)
                                            use_fit = False
                                            break

                    elif gpus:
                        # Can't estimate KV -- file-size-only check; keep the
                        # ceiling at native context (already the default).
                        logger.debug(
                            "Falling back to file-size-only GPU selection",
                            model_size_gb = round(model_size / (1024**3), 2),
                        )
                        # Add the byte-accurate MTP reserve here too when it is
                        # available; otherwise _pin_fraction carries the flat
                        # fallback (the two are mutually exclusive by design).
                        _fs_total = model_size_fit + _mtp_bytes(
                            self._context_length or effective_ctx or 4096
                        )
                        gpu_indices, use_fit = self._select_gpus(
                            _fs_total,
                            gpus,
                            usable_fraction = _pin_fraction,
                            total_by_idx = total_by_idx,
                            per_device_overhead_bytes = _pipeline_overhead_bytes,
                            min_gpus = _layer_min_gpus,
                        )
                        if use_fit and not explicit_ctx:
                            # Weights don't fit on any subset; default UI to 4096
                            # so the slider isn't on an unusable native ctx.
                            effective_ctx = min(4096, effective_ctx) if effective_ctx > 0 else 4096

                    elif _apple_budget_mib > 0 and effective_ctx > 0:
                        # No GPU on Metal: the branches above are skipped and the context
                        # stays at native, over-committing unified memory (#5118, #6529).
                        # Cap with the same fit math (--fit on stays as a backstop); only
                        # auto context shrinks, explicit is honored.
                        native_ctx_for_cap = self._context_length or effective_ctx
                        # Reserve the flat MTP fraction up front like the discrete
                        # _pin_fraction, so an unsized MTP draft (e.g. Qwen3.6-MTP, #6529)
                        # can't over-commit. No-op when MTP is off; exclusive with the
                        # byte-accurate _mtp_bytes reserve.
                        _apple_fit_budget_mib = int(
                            _apple_budget_mib * max(0.0, 1.0 - _flat_mtp_reserve)
                        )
                        if self._can_estimate_kv():
                            cap = self._fit_context_to_vram(
                                native_ctx_for_cap,
                                _apple_fit_budget_mib,
                                model_size_fit,
                                cache_type_kv,
                                n_parallel = n_parallel,
                                mtp_engaged = _mtp_reserves_gpu,
                                mtp_overhead_fn = mtp_overhead_fn,
                                compute_ctx_bytes_fn = _cc_bytes,
                                budget_frac = 1.0,
                                total_mib = None,
                            )
                            _cap_footprint_mib = (
                                model_size_fit
                                + self._estimate_kv_cache_bytes(
                                    cap, cache_type_kv, n_parallel = n_parallel
                                )
                                + _mtp_bytes(cap)
                                + _cc_bytes(cap)
                            ) / (1024 * 1024)
                            # Fit returns the request unchanged when it fits OR weights
                            # exceed budget; only the latter over-commits, so floor to 4096.
                            max_available_ctx = (
                                cap
                                if _cap_footprint_mib <= _apple_fit_budget_mib
                                else min(4096, native_ctx_for_cap)
                            )
                        else:
                            # No KV estimate: mirror the discrete file-size-only fallback
                            # and floor to 4096 rather than launch at native and over-commit.
                            max_available_ctx = min(4096, native_ctx_for_cap)
                        if not explicit_ctx:
                            effective_ctx = max_available_ctx

                    # Prefer fewer serving slots on GPU over --fit on offload: when the extra
                    # --parallel slots push the footprint past the pin budget, llama-server
                    # offloads layers to host and decode collapses ~3x (#6718). Retry the fit
                    # at fewer slots, keeping the largest count that stays fully on GPU and the
                    # chosen context. Skips tensor mode / Metal / KV-inestimable paths.
                    if (
                        use_fit
                        and n_parallel > 1
                        and gpus
                        and self._can_estimate_kv()
                        and effective_ctx > 0
                    ):
                        # Slot-independent footprint (folded compute buffer swapped out so the
                        # helper re-adds a slot-sized one per candidate).
                        _base_footprint = (
                            model_size_fit
                            - _compute_buffer_pipeline
                            + _mtp_bytes(effective_ctx)
                            + _cc_bytes(effective_ctx)
                        )
                        _gi_slots, _uf_slots, _slots = self._slots_that_fit_on_gpu(
                            n_parallel,
                            effective_ctx,
                            gpus,
                            total_by_idx,
                            _base_footprint,
                            cache_type_kv,
                            _pin_fraction,
                            _pipeline_overhead_bytes + _cc_bytes(effective_ctx),
                            _layer_min_gpus,
                            _effective_ubatch,
                        )
                        if not _uf_slots:
                            logger.info(
                                "Serving slots reduced %d -> %d to keep the model on GPU "
                                "(avoid --fit offload) at context %d.",
                                n_parallel,
                                _slots,
                                effective_ctx,
                            )
                            gpu_indices, use_fit, n_parallel = _gi_slots, False, _slots

                    # MTP reserve at the final context, for the logs below.
                    _mtp_reserve_bytes = _mtp_bytes(effective_ctx) if _mtp_will_engage else 0
                    if _mtp_will_engage:
                        _mtp_note = (
                            f"MTP reserve: {_mtp_reserve_bytes / (1024**3):.2f} GB "
                            f"(draft KV @ {effective_ctx} + verify n_max={_mtp_eff_n_max}"
                            + (", flat-frac fallback" if mtp_overhead_fn is None else "")
                            + "), "
                        )
                    else:
                        _mtp_note = ""

                    if effective_ctx < original_ctx:
                        kv_est = self._estimate_kv_cache_bytes(
                            effective_ctx, cache_type_kv, n_parallel = n_parallel
                        )
                        logger.info(
                            f"Context auto-reduced: {original_ctx} -> {effective_ctx} "
                            f"(model: {model_size / (1024**3):.1f} GB, "
                            f"est. KV cache: {kv_est / (1024**3):.1f} GB, "
                            f"{_mtp_note}".rstrip(", ")
                            + ")"
                        )

                    kv_cache_bytes = self._estimate_kv_cache_bytes(
                        effective_ctx, cache_type_kv, n_parallel = n_parallel
                    )
                    mmproj_note = (
                        f"mmproj: {mmproj_size / (1024**3):.1f} GB, " if mmproj_size else ""
                    )
                    logger.info(
                        f"GGUF size: {gguf_size / (1024**3):.1f} GB, "
                        f"{mmproj_note}"
                        f"est. KV cache: {kv_cache_bytes / (1024**3):.1f} GB, "
                        f"{_mtp_note}"
                        f"context: {effective_ctx}, "
                        f"GPUs free: {gpus}, selected: {gpu_indices}, fit: {use_fit}"
                    )
                except Exception as e:
                    logger.warning(f"GPU selection failed ({e}), using --fit on")
                    gpu_indices, use_fit = None, True
                    tp_tensor_split = None
                    effective_ctx = requested_ctx  # fall back to original

                # Unified-memory APUs load weights into system RAM (under WSL the VM
                # cap, not the ROCm-reported VRAM, is the real ceiling); refuse an
                # oversize load the OS would otherwise kill mid-flight. Base model
                # only: an optional MTP drafter is dropped by the MTP-drop fallback.
                if model_size is not None and self._amd_apu_wants_unified_memory(gpu_indices):
                    _ram_msg = self._apu_ram_shortfall_message(
                        model_size, self._available_system_memory_mib()
                    )
                    if _ram_msg:
                        raise RuntimeError(_ram_msg)

                # Audio input straight from the mmproj (clip.has_audio_encoder),
                # independent of token names.
                self._mmproj_has_audio = False
                if launch_mmproj_path:
                    try:
                        from utils.models.gguf_metadata import (
                            read_mmproj_audio_capability,
                        )
                        self._mmproj_has_audio = bool(
                            read_mmproj_audio_capability(launch_mmproj_path)
                        )
                    except Exception as e:
                        logger.debug(f"mmproj audio-capability read failed: {e}")

                cmd = [
                    binary,
                    "-m",
                    model_path,
                    "--port",
                    str(self._port),
                    "-c",
                    str(effective_ctx) if effective_ctx > 0 else "0",
                    "--parallel",
                    str(n_parallel),
                    "--flash-attn",
                    "on",  # Force flash attention for speed
                    # Error out at n_ctx instead of silently rotating the KV cache; frontend catches it and points the user at "Context Length".
                    "--no-context-shift",
                ]

                # Report a clean public model id (matching GET /v1/models) rather
                # than the raw -m path in llama-server's own /v1/models and the
                # "model" field of its chat/completions responses.
                from core.inference.model_ids import public_model_id

                _alias = public_model_id(self._model_identifier or model_path)
                if _alias:
                    cmd.extend(["--alias", _alias])

                fully_gpu_offloaded = False
                if use_fit:
                    cmd.extend(["--fit", "on"])
                elif gpu_indices is not None:
                    # Fits on selected GPU(s) -- force all layers on GPU. --fit off is
                    # required: without it llama.cpp's default --fit on second-guesses
                    # and offloads ~1 GB at --parallel 4 even though the model fits.
                    cmd.extend(["-ngl", "-1", "--fit", "off"])
                    fully_gpu_offloaded = True

                server_caps = self.probe_server_capabilities(binary)
                # Expose Prometheus /metrics for the engine-stats logger, only
                # when the binary advertises it (older/custom binaries may not).
                if server_caps.get("supports_metrics"):
                    cmd.append("--metrics")
                cmd.extend(
                    self._ctx_integrity_flags(
                        n_parallel,
                        use_fit,
                        requested_ctx,
                        effective_ctx,
                        server_caps,
                    )
                )
                offload_overridden = _extra_args_set_any_flag(
                    extra_args, _GPU_OFFLOAD_OVERRIDE_FLAGS
                )
                threads_overridden = _extra_args_set_any_flag(extra_args, _THREAD_OVERRIDE_FLAGS)
                full_offload_tuning_active = fully_gpu_offloaded and not offload_overridden

                # Thread count: an unset --threads makes llama.cpp pick physical
                # cores (common_cpu_get_num_math), but an explicit --threads -1
                # resolves to hardware_concurrency() (every hyperthread), which
                # contends on the memory bus and slows CPU / hybrid decode. So
                # omit the flag when unset and only pin it for an explicit
                # override or the Windows full-offload OpenMP cap. Pass-through
                # thread flags in extra_args still win (appended last). #5692
                if (
                    sys.platform == "win32"
                    and full_offload_tuning_active
                    and not threads_overridden
                ):
                    cmd.extend(["--threads", "2"])
                elif n_threads is not None and n_threads > 0:
                    cmd.extend(["--threads", str(n_threads)])

                # Enable Jinja chat template rendering
                cmd.extend(["--jinja"])

                # KV cache data type
                _valid_cache_types = {
                    "f16",
                    "bf16",
                    "q8_0",
                    "q4_0",
                    "q4_1",
                    "q5_0",
                    "q5_1",
                    "iq4_nl",
                    "f32",
                }
                if (
                    cache_type_kv
                    and cache_type_kv in _valid_cache_types
                    and not _cache_type_from_env
                ):
                    cmd.extend(
                        [
                            "--cache-type-k",
                            cache_type_kv,
                            "--cache-type-v",
                            cache_type_kv,
                        ]
                    )
                    self._cache_type_kv = cache_type_kv
                    logger.info(f"KV cache type: {cache_type_kv}")
                else:
                    # An env-only type is left inherited (untouched) so an
                    # asymmetric K/V env reaches the child as set.
                    self._cache_type_kv = None

                # Tensor parallelism: split the model across GPUs by tensor
                # rather than by layer. Multi-GPU only -- a no-op on a single
                # GPU. Default (layer split) is left implicit by omitting the
                # flag. See llama.cpp --split-mode.
                if tensor_parallel:
                    cmd.extend(["--split-mode", "tensor"])
                    if tp_tensor_split and len(tp_tensor_split) > 1:
                        cmd.extend(
                            [
                                "--tensor-split",
                                ",".join(str(int(x)) for x in tp_tensor_split),
                            ]
                        )
                    self._tensor_parallel = True
                    self._layer_preserves_tensor_intent = False
                    logger.info(
                        "Tensor parallelism: --split-mode tensor, --tensor-split %s",
                        tp_tensor_split,
                    )
                else:
                    self._tensor_parallel = False
                    # > 1 only when a tensor request was downgraded but kept multi-GPU.
                    self._layer_preserves_tensor_intent = _layer_min_gpus > 1

                # Speculative decoding. See _build_speculative_flags for the
                # mode resolution, benchmarks, and llama.cpp references.
                launch_mtp_draft_path = self._resolve_launch_mtp_path(
                    mtp_draft_path = mtp_draft_path,
                )
                spec_flags = self._build_speculative_flags(
                    speculative_type = speculative_type,
                    spec_draft_n_max = spec_draft_n_max,
                    extra_args = extra_args,
                    model_identifier = model_identifier,
                    model_path = model_path,
                    gpus = bool(gpus),
                    binary = binary,
                    mtp_draft_path = launch_mtp_draft_path,
                )
                # Remember where the spec block sits so a drafter-load failure
                # can be retried with these flags swapped out (see below).
                _spec_start = len(cmd)
                cmd.extend(spec_flags)

                # Apply custom chat template override if provided.
                self._chat_template_override = chat_template_override
                if chat_template_override:
                    import tempfile

                    flags = detect_reasoning_flags(
                        chat_template_override,
                        self._model_identifier,
                        log_source = "GGUF chat template override",
                    )
                    self._supports_reasoning = flags["supports_reasoning"]
                    self._reasoning_style = flags["reasoning_style"]
                    self._reasoning_effort_levels = flags.get("reasoning_effort_levels", [])
                    self._reasoning_always_on = flags["reasoning_always_on"]
                    self._supports_preserve_thinking = flags["supports_preserve_thinking"]
                    self._supports_tools = flags["supports_tools"]

                    self._chat_template_file = tempfile.NamedTemporaryFile(
                        mode = "w",
                        encoding = "utf-8",
                        suffix = ".jinja",
                        delete = False,
                        prefix = "unsloth_chat_template_",
                    )
                    self._chat_template_file.write(chat_template_override)
                    self._chat_template_file.close()
                    cmd.extend(["--chat-template-file", self._chat_template_file.name])
                    logger.info(f"Using custom chat template file: {self._chat_template_file.name}")

                # Default thinking mode for reasoning models. Qwen3.5/3.6 below
                # 9B disable thinking by default; 9B+ enable it. Always-on
                # templates ignore the kwarg, so skip.
                if self._supports_reasoning and not self._reasoning_always_on:
                    thinking_default = True
                    mid = (model_identifier or "").lower()
                    if "qwen3.5" in mid or "qwen3.6" in mid:
                        size_val = _extract_model_size_b(mid)
                        if size_val is not None and size_val < 9:
                            thinking_default = False
                    self._reasoning_default = thinking_default
                    reasoning_kw = self._reasoning_kwargs(thinking_default)
                    # preserve_thinking is an independent kwarg. Default it OFF
                    # at launch so direct OpenAI-compatible callers that omit the
                    # field match the UI's default-off behavior (the bundled
                    # gemma-4 template also defaults it false; the frontend sends
                    # preserve_thinking per request once toggled on).
                    if self._supports_preserve_thinking:
                        reasoning_kw["preserve_thinking"] = False
                    cmd.extend(
                        [
                            "--chat-template-kwargs",
                            json.dumps(reasoning_kw),
                        ]
                    )
                    logger.info(f"Reasoning model: {reasoning_kw} by default")

                if launch_mmproj_path and effective_is_vision:
                    cmd.extend(["--mmproj", launch_mmproj_path])
                    logger.info(f"Using mmproj for vision: {launch_mmproj_path}")

                # Option C: --api-key for direct client access when enabled
                import secrets as _secrets

                if os.getenv("UNSLOTH_DIRECT_STREAM", "0") == "1":
                    self._api_key = _secrets.token_urlsafe(32)
                    cmd.extend(["--api-key", self._api_key])
                    logger.info("llama-server started with --api-key for direct streaming")
                else:
                    self._api_key = None

                # Windows + full offload: disable KV checkpoints (WDDM/PCI-E
                # overhead). CPU/partial offload keeps prompt caching. #5692.
                if sys.platform == "win32" and full_offload_tuning_active:
                    unsupported_cache_flags: list[str] = []
                    if server_caps.get("supports_cache_ram"):
                        cmd.extend(["--cache-ram", "0"])
                    else:
                        unsupported_cache_flags.append("--cache-ram")
                    if server_caps.get("supports_ctx_checkpoints"):
                        cmd.extend(["--ctx-checkpoints", "0"])
                    else:
                        unsupported_cache_flags.append("--ctx-checkpoints")
                    if server_caps.get("supports_no_cache_prompt"):
                        cmd.append("--no-cache-prompt")
                    else:
                        unsupported_cache_flags.append("--no-cache-prompt")
                    if unsupported_cache_flags:
                        logger.info(
                            "Skipping unsupported Windows cache flags for llama-server: %s",
                            ", ".join(unsupported_cache_flags),
                        )

                # User pass-through args go last so llama.cpp's last-wins parsing
                # lets the user override Studio's auto-set flags. Already
                # validated by the route via validate_extra_args().
                if extra_args:
                    cmd.extend(str(a) for a in extra_args)
                    logger.info(f"Appending user extra args to llama-server: {list(extra_args)}")

                logger.info(f"Starting llama-server: {' '.join(self._redacted_cmd_for_log(cmd))}")

                # Library paths so llama-server finds its shared libs and CUDA DLLs.
                env = self._llama_server_env_for_binary(binary)
                # Omitting --threads relies on llama.cpp's physical-core default, so
                # drop an inherited LLAMA_ARG_THREADS that would otherwise feed the
                # arg handler and silently force hardware_concurrency(). #5692
                if "--threads" not in cmd:
                    env.pop("LLAMA_ARG_THREADS", None)

                # Reconcile the inherited LLAMA_ARG_* env with Studio's final
                # decision: stripping CLI extras on a tensor->layer downgrade
                # can't remove env vars, so the child could run a mode/KV Studio
                # didn't budget.
                if not tensor_parallel:
                    # Layer split: clear a non-layer inherited split mode (and any
                    # paired tensor-split) so the child can't override the layer plan.
                    _inherited_sm = (env.get("LLAMA_ARG_SPLIT_MODE") or "").strip().lower()
                    if _inherited_sm and _inherited_sm != "layer":
                        env.pop("LLAMA_ARG_SPLIT_MODE", None)
                        env.pop("LLAMA_ARG_TENSOR_SPLIT", None)
                else:
                    # Studio owns the tensor split: it emits --tensor-split when it
                    # picks an uneven one (CLI wins) and nothing when an even split
                    # is safe. Clear any inherited LLAMA_ARG_TENSOR_SPLIT so the even
                    # case can't be overridden by a stale env (the layer branch above
                    # clears it too).
                    env.pop("LLAMA_ARG_TENSOR_SPLIT", None)
                    # Tensor split aborts on a quantized KV; clear an inherited
                    # quantized cache type so the child uses the tensor-safe default.
                    for _ct_var in ("LLAMA_ARG_CACHE_TYPE_K", "LLAMA_ARG_CACHE_TYPE_V"):
                        _ct_raw = (env.get(_ct_var) or "").strip().lower()
                        if _ct_raw and _ct_raw not in self._TENSOR_PARALLEL_KV_TYPES:
                            env.pop(_ct_var, None)

                # Windows + full offload: PASSIVE OMP + 2 threads stop
                # spin-wait burning CPU. CPU/partial offload keeps default
                # OMP parallelism. #5692.
                if sys.platform == "win32" and full_offload_tuning_active:
                    env.setdefault("OMP_WAIT_POLICY", "PASSIVE")
                    if not threads_overridden:
                        env.setdefault("OMP_NUM_THREADS", "2")

                # AMD unified-memory APUs (gfx1150/gfx1151): let llama.cpp use
                # shared system RAM. setdefault so a user value wins.
                if self._amd_apu_wants_unified_memory(gpu_indices):
                    env.setdefault("GGML_CUDA_ENABLE_UNIFIED_MEMORY", "1")
                    logger.info("AMD unified-memory APU: set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1")

                # DC NVIDIA GPUs: FP32 accum (+ P2P / launch queues for multi-GPU).
                # See _apply_datacenter_env; opt out with UNSLOTH_DISABLE_DC_TUNING=1.
                if self._apply_datacenter_env(env, gpu_indices):
                    multi_gpu = self._effective_gpu_count(gpu_indices) > 1
                    logger.info(
                        f"Data-center GPU detected: applied DC llama.cpp env tuning (multi_gpu={multi_gpu})"
                    )

                # Pin to selected GPU(s). On ROCm, narrowing only
                # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full
                # set, so set HIP_VISIBLE_DEVICES too.
                if gpu_indices is not None:
                    pinned = ",".join(str(i) for i in gpu_indices)
                    env["CUDA_VISIBLE_DEVICES"] = pinned
                    try:
                        import torch as _torch
                        if getattr(_torch.version, "hip", None) is not None:
                            env["HIP_VISIBLE_DEVICES"] = pinned
                            # Do NOT also set ROCR_VISIBLE_DEVICES to the same
                            # value. ROCR_VISIBLE_DEVICES filters at the HSA/ROCr
                            # layer and HIP_VISIBLE_DEVICES at the HIP layer, so
                            # setting both with the same physical indices applies
                            # the mask twice: ROCR reduces the visible set and
                            # re-indexes it from 0, then HIP indexes into the
                            # already-reduced set. A single non-zero pin (e.g.
                            # "1") then points out of range at the HIP layer, HIP
                            # enumerates 0 devices, and llama.cpp falls back to
                            # CPU ("ggml_cuda_init: no ROCm-capable device is
                            # detected"). The HIP mask alone narrows correctly;
                            # clear any inherited ROCR mask so it can't double up.
                            env.pop("ROCR_VISIBLE_DEVICES", None)
                    except Exception as e:
                        logger.debug("Failed to set ROCm visibility env vars for child: %s", e)

                # Captured before any text-only fallback strips it from cmd.
                launched_with_mmproj = "--mmproj" in cmd

                # One-shot --fit off retry: recent llama.cpp runs a "fitting
                # params to device memory" step by default (--fit defaults to
                # 'on') even when -ngl is explicit. That step has aborted on
                # some ROCm hosts (ggml-cuda.cu ROCm error during worst-case
                # estimation, e.g. MTP + mmproj models on gfx1151). When
                # Studio's own VRAM math already placed the model
                # (use_fit=False), the step is redundant second-guessing --
                # retry once with --fit off before declaring the load failed.
                # Never retry when fit was requested (use_fit) or the caller
                # passed an explicit fit flag via extra args.
                # Argv actually launched (post --fit off / MTP); text-only retry strips this.
                _last_spawn_cmd = list(cmd)

                def _spawn_and_wait(run_cmd, *, label = ""):
                    """Start llama-server with run_cmd and wait for health.

                    Retries once with --fit off when the first attempt
                    crashes during startup and run_cmd is eligible (see
                    _fit_off_retry_eligible).
                    """
                    nonlocal _last_spawn_cmd
                    _fit_retry_allowed = self._fit_off_retry_eligible(run_cmd, use_fit)
                    for _spawn_attempt in (0, 1):
                        # Defensive kill: drop an orphan Popen a concurrent load may
                        # have stored before we overwrite the reference (#5161).
                        # Also reaps the crashed first attempt on the retry pass.
                        self._kill_process()

                        self._stdout_lines = []
                        # Tee llama-server output to a dedicated log file so a
                        # post-mortem has the full trail even when the parent only
                        # kept the last 50 lines. Path is under the studio home.
                        # ``label`` (MTP fallback) and the attempt index (--fit
                        # off retry) keep a respawn within the same epoch second
                        # from truncating the crash log a retry warning just
                        # pointed the user at.
                        self._llama_log_fh = None
                        try:
                            log_dir = _swa_cache_path().parent / "logs" / "llama-server"
                            log_dir.mkdir(parents = True, exist_ok = True)
                            self._llama_log_path = log_dir / (
                                f"llama-{int(time.time())}{label}-port-{self._port}"
                                f"-try{_spawn_attempt}.log"
                            )
                            self._llama_log_fh = open(
                                self._llama_log_path,
                                "w",
                                encoding = "utf-8",
                                buffering = 1,
                            )
                            logger.info(f"llama-server stdout/stderr -> {self._llama_log_path}")
                        except OSError as e:
                            # Best-effort; never block the load on logging.
                            logger.debug(f"Could not open llama-server log file: {e}")
                            self._llama_log_path = None
                        _last_spawn_cmd = list(run_cmd)
                        self._process = subprocess.Popen(
                            run_cmd,
                            stdout = subprocess.PIPE,
                            stderr = subprocess.STDOUT,
                            text = True,
                            env = env,
                            **_windows_hidden_subprocess_kwargs(),
                            **_child_popen_kwargs(),
                        )
                        self._record_server_pid(self._process.pid)

                        # Background thread to drain stdout (prevents pipe deadlock)
                        self._stdout_thread = threading.Thread(
                            target = self._drain_stdout, daemon = True, name = "llama-stdout"
                        )
                        self._stdout_thread.start()
                        if self._wait_for_health(timeout = 600.0):
                            return True
                        _startup_crashed = (
                            self._process.poll() is not None and self._process.returncode != 0
                        )
                        # A split-axis abort (#6415) is fit-independent: skip the
                        # --fit off retry and let the caller latch it.
                        _split_axis_crash = self._is_tensor_split_assert(
                            "\n".join(self._stdout_lines[-50:])
                        )
                        if (
                            _spawn_attempt == 0
                            and fully_gpu_offloaded
                            and _startup_crashed
                            and not _split_axis_crash
                        ):
                            # We forced --fit off because Studio's (conservative) VRAM
                            # math placed the model fully on GPU. A startup crash here
                            # means that estimate was optimistic, so fall back to --fit
                            # on and let llama.cpp offload rather than fail the load.
                            logger.warning(
                                "llama-server crashed during startup (exit code %s) "
                                "with forced --fit off; the fit estimate was optimistic, "
                                "retrying once with --fit on so it can offload. "
                                "Crash log: %s",
                                self._process.returncode,
                                self._llama_log_path,
                            )
                            # Flip Studio's own --fit off (added first, before any
                            # user extra args) to on; a user's later --fit still wins
                            # by last-arg. Defensive: if absent, the default is already
                            # --fit on, so leave it.
                            _run = list(run_cmd)
                            if "--fit" in _run:
                                _run[_run.index("--fit") + 1] = "on"
                            run_cmd = _run
                            continue
                        if (
                            _spawn_attempt == 0
                            and _fit_retry_allowed
                            and _startup_crashed
                            and not _split_axis_crash
                        ):
                            logger.warning(
                                "llama-server crashed during startup (exit code %s) "
                                "with the default memory-fit step enabled; Studio "
                                "already verified the model fits, retrying once "
                                "with --fit off. Crash log: %s",
                                self._process.returncode,
                                self._llama_log_path,
                            )
                            run_cmd = [*run_cmd, "--fit", "off"]
                            continue
                        return False

                # Store the resolved on-disk path, not the caller's kwarg: in
                # HF mode gguf_path is None and ``model_path`` is what
                # llama-server mmap's, which downstream consumers need. Must be
                # set BEFORE the spawn: load_progress() reads _gguf_path for
                # the mmap progress total while the health wait runs.
                self._gguf_path = model_path
                self._hf_repo = hf_repo
                self._mtp_draft_path = launch_mtp_draft_path
                # For local GGUF files, extract variant from filename if absent
                if hf_variant:
                    self._hf_variant = hf_variant
                elif gguf_path:
                    try:
                        from utils.models.model_config import _extract_quant_label
                        self._hf_variant = _extract_quant_label(gguf_path)
                    except Exception:
                        self._hf_variant = None
                else:
                    self._hf_variant = None
                self._is_vision = effective_is_vision
                self._model_identifier = model_identifier

                # Store the effective (possibly capped) context separately; do
                # NOT overwrite _context_length (the native length for display).
                self._effective_context_length = (
                    effective_ctx if effective_ctx > 0 else self._context_length
                )
                self._reconcile_effective_ctx_with_server()
                self._max_context_length = (
                    max_available_ctx if max_available_ctx > 0 else self._effective_context_length
                )

                healthy = _spawn_and_wait(cmd)
                # #6415 split-mode tensor warmup abort. Latch it on THIS first spawn:
                # the flash-attn-off retry below can't run tensor (needs flash_attn),
                # so its output drops the marker and recording later would miss it,
                # looping every load. Record and raise to the route's layer fallback,
                # skipping the futile flash-attn/MTP retries.
                if not healthy and self._tensor_parallel and not self._cancel_event.is_set():
                    _ts_out = "\n".join(self._stdout_lines[-50:])
                    _ts_rc = self._process.poll() if self._process is not None else None
                    if self._should_record_tensor_split_abort(_ts_rc, _ts_out):
                        LlamaCppBackend._record_tensor_split_abort(binary, model_identifier)
                        self._kill_process()
                        raise RuntimeError(
                            "llama-server aborted on --split-mode tensor "
                            "(split-axis geometry); retrying with layer split."
                        )
                # Flash-attention kernels hard-crash at startup on some ROCm/GPU
                # builds (frequently inside the vision tower). Disabling FA keeps
                # both vision and MTP, so retry that way before dropping either.
                # Only on a hard fault with FA on; a cancel/unload stops respawn.
                if not healthy and not self._cancel_event.is_set():
                    _fa_rc = self._process.poll() if self._process is not None else None
                    _fa_cmd = (
                        self._with_flash_attn_off(_last_spawn_cmd)
                        if self._is_signal_crash(_fa_rc)
                        else None
                    )
                    if _fa_cmd is not None:
                        logger.warning(
                            "llama-server hard-crashed at startup (exit %s) with "
                            "flash attention on; retrying once with --flash-attn "
                            "off (keeps vision and MTP).",
                            _fa_rc,
                        )
                        self._kill_process()
                        cmd = _fa_cmd
                        healthy = _spawn_and_wait(_fa_cmd, label = "-noflash")

                # MTP from Studio's spec flags or the user's (extra_args
                # --spec-type / LLAMA_ARG_SPEC_TYPE). The env reaches the child
                # only when neither emits a spec flag, so consult it only then.
                _launch_spec_env: Mapping[str, str] = (
                    os.environ
                    if (not _extra_args_set_spec_type(extra_args) and not spec_flags)
                    else {}
                )
                _spec_requested_mtp = any(
                    "mtp" in str(t).lower() for t in spec_flags
                ) or _extra_args_requests_mtp(extra_args, env = _launch_spec_env)
                # Is the launched server actually running MTP+tensor? Gates the
                # probe/watchdog/recovery; cleared if the MTP-drop fallback wins.
                _mtp_active_for_launched_server = bool(
                    self._tensor_parallel and _spec_requested_mtp
                )
                # MTP can pass /health then crash the flash-attn kernel on the
                # first decode under tensor; probe one generation so the fallback
                # catches that too. Tensor-only, so ordinary MTP stays probe-free.
                if (
                    healthy
                    and self._tensor_parallel
                    and _spec_requested_mtp
                    and not self._cancel_event.is_set()
                    and not self._probe_mtp_decode()
                ):
                    # A first-decode hard fault is usually the FA kernel: retry
                    # FA-off (keeps MTP) before dropping speculative decoding below.
                    _probe_rc = self._process.poll() if self._process is not None else None
                    _fa_cmd = (
                        self._with_flash_attn_off(_last_spawn_cmd)
                        if self._is_signal_crash(_probe_rc)
                        else None
                    )
                    healthy = False
                    if _fa_cmd is not None:
                        logger.warning(
                            "MTP first-decode hard-crashed (exit %s) with flash "
                            "attention on; retrying with --flash-attn off.",
                            _probe_rc,
                        )
                        self._kill_process()
                        cmd = _fa_cmd
                        healthy = (
                            _spawn_and_wait(_fa_cmd, label = "-noflash-mtp")
                            and self._probe_mtp_decode()
                        )
                    if not healthy:
                        logger.warning(
                            "MTP speculative decoding crashed on the first decode "
                            "under tensor parallelism; retrying without it."
                        )
                # Any MTP request can abort the server: a separate drafter
                # (Gemma) on a binary that predates its arch, or an embedded
                # head (Qwen) the binary cannot build. Retry once with the
                # spec slice replaced by --spec-default so the main model still
                # loads. Gate on the spec block (not the drafter path, which
                # off/ngram local loads also carry) and keep
                # _requested_spec_mode so a duplicate /load doesn't thrash. The
                # cancel check stops an /unload-killed attempt respawning. A
                # decode-probe failure above also routes here.
                if not healthy and _spec_requested_mtp and not self._cancel_event.is_set():
                    # Blame the binary only when the output shows MTP itself
                    # failing (unknown arch / draft or context build); an
                    # unrelated crash (e.g. OOM) gets a neutral message.
                    _lo = "\n".join(self._stdout_lines).lower()
                    # Only an unknown architecture proves the prebuilt predates
                    # this MTP model (an update fixes it). The memory/context
                    # build failures are generic (VRAM / ctx pressure), where an
                    # update may not help, so classify those as runtime_error.
                    _arch_unsupported = "unknown model architecture" in _lo
                    if (
                        _arch_unsupported
                        or "failed to measure draft model memory" in _lo
                        or "failed to measure mtp context memory" in _lo
                        or "failed to create llama_context" in _lo
                    ):
                        _retry_reason = (
                            "the prebuilt may predate it; retrying without "
                            "speculative decoding -- run `unsloth studio "
                            "update` for MTP"
                        )
                        self._spec_fallback_reason = (
                            "binary_outdated" if _arch_unsupported else "runtime_error"
                        )
                    else:
                        _retry_reason = (
                            "retrying without speculative decoding in case MTP is the cause"
                        )
                        self._spec_fallback_reason = "runtime_error"
                    _drafter = (
                        Path(launch_mtp_draft_path).name
                        if launch_mtp_draft_path
                        else "embedded head"
                    )
                    logger.warning(
                        "llama-server failed to start with MTP (%s); %s.",
                        _drafter,
                        _retry_reason,
                    )
                    self._kill_process()
                    fallback_cmd = (
                        cmd[:_spec_start]
                        + ["--spec-default"]
                        + cmd[_spec_start + len(spec_flags) :]
                    )
                    # User/env MTP survives in the tail; llama.cpp takes the last
                    # spec flag, so a trailing --spec-default overrides it too.
                    if _extra_args_requests_mtp(extra_args, env = _launch_spec_env):
                        fallback_cmd.append("--spec-default")
                    healthy = _spawn_and_wait(fallback_cmd, label = "-retry")
                    if healthy:
                        self._speculative_type = "default"
                        _mtp_active_for_launched_server = False

                # A too-old llama.cpp can reject a model's --mmproj projector
                # (format message or a bare SIGSEGV); retry once text-only.
                if not healthy:
                    out = "\n".join(self._stdout_lines[-50:])
                    # Read the crash code before _kill_process() clears _process.
                    _crash_rc = self._process.poll() if self._process is not None else None
                    self._kill_process()
                    # The #6415 split-axis abort is latched earlier (first spawn).
                    # Skip if a cancel/unload is pending (mirrors the MTP guard).
                    if (
                        launched_with_mmproj
                        and not self._cancel_event.is_set()
                        and (
                            self._is_projector_incompatibility(out)
                            or (
                                self._is_signal_crash(_crash_rc)
                                and not self._output_has_nonprojector_diagnostic(out)
                            )
                        )
                    ):
                        logger.warning(
                            "llama-server could not load this model's vision "
                            "projector (--mmproj). The installed llama.cpp build is "
                            "likely too old for it. Loading text-only for this "
                            "session; run 'unsloth studio update' to enable vision."
                        )
                        cmd = self._strip_mmproj_args(_last_spawn_cmd)
                        self._is_vision = False
                        self._mmproj_has_audio = False
                        self._start_llama_process(cmd, env)
                        if not self._wait_for_health(timeout = 600.0):
                            # Read the exit code before _kill_process() clears it, so
                            # an OS-killed text-only retry still gets the OOM message.
                            _retry_rc = self._process.poll() if self._process is not None else None
                            self._kill_process()
                            raise RuntimeError(
                                "Vision projector incompatible with this llama.cpp "
                                "build, and the text-only retry also failed: "
                                + self._classify_llama_start_failure(
                                    "\n".join(self._stdout_lines[-50:]),
                                    gguf_path,
                                    self._model_identifier,
                                    _retry_rc,
                                )
                            )
                    else:
                        raise RuntimeError(
                            self._classify_llama_start_failure(
                                out,
                                gguf_path,
                                self._model_identifier,
                                _crash_rc,
                            )
                        )

                self._healthy = True

                # Commit caller intent only after _healthy=True so a failed start
                # can't poison the next inheritance check. None keeps prior, []
                # clears, list sets. Source records hf_variant for the route's
                # same_source check.
                if extra_args is not None:
                    self._extra_args = list(extra_args)
                    self._extra_args_source = (model_identifier, hf_variant)
                self._requested_n_ctx = int(n_ctx)
                # Commit the known-good snapshot + whether MTP+tensor is live, then
                # watch this load for a mid-generation crash.
                self._last_load_kwargs = _pending_load_kwargs
                self._mtp_runtime_fallback_active = _mtp_active_for_launched_server
                self._start_mtp_crash_watchdog()

                # Catch silent CPU fallback when GPU was intended (#5106).
                self._gpu_offload_active = self._classify_gpu_offload(
                    gpu_indices is not None or use_fit, gpus or []
                )
                if self._gpu_offload_active is False:
                    logger.warning(
                        "llama-server appears to have loaded the model entirely "
                        "on CPU even though Studio detected at least one GPU. "
                        "This usually means the prebuilt binary's GPU backend "
                        "failed to load -- on Windows, cudart64_X.dll / "
                        "cublas64_X.dll could not be resolved. Reinstall the "
                        "Studio llama.cpp prebuilt or install a matching CUDA "
                        "toolkit (issue unslothai/unsloth#5106).",
                    )

                logger.info(
                    f"llama-server ready on port {self._port} for model '{model_identifier}'"
                )
                # Poll llama-server /metrics -> vLLM-style engine_stats logs
                # (only when the binary exposes /metrics).
                if server_caps.get("supports_metrics"):
                    try:
                        from core.inference.llama_stats import maybe_start_stats_logger
                        if self._stats_logger is not None:
                            self._stats_logger.stop()
                        self._stats_logger = maybe_start_stats_logger(self.base_url, logger)
                    except Exception as e:
                        logger.debug(f"engine-stats logger not started: {e}")
                else:
                    self._stats_logger = None

            # Probe outside _lock (interruptible by /unload); init inside.
            self._is_audio = False
            self._audio_type = None
            self._audio_probed = False
            self._has_audio_input = False
            try:
                detected = self._detect_audio_type_strict()
                self._audio_probed = True
            except Exception as exc:
                logger.debug("Audio probe failed: %s", exc)
                detected = None
            if not self._apply_detected_audio(detected):
                return False

            if not self._healthy:
                return False
            return True

    def _build_speculative_flags(
        self,
        *,
        speculative_type: Optional[str],
        spec_draft_n_max: Optional[int],
        extra_args: Optional[List[str]],
        model_identifier: str,
        model_path: Optional[str],
        gpus: bool,
        binary: Optional[str],
        mtp_draft_path: Optional[str] = None,
    ) -> List[str]:
        """Return the llama-server flag list for the requested spec mode.

        Side effects: sets ``self._speculative_type`` (resolved internal
        emit), ``self._requested_spec_mode`` (canonical UI mode for the
        status round-trip), and ``self._spec_draft_n_max`` (user override
        only; None when the platform default applies).

        Speculative decoding (n-gram self-speculation, zero VRAM):
        ngram-mod uses a ~16 MB shared hash pool, constant memory /
        complexity, variable draft lengths. Helps most when the model
        repeats existing text (code refactor, summarisation, reasoning);
        for low-repetition chat, overhead is ~5 ms.

        Benchmarks from upstream llama.cpp speculative-decoding PRs:
          Scenario                        | Without | With    | Speedup
          gpt-oss-120b code refactor      | 181 t/s | 446 t/s | 2.5x
          Qwen3-235B offloaded            |  12 t/s |  21 t/s | 1.8x
          gpt-oss-120b repeat (92% accept)| 181 t/s | 814 t/s | 4.5x
        Refs: https://github.com/ggml-org/llama.cpp/blob/master/docs/speculative.md
              https://github.com/ggml-org/llama.cpp/pull/19164
              https://github.com/ggml-org/llama.cpp/pull/18471
              MTP guide: unsloth.ai/docs/models/qwen3.6#mtp-guide

        Sub-3B dense MTP regresses vs spec-off when the head is baked into the
        main GGUF (Qwen): the draft head's per-token cost exceeds the
        acceptance savings at this scale. Q4_K_XL clean bench (each prompt once
        after an unrelated warmup) on B200 + x86 CPU:
          0.8B GPU: draft-mtp n=2 = 0.58x vs OFF; ngram-only = 1.10x
          2B   GPU: draft-mtp n=2 = 0.82x vs OFF; OFF or ngram = 1.00x
          0.8B CPU: chained n=2   = 0.86x vs OFF; ngram-only = 1.19x
          2B   CPU: chained n=2   = 0.83x vs OFF; ngram-only = 1.01x
          4B+ GPU/CPU: spec on is a net win (1.08x-1.46x).
        A separate drafter (Gemma's root mtp-*.gguf) is a different, cheaper
        mechanism that wins even below 3B, so it is exempt from the sub-3B drop
        (``mtp_draft_path`` set -> not too small). B200 Q4_K_XL bench, draft-mtp
        n=2 vs OFF: gemma-4-E2B (2B) = 1.21x, accept ~0.65 (vs ngram = 1.00x);
        gemma-4-E4B (4B) and 12B engage as usual.
        Auto falls back to ngram-mod (zero-VRAM, near-zero idle cost on
        diverse content) for an embedded sub-3B head; forced MTP on a model
        with no head/drafter defaults back (mtp -> spec-default, mtp+ngram ->
        ngram-mod) since llama-server aborts otherwise; a drafter the binary
        cannot build (older prebuilt, or a CUDA kernel limit) aborts the spawn
        and the load retries once without speculative decoding.
        """
        flags: List[str] = []
        # Reset; emit branches re-set on the resolved emission.
        self._spec_draft_n_max = None
        self._speculative_type = None
        self._spec_fallback_reason = None

        # Canonical UI-facing requested mode (legacy values mapped via
        # _canonicalize_spec_mode).
        canonical_mode = _canonicalize_spec_mode(speculative_type)
        # MTP signals: head baked into the main GGUF (Qwen, via metadata or
        # name), or a separate drafter resolved from the repo (Gemma).
        is_mtp_model = (
            bool(self._nextn_predict_layers)
            or _is_mtp_model_name(model_identifier, model_path)
            or bool(mtp_draft_path)
        )
        user_owns_spec_type = _extra_args_set_spec_type(extra_args)
        _mtp_size_b = _extract_model_size_b(model_identifier)
        # The sub-3B regression is an embedded-head cost; a separate drafter
        # (Gemma) is a cheap standalone model that wins below 3B, so exempt it.
        _mtp_too_small = (
            _mtp_size_b is not None and _mtp_size_b < _MTP_MIN_SIZE_B and not bool(mtp_draft_path)
        )
        # Drafterless Gemma (name-only MTP, no embedded head): emitting MTP
        # would abort llama-server, so every mode below falls back instead.
        _mtp_drafter_missing = (
            _is_gemma_mtp_name(model_identifier, model_path)
            and not mtp_draft_path
            and not self._nextn_predict_layers
        )
        # Embedded MTP head on an MLA model (GLM-5.2/DeepSeek/Kimi, detected by
        # kv_lora_rank): llama.cpp's MLA/DSA MTP path is ~2x slower than no spec,
        # so Auto drops it (override via the Settings dropdown / forced mtp, or
        # UNSLOTH_MLA_MTP_ENABLED=1). Separate drafters (Gemma, mtp_draft_path) and
        # non-MLA embedded heads (Qwen, no kv_lora_rank) are unaffected.
        _auto_mla_embedded_mtp = (
            bool(self._nextn_predict_layers)
            and self._kv_lora_rank is not None
            and not bool(mtp_draft_path)
            and not _mla_mtp_auto_enabled()
        )

        if user_owns_spec_type:
            # User --spec-type wins outright; suppress auto-emit to avoid a
            # duplicate spec block.
            self._requested_spec_mode = None
            return flags

        effective_mode = canonical_mode or "auto"
        self._requested_spec_mode = effective_mode

        def _resolved_draft_n_max() -> int:
            # User override wins; else platform default (the B200 / x86
            # clean-sweep sweet spot from PR #5582 is n=2 GPU, n=3 CPU;
            # past 3 regresses on essay-style low-acceptance prompts).
            if spec_draft_n_max is not None:
                n = int(spec_draft_n_max)
                self._spec_draft_n_max = n
                return n
            return 2 if gpus else 3

        def _emit_mtp(*, chain_ngram: bool) -> bool:
            """Append --spec-type mtp[/draft-mtp][,ngram-mod] + n-max."""
            caps = self.probe_server_capabilities(binary)
            mtp_token = caps.get("mtp_token") if caps else None
            if not mtp_token:
                logger.warning(
                    "Requested MTP speculative decoding but "
                    "llama-server lacks --spec-type mtp/draft-mtp; "
                    "run `unsloth studio update`. Loading without "
                    "speculative decoding."
                )
                # Override an inherited LLAMA_ARG_SPEC_TYPE=draft-mtp (CLI wins
                # over env) so the child matches the binary-capability gate and
                # the no-MTP budget, like the sibling no-head/non-MTP fallbacks.
                flags.append("--spec-default")
                self._speculative_type = "default"
                self._spec_fallback_reason = "binary_no_mtp"
                return False
            draft_n_max = _resolved_draft_n_max()
            n_max_flag = caps.get("spec_draft_n_max_flag") or "--spec-draft-n-max"
            # Separate-file drafter (Gemma): point llama-server at it. Baked-in
            # heads (Qwen) pass no path -- llama-server reads them from the
            # main GGUF.
            if mtp_draft_path:
                flags.extend(["--model-draft", mtp_draft_path])
                logger.info(f"Using separate MTP drafter: {mtp_draft_path}")
            spec_value = mtp_token
            ngram_knobs: list[str] = []
            if chain_ngram:
                ngram_knobs = _build_ngram_mod_flags(caps)
                if ngram_knobs:
                    spec_value = f"ngram-mod,{mtp_token}"
                else:
                    logger.warning(
                        "llama-server lacks ngram-mod tuning "
                        "flags; loading MTP only (no ngram chain)"
                    )
            flags.extend(["--spec-type", spec_value, n_max_flag, str(draft_n_max)])
            flags.extend(ngram_knobs)
            self._speculative_type = "draft-mtp"
            chain_label = "chained ngram-mod" if chain_ngram else "MTP-only"
            logger.info(f"Spec decoding: {mtp_token} ({chain_label})")
            return True

        def _emit_ngram_mod() -> bool:
            """Append --spec-type ngram-mod + flag-set knobs."""
            ngram_caps = self.probe_server_capabilities(binary)
            ngram_knobs = _build_ngram_mod_flags(ngram_caps)
            flags.extend(["--spec-type", "ngram-mod"])
            if not ngram_knobs:
                logger.warning(
                    "llama-server lacks ngram-mod tuning "
                    "flags; loading without --spec-ngram-mod-* knobs"
                )
            flags.extend(ngram_knobs)
            self._speculative_type = "ngram-mod"
            logger.info("Spec decoding: ngram-mod")
            return True

        def _fallback_drafter_not_found() -> None:
            """Drafterless Gemma: use ngram-mod (or spec-default) and record why."""
            logger.warning(
                "Model %s is MTP-capable but no drafter or head was found; "
                "falling back. Check network or run `unsloth studio update`.",
                model_identifier,
            )
            if self.probe_server_capabilities(binary).get("supports_ngram_mod"):
                _emit_ngram_mod()
            else:
                flags.append("--spec-default")
                self._speculative_type = "default"
            self._spec_fallback_reason = "drafter_not_found"

        if effective_mode == "off":
            return flags  # nothing to emit
        if effective_mode == "ngram-simple":
            flags.extend(["--spec-type", "ngram-simple"])
            self._speculative_type = "ngram-simple"
            return flags
        if effective_mode == "ngram":
            _emit_ngram_mod()
            return flags
        if effective_mode == "mtp":
            if not is_mtp_model:
                # No head and no drafter: llama-server aborts on draft-mtp
                # instead of no-op'ing, so default back.
                logger.warning(
                    "MTP requested but this GGUF has no MTP head or drafter; "
                    "loading without speculative decoding."
                )
                flags.append("--spec-default")
                self._speculative_type = "default"
                return flags
            if _mtp_drafter_missing:
                # Drafterless: draft-mtp would abort llama-server, so fall back.
                _fallback_drafter_not_found()
                return flags
            if _mtp_too_small:
                logger.warning(
                    f"Forcing MTP on a {_mtp_size_b:.1f}B model; "
                    "the bench shows draft-mtp regresses below 3B. "
                    "Engaging anyway (user override)."
                )
            _emit_mtp(chain_ngram = False)
            return flags
        if effective_mode == "mtp+ngram":
            if not is_mtp_model:
                # No head/drafter: keep the ngram half (needs no head),
                # drop the draft-mtp chain that would abort the server.
                logger.warning(
                    "MTP+Ngram requested but this GGUF has no MTP head or "
                    "drafter; loading ngram-mod only."
                )
                _emit_ngram_mod()
                return flags
            if _mtp_drafter_missing:
                # No head/drafter: keep ngram-mod, drop the draft-mtp chain.
                _fallback_drafter_not_found()
                return flags
            if _mtp_too_small:
                logger.warning(
                    f"Forcing MTP+Ngram on a {_mtp_size_b:.1f}B model; "
                    "the bench shows the chain regresses below 3B. "
                    "Engaging anyway (user override)."
                )
            _emit_mtp(chain_ngram = True)
            return flags

        # effective_mode == "auto": the promotion path. llama.cpp #22673:
        # MTP is compatible with mmproj, so there's no vision gate.
        if _auto_mla_embedded_mtp:
            # MLA embedded-MTP (GLM-5.2 et al.): the MTP path regresses vs spec-off
            # on llama.cpp today, so Auto drops it and falls back to ngram-mod (or
            # spec-off if unsupported), mirroring the sub-3B branch. Forced mtp /
            # mtp+ngram (handled above) still engage; UNSLOTH_MLA_MTP_ENABLED=1
            # re-enables this promotion once upstream optimizes the path.
            self._spec_fallback_reason = "mla_mtp_disabled"
            _mla_caps = self.probe_server_capabilities(binary)
            if _mla_caps.get("supports_ngram_mod"):
                logger.info(
                    "Auto: MLA embedded-MTP model detected; llama.cpp's MLA/DSA "
                    "MTP path is slower than no speculation, so using ngram-mod "
                    "instead. Override via the Studio Speculative Decoding "
                    "dropdown or UNSLOTH_MLA_MTP_ENABLED=1."
                )
                _emit_ngram_mod()
            else:
                logger.info(
                    "Auto: MLA embedded-MTP model detected; disabling speculative "
                    "decoding (this llama-server does not advertise ngram-mod). "
                    "Override via the dropdown or UNSLOTH_MLA_MTP_ENABLED=1."
                )
                # spec-off: emit nothing, mirroring the sub-3B no-ngram path.
        elif is_mtp_model and not _mtp_too_small:
            if _mtp_drafter_missing:
                # Name-only MTP, drafter did not resolve (download failed/absent).
                _fallback_drafter_not_found()
            else:
                # GPU: MTP-only. CPU/Mac: chain ngram-mod + MTP.
                _emit_mtp(chain_ngram = not gpus)
        elif is_mtp_model and _mtp_too_small:
            # Sub-3B fallback: drop the MTP draft head, keep ngram-mod when
            # the binary supports it.
            if _mtp_drafter_missing:
                _fallback_drafter_not_found()
            elif self.probe_server_capabilities(binary).get("supports_ngram_mod"):
                logger.info(
                    f"MTP GGUF detected but model size {_mtp_size_b:.1f}B "
                    "is below the 3B speedup threshold; using ngram-mod "
                    "only (zero-VRAM, no draft head). Override via "
                    "--spec-type or the Studio Speculative Decoding "
                    "dropdown."
                )
                _emit_ngram_mod()
            else:
                logger.info(
                    f"MTP GGUF detected but model size {_mtp_size_b:.1f}B "
                    "is below the 3B speedup threshold and the bundled "
                    "llama-server does not advertise ngram-mod; "
                    "auto-disabling speculative decoding."
                )
        else:
            # Non-MTP model: let llama-server choose its default strategy.
            flags.append("--spec-default")
            self._speculative_type = "default"
        return flags

    def _already_in_target_state(
        self,
        *,
        model_identifier: str,
        hf_variant: Optional[str],
        n_ctx: int,
        cache_type_kv: Optional[str],
        speculative_type: Optional[str],
        chat_template_override: Optional[str],
        extra_args: Optional[List[str]],
        is_vision: bool,
        gguf_path: Optional[str] = None,
        spec_draft_n_max: Optional[int] = None,
        tensor_parallel: bool = False,
        mtp_draft_path: Optional[str] = None,
        preserve_multi_gpu_on_layer: bool = False,
    ) -> bool:
        """True iff the live server already satisfies these load kwargs.

        Mirrors ``routes/inference.py:_request_matches_loaded_settings`` but
        compares raw kwargs so ``load_model`` can short-circuit a duplicate
        /load that raced past the route-level check (#5401).
        """
        if not self.is_loaded:
            return False
        if (self._model_identifier or "").lower() != (model_identifier or "").lower():
            return False
        # Direct-file loads pass hf_variant=None while the backend stores an
        # extracted filename label; compare paths to keep the guard symmetric.
        if gguf_path is not None and self._gguf_path:
            try:
                if Path(self._gguf_path).resolve() != Path(gguf_path).resolve():
                    return False
            except OSError:
                return False
        elif (self._hf_variant or "").lower() != (hf_variant or "").lower():
            return False
        if self._requested_n_ctx != int(n_ctx):
            return False

        def _norm(value):
            if value is None:
                return None
            if isinstance(value, str):
                stripped = value.strip().lower()
                return stripped or None
            return value

        if _norm(self._cache_type_kv) != _norm(cache_type_kv):
            return False

        # Reconcile a user --split-mode in extras AND an inherited tensor
        # LLAMA_ARG_SPLIT_MODE env, but only against a server that actually
        # launched tensor: if load_model downgraded to layer split it scrubbed
        # the child env, so the env must not force an endless reload of a healthy
        # server. An identical request would downgrade the same way.
        if not _tensor_parallel_matches_loaded(extra_args, tensor_parallel, self._tensor_parallel):
            return False
        # Preserved tensor->layer fallback + an EXPLICIT tensor drop: reload so
        # placement re-selects instead of keeping the all-GPU mask (mirrors the route,
        # #6659). preserve_multi_gpu_on_layer carries the route's carry-forward decision
        # (True for an implicit same-settings reload), so those still dedupe -- the HF
        # auto-pick / local-dir flows skip the route guard and only reach here.
        if (
            self._layer_preserves_tensor_intent
            and not _effective_tensor_parallel(extra_args, tensor_parallel)
            and not preserve_multi_gpu_on_layer
        ):
            return False

        # Compare on the canonical requested mode. With --spec-type in
        # extra_args the backend stores None; mirror that here.
        if _extra_args_set_spec_type(extra_args):
            req_mode = None
        else:
            req_mode = _canonicalize_spec_mode(speculative_type) or "auto"
        backend_mode = self._requested_spec_mode
        if req_mode != backend_mode:
            return False

        # Prior HF load fell back with drafter_not_found; a same-settings reload
        # must retry the download in load_model, not dedupe to the stale fallback
        # (HF loads resolve the drafter there, so gguf_path is None here).
        if (
            self._spec_fallback_reason == "drafter_not_found"
            and gguf_path is None
            and req_mode in ("auto", "mtp", "mtp+ngram")
        ):
            return False

        # spec_draft_n_max only matters when an MTP variant is engaged. Compare
        # on the resolved spec so an Auto request promoted to draft-mtp still
        # bounces a reload when n_max changes.
        if (
            self._speculative_type == "draft-mtp"
            and spec_draft_n_max is not None
            and int(spec_draft_n_max) != (self._spec_draft_n_max or 0)
        ):
            return False

        if (self._chat_template_override or None) != (chat_template_override or None):
            return False

        # A drafter appearing/disappearing next to a local GGUF changes the
        # launch command (--model-draft) when the mode can use it; without
        # this, adding mtp-*.gguf after a load is deduped away and MTP can't
        # engage short of an unload. HF loads resolve the drafter inside
        # load_model (gguf_path is None here), so only local paths compare;
        # the route-level probe covers HF cache repos. No sub-3B gate: both
        # sides come from the same config detection, so a sub-3B mismatch
        # only happens when a drafter genuinely appeared (one benign reload,
        # then the stored path converges).
        if (
            gguf_path is not None
            and req_mode in ("auto", "mtp", "mtp+ngram")
            and (mtp_draft_path or None) != (self._mtp_draft_path or None)
        ):
            return False

        # extra_args=None means "no opinion" (inherit handled at the route
        # layer); only an explicit list forces equality.
        if extra_args is not None:
            current = list(self._extra_args) if self._extra_args is not None else []
            if list(extra_args) != current:
                return False
        return True

    def _classify_gpu_offload(
        self, expected_gpu: bool, detected_gpus: list[tuple[int, int]]
    ) -> Optional[bool]:
        """True if the model landed on a GPU, False if only CPU buffers landed
        despite GPU intent, None when there's no signal. Delegates to the shared
        classifier so it tracks current llama.cpp logs (offloaded-layer counts /
        device_info), not just the older "model buffer size" lines."""
        if not detected_gpus or not expected_gpu:
            return None
        return classify_gpu_offload_lines(self._stdout_lines)

    def load_cancelled(self) -> bool:
        """True if a load was cancelled (e.g. via unload/_cancel_event) and not
        yet consumed by the next load_model. Lets the tensor->layer fallback
        avoid restarting a load the user just cancelled."""
        return self._cancel_event.is_set()

    def unload_model(self) -> bool:
        """Terminate the subprocess and cancel any in-flight download."""
        self._cancel_event.set()
        with self._lock:
            self._kill_process()
            logger.info(f"Unloaded GGUF model: {self._model_identifier}")
            self._model_identifier = None
            self._gguf_path = None
            self._hf_repo = None
            self._mtp_draft_path = None
            self._spec_fallback_reason = None
            self._last_load_kwargs = None
            self._mtp_runtime_fallback_active = False
            self._hf_variant = None
            self._is_vision = False
            self._is_audio = False
            self._audio_type = None
            self._audio_probed = False
            self._has_audio_input = False
            self._mmproj_has_audio = False
            self._port = None
            self._healthy = False
            self._context_length = None
            self._effective_context_length = None
            self._max_context_length = None
            self._chat_template = None
            self._chat_template_override = None
            self._supports_reasoning = False
            self._reasoning_always_on = False
            self._reasoning_style = "enable_thinking"
            self._reasoning_effort_levels = []
            self._reasoning_default = True
            self._supports_preserve_thinking = False
            self._supports_tools = False
            self._cache_type_kv = None
            self._tensor_parallel = False
            self._layer_preserves_tensor_intent = False
            self._speculative_type = None
            self._requested_spec_mode = None
            self._spec_draft_n_max = None
            self._n_layers = None
            self._n_kv_heads = None
            self._n_kv_heads_by_layer = None
            self._n_heads = None
            self._embedding_length = None
            self._kv_key_length = None
            self._kv_value_length = None
            self._sliding_window = None
            self._sliding_window_pattern = None
            self._full_attention_interval = None
            self._kv_lora_rank = None
            self._key_length_mla = None
            self._kv_key_length_swa = None
            self._kv_value_length_swa = None
            self._ssm_inner_size = None
            self._ssm_state_size = None
            self._shared_kv_layers = None
            self._nextn_predict_layers = None
            # Clean up temp chat template file.
            if hasattr(self, "_chat_template_file") and self._chat_template_file:
                try:
                    os.unlink(self._chat_template_file.name)
                except Exception:
                    pass
                self._chat_template_file = None
            # Free audio codec GPU memory.
            if LlamaCppBackend._codec_mgr is not None:
                LlamaCppBackend._codec_mgr.unload()
                LlamaCppBackend._codec_mgr = None
                import torch

                if torch.cuda.is_available():
                    torch.cuda.empty_cache()
            return True

    def _kill_process(self):
        """Terminate the subprocess if running."""
        # Stop the watchdog before a deliberate kill so a planned reload/unload
        # isn't seen as a crash; a real crash never routes through here.
        self._stop_mtp_crash_watchdog()
        if self._process is None:
            return
        try:
            self._process.terminate()
            self._process.wait(timeout = 5)
        except subprocess.TimeoutExpired:
            logger.warning("llama-server did not exit on SIGTERM, sending SIGKILL")
            self._process.kill()
            self._process.wait(timeout = 5)
        except Exception as e:
            logger.warning(f"Error killing llama-server process: {e}")
        finally:
            # getattr: teardown must tolerate a partially-built backend (failed
            # __init__ or a __new__-built instance), as with _llama_log_fh below.
            if getattr(self, "_stats_logger", None) is not None:
                self._stats_logger.stop()
                self._stats_logger = None
            self._process = None
            self._clear_server_pid()
            # Clear healthy so a /load during the replacement's warm-up can't
            # short-circuit against the previous server's health (#5401).
            self._healthy = False
            # Drives _wait_for_vram_settle in the next load_model; set in finally
            # so both in-process and frontend Apply paths record the kill.
            self._last_kill_monotonic = time.monotonic()
            stdout_thread = getattr(self, "_stdout_thread", None)
            if stdout_thread is not None:
                stdout_thread.join(timeout = 2)
                self._stdout_thread = None
            fh = getattr(self, "_llama_log_fh", None)
            if fh is not None:
                try:
                    fh.close()
                except Exception:
                    pass
                self._llama_log_fh = None

    @staticmethod
    def _server_pidfile_path() -> Optional[Path]:
        """Pidfile recording the live llama-server PID, under the active studio root
        (per-root, so concurrent Studios with distinct UNSLOTH_STUDIO_HOME stay
        isolated, mirroring the reaper's custom-root isolation)."""
        try:
            from utils.paths.storage_roots import studio_root  # noqa: WPS433
            return studio_root() / "llama-server.pid"
        except Exception:
            return None

    @classmethod
    def _record_server_pid(cls, pid: int) -> None:
        """Best-effort record of the spawned llama-server PID for orphan reaping.

        Stores ``pid:starttime`` so a later startup can reject a PID that has
        since been recycled to a different process (see ``_pid_start_identity``).
        A bare ``pid`` (no identity) is still accepted on read for compatibility.
        """
        path = cls._server_pidfile_path()
        if path is None:
            return
        try:
            path.parent.mkdir(parents = True, exist_ok = True)
            path.write_text(f"{pid}:{cls._pid_start_identity(pid)}")
        except Exception as e:
            logger.debug(f"Could not write llama-server pidfile: {e}")

    @classmethod
    def _clear_server_pid(cls) -> None:
        """Best-effort removal of the llama-server pidfile."""
        path = cls._server_pidfile_path()
        if path is None:
            return
        try:
            path.unlink(missing_ok = True)
        except Exception as e:
            logger.debug(f"Could not remove llama-server pidfile: {e}")

    @staticmethod
    def _pid_is_llama_server(pid: int) -> bool:
        """True only if pid is a live process whose binary is a llama-server. Guards
        against PID reuse before killing a recorded orphan; returns False on any
        uncertainty so an unrelated process is never killed."""
        try:
            import psutil
            try:
                proc = psutil.Process(pid)
                if (proc.name() or "").lower().startswith("llama-server"):
                    return True
                return Path(proc.exe() or "").name.lower().startswith("llama-server")
            except (psutil.NoSuchProcess, psutil.AccessDenied, psutil.ZombieProcess):
                return False
        except ImportError:
            pass
        if sys.platform != "linux":
            return False
        try:
            if Path(os.readlink(f"/proc/{pid}/exe")).name.lower().startswith("llama-server"):
                return True
        except OSError:
            pass
        try:
            with open(f"/proc/{pid}/cmdline", "rb") as fh:
                tokens = fh.read().split(b"\x00")
            first = tokens[0].decode("utf-8", "replace") if tokens else ""
            return Path(first).name.lower().startswith("llama-server")
        except OSError:
            return False

    @staticmethod
    def _pid_start_identity(pid: int) -> str:
        """Stable per-PID identity (process start time) guarding against PID reuse.

        Returns a token string, or "" when it cannot be determined (the caller
        then falls back to the llama-server name check only)."""
        try:
            import psutil
            try:
                return str(psutil.Process(pid).create_time())
            except (psutil.NoSuchProcess, psutil.AccessDenied, psutil.ZombieProcess):
                return ""
        except ImportError:
            pass
        if sys.platform == "linux":
            try:
                with open(f"/proc/{pid}/stat", "rb") as fh:
                    data = fh.read()
                # field 22 (starttime), counted from after the ")" that closes comm.
                return data[data.rfind(b")") + 2 :].split()[19].decode()
            except (OSError, IndexError):
                return ""
        return ""

    @staticmethod
    def _pid_parent_is_alive(pid: int) -> bool:
        """True if the recorded server's parent is still running, i.e. the server is
        NOT orphaned. Lets the cross-session reap kill only a true orphan (parent
        gone) and never a live server owned by a running Studio, regardless of which
        process performs the sweep. Biased toward "alive" on uncertainty so a live
        server is never mistakenly reaped."""
        try:
            import psutil

            try:
                ppid = psutil.Process(pid).ppid()
            except psutil.NoSuchProcess:
                return False  # the recorded server itself is gone
            except psutil.Error:
                return True  # cannot tell -- never risk killing a live server
            if ppid <= 1:
                return False  # reparented to init -> orphan
            return psutil.pid_exists(ppid)
        except ImportError:
            pass
        if sys.platform == "linux":
            try:
                with open(f"/proc/{pid}/stat", "rb") as fh:
                    data = fh.read()
                ppid = int(data[data.rfind(b")") + 2 :].split()[1])
            except (OSError, IndexError, ValueError):
                return False
            if ppid <= 1:
                return False
            return Path(f"/proc/{ppid}").exists()
        return False

    @staticmethod
    def _unlink_pidfile(path: Path) -> None:
        """Best-effort removal of a resolved pidfile path."""
        try:
            path.unlink(missing_ok = True)
        except Exception:
            pass

    @classmethod
    def _reap_recorded_pid(cls) -> int:
        """Kill the exact llama-server PID recorded at spawn, but only when it is a
        genuine orphan -- its parent (the Studio that spawned it) is gone. This is
        the cross-session backstop the parent-death reaper (Job Object /
        PR_SET_PDEATHSIG) cannot cover: an orphan left by an already-dead Studio
        (macOS, a best-effort failure, or a pre-existing orphan). Path-independent,
        so it also catches an orphan the install-root match would miss.

        A live server whose parent is still running is never reaped, so constructing
        a second backend in-process (the helper / advisor paths each build a
        LlamaCppBackend) cannot kill the active chat server. A recorded PID that has
        been recycled to a different process is rejected by the start-time identity
        and the llama-server name check, so unrelated user processes are never
        touched. SIGKILL falls back to SIGTERM on Windows, where os.kill maps it to
        TerminateProcess and SIGKILL is undefined."""
        path = cls._server_pidfile_path()
        if path is None or not path.exists():
            return 0

        pid = -1
        identity = ""
        try:
            pid_str, _, identity = path.read_text().strip().partition(":")
            pid = int(pid_str)
        except Exception:
            pid = -1

        if pid <= 0:
            cls._unlink_pidfile(path)  # garbage record
            return 0
        if pid == os.getpid():
            return 0  # never our own pid; leave the record alone

        if cls._pid_parent_is_alive(pid):
            # Live server with a running parent -> not an orphan; keep the record so
            # a later startup can still reap it if that parent later dies abnormally.
            return 0

        # Parent is gone: candidate orphan. Reject a PID recycled to something else.
        if identity and cls._pid_start_identity(pid) != identity:
            cls._unlink_pidfile(path)
            return 0

        killed = 0
        if cls._pid_is_llama_server(pid):
            try:
                os.kill(pid, getattr(signal, "SIGKILL", signal.SIGTERM))
                killed = 1
                logger.info(f"Killed orphaned llama-server from pidfile (pid={pid})")
            except (ProcessLookupError, PermissionError):
                pass
            except Exception as e:
                logger.debug(f"Could not kill recorded llama-server pid {pid}: {e}")
        cls._unlink_pidfile(path)
        return killed

    @staticmethod
    def _kill_orphaned_servers() -> int:
        """Kill orphaned llama-server processes started by studio.

        Only kills processes whose resolved binary lives under a known
        Studio install dir (or matches an exact env-var override), to avoid
        terminating unrelated llama-server instances. Mirrors every location
        _find_llama_server_binary() can return, so orphans from any
        supported install path are cleaned up.

        Uses psutil for cross-platform support (Linux, macOS, Windows);
        falls back to pgrep + /proc/<pid>/exe on Linux when psutil is
        absent.

        Returns the count of processes killed; callers arm the VRAM-settle
        wait on a positive count.
        """
        # Cross-session backstop first: reap the exact PID we recorded at spawn,
        # but only if it is a true orphan whose parent is gone (so a helper backend
        # built while a chat server is live can never kill it). The root-gated
        # enumeration below stays as a fallback.
        killed = LlamaCppBackend._reap_recorded_pid()
        try:
            # -- Build the ownership allowlist --------------------------------
            # exact_binaries -- env var overrides (exact path match).
            # install_roots  -- Studio-owned dir trees (binary must be under one).
            install_roots: list[Path] = []

            # Env-mode custom root (mirrors _find_llama_server_binary).
            _resolved_sr, _is_legacy = LlamaCppBackend._resolved_studio_root_and_is_legacy()
            _is_custom_root = not _is_legacy
            if _is_custom_root:
                install_roots.append(_resolved_sr / "llama.cpp")

            # Primary install dir (default mode only). Env-mode skips this so a
            # custom-root Studio can't kill a default-install Studio's server.
            if not _is_custom_root:
                install_roots.append(Path.home() / ".unsloth" / "llama.cpp")

            # Legacy in-tree build dirs (older setup.sh)
            project_root = Path(__file__).resolve().parents[4]
            install_roots.append(project_root / "llama.cpp")

            # Legacy: extracted binary
            install_roots.append(project_root / "bin")

            # UNSLOTH_LLAMA_CPP_PATH env var (custom install dir)
            custom_dir = os.environ.get("UNSLOTH_LLAMA_CPP_PATH")
            if custom_dir:
                install_roots.append(Path(custom_dir))

            # LLAMA_SERVER_PATH env var (exact binary path)
            exact_binaries: list[Path] = []
            env_binary = os.environ.get("LLAMA_SERVER_PATH")
            if env_binary:
                try:
                    exact_binaries.append(Path(env_binary).resolve())
                except OSError:
                    pass

            # Resolve all roots so is_relative_to works reliably.
            resolved_roots: list[Path] = []
            for root in install_roots:
                try:
                    # A --with-llama-cpp-dir local link (symlink/junction)
                    # resolves into the user's own checkout. Adding it would let
                    # us treat the user's externally-launched llama-server as our
                    # orphan and kill it, so leave such roots out of the
                    # allowlist (we forgo orphan-reaping for local-link installs).
                    if _is_external_link(root):
                        continue
                    resolved_roots.append(root.resolve())
                except OSError:
                    pass

            my_pid = os.getpid()

            # -- Enumerate processes -------------------------------------------
            # Prefer psutil (cross-platform); fall back to pgrep + /proc on
            # Linux when psutil is absent.
            try:
                import psutil
                has_psutil = True
            except ImportError:
                has_psutil = False

            if has_psutil:
                for proc in psutil.process_iter(["pid", "name", "exe"]):
                    try:
                        if proc.info["pid"] == my_pid:
                            continue

                        name = proc.info.get("name") or ""
                        if not name.lower().startswith("llama-server"):
                            continue

                        exe = proc.info.get("exe")
                        if not exe:
                            continue

                        exe_path = Path(exe).resolve()

                        # Ownership: exact match OR binary under a known root.
                        is_ours = exe_path in exact_binaries or any(
                            exe_path.is_relative_to(root) for root in resolved_roots
                        )
                        if not is_ours:
                            continue

                        proc.kill()
                        killed += 1
                        logger.info(
                            f"Killed orphaned llama-server process (pid={proc.info['pid']})"
                        )
                    except (
                        psutil.NoSuchProcess,
                        psutil.AccessDenied,
                        psutil.ZombieProcess,
                    ):
                        pass
            else:
                # -- Fallback: pgrep + /proc/<pid>/exe (Linux only) -----------
                if sys.platform != "linux":
                    return killed
                result = subprocess.run(
                    ["pgrep", "-a", "-f", "llama-server"],
                    capture_output = True,
                    text = True,
                    timeout = 5,
                    env = child_env_without_native_path_secret(),
                )
                if result.returncode != 0:
                    return killed

                for line in result.stdout.strip().splitlines():
                    parts = line.strip().split(None, 1)
                    if len(parts) < 2:
                        continue
                    pid = int(parts[0])
                    if pid == my_pid:
                        continue

                    # /proc/<pid>/exe symlinks the real binary, avoiding
                    # cmdline-parsing ambiguities; fall back to the first
                    # cmdline token when /proc is unavailable.
                    proc_exe = Path(f"/proc/{pid}/exe")
                    try:
                        binary = proc_exe.resolve(strict = True)
                    except (OSError, ValueError):
                        cmdline = parts[1]
                        token = cmdline.split()[0] if cmdline.strip() else ""
                        if not token:
                            continue
                        binary = Path(token).resolve(strict = False)

                    owned = binary in exact_binaries or any(
                        binary.is_relative_to(root) for root in resolved_roots
                    )
                    if not owned:
                        continue

                    try:
                        os.kill(pid, signal.SIGKILL)
                        killed += 1
                        logger.info(f"Killed orphaned llama-server process (pid={pid})")
                    except ProcessLookupError:
                        pass
                    except PermissionError:
                        pass
        except Exception:
            logger.warning("Error during orphan server cleanup", exc_info = True)
        return killed

    def _cleanup(self):
        """atexit handler to ensure llama-server is terminated."""
        self._kill_process()

    @staticmethod
    def _fit_off_retry_eligible(cmd: "list[str]", use_fit: bool) -> bool:
        """Whether a llama-server startup crash may be retried with --fit off.

        Only when Studio's own VRAM math placed the model (use_fit=False)
        and nothing on the command line set the fit mode explicitly
        (-fit / --fit, space- or equals-form). --fit-ctx / --fit-target /
        -fitc / -fitt tune the fit step but do not select the mode, so
        they do not block the retry.
        """
        if use_fit:
            return False
        for a in cmd:
            if a in ("-fit", "--fit") or a.startswith(("-fit=", "--fit=")):
                return False
        return True

    def _probe_mtp_decode(self, timeout: float = 60.0) -> bool:
        """One tiny /completion to confirm MTP survives the first decode.

        MTP-draft can pass /health yet crash the flash-attn kernel only once
        tokens generate (e.g. under --split-mode tensor). False on any error so
        the caller can drop MTP and retry.
        """
        url = f"{self.base_url}/completion"
        payload = {"prompt": "Hi", "n_predict": 4, "temperature": 0.0, "stream": False}
        try:
            resp = httpx.post(
                url,
                json = payload,
                timeout = timeout,
                headers = self._auth_headers,
                trust_env = False,
            )
        except Exception as e:
            logger.debug(f"MTP decode probe failed: {e}")
            return False
        if resp.status_code != 200:
            logger.debug(f"MTP decode probe returned HTTP {resp.status_code}")
            return False
        # A crash can drop the connection or kill the process right after a reply.
        if self._process is not None and self._process.poll() is not None:
            return False
        return True

    def _maybe_recover_from_mtp_crash(self, exc: Optional[BaseException] = None) -> bool:
        """Schedule one background reload without MTP after a mid-generation death.

        MTP+tensor can crash the flash-attn kernel on a later request, after
        load_model returned, past the load-time fallback and decode probe. Not a
        persistent ban: a fresh load re-tries MTP. Returns True if scheduled.
        """
        # Cheap async-safe gate: only our live MTP+tensor launch, not cancelled,
        # with a snapshot to replay.
        if self._cancel_event.is_set():
            return False
        if not self._mtp_runtime_fallback_active:
            return False
        if not self._last_load_kwargs or self._process is None:
            return False
        # Single-flight: the first failure claims the reload.
        with self._mtp_runtime_fallback_lock:
            if self._mtp_runtime_fallback_in_progress:
                return False
            self._mtp_runtime_fallback_in_progress = True
        snapshot = dict(self._last_load_kwargs)
        proc = self._process

        def _recover():
            try:
                # Confirm the process really exited (the error can arrive a beat
                # early) so a transient stream error can't disable MTP.
                deadline = time.monotonic() + 5.0
                while proc.poll() is None and time.monotonic() < deadline:
                    time.sleep(0.1)
                if proc.poll() is None:
                    logger.debug("Generation error but llama-server is alive; keeping MTP.")
                    return
                logger.warning(
                    "llama-server exited mid-generation with MTP under tensor "
                    "parallelism (%s); reloading without speculative decoding.",
                    type(exc).__name__ if exc is not None else "server exited",
                )
                # Re-check under the load lock (RLock allows the nested
                # load_model) so a newer load isn't clobbered by this stale replay.
                requested_mode = snapshot.get("speculative_type")
                with self._serial_load_lock:
                    if self._cancel_event.is_set():
                        logger.info("MTP-crash reload skipped: load was cancelled/unloaded.")
                        return
                    if self._process is not proc:
                        logger.info("MTP-crash reload skipped: a newer load is already active.")
                        return
                    if self._last_load_kwargs != snapshot:
                        logger.info("MTP-crash reload skipped: load settings changed.")
                        return
                    snapshot["speculative_type"] = "off"
                    # Drop user/env MTP too: append a last-wins --spec-default.
                    _ea = list(snapshot.get("extra_args") or [])
                    if _extra_args_requests_mtp(_ea, env = os.environ):
                        _ea.append("--spec-default")
                        snapshot["extra_args"] = _ea
                    self.load_model(**snapshot)
                    # Restore the requested mode + reason load_model("off") cleared,
                    # so /status shows the user's mode + note (like the startup fallback).
                    self._requested_spec_mode = _canonicalize_spec_mode(requested_mode)
                    self._spec_fallback_reason = "runtime_error"
                logger.info("Reloaded without MTP after the tensor-parallel crash.")
            except Exception as e:
                logger.error(f"Reload without MTP failed: {e}")
            finally:
                with self._mtp_runtime_fallback_lock:
                    self._mtp_runtime_fallback_in_progress = False

        threading.Thread(target = _recover, daemon = True, name = "mtp-crash-reload").start()
        return True

    def _start_mtp_crash_watchdog(self) -> None:
        """Background poll that recovers on an MTP+tensor crash even when no
        request observes it (direct proxy endpoints, or nothing in flight).

        Armed only for a live MTP+tensor launch; the no-MTP reload disarms it, so
        it can't loop.
        """
        if not self._mtp_runtime_fallback_active:
            return
        proc = self._process
        if proc is None:
            return
        # Replace any prior watchdog (loads are serialised, so at most one).
        self._stop_mtp_crash_watchdog()
        stop = threading.Event()
        self._mtp_watchdog_stop = stop

        def _watch():
            # Exit on stop or process death. _kill_process sets stop before
            # terminating, so re-check it: only a real crash (stop unset) recovers.
            while not stop.wait(1.0):
                if proc.poll() is not None:
                    if not stop.is_set():
                        self._maybe_recover_from_mtp_crash()
                    return

        t = threading.Thread(target = _watch, daemon = True, name = "mtp-crash-watchdog")
        self._mtp_watchdog_thread = t
        t.start()

    def _stop_mtp_crash_watchdog(self) -> None:
        """Signal the crash watchdog to exit; called before any deliberate kill."""
        stop = getattr(self, "_mtp_watchdog_stop", None)
        if stop is not None:
            stop.set()
        self._mtp_watchdog_thread = None

    def _wait_for_health(
        self,
        timeout: float = 120.0,
        interval: float = 0.5,
    ) -> bool:
        """Poll llama-server's /health until 200; also detect early exit/crash."""
        deadline = time.monotonic() + timeout
        url = f"{self.base_url}/health"

        while time.monotonic() < deadline:
            # Process crashed?
            if self._process.poll() is not None:
                # Let the drain thread collect final output.
                if self._stdout_thread is not None:
                    self._stdout_thread.join(timeout = 2)
                output = "\n".join(self._stdout_lines[-50:])
                # Keep the TAIL: crash details (abort reason, ROCm/CUDA error
                # text) print last, after the long startup banner. Head
                # truncation has cut off exactly the diagnostic line before.
                _log_hint = (
                    f" Full log: {self._llama_log_path}"
                    if getattr(self, "_llama_log_path", None)
                    else ""
                )
                logger.error(
                    f"llama-server exited with code {self._process.returncode}. "
                    f"Output (tail): {output[-2000:]}{_log_hint}"
                )
                return False

            try:
                # trust_env=False: skip ambient HTTP(S)_PROXY, which if it 503s
                # for 127.0.0.1 loops the probe until timeout and hangs load.
                resp = httpx.get(url, timeout = 2.0, trust_env = False)
                if resp.status_code == 200:
                    return True
            except (
                httpx.ConnectError,
                httpx.TimeoutException,
                # ReadError covers TCP RST mid-read while still binding the port
                # (Windows: WinError 10054); the crash branch catches real exits.
                httpx.ReadError,
                httpx.RemoteProtocolError,
                httpx.WriteError,
            ):
                pass

            time.sleep(interval)

        # Leave a marker so _classify_llama_start_failure tells a live but
        # never-healthy load (too large, or a proxy hijacking the loopback
        # probe) apart from a bad GGUF (#5740).
        self._stdout_lines.append(f"llama-server health check timed out after {timeout}s")
        logger.error(f"llama-server health check timed out after {timeout}s")
        return False

    @staticmethod
    def _ctx_integrity_flags(
        n_parallel: int, use_fit: bool, requested_ctx: int, effective_ctx: int, caps: dict
    ) -> list[str]:
        """Flags that keep the per-request window equal to the advertised ctx.

        Explicit ``--parallel`` disables llama-server's auto-slots
        ``--kv-unified`` default, silently splitting ``-c`` into per-slot
        windows of ``-c / N``; restore the shared pool so one request can use
        the full context. With ``--fit on``, ``--fit-ctx`` floors the fit step
        at an explicitly requested ctx (default floor is 4096) so it offloads
        or fails instead of silently shrinking the window.
        """
        flags: list[str] = []
        if n_parallel > 1 and caps.get("supports_kv_unified"):
            flags.append("--kv-unified")
        if use_fit and requested_ctx > 0 and effective_ctx > 0 and caps.get("supports_fit_ctx"):
            flags.extend(["--fit-ctx", str(effective_ctx)])
        return flags

    def _query_server_n_ctx(self) -> Optional[int]:
        """Per-slot context llama-server actually allocated, from ``/props``.

        The memory-fit step or ``--parallel`` slot split can leave this below
        the requested ``-c``; requests are validated against this value.
        """
        url = f"{self.base_url}/props"
        try:
            resp = httpx.get(url, timeout = 5.0, trust_env = False)
            if resp.status_code != 200:
                return None
            settings = resp.json().get("default_generation_settings") or {}
            n_ctx = settings.get("n_ctx")
            return int(n_ctx) if n_ctx else None
        except Exception:
            return None

    def _reconcile_effective_ctx_with_server(self) -> None:
        """Adopt the server's real ``n_ctx`` when it is below Studio's value.

        Keeps ``context_length`` (load response, status route, passthrough
        ``max_tokens`` ceiling) honest; clients sized to the requested value
        would otherwise hit ``exceed_context_size_error`` 400s early.
        """
        actual_n_ctx = self._query_server_n_ctx()
        if not actual_n_ctx or actual_n_ctx <= 0:
            return
        if self._effective_context_length and actual_n_ctx < self._effective_context_length:
            logger.warning(
                "llama-server allocated a smaller per-request context than "
                f"requested ({self._effective_context_length} -> {actual_n_ctx}; "
                "memory fit or --parallel slot split); clients must treat "
                f"{actual_n_ctx} as the real context window."
            )
            self._effective_context_length = actual_n_ctx
        elif not self._effective_context_length:
            self._effective_context_length = actual_n_ctx

    # ── Message building (OpenAI format) ──────────────────────────

    @staticmethod
    def _parse_tool_calls_from_text(
        content: str,
        *,
        allow_incomplete: bool = True,
        enabled_tool_names: Optional[set] = None,
    ) -> list[dict]:
        """Wrapper around the shared parser; ``enabled_tool_names`` gates the markerless bare-JSON form."""
        return _shared_parse_tool_calls_from_text(
            content,
            allow_incomplete = allow_incomplete,
            enabled_tool_names = enabled_tool_names,
        )

    @staticmethod
    def _build_openai_messages(messages: list[dict], image_b64: Optional[str] = None) -> list[dict]:
        """Build OpenAI-format messages, optionally injecting an image_url part
        into the last user message for vision models. As-is if no image."""
        if not image_b64:
            return messages

        # Convert the last user message to multimodal content parts
        result = [msg.copy() for msg in messages]
        last_user_idx = None
        for i, msg in enumerate(result):
            if msg["role"] == "user":
                last_user_idx = i

        if last_user_idx is not None:
            text_content = result[last_user_idx].get("content", "")
            result[last_user_idx]["content"] = [
                {"type": "text", "text": text_content},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/png;base64,{image_b64}",
                    },
                },
            ]

        return result

    # ── Generation (proxy to llama-server) ────────────────────────

    @contextlib.contextmanager
    def _open_stream(self, url: str, payload: dict, cancel_event):
        """Open a streaming POST to llama-server, retrying through prefill, and
        yield ``(response, first_token_deadline)`` once a 200 lands. Owns the
        httpx.Client + auth headers for the stream's lifetime; raises
        RuntimeError on a non-200. Shared scaffold for the streaming consumers,
        which differ only in how they parse the SSE body."""
        stream_timeout = httpx.Timeout(connect = 10, read = 0.5, write = 10, pool = 10)
        with httpx.Client(
            timeout = stream_timeout,
            limits = httpx.Limits(max_keepalive_connections = 0),
            trust_env = False,
        ) as client:
            first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
            with self._stream_with_retry(
                client,
                url,
                payload,
                cancel_event,
                headers = self._auth_headers,
                first_token_deadline = first_token_deadline,
            ) as response:
                if response.status_code != 200:
                    error_body = response.read().decode()
                    raise RuntimeError(
                        f"llama-server returned {response.status_code}: {error_body}"
                    )
                yield response, first_token_deadline

    @staticmethod
    def _iter_text_cancellable(
        response: "httpx.Response",
        cancel_event: Optional[threading.Event] = None,
        stall_timeout_s: float = _DEFAULT_STREAM_STALL_TIMEOUT_S,
        first_token_deadline: Optional[float] = None,
        post_first_chunk_read_timeout_s: Optional[float] = _DEFAULT_STREAM_STALL_TIMEOUT_S,
    ) -> Generator[str, None, None]:
        """Iterate a stream while polling cancel and stall timeouts."""
        text_iter = response.iter_text()
        if first_token_deadline is None:
            first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
        last_chunk_at: Optional[float] = None
        while True:
            if cancel_event is not None and cancel_event.is_set():
                response.close()
                return
            try:
                if last_chunk_at is None:
                    remaining_s = first_token_deadline - time.monotonic()
                    if remaining_s <= 0:
                        raise httpx.ReadTimeout("The model did not produce a first token in time.")
                    LlamaCppBackend._set_stream_read_timeout(response, remaining_s)
                chunk = next(text_iter)
                if chunk:
                    if last_chunk_at is None and post_first_chunk_read_timeout_s is not None:
                        LlamaCppBackend._set_stream_read_timeout(
                            response,
                            post_first_chunk_read_timeout_s,
                        )
                    last_chunk_at = time.monotonic()
                yield chunk
            except StopIteration:
                return
            except httpx.ReadTimeout:
                now = time.monotonic()
                if last_chunk_at is None:
                    if now >= first_token_deadline:
                        raise
                elif now - last_chunk_at >= stall_timeout_s:
                    raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
                continue

    @staticmethod
    def _set_stream_read_timeout(response: "httpx.Response", read_timeout_s: float) -> None:
        """Lower only post-header stream reads; keep prefill timeout long."""
        try:
            timeout_ext = response.request.extensions.get("timeout")
            if isinstance(timeout_ext, dict):
                timeout_ext["read"] = read_timeout_s
        except Exception:
            logger.debug("Could not lower response read timeout", exc_info = True)

    @staticmethod
    def _shutdown_active_httpx_sockets(client: "httpx.Client") -> None:
        """Best-effort interrupt for a sync httpx request blocked before headers."""
        try:
            pool = getattr(getattr(client, "_transport", None), "_pool", None)
            connections = list(getattr(pool, "_connections", []) or [])
            for connection in connections:
                inner = getattr(connection, "_connection", None)
                stream = getattr(inner, "_network_stream", None)
                sock = getattr(stream, "_sock", None)
                if sock is None:
                    continue
                try:
                    sock.shutdown(socket.SHUT_RDWR)
                except OSError:
                    pass
                try:
                    sock.close()
                except OSError:
                    pass
        except Exception:
            logger.debug("Could not shutdown active httpx socket", exc_info = True)
        try:
            client.close()
        except Exception:
            logger.debug("Could not close httpx client", exc_info = True)

    @staticmethod
    @contextlib.contextmanager
    def _stream_with_retry(
        client: "httpx.Client",
        url: str,
        payload: dict,
        cancel_event: Optional[threading.Event] = None,
        headers: Optional[dict] = None,
        first_token_deadline: Optional[float] = None,
    ):
        """Open one streaming POST and let cancel interrupt prefill or reads."""
        if cancel_event is not None and cancel_event.is_set():
            raise GeneratorExit

        _cancel_closed = threading.Event()
        _response_ref: list = [None]

        def _cancel_watcher():
            while not _cancel_closed.is_set():
                if cancel_event.wait(timeout = 0.3):
                    while not _cancel_closed.is_set():
                        r = _response_ref[0]
                        try:
                            if r is not None:
                                r.close()
                            else:
                                LlamaCppBackend._shutdown_active_httpx_sockets(client)
                            return
                        except Exception as e:
                            logger.debug(f"Error closing request in cancel watcher: {e}")
                        _cancel_closed.wait(timeout = 0.1)
                    return

        watcher = None
        if cancel_event is not None:
            watcher = threading.Thread(target = _cancel_watcher, daemon = True, name = "prefill-cancel")
            watcher.start()

        try:
            if first_token_deadline is None:
                first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
            prefill_read_timeout = max(0.1, first_token_deadline - time.monotonic())
            prefill_timeout = httpx.Timeout(
                connect = 30,
                read = prefill_read_timeout,
                write = 10,
                pool = 10,
            )
            with client.stream(
                "POST",
                url,
                json = payload,
                timeout = prefill_timeout,
                headers = headers,
            ) as response:
                _response_ref[0] = response
                if cancel_event is not None and cancel_event.is_set():
                    raise GeneratorExit
                yield response
                return
        except (httpx.RequestError, RuntimeError):
            # Response was closed by the cancel watcher
            if cancel_event is not None and cancel_event.is_set():
                raise GeneratorExit
            raise
        finally:
            _cancel_closed.set()

    def _respawn_if_dead(self) -> bool:
        """Relaunch the llama-server if its process has exited.

        A loaded chat model can be SIGKILL'd mid-session (usually GPU/RAM pressure
        from a training run on the same box), leaving a defunct process while
        ``is_loaded`` still reads True. Replay the last ``load_model`` call to
        recover, returning True once healthy. Serialised on ``_respawn_lock`` so
        many generations hitting the dead server trigger at most one reload.
        """
        with self._respawn_lock:
            proc = self._process
            if proc is None:
                return False
            if proc.poll() is None:
                # Process is alive: either a concurrent caller already respawned
                # it (healthy), or this connection error wasn't a dead server.
                return self._healthy
            kwargs = self._last_load_kwargs
            if not kwargs:
                return False
            logger.warning(
                f"llama-server for '{self._model_identifier}' exited "
                f"(code {proc.returncode}); respawning to recover the session"
            )
            with self._lock:
                self._healthy = False
            try:
                return bool(self.load_model(**kwargs))
            except Exception as exc:
                logger.error(f"Failed to respawn llama-server: {exc}")
                return False

    def generate_chat_completion(
        self,
        messages: list[dict],
        image_b64: Optional[str] = None,
        temperature: float = 0.6,
        top_p: float = 0.95,
        top_k: int = 20,
        min_p: float = 0.01,
        max_tokens: Optional[int] = None,
        repetition_penalty: float = 1.0,
        presence_penalty: float = 0.0,
        stop: Optional[list[str]] = None,
        cancel_event: Optional[threading.Event] = None,
        enable_thinking: Optional[bool] = None,
        reasoning_effort: Optional[str] = None,
        preserve_thinking: Optional[bool] = None,
        seed: Optional[int] = None,
        _allow_respawn_retry: bool = True,
    ) -> Generator[Union[str, dict], None, None]:
        """
        Send a chat completion to llama-server and stream tokens back.

        Uses /v1/chat/completions -- llama-server applies the chat template
        and handles vision (multimodal image_url parts) natively.

        Yields cumulative text (matching InferenceBackend's convention).
        """
        if not self.is_loaded:
            raise RuntimeError("llama-server is not loaded")

        openai_messages = self._build_openai_messages(messages, image_b64)

        payload = {
            "messages": openai_messages,
            "stream": True,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k if top_k >= 0 else 0,
            "min_p": min_p,
            "repeat_penalty": repetition_penalty,
            "presence_penalty": presence_penalty,
        }
        # Per-request enable_thinking / reasoning_effort / preserve_thinking
        _reasoning_kw = self._request_reasoning_kwargs(
            enable_thinking, reasoning_effort, preserve_thinking
        )
        if _reasoning_kw is not None:
            payload["chat_template_kwargs"] = _reasoning_kw
        # Default cap to the model context when known.
        payload["max_tokens"] = (
            max_tokens
            if max_tokens is not None
            else (self._effective_context_length or _DEFAULT_MAX_TOKENS_FLOOR)
        )
        if stop:
            payload["stop"] = stop
        if seed is not None:
            payload["seed"] = seed
        payload["stream_options"] = {"include_usage": True}

        url = f"{self.base_url}/v1/chat/completions"
        cumulative = ""
        in_thinking = False
        _stream_done = False
        _metadata_usage = None
        _metadata_timings = None
        _metadata_finish_reason = None

        try:
            with self._open_stream(url, payload, cancel_event) as (
                response,
                first_token_deadline,
            ):
                buffer = ""
                has_content_tokens = False
                reasoning_text = ""
                for raw_chunk in self._iter_text_cancellable(
                    response,
                    cancel_event,
                    first_token_deadline = first_token_deadline,
                ):
                    buffer += raw_chunk
                    while "\n" in buffer:
                        line, buffer = buffer.split("\n", 1)
                        line = line.strip()

                        if not line:
                            continue
                        if line == "data: [DONE]":
                            if in_thinking:
                                if has_content_tokens:
                                    # Real thinking + content: close the tag
                                    cumulative += "</think>"
                                    yield cumulative
                                else:
                                    # Only reasoning_content, no content:
                                    # model put its whole reply in reasoning
                                    # (e.g. Qwen3 always-think). Show it as
                                    # the main response, not a thinking block.
                                    cumulative = reasoning_text
                                    yield cumulative
                            _stream_done = True
                            break  # exit inner while
                        if not line.startswith("data: "):
                            continue

                        try:
                            data = json.loads(line[6:])
                            # Diffusion frame (per-step canvas) from the shim: forward untouched so
                            # the frontend renders it in place. No assistant text, so it never enters
                            # the cumulative content.
                            if data.get("type") == "diffusion_frame":
                                yield data
                                continue
                            # Capture server timings/usage from final chunks.
                            _chunk_timings = data.get("timings")
                            if _chunk_timings:
                                _metadata_timings = _chunk_timings
                            _chunk_usage = data.get("usage")
                            if _chunk_usage:
                                _metadata_usage = _chunk_usage
                            choices = data.get("choices", [])
                            if choices:
                                delta = choices[0].get("delta", {})
                                _fr = choices[0].get("finish_reason")
                                if _fr:
                                    _metadata_finish_reason = _fr

                                # Reasoning/thinking tokens: llama-server
                                # sends these as "reasoning_content"; wrap
                                # in <think> tags for the frontend parser.
                                reasoning = delta.get("reasoning_content", "")
                                if reasoning:
                                    reasoning_text += reasoning
                                    if not in_thinking:
                                        cumulative += "<think>"
                                        in_thinking = True
                                    cumulative += reasoning
                                    yield cumulative

                                token = delta.get("content", "")
                                if token:
                                    has_content_tokens = True
                                    if in_thinking:
                                        cumulative += "</think>"
                                        in_thinking = False
                                    cumulative += token
                                    yield cumulative
                        except json.JSONDecodeError:
                            logger.debug(f"Skipping malformed SSE line: {line[:100]}")
                    if _stream_done:
                        break  # exit outer for
                if _metadata_usage or _metadata_timings or _metadata_finish_reason:
                    _metadata_usage = _backfill_usage_from_timings(
                        _metadata_usage, _metadata_timings
                    )
                    yield {
                        "type": "metadata",
                        # Never None: a finish-only metadata event (no usage,
                        # no timings) would otherwise crash consumers that do
                        # usage.get(...) on the non-streaming paths.
                        "usage": _metadata_usage or {},
                        "timings": _metadata_timings,
                        "finish_reason": _metadata_finish_reason,
                    }

        except httpx.ConnectError as e:
            # Server already down. If this was an MTP+tensor crash, recover by
            # reloading without MTP (scheduled in the background) and fail this
            # request. Otherwise the server was likely SIGKILL'd by GPU pressure
            # from a concurrent training run: respawn the same config and retry the
            # generation once (bounded by the private flag, no duplicate output).
            if self._maybe_recover_from_mtp_crash(e):
                raise RuntimeError("Lost connection to llama-server")
            if _allow_respawn_retry and not cumulative and self._respawn_if_dead():
                logger.warning(
                    "llama-server was unreachable; respawned it and retrying the generation"
                )
                yield from self.generate_chat_completion(
                    messages,
                    image_b64 = image_b64,
                    temperature = temperature,
                    top_p = top_p,
                    top_k = top_k,
                    min_p = min_p,
                    max_tokens = max_tokens,
                    repetition_penalty = repetition_penalty,
                    presence_penalty = presence_penalty,
                    stop = stop,
                    cancel_event = cancel_event,
                    enable_thinking = enable_thinking,
                    reasoning_effort = reasoning_effort,
                    preserve_thinking = preserve_thinking,
                    seed = seed,
                    _allow_respawn_retry = False,
                )
                return
            raise RuntimeError("Lost connection to llama-server")
        except Exception as e:
            if cancel_event is not None and cancel_event.is_set():
                return
            # Died mid-generation: recover MTP, re-raise unchanged for this request.
            self._maybe_recover_from_mtp_crash(e)
            raise

    # ── Tool-calling agentic loop ──────────────────────────────

    def generate_chat_completion_with_tools(
        self,
        messages: list[dict],
        tools: list[dict],
        temperature: float = 0.6,
        top_p: float = 0.95,
        top_k: int = 20,
        min_p: float = 0.01,
        max_tokens: Optional[int] = None,
        repetition_penalty: float = 1.0,
        presence_penalty: float = 0.0,
        stop: Optional[list[str]] = None,
        cancel_event: Optional[threading.Event] = None,
        enable_thinking: Optional[bool] = None,
        reasoning_effort: Optional[str] = None,
        preserve_thinking: Optional[bool] = None,
        max_tool_iterations: int = 25,
        auto_heal_tool_calls: bool = True,
        nudge_tool_calls: Optional[bool] = None,
        tool_call_timeout: int = 300,
        session_id: Optional[str] = None,
        rag_scope: Optional[dict] = None,
        seed: Optional[int] = None,
        disable_parallel_tool_use: bool = False,
        confirm_tool_calls: bool = False,
        bypass_permissions: bool = False,
    ) -> Generator[dict, None, None]:
        """
        Agentic loop: let the model call tools, execute them, and continue.

        Yields dicts:
          {"type": "status", "text": "Searching: ..."/"Reading: ..."}   -- tool status updates
          {"type": "content", "text": "token"}            -- streamed content tokens (cumulative)
          {"type": "reasoning", "text": "token"}          -- streamed reasoning tokens (cumulative)
        """
        from core.inference.tools import build_rag_autoinject, execute_tool

        if not self.is_loaded:
            raise RuntimeError("llama-server is not loaded")

        conversation = list(messages)

        # Forced first-pass RAG so a doc question doesn't lose to web_search. Emits
        # the same tool card + citations a real call would.
        _auto = None if confirm_tool_calls else build_rag_autoinject(conversation, rag_scope)
        if _auto:
            for _ev in _auto["events"]:
                yield _ev
            conversation.extend(_auto["messages"])

        url = f"{self.base_url}/v1/chat/completions"
        _accumulated_completion_tokens = 0
        _accumulated_predicted_ms = 0.0
        _accumulated_predicted_n = 0
        # GGUF buffers reasoning; emit server-side timing before answer text.
        _reasoning_started_at: Optional[float] = None
        _reasoning_summary_emitted = False

        # Gate telling a genuine NAME[ARGS] rehearsal from inactive-name prose; built from the
        # ORIGINAL tools list so a spent one-shot still reads as a tool name. None = no gate.
        _enabled_names_gate = set(_gguf_active_tool_names(tools)) if tools else None
        # Detection must see the same names as the strip gate (ORIGINAL list, incl. a spent
        # one-shot), else its repeat is stripped but never drained and the turn ends blank.
        _detect_tools = list(tools or [])

        def _reasoning_summary_event(started_at: float) -> dict:
            return {
                "type": "reasoning_summary",
                "duration_ms": round((time.monotonic() - started_at) * 1000.0),
            }

        # Enabled-name gate for the markerless Gemma strip (disabled/example
        # names stay visible). Set per iteration; None = pre-loop name-agnostic.
        _enabled_tool_names = None

        def _strip_tool_markup(
            text: str,
            *,
            final: bool = False,
            force: bool = False,
        ) -> str:
            if not (auto_heal_tool_calls or force):
                return text
            # Delegate to the shared parser-side strip so the GGUF cleanup covers every family the
            # parser promotes (Llama <|python_tag|>, Mistral [TOOL_CALLS], bare rehearsal, function
            # XML, Gemma) and stays aligned with detection; tool_healing's strip omits the loop-only
            # forms (python_tag / Mistral name) and would leak them into display.
            return _shared_strip_tool_markup(
                text, final = final, enabled_tool_names = _enabled_names_gate
            )

        def _strip_tool_markup_streaming(text: str, *, force: bool = False) -> str:
            if not (auto_heal_tool_calls or force):
                return text

            def _seg(segment: str, is_last: bool) -> str:
                # Same scan order as the parser's _strip_segment (seg_final -> is_last): balanced
                # strips first (nested JSON removed whole; literal markup inside a value is that
                # call's data), then the guarded function-XML / GLM scans, then the regex arms
                # (DeepSeek / Kimi / closed forms). EOS-anchored tail arms run only on the last
                # segment (a bare ``foo[ARGS]`` before <think> is prose). Rehearsal + markerless
                # strips are name-gated on the ORIGINAL list (strip/detect aligned).
                seg = _strip_mistral_closed_calls(segment)
                seg = _strip_bracket_tag_calls(seg, enabled_tool_names = _enabled_names_gate)
                if is_last:
                    seg = _strip_gemma_wrapperless_calls(seg, _enabled_names_gate)
                seg = _strip_function_xml_calls(seg, final = is_last)
                seg = _strip_glm_calls(seg, final = is_last)
                pats = _PARSER_TOOL_ALL_PATS if is_last else _PARSER_TOOL_CLOSED_PATS
                for pat in pats:
                    seg = pat.sub("", seg)
                if is_last:
                    seg = apply_tool_strip_patterns(
                        seg, [_REHEARSAL_TAIL_STRIP_RE], enabled_tool_names = _enabled_names_gate
                    )
                return seg

            # Preserve think blocks verbatim (a rehearsed call inside one must not be deleted).
            return strip_outside_think(text, _seg)

        def _build_metadata_event(usage, timings, finish_reason):
            """Final usage+timings metadata event for the given pass, merging its
            usage/timings with the running cross-iteration accumulators. None when
            there is nothing to report."""
            _fu = _backfill_usage_from_timings(usage, timings) or {}
            _fp = _fu.get("prompt_tokens", 0)
            _tc = _fu.get("completion_tokens", 0) + _accumulated_completion_tokens
            if not (usage or timings or _accumulated_completion_tokens or finish_reason):
                return None
            _mt = dict(timings) if timings else {}
            if _accumulated_predicted_ms or _accumulated_predicted_n:
                _mt["predicted_ms"] = _mt.get("predicted_ms", 0) + _accumulated_predicted_ms
                _mt["predicted_n"] = _mt.get("predicted_n", 0) + _accumulated_predicted_n
                if _mt["predicted_ms"] > 0:
                    _mt["predicted_per_second"] = _mt["predicted_n"] / (
                        _mt["predicted_ms"] / 1000.0
                    )
            _usage = {
                "prompt_tokens": _fp,
                "completion_tokens": _tc,
                "total_tokens": _fp + _tc,
            }
            # Preserve KV-cache hit details (cached_tokens) so the tool path
            # reports them like the standard non-tool path does, not always 0.
            if _fu.get("prompt_tokens_details"):
                _usage["prompt_tokens_details"] = _fu["prompt_tokens_details"]
            return {
                "type": "metadata",
                "usage": _usage,
                "timings": _mt,
                "finish_reason": finish_reason,
            }

        def _flush_reasoning_and_buffer():
            """Close a live-streamed <think> block (or emit the buffered reasoning
            as one block if it never streamed), then append the held
            content_buffer to the cumulative display text."""
            nonlocal cumulative_display, in_thinking
            if in_thinking:
                cumulative_display += "</think>"
                in_thinking = False
            elif reasoning_accum:
                cumulative_display += "<think>" + reasoning_accum + "</think>"
            cumulative_display += content_buffer

        def _close_streamed_think() -> bool:
            """Close a live-streamed <think> before a tool call drains, so
            consumers without a reasoning extractor (Anthropic) get a balanced
            block. Returns True when the caller should yield the result."""
            nonlocal cumulative_display, in_thinking, _last_emitted
            if not in_thinking:
                return False
            cumulative_display += "</think>"
            in_thinking = False
            if len(cumulative_display) > len(_last_emitted) and not _suppress_visible_output:
                _last_emitted = cumulative_display
                return True
            return False

        def _looks_like_enabled_bare_json(text: str, enabled_tool_names: set) -> bool:
            """True when ``text`` opens with an ENABLED markerless bare-JSON call; an ordinary JSON answer returns False."""
            probe = strip_llama3_leading_sentinels(text.lstrip())
            if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
                return False
            return strip_leading_bare_json_call(probe, enabled_tool_names) != probe

        tool_controller = ToolLoopController(
            tools = tools,
            auto_heal_tool_calls = auto_heal_tool_calls,
        )

        def _tool_succeeded(tool_name: str) -> bool:
            key_prefix = f"{tool_name}:"
            return any(
                record.executed and not record.is_error and record.key.startswith(key_prefix)
                for record in tool_controller.history
            )

        _MAX_BUFFER_CHARS = 32
        # Hold a leading ``{`` well past the 32-char XML cap until it balances (mirrors safetensors).
        _MAX_BARE_JSON_BUFFER = 16384
        _append_budget_exhausted_nudge = True
        # RAG: cap knowledge-base searches per assistant turn. The controller is
        # tool-agnostic, so this gate stays in the loop.
        _kb_search_count = 0

        # ── Re-prompt on plan-without-action ─────────────────
        # Model describes intent without calling a tool: re-prompt once. A
        # direct answer ("4", "Hello!") won't match. Pattern shared with the
        # safetensors loop (tool_call_parser.INTENT_SIGNAL).
        _reprompt_count = 0
        # Gates ``max_tool_iterations`` on real tool turns (not the enlarged range) so reserved
        # re-prompt slots don't extend the budget. Mirrors the safetensors guard.
        _tool_iters_done = 0
        _forced_tool_call_pending = False

        # Reserve extra iterations for re-prompts so they don't consume the
        # caller's tool-call budget; only when tool iterations are allowed.
        _extra = _MAX_REPROMPTS if max_tool_iterations > 0 else 0
        for iteration in range(max_tool_iterations + _extra):
            if cancel_event is not None and cancel_event.is_set():
                return
            # Whether this turn ran a tool; a no-op-only turn stays False and doesn't consume budget.
            _turn_executed_real_tool = False

            active_tools = tool_controller.active_tools()
            if not active_tools:
                _append_budget_exhausted_nudge = False
                break
            # Gate the markerless bare-JSON form on enabled names so an ordinary JSON answer isn't misread as a call.
            _enabled_tool_names = {
                (tool.get("function") or {}).get("name")
                for tool in active_tools
                if (tool.get("function") or {}).get("name")
            }
            # Shared signal tuple so GGUF BUFFERING wakes on every format the parser knows (like safetensors).
            _tool_xml_signals = _SHARED_TOOL_XML_SIGNALS

            # Build payload -- stream: True so we detect tool signals
            # in the first 1-2 chunks without a non-streaming penalty.
            payload = {
                "messages": conversation,
                "stream": True,
                "stream_options": {"include_usage": True},
                "temperature": temperature,
                "top_p": top_p,
                "top_k": top_k if top_k >= 0 else 0,
                "min_p": min_p,
                "repeat_penalty": repetition_penalty,
                "presence_penalty": presence_penalty,
                "tools": active_tools,
                "tool_choice": "auto",
            }
            _reasoning_kw = self._request_reasoning_kwargs(
                enable_thinking, reasoning_effort, preserve_thinking
            )
            if _reasoning_kw is not None:
                payload["chat_template_kwargs"] = _reasoning_kw
            payload["max_tokens"] = (
                max_tokens
                if max_tokens is not None
                else (self._effective_context_length or _DEFAULT_MAX_TOKENS_FLOOR)
            )
            if stop:
                payload["stop"] = stop
            if seed is not None:
                payload["seed"] = seed

            try:
                # ── Speculative buffer state machine ──────────────────
                # BUFFERING: accumulate content, check for tool signals
                # STREAMING: no tool detected, yield tokens to caller
                # DRAINING:  tool signal found, silently consume rest
                _S_BUFFERING = 0
                _S_STREAMING = 1
                _S_DRAINING = 2

                detect_state = _S_BUFFERING
                content_buffer = ""  # Raw content held during BUFFERING
                content_accum = ""  # All content tokens (for tool parsing)
                reasoning_accum = ""
                # Time each reasoning pass so final answers can replace tool timing.
                _reasoning_started_at = None
                _reasoning_summary_emitted = False
                cumulative_display = ""  # Cumulative yielded text (with <think>)
                in_thinking = False
                has_content_tokens = False
                tool_calls_acc = {}  # Structured delta.tool_calls fragments
                has_structured_tc = False
                _iter_usage = None
                _iter_timings = None
                _iter_finish_reason = None
                _stream_done = False
                _last_emitted = ""
                # Provisional tool_start cards already shown, keyed by tool_call_id.
                provisional_started_tool_calls: dict[str, str] = {}
                resolved_provisional_tool_call_ids: set[str] = set()
                _suppress_visible_output = _forced_tool_call_pending

                with self._open_stream(url, payload, cancel_event) as (
                    response,
                    first_token_deadline,
                ):
                    raw_buf = ""
                    for raw_chunk in self._iter_text_cancellable(
                        response,
                        cancel_event,
                        first_token_deadline = first_token_deadline,
                    ):
                        raw_buf += raw_chunk
                        while "\n" in raw_buf:
                            line, raw_buf = raw_buf.split("\n", 1)
                            line = line.strip()

                            if not line:
                                continue
                            if line == "data: [DONE]":
                                # Flush thinking state for STREAMING
                                if detect_state == _S_STREAMING and in_thinking:
                                    if has_content_tokens:
                                        cumulative_display += "</think>"
                                        if not _suppress_visible_output:
                                            yield {
                                                "type": "content",
                                                "text": _strip_tool_markup(
                                                    cumulative_display,
                                                    final = True,
                                                ),
                                            }
                                    else:
                                        cumulative_display = reasoning_accum
                                        if not _suppress_visible_output:
                                            yield {
                                                "type": "content",
                                                "text": cumulative_display,
                                            }
                                _stream_done = True
                                break  # exit inner while
                            if not line.startswith("data: "):
                                continue

                            try:
                                chunk_data = json.loads(line[6:])
                                _ct = chunk_data.get("timings")
                                if _ct:
                                    _iter_timings = _ct
                                _cu = chunk_data.get("usage")
                                if _cu:
                                    _iter_usage = _cu

                                choices = chunk_data.get("choices", [])
                                if not choices:
                                    continue

                                delta = choices[0].get("delta", {})
                                _fr = choices[0].get("finish_reason")
                                if _fr:
                                    _iter_finish_reason = _fr

                                # ── Structured tool_calls ──
                                tc_deltas = delta.get("tool_calls")
                                if tc_deltas:
                                    # Preserve any visible preface before draining
                                    # the structured tool call.
                                    has_structured_tc = True
                                    detect_state = _S_DRAINING
                                    # Close the reasoning prefix before the tool card
                                    # (mirrors the is_match path).
                                    if _close_streamed_think():
                                        yield {"type": "content", "text": cumulative_display}
                                    for tc_d in tc_deltas:
                                        idx = tc_d.get("index", 0)
                                        if idx not in tool_calls_acc:
                                            tool_calls_acc[idx] = {
                                                "id": tc_d.get("id", f"call_{idx}"),
                                                "type": "function",
                                                "function": {
                                                    "name": "",
                                                    "arguments": "",
                                                },
                                            }
                                        elif tc_d.get("id"):
                                            # Update ID if a real one
                                            # arrives on a later delta.
                                            tool_calls_acc[idx]["id"] = tc_d["id"]
                                        func = tc_d.get("function", {})
                                        if func.get("name"):
                                            tool_calls_acc[idx]["function"]["name"] += func["name"]
                                        if func.get("arguments"):
                                            tool_calls_acc[idx]["function"]["arguments"] += func[
                                                "arguments"
                                            ]
                                        current_name = tool_calls_acc[idx]["function"].get(
                                            "name", ""
                                        )
                                        fallback_id = f"call_{idx}"
                                        current_id = tool_calls_acc[idx].get("id", fallback_id)
                                        already_started = (
                                            current_id in provisional_started_tool_calls
                                        )
                                        # Empty/synthetic ids cannot reconcile with real starts.
                                        has_real_id = bool(current_id) and current_id != fallback_id
                                        # Show one early card per eligible streamed tool call.
                                        _is_completed_one_shot = (
                                            current_name == "render_html"
                                            and _tool_succeeded("render_html")
                                        )
                                        # render_html is one-shot.
                                        _one_shot_already_provisional = (
                                            current_name == "render_html"
                                            and "render_html"
                                            in provisional_started_tool_calls.values()
                                        )
                                        # Later parallel cards only reconcile when parallel use is enabled.
                                        _confirm_gated = (
                                            confirm_tool_calls and not bypass_permissions
                                        )
                                        # Keep small-argument tools on the normal path.
                                        _args_len = len(
                                            tool_calls_acc[idx]["function"].get("arguments", "")
                                        )
                                        _payload_is_large = (
                                            current_name == "render_html"
                                            or _args_len >= _PROVISIONAL_ARGS_MIN_CHARS
                                        )
                                        if (
                                            current_name
                                            and (idx == 0 or not disable_parallel_tool_use)
                                            and has_real_id
                                            and not already_started
                                            and not _is_completed_one_shot
                                            and not _one_shot_already_provisional
                                            and not _confirm_gated
                                            and _payload_is_large
                                            and any(
                                                (tool.get("function") or {}).get("name")
                                                == current_name
                                                for tool in active_tools
                                            )
                                        ):
                                            provisional_started_tool_calls[current_id] = (
                                                current_name
                                            )
                                            yield {
                                                "type": "tool_start",
                                                "tool_name": current_name,
                                                "tool_call_id": current_id,
                                                "arguments": {},
                                                "provenance": tool_event_provenance(
                                                    provisional = True,
                                                ),
                                            }
                                    continue

                                # ── Reasoning tokens ──
                                # Stream live except while DRAINING: reasoning is
                                # orthogonal to tool detection (content_buffer
                                # only), and the route resets prev_text on
                                # tool_start, so the <think> block stays a
                                # monotonic prefix like the no-tool path.
                                reasoning = delta.get("reasoning_content", "")
                                if reasoning:
                                    if _reasoning_started_at is None:
                                        _reasoning_started_at = time.monotonic()
                                    reasoning_accum += reasoning
                                    if detect_state != _S_DRAINING:
                                        if not in_thinking:
                                            cumulative_display += "<think>"
                                            in_thinking = True
                                        cumulative_display += reasoning
                                        if not _suppress_visible_output:
                                            yield {
                                                "type": "content",
                                                "text": cumulative_display,
                                            }

                                # ── Content tokens ──
                                token = delta.get("content", "")
                                if token:
                                    # First answer token ends reasoning.
                                    if (
                                        _reasoning_started_at is not None
                                        and not _reasoning_summary_emitted
                                    ):
                                        _reasoning_summary_emitted = True
                                        yield _reasoning_summary_event(_reasoning_started_at)
                                    has_content_tokens = True
                                    content_accum += token

                                    if detect_state == _S_DRAINING:
                                        pass  # accumulate silently

                                    elif detect_state == _S_STREAMING:
                                        if in_thinking:
                                            cumulative_display += "</think>"
                                            in_thinking = False
                                        cumulative_display += token
                                        cleaned = _strip_tool_markup_streaming(cumulative_display)
                                        # Hold a trailing bare active-tool-name (split rehearsal)
                                        # until [ARGS] arrives; released by later prose or stream end.
                                        _hold = _held_rehearsal_tail_len(cleaned, _detect_tools)
                                        _emit = (
                                            cleaned[: len(cleaned) - _hold] if _hold else cleaned
                                        )
                                        if len(_emit) > len(_last_emitted):
                                            _last_emitted = _emit
                                            if not _suppress_visible_output:
                                                yield {
                                                    "type": "content",
                                                    "text": _emit,
                                                }

                                    elif detect_state == _S_BUFFERING:
                                        content_buffer += token
                                        stripped_buf = content_buffer.lstrip()
                                        if not stripped_buf:
                                            continue

                                        # Bracket tags arrive mid-buffer, so substring-check too;
                                        # ``[ARGS]`` counts only as a regex-matched NAME[ARGS].
                                        is_prefix = False
                                        is_match = False
                                        for sig in _tool_xml_signals:
                                            if stripped_buf.startswith(sig):
                                                is_match = True
                                                break
                                            if sig.startswith(stripped_buf):
                                                is_prefix = True
                                                break
                                            if sig == "[ARGS]":
                                                # Active NAME[ARGS] only; inactive-name prose
                                                # is gated out, not drained/parsed.
                                                if (
                                                    _gguf_rehearsal_signal_pos(
                                                        stripped_buf, _detect_tools
                                                    )
                                                    >= 0
                                                ):
                                                    is_match = True
                                                    break
                                            elif sig.startswith("[") and sig in stripped_buf:
                                                is_match = True
                                                break

                                        # Split rehearsal: hold the bare name until
                                        # its [ARGS] arrives and matches above.
                                        is_rehearsal_prefix = False
                                        if (
                                            not is_match
                                            and not is_prefix
                                            and _is_rehearsal_prefix(stripped_buf, _detect_tools)
                                        ):
                                            is_prefix = True
                                            is_rehearsal_prefix = True

                                        # Signal-less call shapes (mirror the safetensors
                                        # loop): Llama-3.2 bare {"name":..} and Gemma
                                        # call:NAME{...} would otherwise stream raw.
                                        _hold_buffer = False
                                        # Whole buffer is the call (no visible prefix) -- drain silently.
                                        _drain_silently = False
                                        if not is_match and not is_prefix:
                                            _bare = strip_llama3_leading_sentinels(stripped_buf)
                                            if _bare.startswith("{"):
                                                if _balanced_brace_end(_bare, 0) is None:
                                                    if len(stripped_buf) < _MAX_BARE_JSON_BUFFER:
                                                        _hold_buffer = True
                                                    elif _looks_like_enabled_bare_json(
                                                        _bare, _enabled_tool_names
                                                    ):
                                                        # Oversized still-open enabled call: drain
                                                        # rather than leak; a giant ordinary JSON
                                                        # answer still streams.
                                                        _drain_silently = True
                                                elif self._parse_tool_calls_from_text(
                                                    content_buffer,
                                                    allow_incomplete = auto_heal_tool_calls,
                                                    enabled_tool_names = _enabled_tool_names,
                                                ):
                                                    _drain_silently = True
                                            elif (
                                                "call:".startswith(stripped_buf)
                                                or _GEMMA_BARE_TC_PREFIX_RE.match(stripped_buf)
                                                is not None
                                                or _GEMMA_BARE_TC_RE.match(stripped_buf) is not None
                                            ):
                                                # Whitespace-tolerant like the parser.
                                                if _GEMMA_BARE_TC_RE.match(stripped_buf):
                                                    _drain_silently = True
                                                elif len(stripped_buf) < _MAX_BUFFER_CHARS:
                                                    _hold_buffer = True

                                        if _drain_silently:
                                            # The buffered content IS the call; drain it
                                            # without yielding. A live <think> prefix is
                                            # separate from it -- close that.
                                            detect_state = _S_DRAINING
                                            if _close_streamed_think():
                                                yield {
                                                    "type": "content",
                                                    "text": cumulative_display,
                                                }
                                        elif is_match:
                                            # Tool signal -- flush any visible
                                            # prefix before DRAINING so the
                                            # route sends it before tool_start.
                                            # Use the final strip (all families incl. Llama
                                            # <|python_tag|> / Mistral name): the buffer holds
                                            # the whole call, so a streaming closed-only strip
                                            # would leak its open-ended markup as display text.
                                            _flush_reasoning_and_buffer()
                                            cleaned = _strip_tool_markup(
                                                cumulative_display,
                                                final = True,
                                                force = True,
                                            )
                                            if len(cleaned) > len(_last_emitted):
                                                _last_emitted = cleaned
                                                if not _suppress_visible_output:
                                                    yield {
                                                        "type": "content",
                                                        "text": cleaned,
                                                    }
                                            detect_state = _S_DRAINING
                                        elif _hold_buffer or (
                                            is_prefix
                                            and (
                                                is_rehearsal_prefix
                                                or len(stripped_buf) < _MAX_BUFFER_CHARS
                                            )
                                        ):
                                            # A rehearsal prefix is self-bounded; the buffer
                                            # cap must not cut long MCP names short.
                                            pass  # keep buffering
                                        else:
                                            # Not a tool -- flush buffer
                                            detect_state = _S_STREAMING
                                            # Flush reasoning accumulated
                                            # during BUFFERING.
                                            _flush_reasoning_and_buffer()
                                            cleaned = _strip_tool_markup(
                                                cumulative_display,
                                            )
                                            # Same trailing-name hold as STREAMING for this
                                            # first flush out of BUFFERING.
                                            _hold = _held_rehearsal_tail_len(cleaned, _detect_tools)
                                            _emit = (
                                                cleaned[: len(cleaned) - _hold]
                                                if _hold
                                                else cleaned
                                            )
                                            if len(_emit) > len(_last_emitted):
                                                _last_emitted = _emit
                                                if not _suppress_visible_output:
                                                    yield {
                                                        "type": "content",
                                                        "text": _emit,
                                                    }

                            except json.JSONDecodeError:
                                logger.debug(f"Skipping malformed SSE line: {line[:100]}")
                        if _stream_done:
                            break  # exit outer for

                # ── Resolve BUFFERING at stream end ──
                if detect_state == _S_BUFFERING:
                    stripped_buf = content_buffer.lstrip()
                    # A held bare-JSON fragment has no XML signal; route it to DRAINING (the signal-only
                    # gate below would flush the raw JSON to the user).
                    _bare_eos = strip_llama3_leading_sentinels(stripped_buf)
                    # Gate on enabled names so an ordinary JSON answer isn't routed to DRAINING and dropped.
                    _is_bare_tc = bool(active_tools) and _looks_like_enabled_bare_json(
                        _bare_eos, _enabled_tool_names
                    )
                    if stripped_buf and _gguf_has_genuine_tool_signal(
                        stripped_buf, _tool_xml_signals, _detect_tools
                    ):
                        detect_state = _S_DRAINING
                    elif _is_bare_tc:
                        detect_state = _S_DRAINING
                    elif content_accum or reasoning_accum:
                        detect_state = _S_STREAMING
                        if content_buffer:
                            # Flush reasoning first.
                            _flush_reasoning_and_buffer()
                            if not _suppress_visible_output:
                                yield {
                                    "type": "content",
                                    "text": _strip_tool_markup(
                                        cumulative_display,
                                        final = True,
                                    ),
                                }
                        elif reasoning_accum and not has_content_tokens:
                            # Reasoning-only reply: show it as the main response,
                            # not a thinking block (mirrors the no-tool path; the
                            # route's extractor closes the streamed <think>).
                            if _reasoning_started_at is not None and not _reasoning_summary_emitted:
                                _reasoning_summary_emitted = True
                                yield _reasoning_summary_event(_reasoning_started_at)
                            cumulative_display = reasoning_accum
                            if not _suppress_visible_output:
                                yield {
                                    "type": "content",
                                    "text": cumulative_display,
                                }
                    else:
                        # Held buffer was no tool signal and no enabled bare-JSON call: a leading ``{`` is an
                        # ordinary JSON answer and must be shown; any other partial-markup prefix is dropped.
                        _held = strip_llama3_leading_sentinels(content_buffer.lstrip())
                        if _held.startswith("{") and not _suppress_visible_output:
                            yield {"type": "content", "text": _held}
                        return

                # ── STREAMING path: no tool call ──
                if detect_state == _S_STREAMING:
                    # Safety net: re-parse the full content for tool calls. The
                    # route layer resets prev_text on tool_start, so post-tool
                    # synthesis streams correctly even if content was emitted
                    # before the tool XML.
                    # Unconditional (not gated on _tool_xml_signals): bare-JSON and Gemma wrapper-less
                    # calls carry no XML signal, so a signal gate would let them slip past.
                    _safety_tc = self._parse_tool_calls_from_text(
                        content_accum,
                        allow_incomplete = auto_heal_tool_calls,
                        enabled_tool_names = _enabled_tool_names,
                    )
                    if not _safety_tc:
                        # ── Re-prompt on plan-without-action ──
                        # If the model described its intent (forward-looking
                        # language) without calling a tool, nudge it to act.
                        # Fires at most once per request, only on short
                        # responses with intent signals -- "4" or "Hello!"
                        # won't trigger it. Use content if available, else
                        # fall back to reasoning text (reasoning-only stalls).
                        _stripped = content_accum.strip()
                        if not _stripped:
                            _stripped = reasoning_accum.strip()
                        _render_html_already_done_intent = _tool_succeeded(
                            "render_html"
                        ) and re.search(
                            r"(?i)\brender[_\s-]?html\b",
                            _stripped,
                        )
                        # None keeps the default-on re-prompt; False disables it.
                        if (
                            auto_heal_tool_calls
                            and (nudge_tool_calls is None or nudge_tool_calls)
                            and active_tools
                            and not _render_html_already_done_intent
                            and _reprompt_count < _MAX_REPROMPTS
                            and _is_short_intent_without_action(_stripped)
                        ):
                            _reprompt_count += 1
                            logger.info(
                                f"Re-prompt {_reprompt_count}/{_MAX_REPROMPTS}: "
                                f"model responded without calling tools "
                                f"({len(_stripped)} chars)"
                            )
                            conversation.append(
                                {
                                    "role": "assistant",
                                    "content": _stripped,
                                }
                            )
                            available_tool_names = [
                                (tool.get("function") or {}).get("name")
                                for tool in active_tools
                                if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
                            ]
                            available_tool_names = [name for name in available_tool_names if name]
                            tool_hint = " or ".join(available_tool_names) or "an available tool"
                            _forced_tool_call_pending = True
                            conversation.append(
                                {
                                    "role": "user",
                                    "content": _reprompt_to_act_message(tool_hint),
                                }
                            )
                            # Accumulate tokens and timing from this iteration.
                            _fu_r = _backfill_usage_from_timings(_iter_usage, _iter_timings) or {}
                            _accumulated_completion_tokens += _fu_r.get("completion_tokens", 0)
                            _it_r = _iter_timings or {}
                            _accumulated_predicted_ms += _it_r.get("predicted_ms", 0)
                            _accumulated_predicted_n += _it_r.get("predicted_n", 0)
                            yield {"type": "status", "text": ""}
                            continue

                        if _forced_tool_call_pending:
                            _forced_tool_call_pending = False
                            if not _should_suppress_forced_no_tool_output(_stripped):
                                if cumulative_display:
                                    forced_visible_text = _strip_tool_markup(
                                        cumulative_display,
                                        final = True,
                                    )
                                elif content_accum:
                                    forced_visible_text = _strip_tool_markup(
                                        content_accum,
                                        final = True,
                                    )
                                else:
                                    forced_visible_text = reasoning_accum
                                if forced_visible_text:
                                    yield {
                                        "type": "content",
                                        "text": forced_visible_text,
                                    }
                        elif not _suppress_visible_output:
                            # Turn ended as a plain answer (no [ARGS] followed): the held
                            # rehearsal tail is real prose, release it.
                            _final_clean = _strip_tool_markup_streaming(cumulative_display)
                            if len(_final_clean) > len(_last_emitted):
                                yield {"type": "content", "text": _final_clean}

                        # Content was already streamed.  Yield metadata.
                        yield {"type": "status", "text": ""}
                        _meta = _build_metadata_event(
                            _iter_usage, _iter_timings, _iter_finish_reason
                        )
                        if _meta is not None:
                            yield _meta
                        return

                    # Safety net caught tool XML -- treat as tool call.
                    tool_calls = _safety_tc
                    content_text = _strip_tool_markup(
                        content_accum,
                        final = True,
                        force = True,
                    )
                    logger.info(
                        f"Safety net: parsed {len(tool_calls)} tool call(s) from streamed content"
                    )
                else:
                    # ── DRAINING path: assemble tool_calls ──
                    tool_calls = None
                    content_text = content_accum
                    if has_structured_tc:
                        # Drop incomplete fragments (e.g. from max_tokens
                        # truncation or disconnect).
                        tool_calls = [
                            tool_calls_acc[i]
                            for i in sorted(tool_calls_acc)
                            if (tool_calls_acc[i].get("function", {}).get("name", "").strip())
                        ] or None
                    if not tool_calls:
                        # Unconditional re-parse: we only reach DRAINING when the buffer looked like a
                        # call, and bare-JSON / Gemma wrapper-less calls carry no XML signal to gate on.
                        tool_calls = self._parse_tool_calls_from_text(
                            content_accum,
                            allow_incomplete = auto_heal_tool_calls,
                            enabled_tool_names = _enabled_tool_names,
                        )
                    if tool_calls and not has_structured_tc:
                        content_text = _strip_tool_markup(
                            content_text,
                            final = True,
                            force = True,
                        )
                        # ``_strip_tool_markup`` only knows XML; also drop a leading bare-JSON call so the
                        # executed call isn't replayed as text or next-turn history.
                        content_text = strip_leading_bare_json_call(
                            content_text, _enabled_tool_names
                        )
                    if tool_calls:
                        logger.info(
                            f"Parsed {len(tool_calls)} tool call(s) from "
                            f"{'structured delta' if has_structured_tc else 'content text'}"
                        )
                    if not tool_calls:
                        # DRAINING but no tool calls (false positive). Merge
                        # accumulated metrics from prior tool iterations so
                        # they aren't silently dropped.
                        yield {"type": "status", "text": ""}
                        if content_accum:
                            # Strip leaked tool-call XML before yielding.
                            content_accum = _strip_tool_markup(content_accum, final = True)
                        # A truncated bare-JSON call has no XML markup to strip and didn't parse. With
                        # Auto-Heal on, drop a leading ENABLED-tool fragment (ordinary JSON answers untouched);
                        # off keeps it visible per the strict contract.
                        if content_accum and active_tools and auto_heal_tool_calls:
                            content_accum = strip_leading_bare_json_call(
                                content_accum, _enabled_tool_names
                            )
                        if content_accum:
                            yield {"type": "content", "text": content_accum}
                        _meta = _build_metadata_event(
                            _iter_usage, _iter_timings, _iter_finish_reason
                        )
                        if _meta is not None:
                            yield _meta
                        return

                # ── Execute tool calls ──
                _accumulated_completion_tokens += (
                    _backfill_usage_from_timings(_iter_usage, _iter_timings) or {}
                ).get("completion_tokens", 0)
                _it = _iter_timings or {}
                _accumulated_predicted_ms += _it.get("predicted_ms", 0)
                _accumulated_predicted_n += _it.get("predicted_n", 0)

                # Collapse exact-duplicate calls and cap the count for the TEXTUAL
                # fallback (mirrors the safetensors loop; see _MAX_TOOL_CALLS_PER_TURN).
                if tool_calls and not has_structured_tc and len(tool_calls) > 1:
                    _seen_keys: set = set()
                    _deduped: list = []
                    for _tc in tool_calls:
                        _fn = _tc.get("function", {}) or {}
                        _key = (_fn.get("name", ""), str(_fn.get("arguments", "")))
                        if _key in _seen_keys:
                            continue
                        _seen_keys.add(_key)
                        _deduped.append(_tc)
                        if len(_deduped) >= _MAX_TOOL_CALLS_PER_TURN:
                            break
                    if len(_deduped) != len(tool_calls):
                        logger.info(
                            "GGUF textual fallback: collapsed %d repeated tool call(s) "
                            "in one turn to %d",
                            len(tool_calls),
                            len(_deduped),
                        )
                    tool_calls = _deduped

                # disable_parallel_tool_use: execute only the first tool call
                # this turn. Truncate before building assistant_msg so the
                # conversation stays consistent and extra calls are never executed.
                if disable_parallel_tool_use and tool_calls and len(tool_calls) > 1:
                    tool_calls = tool_calls[:1]

                assistant_msg: dict = {"role": "assistant", "content": content_text}
                assistant_appended = False

                for tc in tool_calls or []:
                    func = tc.get("function", {})
                    tool_name = func.get("name", "")
                    provisional_match = tc.get("id") in provisional_started_tool_calls
                    decision = tool_controller.prepare_call(
                        tc,
                        forced = _forced_tool_call_pending,
                        provisional = provisional_match,
                    )

                    if not decision.should_execute:
                        if content_text and not assistant_appended:
                            conversation.append(assistant_msg)
                            assistant_appended = True
                        if provisional_match:
                            # A provisional tool card is already on screen for this
                            # id; close it so it never dangles when the controller
                            # turns the call into an internal no-op (duplicate /
                            # disabled / render_html_repeat).
                            resolved_provisional_tool_call_ids.add(decision.tool_call_id)
                            yield {
                                "type": "tool_end",
                                "tool_name": decision.tool_name,
                                "tool_call_id": decision.tool_call_id,
                                "result": "",
                                "provenance": decision.provenance,
                            }
                        completion = tool_controller.record_noop(decision)
                        conversation.append(completion.model_message())
                        if _forced_tool_call_pending:
                            _forced_tool_call_pending = False
                        logger.info(
                            "Suppressed local GGUF tool call as internal no-op: "
                            f"action={decision.action} tool={decision.tool_name}"
                        )
                        break

                    if not assistant_appended:
                        assistant_msg["tool_calls"] = [decision.as_assistant_tool_call()]
                        conversation.append(assistant_msg)
                        assistant_appended = True
                    else:
                        assistant_msg.setdefault("tool_calls", []).append(
                            decision.as_assistant_tool_call()
                        )

                    # Bypass wins over the confirm gate at the loop level too,
                    # so a direct internal caller with both flags never prompts.
                    needs_confirm = bool(confirm_tool_calls) and not bypass_permissions
                    approval_id = new_approval_id() if needs_confirm else ""
                    decision_slot = (
                        begin_tool_decision(session_id, approval_id) if needs_confirm else None
                    )
                    start_event = decision.tool_start_event()
                    start_event["approval_id"] = approval_id
                    start_event["awaiting_confirmation"] = needs_confirm

                    try:
                        yield {"type": "status", "text": decision.status_text}
                        yield start_event

                        if (
                            decision_slot is not None
                            and wait_tool_decision(
                                decision_slot,
                                approval_id,
                                cancel_event = cancel_event,
                            )
                            == "deny"
                        ):
                            decision_slot = None
                            resolved_provisional_tool_call_ids.add(decision.tool_call_id)
                            yield {
                                "type": "tool_end",
                                "tool_name": decision.tool_name,
                                "tool_call_id": decision.tool_call_id,
                                "result": TOOL_REJECTED_MESSAGE,
                                "provenance": decision.provenance,
                            }
                            denied_message = {
                                "role": "tool",
                                "name": decision.tool_name,
                                "content": TOOL_REJECTED_MESSAGE,
                            }
                            if decision.tool_call_id:
                                denied_message["tool_call_id"] = decision.tool_call_id
                            conversation.append(denied_message)
                            if _forced_tool_call_pending:
                                _forced_tool_call_pending = False
                            continue
                        decision_slot = None
                    finally:
                        if decision_slot is not None:
                            abort_tool_decision(decision_slot, approval_id)

                    _effective_timeout = None if tool_call_timeout >= 9999 else tool_call_timeout
                    # RAG: cap paraphrased KB re-searches that slip past the dup guard.
                    if (
                        decision.tool_name == "search_knowledge_base"
                        and _kb_search_count >= RAG_MAX_SEARCHES_PER_TURN
                    ):
                        result = RAG_SEARCH_CAP_NUDGE
                    else:
                        result = execute_tool(
                            decision.tool_name,
                            decision.arguments,
                            cancel_event = cancel_event,
                            timeout = _effective_timeout,
                            session_id = session_id,
                            rag_scope = rag_scope,
                            disable_sandbox = bypass_permissions,
                        )
                        if decision.tool_name == "search_knowledge_base":
                            _kb_search_count += 1
                    completion = tool_controller.record_result(decision, result)
                    resolved_provisional_tool_call_ids.add(decision.tool_call_id)
                    # A tool ran this turn, so it counts against the caller's budget.
                    _turn_executed_real_tool = True
                    yield completion.tool_end_event()
                    conversation.append(completion.tool_message())

                    if _forced_tool_call_pending:
                        _forced_tool_call_pending = False

                # Close provisional cards not resolved by execution/no-op handling.
                for _pid, _pname in provisional_started_tool_calls.items():
                    if _pid not in resolved_provisional_tool_call_ids:
                        resolved_provisional_tool_call_ids.add(_pid)
                        yield {
                            "type": "tool_end",
                            "tool_name": _pname,
                            "tool_call_id": _pid,
                            "result": "",
                            "provenance": tool_event_provenance(provisional = True),
                        }

                # Clear tool status badge before next generation/final pass.
                yield {"type": "status", "text": ""}
                if tool_controller.force_final_answer or not tool_controller.active_tools():
                    _append_budget_exhausted_nudge = False
                    break
                # Count only real tool turns against the cap so reserved re-prompt slots can't become
                # extra tool rounds; a no-op correction turn doesn't consume budget (GGUF parity).
                if _turn_executed_real_tool:
                    _tool_iters_done += 1
                    if _tool_iters_done >= max_tool_iterations:
                        break
                continue

            except httpx.ConnectError:
                # Mark unresolved provisional cards as failed before raising.
                for _pid, _pname in provisional_started_tool_calls.items():
                    if _pid not in resolved_provisional_tool_call_ids:
                        resolved_provisional_tool_call_ids.add(_pid)
                        yield {
                            "type": "tool_end",
                            "tool_name": _pname,
                            "tool_call_id": _pid,
                            "result": "Error: lost connection to llama-server before the tool call completed.",
                            "provenance": tool_event_provenance(provisional = True),
                        }
                raise RuntimeError("Lost connection to llama-server")
            except Exception as e:
                if cancel_event is not None and cancel_event.is_set():
                    return
                # Same cleanup for other mid-iteration failures.
                for _pid, _pname in provisional_started_tool_calls.items():
                    if _pid not in resolved_provisional_tool_call_ids:
                        resolved_provisional_tool_call_ids.add(_pid)
                        yield {
                            "type": "tool_end",
                            "tool_name": _pname,
                            "tool_call_id": _pid,
                            "result": "Error: the tool call was interrupted before it completed.",
                            "provenance": tool_event_provenance(provisional = True),
                        }
                raise

        # ── Tool iteration cap reached -- synthesize final answer ──
        # The model used all iterations without a final text response. Nudge
        # the final streaming pass to produce a useful answer instead of
        # continuing to request tools.
        if max_tool_iterations > 0 and _append_budget_exhausted_nudge:
            conversation.append(
                {
                    "role": "user",
                    "content": (
                        "You have used all available tool calls. Based on "
                        "everything you have found so far, provide your final "
                        "answer now. Do not call any more tools."
                    ),
                }
            )

        # Clear status.
        yield {"type": "status", "text": ""}

        # Final streaming pass with the full conversation context.
        stream_payload = {
            "messages": conversation,
            "stream": True,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k if top_k >= 0 else 0,
            "min_p": min_p,
            "repeat_penalty": repetition_penalty,
            "presence_penalty": presence_penalty,
        }
        _reasoning_kw = self._request_reasoning_kwargs(
            enable_thinking, reasoning_effort, preserve_thinking
        )
        if _reasoning_kw is not None:
            stream_payload["chat_template_kwargs"] = _reasoning_kw
        stream_payload["max_tokens"] = (
            max_tokens
            if max_tokens is not None
            else (self._effective_context_length or _DEFAULT_MAX_TOKENS_FLOOR)
        )
        if stop:
            stream_payload["stop"] = stop
        if seed is not None:
            stream_payload["seed"] = seed
        stream_payload["stream_options"] = {"include_usage": True}

        cumulative = ""
        _last_emitted = ""
        in_thinking = False
        has_content_tokens = False
        reasoning_text = ""
        _final_reasoning_started_at: Optional[float] = None
        _final_reasoning_summary_emitted = False
        _metadata_usage = None
        _metadata_timings = None
        _metadata_finish_reason = None
        _stream_done = False

        try:
            with self._open_stream(url, stream_payload, cancel_event) as (
                response,
                first_token_deadline,
            ):
                buffer = ""
                for raw_chunk in self._iter_text_cancellable(
                    response,
                    cancel_event,
                    first_token_deadline = first_token_deadline,
                ):
                    buffer += raw_chunk
                    while "\n" in buffer:
                        line, buffer = buffer.split("\n", 1)
                        line = line.strip()

                        if not line:
                            continue
                        if line == "data: [DONE]":
                            if in_thinking:
                                if (
                                    _final_reasoning_started_at is not None
                                    and not _final_reasoning_summary_emitted
                                ):
                                    _final_reasoning_summary_emitted = True
                                    yield _reasoning_summary_event(_final_reasoning_started_at)
                                if has_content_tokens:
                                    cumulative += "</think>"
                                    yield {
                                        "type": "content",
                                        "text": _strip_tool_markup(cumulative, final = True),
                                    }
                                else:
                                    cumulative = reasoning_text
                                    yield {"type": "content", "text": cumulative}
                            _stream_done = True
                            break  # exit inner while
                        if not line.startswith("data: "):
                            continue

                        try:
                            chunk_data = json.loads(line[6:])
                            # Capture server timings/usage from final chunks.
                            _chunk_timings = chunk_data.get("timings")
                            if _chunk_timings:
                                _metadata_timings = _chunk_timings
                            _chunk_usage = chunk_data.get("usage")
                            if _chunk_usage:
                                _metadata_usage = _chunk_usage
                            choices = chunk_data.get("choices", [])
                            if choices:
                                delta = choices[0].get("delta", {})
                                _fr = choices[0].get("finish_reason")
                                if _fr:
                                    _metadata_finish_reason = _fr

                                reasoning = delta.get("reasoning_content", "")
                                if reasoning:
                                    if _final_reasoning_started_at is None:
                                        _final_reasoning_started_at = time.monotonic()
                                    reasoning_text += reasoning
                                    if not in_thinking:
                                        cumulative += "<think>"
                                        in_thinking = True
                                    cumulative += reasoning
                                    yield {"type": "content", "text": cumulative}

                                token = delta.get("content", "")
                                if token:
                                    if (
                                        _final_reasoning_started_at is not None
                                        and not _final_reasoning_summary_emitted
                                    ):
                                        _final_reasoning_summary_emitted = True
                                        yield _reasoning_summary_event(_final_reasoning_started_at)
                                    has_content_tokens = True
                                    if in_thinking:
                                        cumulative += "</think>"
                                        in_thinking = False
                                    cumulative += token
                                    cleaned = _strip_tool_markup(cumulative)
                                    # Emit only when cleaned text grows (monotonic).
                                    if len(cleaned) > len(_last_emitted):
                                        _last_emitted = cleaned
                                        yield {"type": "content", "text": cleaned}
                        except json.JSONDecodeError:
                            logger.debug(f"Skipping malformed SSE line: {line[:100]}")
                    if _stream_done:
                        break  # exit outer for
                _meta = _build_metadata_event(
                    _metadata_usage, _metadata_timings, _metadata_finish_reason
                )
                if _meta is not None:
                    yield _meta

        except httpx.ConnectError:
            raise RuntimeError("Lost connection to llama-server")
        except Exception as e:
            if cancel_event is not None and cancel_event.is_set():
                return
            raise

    # ── Prompt token counting ──────────────────────────────────

    def count_chat_tokens(
        self,
        messages,
        system = None,
        tools = None,
        strict: bool = False,
    ) -> int:
        """Count prompt tokens for a chat request via llama-server.

        Non-strict callers keep the historical best-effort behavior and receive
        0 when a count cannot be determined. Strict callers (public count_tokens
        endpoints) get an exception instead of a successful-looking zero when
        tokenizer/template calls fail or a multimodal prompt would fall back to a
        text-only approximation.
        """
        if not self.is_loaded:
            if strict:
                raise RuntimeError("llama-server is not loaded")
            return 0

        def _has_non_text_content(content) -> bool:
            if isinstance(content, list):
                for block in content:
                    if isinstance(block, str):
                        continue
                    if not isinstance(block, dict):
                        return True
                    if isinstance(block.get("text"), str):
                        continue
                    return True
            return False

        def _has_non_text_prompt_parts() -> bool:
            if _has_non_text_content(system):
                return True
            for msg in messages or []:
                if isinstance(msg, dict) and _has_non_text_content(msg.get("content", "")):
                    return True
            return False

        def _block_text(content) -> str:
            if isinstance(content, str):
                return content
            if isinstance(content, list):
                parts = []
                for block in content:
                    if isinstance(block, dict):
                        if isinstance(block.get("text"), str):
                            parts.append(block["text"])
                    elif isinstance(block, str):
                        parts.append(block)
                return "".join(parts)
            return ""

        # Normalize system into a leading message / plain text.
        system_text = ""
        if isinstance(system, str):
            system_text = system
        elif isinstance(system, list):
            system_text = _block_text(system)

        try:
            with httpx.Client(timeout = 10, headers = self._auth_headers, trust_env = False) as client:

                def _tokenize(text: str) -> int:
                    r = client.post(
                        f"{self.base_url}/tokenize",
                        json = {"content": text, "add_special": True},
                    )
                    if r.status_code != 200:
                        if strict:
                            raise RuntimeError("llama-server tokenizer failed")
                        return 0
                    tokens = r.json().get("tokens", [])
                    if not isinstance(tokens, list):
                        if strict:
                            raise RuntimeError("llama-server tokenizer returned invalid tokens")
                        return 0
                    return len(tokens)

                # 1. Try /apply-template to render the real chat prompt.
                template_messages = list(messages) if messages else []
                if system_text:
                    template_messages = [
                        {"role": "system", "content": system_text}
                    ] + template_messages
                apply_template_failed = False
                try:
                    # llama-server's /apply-template renders tool declarations
                    # into the prompt when ``tools`` is supplied, so pass them
                    # through, otherwise tool-schema tokens go uncounted.
                    template_body = {"messages": template_messages}
                    if tools:
                        template_body["tools"] = tools
                    resp = client.post(
                        f"{self.base_url}/apply-template",
                        json = template_body,
                    )
                    if resp.status_code == 200:
                        prompt = resp.json().get("prompt", "")
                        if isinstance(prompt, str):
                            return _tokenize(prompt)
                    apply_template_failed = True
                except Exception:
                    apply_template_failed = True

                if strict and apply_template_failed and _has_non_text_prompt_parts():
                    raise RuntimeError(
                        "cannot fall back to text-only token counting for multimodal messages"
                    )

                # 2. Fallback: concatenate plain text and tokenize. Append a
                # serialized form of the tools so they still contribute to the
                # count when /apply-template is unavailable.
                parts = []
                if system_text:
                    parts.append(system_text)
                for msg in messages or []:
                    if isinstance(msg, dict):
                        parts.append(_block_text(msg.get("content", "")))
                if tools:
                    try:
                        parts.append(json.dumps(tools, ensure_ascii = False))
                    except Exception:
                        pass
                return _tokenize("\n".join(p for p in parts if p))
        except Exception:
            if strict:
                raise
            return 0

    # ── TTS support ────────────────────────────────────────────

    def detect_audio_type(self) -> Optional[str]:
        """Detect audio/TTS codec; swallows errors (use _strict to distinguish)."""
        try:
            return self._detect_audio_type_strict()
        except Exception as e:
            logger.debug(f"Audio type detection failed: {e}")
            return None

    def _apply_detected_audio(self, detected: Optional[str]) -> bool:
        """Apply a probed audio codec under self._lock. Returns True to continue
        the load (codec inited OK, or nothing to init), False to abort (server
        unhealthy or codec init failed). Shared by the fast-path retry and the
        main load path."""
        if detected in ("snac", "bicodec", "dac"):
            with self._lock:
                if not self._healthy:
                    return False
                try:
                    self.init_audio_codec(detected)
                    self._is_audio = True
                    self._audio_type = detected
                except Exception as exc:
                    # Surface as HTTP 500 (matches pre-PR contract).
                    logger.warning("Failed to init audio codec '%s': %s", detected, exc)
                    self._audio_probed = False
                    return False
        elif detected:
            # csm / whisper / audio_vlm: track type but keep _is_audio False --
            # GGUF TTS routing only fires for snac/bicodec/dac.
            with self._lock:
                if not self._healthy:
                    return False
                self._audio_type = detected
        # Audio input = token probe (audio_vlm/whisper) OR mmproj encoder.
        from utils.models.model_config import is_audio_input_type

        self._has_audio_input = bool(is_audio_input_type(self._audio_type)) or bool(
            self._mmproj_has_audio
        )
        return True

    def _detect_audio_type_strict(self) -> Optional[str]:
        """Codec name on match, None on non-audio, raises on transport/JSON errors."""
        if not self.is_loaded:
            return None
        with httpx.Client(timeout = 10, headers = self._auth_headers, trust_env = False) as client:

            def _detok(tid: int) -> str:
                # Non-200 means "marker not in vocab" -- keep probing.
                # Transport / JSON errors still raise.
                r = client.post(f"{self.base_url}/detokenize", json = {"tokens": [tid]})
                if r.status_code != 200:
                    return ""
                return r.json().get("content", "")

            def _tok(text: str) -> list[int]:
                r = client.post(
                    f"{self.base_url}/tokenize",
                    json = {"content": text, "add_special": False},
                )
                if r.status_code != 200:
                    return []
                return r.json().get("tokens", [])

            # Codec-specific tokens (not generic ones that non-audio models may have)
            if "<custom_token_" in _detok(128258) and "<custom_token_" in _detok(128259):
                return "snac"
            if len(_tok("<|AUDIO|>")) == 1 and len(_tok("<|audio_eos|>")) == 1:
                return "csm"
            if len(_tok("<|startoftranscript|>")) == 1:
                return "whisper"
            # Gemma 3n: <audio_soft_token>; Gemma 4: <|audio|> (not csm's <|AUDIO|>).
            if len(_tok("<audio_soft_token>")) == 1 or len(_tok("<|audio|>")) == 1:
                return "audio_vlm"
            if len(_tok("<|bicodec_semantic_0|>")) == 1 and len(_tok("<|bicodec_global_0|>")) == 1:
                return "bicodec"
            if len(_tok("<|c1_0|>")) == 1 and len(_tok("<|c2_0|>")) == 1:
                return "dac"
        return None

    # Prompt format per codec: (template, stop_tokens, needs_token_ids).
    # Matches InferenceBackend._generate_snac/bicodec/dac.
    _TTS_PROMPTS = {
        "snac": (
            "<custom_token_3>{text}<|eot_id|><custom_token_4>",
            ["<custom_token_2>"],
            True,
        ),
        "bicodec": (
            "<|task_tts|><|start_content|>{text}<|end_content|><|start_global_token|>",
            ["<|im_end|>", "</s>"],
            False,
        ),
        "dac": (
            "<|im_start|>\n<|text_start|>{text}<|text_end|>\n<|audio_start|><|global_features_start|>\n",
            ["<|im_end|>", "<|audio_end|>"],
            False,
        ),
    }

    _codec_mgr = None  # Shared AudioCodecManager instance

    def init_audio_codec(self, audio_type: str) -> None:
        """Load the audio codec at model load time (mirrors the non-GGUF path)."""
        import torch
        from core.inference.audio_codecs import AudioCodecManager

        if LlamaCppBackend._codec_mgr is None:
            LlamaCppBackend._codec_mgr = AudioCodecManager()

        device = "cuda" if torch.cuda.is_available() else "cpu"
        model_repo_path = None

        # BiCodec needs a repo with BiCodec/ weights -- download canonical SparkTTS
        if audio_type == "bicodec":
            from huggingface_hub import snapshot_download
            import os

            repo_path = snapshot_download("unsloth/Spark-TTS-0.5B", local_dir = "Spark-TTS-0.5B")
            model_repo_path = os.path.abspath(repo_path)

        LlamaCppBackend._codec_mgr.load_codec(audio_type, device, model_repo_path = model_repo_path)
        logger.info(f"Loaded audio codec for GGUF TTS: {audio_type}")

    def generate_audio_response(
        self,
        text: str,
        audio_type: str,
        temperature: float = 0.6,
        top_p: float = 0.95,
        top_k: int = 50,
        min_p: float = 0.0,
        max_new_tokens: int = 2048,
        repetition_penalty: float = 1.1,
    ) -> tuple:
        """
        Generate TTS audio via llama-server /completion + codec decode.
        Returns (wav_bytes, sample_rate).
        """
        if audio_type not in self._TTS_PROMPTS:
            raise RuntimeError(f"GGUF TTS does not support '{audio_type}' codec.")

        tpl, stop, need_ids = self._TTS_PROMPTS[audio_type]

        payload: dict = {
            "prompt": tpl.format(text = text),
            "stream": False,
            "n_predict": max_new_tokens,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k if top_k >= 0 else 0,
            "min_p": min_p,
            "repeat_penalty": repetition_penalty,
        }
        if stop:
            payload["stop"] = stop
        if need_ids:
            payload["n_probs"] = 1

        with httpx.Client(
            timeout = httpx.Timeout(300, connect = 10),
            headers = self._auth_headers,
            trust_env = False,
        ) as client:
            resp = client.post(f"{self.base_url}/completion", json = payload)
            if resp.status_code != 200:
                raise RuntimeError(f"llama-server returned {resp.status_code}: {resp.text}")

        data = resp.json()
        token_ids = (
            [p["id"] for p in data.get("completion_probabilities", []) if "id" in p]
            if need_ids
            else None
        )

        import torch

        device = "cuda" if torch.cuda.is_available() else "cpu"
        return LlamaCppBackend._codec_mgr.decode(
            audio_type, device, token_ids = token_ids, text = data.get("content", "")
        )