fix: keep LoRA reloads working with PEFT 0.19 (#6748)
Some checks are pending
Core / Core (HF=default + TRL=default) (push) Waiting to run
Core / Core (HF=4.57.6 + TRL<1) (push) Waiting to run
Core / Core (HF=latest + TRL=latest) (push) Waiting to run
Core / llama.cpp build + smoke (push) Waiting to run
Lint CI / Source lint (Python + shell + YAML + JSON + safety nets) (push) Waiting to run
MLX CI on Mac M1 / dispatch (push) Waiting to run
Security audit / advisory audit (pip + npm + cargo) (push) Waiting to run
Security audit / pip scan-packages :: extras (push) Waiting to run
Security audit / pip scan-packages :: studio (push) Waiting to run
Security audit / pip scan-packages :: hf-stack (push) Waiting to run
Security audit / npm scan-packages (Studio frontend tarballs) (push) Waiting to run
Security audit / workflow-trigger lint (pull_request_target / cache-poisoning) (push) Waiting to run
Security audit / pytest tests/security (push) Waiting to run
Security audit / npm provenance + new install-script diff (push) Waiting to run
Studio API CI / Studio API & Auth Tests (push) Waiting to run
Backend CI / (Python 3.10) (push) Waiting to run
Backend CI / (Python 3.11) (push) Waiting to run
Backend CI / (Python 3.12) (push) Waiting to run
Backend CI / (Python 3.13) (push) Waiting to run
Backend CI / Repo tests (CPU) (push) Waiting to run
Frontend CI / Frontend build + bundle sanity (push) Waiting to run
Studio GGUF CI / OpenAI, Anthropic API tests (push) Waiting to run
Studio GGUF CI / Tool calling Tests (push) Waiting to run
Studio GGUF CI / JSON, images (push) Waiting to run
Studio load-orchestrator CI / test (push) Waiting to run
Mac Studio API CI / Studio API & Auth Tests (push) Waiting to run
Mac Studio GGUF CI / OpenAI, Anthropic API tests (push) Waiting to run
Mac Studio GGUF CI / Tool calling Tests (push) Waiting to run
Mac Studio GGUF CI / JSON, images (push) Waiting to run
Mac Studio Install Matrix CI / Install + load (macos-26) (push) Waiting to run
Mac Studio Install Matrix CI / Install + load (macos-14) (push) Waiting to run
Mac Studio Install Matrix CI / Install + load (macos-15) (push) Waiting to run
Mac Studio Install Matrix CI / Install + load (macos-15-intel) (push) Waiting to run
Mac Studio Install Matrix CI / Install + load (macos-26-intel) (push) Waiting to run
Mac Studio UI CI / Chat UI Tests (push) Waiting to run
Mac Studio Update CI / Studio Updating Tests (push) Waiting to run
Studio Tauri CI / Tauri Linux debug build (no codesign) (push) Waiting to run
Studio UI CI / Chat UI Tests (push) Waiting to run
Studio Update CI / Studio Updating Tests (push) Waiting to run
Windows Studio API CI / Studio API & Auth Tests (push) Waiting to run
Windows Studio GGUF CI / OpenAI, Anthropic API tests (push) Waiting to run
Windows Studio GGUF CI / Tool calling Tests (push) Waiting to run
Windows Studio GGUF CI / JSON, images (push) Waiting to run
Windows Studio GGUF CI / Studio install + inference without Visual Studio (push) Waiting to run
Windows Studio GGUF CI / GPU prebuilt resolves without Visual Studio (push) Waiting to run
Windows Studio GGUF CI / setup.ps1 unit tests (VS 2026 / CMake guard) (push) Waiting to run
Windows Studio GGUF CI / real-VS detection (VS 2022) (push) Waiting to run
Windows Studio GGUF CI / real-VS detection (VS 2026) (push) Waiting to run
Windows Studio GGUF CI / VC++ runtime detect + install round-trip (windows-2025-vs2026) (push) Waiting to run
Windows Studio GGUF CI / VC++ runtime detect + install round-trip (windows-latest) (push) Waiting to run
Windows Studio UI CI / Chat UI Tests (push) Waiting to run
Windows Studio Update CI / Studio Updating Tests (push) Waiting to run
Wheel CI / Wheel build + content sanity + import smoke (push) Waiting to run

* fix: keep LoRA reloads working with PEFT 0.19

* test: exercise the PEFT tensor-parallel symbol extractor

* test: prove the full PEFT tensor-parallel seam

* fix: harden PEFT tensor-parallel shims

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: fall back when PEFT tensor-parallel source inspection fails

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
This commit is contained in:
Rod Boev 2026-06-30 15:26:57 -04:00 committed by GitHub
parent d915a139eb
commit 2246a6c9ae
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 402 additions and 4 deletions

View file

@ -23,6 +23,7 @@ import inspect
import os
import re
import sys
from pathlib import Path
from importlib.metadata import version as importlib_version
import pytest
@ -593,6 +594,7 @@ def test_accelerate_gather_empty_logits_debug_mode_patch():
orig_debug = state.debug
orig_dist_type = state.distributed_type
orig_num_processes = state.num_processes
orig_device = state.device
state.debug = True
state.distributed_type = DistributedType.MULTI_GPU
@ -624,6 +626,8 @@ def test_accelerate_gather_empty_logits_debug_mode_patch():
side_effect = mock_gpu_broadcast,
),
):
state.device = torch.device("cpu")
# Top-level EmptyLogits gathers to itself
res = acc_ops.gather(e)
assert res is e
@ -656,6 +660,7 @@ def test_accelerate_gather_empty_logits_debug_mode_patch():
state.debug = orig_debug
state.distributed_type = orig_dist_type
state.num_processes = orig_num_processes
state.device = orig_device
def test_accelerate_patch_is_idempotent():
@ -698,10 +703,8 @@ def test_accelerate_find_device_skips_empty_logits():
def test_accelerate_patch_wired_into_gpu_init():
"""The patch must be installed at startup, not only importable."""
import pathlib
import unsloth.import_fixes as import_fixes
source = pathlib.Path(import_fixes.__file__).with_name("_gpu_init.py").read_text()
source = Path(__file__).resolve().parent.parent / "unsloth" / "_gpu_init.py"
source = source.read_text()
assert "patch_accelerate_recursively_apply()" in source, (
"DRIFT DETECTED: patch_accelerate_recursively_apply is defined but "
"never called in _gpu_init.py, so real imports never install it."

View file

@ -0,0 +1,252 @@
import importlib.machinery
import importlib.util
from pathlib import Path
import sys
import types
import pytest
REPO_ROOT = Path(__file__).resolve().parents[1]
IMPORT_FIXES = REPO_ROOT / "unsloth" / "import_fixes.py"
def _load_import_fixes():
spec = importlib.util.spec_from_file_location("_unsloth_import_fixes_peft_tp", IMPORT_FIXES)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
def _install_fake_module(
name,
*,
is_package = False,
attrs = None,
):
module = types.ModuleType(name)
module.__spec__ = importlib.machinery.ModuleSpec(name, loader = None, is_package = is_package)
if is_package:
module.__path__ = []
if attrs:
for k, v in attrs.items():
setattr(module, k, v)
sys.modules[name] = module
return module
@pytest.fixture(autouse = True)
def _restore_import_fixtures():
keep = {
"transformers",
"transformers.integrations",
"transformers.integrations.tensor_parallel",
"peft",
"peft.utils",
"peft.utils.save_and_load",
}
snapshot = {name: sys.modules.get(name) for name in keep}
yield
for name, value in snapshot.items():
if value is None:
sys.modules.pop(name, None)
else:
sys.modules[name] = value
def _install_fake_transformers_tensor_parallel(existing):
transformers = _install_fake_module("transformers", is_package = True)
integrations = _install_fake_module("transformers.integrations", is_package = True)
setattr(transformers, "integrations", integrations)
tp = _install_fake_module(
"transformers.integrations.tensor_parallel",
attrs = existing,
)
setattr(integrations, "tensor_parallel", tp)
return tp
def _fake_peft_shard_state_dict_for_tp():
from transformers.integrations.tensor_parallel import (
ALL_PARALLEL_STYLES,
ColwiseParallel,
EmbeddingParallel,
RowwiseParallel,
)
return (
ALL_PARALLEL_STYLES,
ColwiseParallel,
EmbeddingParallel,
RowwiseParallel,
)
def _install_fake_peft_tensor_parallel_import():
peft = _install_fake_module("peft", is_package = True)
utils = _install_fake_module("peft.utils", is_package = True)
setattr(peft, "utils", utils)
save_and_load = _install_fake_module("peft.utils.save_and_load")
save_and_load._maybe_shard_state_dict_for_tp = _fake_peft_shard_state_dict_for_tp
setattr(utils, "save_and_load", save_and_load)
return save_and_load
def test_missing_tensor_parallel_symbol_import_succeeds_after_fix(monkeypatch):
module = _load_import_fixes()
_install_fake_peft_tensor_parallel_import()
_install_fake_transformers_tensor_parallel(
{
"ColwiseParallel": object,
"RowwiseParallel": object,
}
)
with pytest.raises(ImportError):
_fake_peft_shard_state_dict_for_tp()
assert module.fix_peft_transformers_tensor_parallel_import_compat() is True
import transformers.integrations.tensor_parallel as patched
all_parallel_styles, _, embedding_parallel, _ = _fake_peft_shard_state_dict_for_tp()
assert patched == sys.modules["transformers.integrations.tensor_parallel"]
assert embedding_parallel is getattr(patched, "EmbeddingParallel")
assert getattr(embedding_parallel, "__unsloth_stub__", False)
with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
all_parallel_styles["rowwise"]
with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
"rowwise" in all_parallel_styles
with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
iter(all_parallel_styles)
with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
len(all_parallel_styles)
def test_existing_embedding_parallel_is_not_replaced(monkeypatch):
module = _load_import_fixes()
class RealEmbeddingParallel:
pass
tp_mod = _install_fake_transformers_tensor_parallel(
{
"EmbeddingParallel": RealEmbeddingParallel,
"ColwiseParallel": object,
}
)
monkeypatch.setattr(
module,
"_extract_peft_tensor_parallel_imported_symbols",
lambda: ("ALL_PARALLEL_STYLES", "ColwiseParallel", "EmbeddingParallel", "RowwiseParallel"),
)
assert module.fix_peft_transformers_tensor_parallel_import_compat() is True
assert tp_mod.EmbeddingParallel is RealEmbeddingParallel
assert not getattr(tp_mod.EmbeddingParallel, "__unsloth_stub__", False)
def test_missing_tensor_parallel_module_is_not_created(monkeypatch):
module = _load_import_fixes()
previous_tp_module = sys.modules.pop("transformers.integrations.tensor_parallel", None)
original_find_spec = module.importlib.util.find_spec
def fake_find_spec(name):
if name == "transformers.integrations.tensor_parallel":
return None
return original_find_spec(name)
monkeypatch.setattr(module.importlib.util, "find_spec", fake_find_spec)
monkeypatch.setattr(
module,
"_extract_peft_tensor_parallel_imported_symbols",
lambda: ("EmbeddingParallel",),
)
assert module.fix_peft_transformers_tensor_parallel_import_compat() is None
try:
spec = importlib.util.find_spec("transformers.integrations.tensor_parallel")
except ModuleNotFoundError as exc:
assert exc.name == "transformers"
else:
assert spec is None
assert "transformers.integrations.tensor_parallel" not in sys.modules
def test_placeholder_raises_on_real_use(monkeypatch):
module = _load_import_fixes()
tp_mod = _install_fake_transformers_tensor_parallel({})
monkeypatch.setattr(
module,
"_extract_peft_tensor_parallel_imported_symbols",
lambda: ("EmbeddingParallel",),
)
assert module.fix_peft_transformers_tensor_parallel_import_compat() is True
with pytest.raises(NotImplementedError, match = "EmbeddingParallel"):
tp_mod.EmbeddingParallel()
assert getattr(tp_mod.EmbeddingParallel, "__unsloth_stub__", False)
def test_symbol_extractor_falls_back_when_parse_returns_no_identifiers(monkeypatch):
module = _load_import_fixes()
_install_fake_peft_tensor_parallel_import()
monkeypatch.setattr(
module.inspect,
"getsource",
lambda _: "from transformers.integrations.tensor_parallel import ()",
)
assert module._extract_peft_tensor_parallel_imported_symbols() == (
"ALL_PARALLEL_STYLES",
"ColwiseParallel",
"EmbeddingParallel",
"RowwiseParallel",
)
def test_symbol_extractor_falls_back_when_getsource_fails(monkeypatch):
module = _load_import_fixes()
_install_fake_peft_tensor_parallel_import()
monkeypatch.setattr(
module.inspect,
"getsource",
lambda _: (_ for _ in ()).throw(ValueError("no source")),
)
assert module._extract_peft_tensor_parallel_imported_symbols() == (
"ALL_PARALLEL_STYLES",
"ColwiseParallel",
"EmbeddingParallel",
"RowwiseParallel",
)
def test_tensor_parallel_import_module_not_found_returns_none(monkeypatch):
module = _load_import_fixes()
_install_fake_peft_tensor_parallel_import()
monkeypatch.setattr(
module.importlib.util,
"find_spec",
lambda name: object() if name == "transformers.integrations.tensor_parallel" else None,
)
def _raise_missing(name):
exc = ModuleNotFoundError(name)
exc.name = name
raise exc
monkeypatch.setattr(module.importlib, "import_module", _raise_missing)
assert module.fix_peft_transformers_tensor_parallel_import_compat() is None

View file

@ -188,6 +188,7 @@ from .import_fixes import (
disable_torchcodec_if_broken,
disable_broken_wandb,
fix_trl_vllm_ascend,
fix_peft_transformers_tensor_parallel_import_compat,
fix_peft_transformers_weight_conversion_import,
patch_peft_weight_converter_compatibility,
patch_accelerate_recursively_apply,
@ -219,6 +220,7 @@ disable_broken_wandb()
# Must run before patch_peft_weight_converter_compatibility: stubs the
# transformers v5 submodules peft 0.19.x imports, so the next patch can wrap
# build_peft_weight_mapping instead of being swallowed by its ImportError.
fix_peft_transformers_tensor_parallel_import_compat()
fix_peft_transformers_weight_conversion_import()
patch_peft_weight_converter_compatibility()
patch_accelerate_recursively_apply()
@ -244,6 +246,7 @@ del patch_vllm_for_notebooks
del patch_torchcodec_audio_decoder
del disable_torchcodec_if_broken
del disable_broken_wandb
del fix_peft_transformers_tensor_parallel_import_compat
del fix_peft_transformers_weight_conversion_import
del patch_peft_weight_converter_compatibility
del patch_accelerate_recursively_apply

View file

@ -1492,6 +1492,146 @@ def disable_broken_wandb():
# Stamped on stub modules so a second call is a strict no-op and so third
# parties can introspect ``__unsloth_stub__`` to detect our patch.
_UNSLOTH_STUB_SENTINEL = "__unsloth_stub__"
_PEFT_TENSOR_PARALLEL_FALLBACK_SYMBOLS = (
"ALL_PARALLEL_STYLES",
"ColwiseParallel",
"EmbeddingParallel",
"RowwiseParallel",
)
def _extract_peft_tensor_parallel_imported_symbols():
"""Return names PEFT expects from ``transformers.integrations.tensor_parallel``.
The supported PEFT import line is the one in
``peft.utils.save_and_load._maybe_shard_state_dict_for_tp`` for fast
LoRA adapter checkpoints. Parse that source to avoid stale hard-coded
symbol lists.
"""
try:
import peft.utils.save_and_load as _save_and_load
except Exception:
return ()
try:
sharding_fn = _save_and_load._maybe_shard_state_dict_for_tp
except AttributeError:
return ()
try:
source = inspect.getsource(sharding_fn)
except Exception as exc:
logger.debug("Failed to inspect PEFT tensor-parallel imports: %r", exc)
return _PEFT_TENSOR_PARALLEL_FALLBACK_SYMBOLS
import_pattern = re.compile(
r"from\s+transformers\.integrations\.tensor_parallel\s+import\s*\((.*?)\)",
re.S,
)
import_pattern_single = re.compile(
r"from\s+transformers\.integrations\.tensor_parallel\s+import\s+([A-Za-z_][A-Za-z0-9_\s,]*)",
re.S,
)
matches = import_pattern.findall(source)
if not matches:
matches = import_pattern_single.findall(source)
symbols = []
seen = set()
for match in matches:
pieces = re.split(r"[,\n]", match)
for piece in pieces:
candidate = piece.strip()
if not candidate:
continue
if candidate.endswith(")"):
candidate = candidate[:-1].strip()
if not candidate.isidentifier():
continue
if candidate in seen:
continue
symbols.append(candidate)
seen.add(candidate)
return tuple(symbols) or _PEFT_TENSOR_PARALLEL_FALLBACK_SYMBOLS
def _raise_on_peft_tensor_parallel_symbol_use(symbol_name):
raise NotImplementedError(
f"Unsloth: cannot use unsupported "
f"`transformers.integrations.tensor_parallel.{symbol_name}` on this "
f"transformers installation. Please upgrade transformers before "
f"using PEFT tensor-parallel adapter sharding features."
)
def fix_peft_transformers_tensor_parallel_import_compat():
"""Preserve existing ``transformers.integrations.tensor_parallel`` objects, then add
lightweight placeholders for symbols that PEFT expects but this transformers
build omits.
Returns ``True`` when patched, ``False`` when no patch is needed, and
``None`` when transformers / PEFT context is absent.
"""
try:
tensor_parallel_spec = importlib.util.find_spec("transformers.integrations.tensor_parallel")
except ModuleNotFoundError:
return None
if tensor_parallel_spec is None:
return None
required_symbols = _extract_peft_tensor_parallel_imported_symbols()
if not required_symbols:
return None
try:
tp_mod = importlib.import_module("transformers.integrations.tensor_parallel")
except ModuleNotFoundError as exc:
if exc.name not in {
"transformers",
"transformers.integrations",
"transformers.integrations.tensor_parallel",
}:
raise
return None
missing = [symbol for symbol in required_symbols if not hasattr(tp_mod, symbol)]
if not missing:
return False
def _install_symbol_placeholder(symbol_name):
if symbol_name == "ALL_PARALLEL_STYLES":
class _UnslothTensorParallelStyles(dict):
def __getitem__(self, key):
_raise_on_peft_tensor_parallel_symbol_use(symbol_name)
def get(self, *args, **kwargs):
_raise_on_peft_tensor_parallel_symbol_use(symbol_name)
def __contains__(self, key):
_raise_on_peft_tensor_parallel_symbol_use(symbol_name)
def __iter__(self):
_raise_on_peft_tensor_parallel_symbol_use(symbol_name)
def __len__(self):
_raise_on_peft_tensor_parallel_symbol_use(symbol_name)
value = _UnslothTensorParallelStyles()
else:
class _UnslothTensorParallelPlaceholder:
def __init__(self, *args, **kwargs):
_raise_on_peft_tensor_parallel_symbol_use(symbol_name)
value = _UnslothTensorParallelPlaceholder
value.__name__ = f"UnslothTensorParallelPlaceholder{symbol_name}"
setattr(value, _UNSLOTH_STUB_SENTINEL, True)
setattr(tp_mod, symbol_name, value)
for symbol in missing:
_install_symbol_placeholder(symbol)
return True
def _peft_stub_module_importable(name):