* test-backend-sampler: skip multi_output_sampling_chain on HIP
The new multi_output_sampling_chain test uses top_k, whose backend probs
path needs CUB (unavailable on HIP), so sampled_probs is null and the test
aborts. Add it to the existing HIP skip list alongside the other TOP_K tests.
* ci: keep gpu-rocm logs in a per-run dir keyed by GitHub run id
The self-hosted gpu-rocm runner can't upload logs to Azure blob (egress
firewalled), so a run's logs were wiped by the next run. Write each run's
logs to $OUT/run-<run_id>-<attempt>/ so an Actions run URL maps to its logs.
* test-backend-sampler: also skip multi_output_cpu on HIP
Like the other TOP_K-based subtests, multi_output_cpu's backend sampler
never initializes on HIP (no CUB TOP_K), so it aborts. Add it to the skip list.
---------
Co-authored-by: Jim Wu <ywu@xilinx.com>
* Enable backend sampling with token speculation
* Clamp the mask sum before converting it into the sampled index
* Add a numeric context parameter declaring the maximum outputs one sequence
* More fixes
* Don't reuse memory for output views.
* Match dist between CPU and GPU
* Fix CPU and backend sampling mismatches
* Simpify some of the changes
* Fix tests on Vulkan
* More test fixes
* Rebase changes
* Rebase and address review comments
* Address review comments
* Address review comments
* Update src/llama-sampler.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
ggml_roll only asserts nb[0] == ggml_type_size, so a permuted src is a
valid input, but the CUDA and Metal roll kernels index by ne alone and
never read the nb strides. A non-contiguous src therefore produced
silently wrong results. Neither backend declared a contiguity
requirement in supports_op, so the scheduler did not fall back to the
CPU implementation, which does handle strides correctly.
Add the requirement to both backends, matching the existing
GGML_OP_ROPE guard, and add a permuted test_roll case.
* Get started with Onyx
* Add architecture
* Skip keys handled in super()
* Loading tensors
* Shorten
* Graph
* Apply suggestion from @pcuenca
* Remove norm now embedding in transformers weights
* Add eot
* Explicit output_multiplier
* Handle post_norm_eps
* No super call; unhardcode eot.
The pattern `self._set_vocab_gpt2()` seems preferred throughout the
codebase, and it allows `set_vocab()` to be called from a different part
of the Python class hierarchy: the drafter model converter that we may
need eventually.
* Register for drafting
* DFlash: inherit rope type from the linked target.
Another option would be to store it in the gguf file itself.
* mmproj conversion
Note: some fields to be renamed after the implementation works. We are
keeping compatibility with the reference Meta gguf for testing purposes.
* "clip" header declarations
* Load mmproj
* Pre-processing
* Graph
* Go back to using delimiters.
Otherwise our generations are worse.
Transformers does not use them. We need to trace inputs to verify
whether they are equivalent.
* downsample_factor -> merge_size
* Add vision graph
lol, forgot from a previous commit
* Additional renames, align with llama.cpp / transformers
* Prefer _size instead of independent _h and _w
* Fix token layout
Co-authored-by: Young Han <younghan@fb.com>
* onyx: bring the chat parser onto the onyx branch
common/chat.cpp on this branch has no Onyx handling, so a converted model
serves malformed chat: the assistant preamble leaks into content
("to=self<|message|>...") and tool calls fail with
HTTP 500 "The model produced output that does not match the expected
peg-native format"
common_chat_params_init_onyx exists on onyx-fair-patch, added there by
8bb73dd3d. It was never on this branch, so this is not a regression --
the two lines developed independently.
The code here is taken verbatim from that commit. It is the clean side of
`git merge origin/onyx-fair-patch`: chat.cpp is one of the files that
merges without conflict. The full merge is not viable -- it produces 13
conflicts, including add/add on conversion/onyx.py and src/models/onyx.cpp
where the q_norm-folding and metadata-scale approaches contradict each
other, and #4/#7 are stacked on this branch's side of that.
Verified on this branch: builds with 0 errors, converts an Onyx checkpoint,
and serving it gives "4" for "What is 2+2?" plus a correct
get_weather {"city":"Paris"} tool call, where the unported branch gives the
two failures above.
No converter or runtime changes are included, so this should not interact
with the q_norm work.
Co-authored-by: Beto de Paola <betodepaola@meta.com>
* Less params, bilinear pos-emb interpolation as a graph op instead of CPU
* Map to symbolic V_MMPROJ instead of strings
* Make a couple params explicit
* Patchify via build_inp()
* No param for rope_theta
* Small cleanup
* Restore blank line
* Unpermute, to adapt to the latest transformers checkpoint
* Apply norm after token embeddings
This follows the latest transformers approach.
* Remove duplicated function
* build_vit
* onyx: use the model rope theta on sliding-window layers
* DFlash: conversion from transformers drafter
* Revert rope_type derivation from target
NOTE: this breaks compatibility with Meta's distributed DFlash GGUFs, as
the Q/K are stored in "NEOX" (rotated half) format, like in
transformers.
* Apply suggestion from @pcuenca
* Set model type
* Remove comment that will become obsolete
* Hardcode post_norm_rms_eps instead of new param
* Derive SWA+RoPE pattern from gguf array or scalar
* Fix model type <-> number of layers
* Reorder
* Rename
* Fix typo
* DFlash: seed the draft KV cache from multimodal embedding batches
`common_speculative_impl_draft_dflash::process()` returned early on any batch carrying embeddings, so an image prefill never had its target-layer features fused through the DFlash encoder and injected into the draft's KV cache. That left a hole spanning the image's positions, and the next injection at a post-image position failed to initialize its batch:
```
decoding image batch 1/1, n_tokens_batch = 256
decode: failed to initialize batch
llama_decode: failed to decode, ret = -1
process: llama_decode(ctx_dft) failed rc=-1 (n_tokens=17, offset=0)
srv decode: failed to process speculative batch
```
Every image request with `--spec-type draft-dflash` failed with HTTP 500. Text-only was unaffected, since those batches carry token ids and were let through.
Restore the earlier condition, which admits a batch that is either tokens or embeddings and skips only the degenerate neither/both cases. The rest of `process()` is already layout-agnostic -- it gathers features via `llama_get_embeddings_layer_inp()` and indexes `batch_in.pos[]` / `batch_in.seq_id[]`, none of which assume token ids -- so this is the whole fix.
Validated against `muse-glimmer-30B-bf16.gguf` + `mmproj-muse-glimmer-30B-bf16.gguf` + a DFlash draft head, on an image describe-the-shapes request:
- before: HTTP 500, `failed to process speculative batch`
- after: HTTP 200, draft acceptance 0.34012 (167 accepted / 491 generated), mean len 3.04
Output equivalence holds, which is the property that matters: at temperature 0 the drafted response is byte-identical to the same request served with no draft attached (1213/1213 chars), so the draft is drafting correctly through the image context rather than merely not crashing.
* Conversion: prefer rewrite to mapping
* Revert "Conversion: prefer rewrite to mapping"
This reverts commit a92d0ac584d315e876741e85b6dad3dbc8b23bf7.
* fix lint
* sliding_window metadata is not optional
* disable state save/load
* Apply suggestion from @pcuenca
---------
Co-authored-by: Young Han <younghan@fb.com>
Co-authored-by: Beto de Paola <betodepaola@meta.com>
Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
Co-authored-by: ruanrms <ruanslv@gmail.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* granite-switch: add llama.cpp backend (POC, CPU)
New "granite-switch" architecture: a dense, all-attention Granite-4.1
model with N embedded LoRA adapters selected per-token by control tokens.
- gguf-py schema (arch, KV keys, stacked LoRA tensor names) + writer helpers
- conversion/granite.py: GraniteSwitchModel converter (stacks N adapters +
zero base slot into per-projection A/B tensors; emits switch metadata)
- C++ arch registration (llama-arch.{h,cpp}, llama-model.{h,cpp})
- src/models/granite_switch.cpp: load + per-token switched-LoRA graph via
ggml_mul_mat_id over stacked tensors; sticky per-token index + control-token
substitution in llm_graph_input_switch::set_input
- llm_graph_input_switch in src/models/models.h
Runs end-to-end on CPU: convert 3b checkpoint (842 tensors, stacked dim 13)
and generate on both base and control-token paths. Sticky switch state is
single-sequence (POC); full multi-sequence machinery is a follow-up.
* granite-switch: add Mac (Metal) build + mid-sequence switch demo script
Self-contained script to build llama.cpp on Apple Silicon (Metal),
convert the composed 3b checkpoint, and run the crisp mid-sequence
adapter-switch demos verified on Vela:
- answerability: <|answerability|> mid-seq -> "unanswerable"
- query_rewrite: <|query_rewrite|> mid-seq -> {"rewritten_question": ...}
Each demo runs the same prompt twice, differing only by a control token
placed before the assistant turn, so the per-token switch is visible.
* granite-switch mac demo: add -no-cnv so each run is one-shot
The composed model ships a chat template, so llama-completion auto-enables
interactive conversation mode and halts at a `>` prompt after generating,
stalling the script. -no-cnv disables conversation mode: generate once from
the raw prompt and exit (also prints special tokens, making the switch visible).
* granite-switch: replace global sticky index with in-graph router attention
The POC computed the per-token adapter index on the CPU and carried it
across ubatches in ONE global `mutable int32_t poc_sticky_index`, reset
only when a ubatch contained sequence position 0. That global had two
problems:
1. Concurrency: with multiple sequences in a batch it was last-writer-
wins — one sequence's adapter leaked into the others.
2. Multi-turn: an interactive `ollama run` chat continues one KV cache,
so turn 2 never saw position 0 and the index never reset — the
adapter stayed stuck on across turns.
Port the vLLM/HF backend mechanism faithfully: a single-head causal
"router" attention recovers the adapter index in-graph. Per token, only
dim 0 carries signal — Q[0]=1, K[0]=+gain for a control token / -gain
otherwise, V[0]=adapter slot / 0 — and the causal softmax over the single
visible control token recovers that adapter's slot (readback =
clamp(round(V[0]), 0, n_adapters)). gain=15 matches config.py and is
F16-safe (no F32 cache).
The router's K/V live in the model KV cache at an extra layer
R == hparams.router_layer (== n_layer). We bump n_layer_all to n_real+1
so the cache allocator gives the router its own per-sequence slot, and
set n_layer_nextn=1 so n_layer() stays n_real — the decoder loop and
tensor loading are untouched and never reference layer R. The router K is
exempted from the k-shift RoPE loop (its dim-0 value is a literal
magnitude, not a rotation).
Because the selection now lives in the per-sequence KV cache, CONCURRENT
requests are isolated for free (problem 1 fixed; verified by
scratch/concurrent_switch_test.cpp). set_input becomes stateless pure
per-token maps; the global is gone.
Single-switch contract / known limitation, identical to vLLM & HF: the
gain is flat (no recency), so within one sequence there is no mechanism to
revert to base mid-sequence — once an adapter fires it stays on until that
sequence ends (problem 2 is therefore NOT fixed by a faithful copy; vLLM/HF
avoid it only because each served request is a fresh sequence). A client
continuing one KV cache across turns must start a fresh sequence per turn,
or opt into a recency-biased router (a deliberate divergence, not done
here). Documented in granite_switch.cpp and asserted by
scratch/multiturn_leak_test.cpp.
Verified (CPU): both demos unchanged (answerability -> "unanswerable",
query_rewrite -> rewritten query); concurrent two-sequence isolation
passes; multi-turn carry-over matches the vLLM/HF contract.
* granite-switch: drop scratch tests and mac demo for upstream PR
Remove the local-only development artifacts that should not ship in the
upstream PR:
- granite-switch-mac-demo.sh (local Metal build + demo driver)
- scratch/concurrent_switch_test.cpp
- scratch/multiturn_leak_test.cpp
Also drop the now-dangling reference to the scratch tests from the
granite_switch.cpp header comment. Leaves only the core architecture
support (conversion, gguf constants, llama-arch/model/kv-cache, and the
granite_switch graph).
* granite-switch: trim comments to match native llama.cpp style
* granite-switch: trim conversion comments to match native style
* granite-switch: drop unused adapter_ranks metadata
* granite-switch: rename arch to graniteswitch and drop obid alias
* granite-switch: fix non-ASCII comments and document router gain assumption
* granite-switch: drop section comments from constants.py to match native style
* granite-switch: add functional tensor block comments matching Granite4 Vision style
* granite-switch: clarify n_expert_used comment
State the actual constraint: mul_mat_id needs n_expert_used == 1, and
since the GGUF carries expert_count = 0 the generic loader's
n_expert == 0 => n_expert_used == 0 assertion has already passed by the
time load_arch_hparams runs, so it is forced to 1 here.
* granite-switch: note n_layer_nextn reuse has no MTP
The router carving reuses n_layer_nextn, normally the MTP/next-token
count. Clarify in the comment that it is borrowed here purely as the
trailing-layers lever and that there is no MTP head, to spare readers
the double-take.
* granite-switch: rename source file and apply review nits
* granite-switch: don't force LoRA tensors to F16, follow --outtype instead
* granite-switch: drop redundant _permute_qk wrapper, call LlamaModel.permute directly
* granite-switch: read router gain from GGUF (control_token_gain) instead of hardcoding 15.0
* granite-switch: derive n_slots()
* granite-switch: move llm_graph_input_switch into granite-switch.cpp
* granite-switch: cut AI-style narration comments
* granite-switch: collapse multi-line comments
* granite-switch: rename control_token_* maps to adapter_token_*
* granite-switch: cut noise comments
* granite-switch: rename embedded LoRA tensors to <base>.lora_a/lora_b
* granite-switch: GGML_ASSERT token input to avoid UB on embeddings
* granite-switch: TODO for raw embedding input support
* granite-switch: collapse LoRA tensor constants to .lora_a/.lora_b suffix
* granite-switch: drop n_expert_used hack, guard mul_mat_id buft probe
* granite-switch: stop forcing dense expert counts, read from config
* granite-switch: renamed control_token_gain metadata key to router_gain
* granite-switch: trim header comments to match native style
* granite-switch: collapse LoRA tensors to base name + suffix
* granite-switch: inline suffix checks in tensor op resolution
* granite-switch: drop switch-lora struct comment
* granite-switch: guard router layer index and inline n_slots
* granite-switch: group adapter metadata under {arch}.adapters.* namespace
* granite-switch: add hparams.has_rope(il) for KV-shift rope skipping
* granite-switch: skip arch in test-llama-archs (adapter fixture missing, TODO)
* granite-switch: Keys.Adapters namespace + simplify n_slots
* granite-switch: validate substitute token ids against n_vocab
* granite-switch: bound adapter count and lora rank from GGUF
* granite-switch: reject MTP context type when router_layer is set
* granite-switch: throw on bad adapter metadata instead of GGML_ASSERT
* granite-switch: use ASCII +/- in router K signal comment
* granite-switch: document n_layer_nextn repurpose and its leak points
* granite-switch: gate lora_a/lora_b op mapping on router_layer
* granite-switch: label all three preview model sizes
The saver called add_kv with LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH twice, the
second time passing n_ff_chexp. gguf_set_val_u32 removes-then-appends, so the second
call clobbers the first: the saved shared_feed_forward_length ends up as n_ff_chexp
(0 for every arch except GroveMoE), and expert_chunk_feed_forward_length is never
written at all.
So a save->load roundtrip of any MoE model with a shared expert loses n_ff_shexp. On
reload the arch falls back to n_ff for the shexp tensor shape, that no longer matches
the saved tensor, and the model FAILS to load. Hits qwen2moe, qwen3-next, granite-moe,
hunyuan-moe, ernie4.5, bailingmoe2, nemotron-h, and the other shared-expert MoEs.
Fix: the second call writes LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH.
test-llama-archs: set expert_shared_feed_forward_length to a value distinct from n_ff
in the MoE setup so the roundtrip exercises it. Without the fix the reload fails on a
shexp tensor-shape mismatch; with it, every arch roundtrips clean.
ggml_metal_op_norm sized the threadgroup with
`nth = std::min(nth, args.ne00_t)`, which can leave nth not a multiple of
the simdgroup size. The kernels finish their row reduction with a
cross-simdgroup step where each lane of the last simdgroup reads one
per-simdgroup partial sum out of shmem_f32:
if (tiisg == 0) { shmem_f32[sgitg] = sumf; }
threadgroup_barrier(mem_flags::mem_threadgroup);
sumf = shmem_f32[tiisg];
sumf = simd_sum(sumf);
When the last simdgroup is partial it has fewer lanes than the
threadgroup has simdgroups, so the tail of the partial sums is never
read and the row sum is too small. For ne00_t = 33 nth becomes 33: two
simdgroups, but only one lane in the second, so one of the two partial
sums is dropped. The mean and variance are then wrong for the whole row.
Round ne00_t up to a whole number of simdgroups instead. Rounding up
rather than dropping the clamp keeps the threadgroup as small as
possible: deleting the line would raise nth to the next power of two
(ne00_t = 544 -> 1024 instead of 544), which costs idle lanes on 26 row
lengths below 8192 that were already correct, including 1536 and 3584.
GGML_OP_NORM is affected as well as GGML_OP_RMS_NORM - both dispatch
through ggml_metal_op_norm.
No mainstream LLM hidden size hits this: ne00_t is ne00/4 on the
vectorized path, so 4096, 8192, 2048 and friends all give a multiple of
32. It is reachable from other norm shapes, e.g. 320-channel norms.
Add NORM and RMS_NORM cases for ne0 = 33, 132 and 260 across the
existing eps values. 33 exercises the scalar path and 132/260 the
vectorized one, since only those divide by 4.
Before, on M3 Pro:
test-backend-ops test -b MTL0 -o NORM 25/50
test-backend-ops test -b MTL0 -o RMS_NORM 26/51
After:
test-backend-ops test -b MTL0 -o NORM 50/50
test-backend-ops test -b MTL0 -o RMS_NORM 51/51
test-backend-ops test -b MTL0 13943/13943
* tests: add SWIGLU perf cases
perf mode had no GLU coverage. Adds SWIGLU at 17408 columns, 512 and
2048 tokens, f16 and f32, with the operands both fused and split.
* sycl: consolidate fused-GLU kernels
They differed only in which op_* they called, so take the op as an argument and share a common launcher.
Their block sizes were all 256, so launch geometry is unchanged;
SYCL_GELU_BLOCK_SIZE and SYCL_SILU_BLOCK_SIZE lose their last users so are dropped.
* sycl: contiguous fast path for the fused GLU ops
o0 == n and o1 == n collapse the de-interleave index math to the
identity, so dispatch a flat kernel in that case. It fires for
ggml_glu_split with packed operands; a fused [gate|up] tensor keeps the
strided path. test-backend-ops perf -o SWIGLU on an Arc Pro B70: split
+14% f16 and +4% f32, fused unchanged.
* ci: prepare for amd rocm ci
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
* ci: fix editorconfig-checker
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
* ci: fix device not recognised
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
* ci: rename gpu-amd to gpu-hip
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
* ci: gpu-hip to gpu-rocm
haha
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
* CUDA: allow integrated-GPU host output buffer in debug assert
On integrated GPUs (APUs), the scheduler can legitimately place a graph
node's output on the host-visible buffer, which ggml_cuda_compute_forward
already handles. The debug assert in ggml_cuda_graph_evaluate_and_capture
required every node output to be on the device buffer, so a debug build
aborts on such a node (e.g. attn_residual ADD -> ROCm_Host on RDNA3.5).
The source-tensor assert directly below already permits this via the
integrated + cuda_host exception; apply the same exception to the node's
own output buffer. Debug-only; no effect on release/compute.
Fixes test-recurrent-state-rollback on gfx1151 (Strix Halo).
* ci: enable unified memory for ROCm gfx1151 job
Work around a coherence issue on integrated RDNA3.5 (gfx1151) where GPU
kernels reading mmap-loaded weights can return incorrect output, which
makes test-llama-archs (and real inference) intermittently wrong.
GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 uses managed memory, which restores
coherence. Remove once the underlying ROCm/HIP issue is fixed.
* test-llama-archs: skip jamba on HIP backend
jamba produces incorrect output (~0.55 NMSE vs CPU) on the HIP backend on
RDNA3.5 (gfx1151); the SSM kernels need separate investigation. Skip it
for now, matching the existing per-backend carve-outs (WebGPU), so the
ROCm CI can run the test for the remaining architectures.
* ci: use HIP_LAUNCH_BLOCKING for ROCm gfx1151 job
The gfx1151 ROCm CI job produced incorrect inference output (qwen3 perplexity ~88 vs ~9.4) due to an async-execution correctness issue in the HIP path. Serializing kernel launches with HIP_LAUNCH_BLOCKING=1 restores correctness. This replaces the earlier GGML_CUDA_ENABLE_UNIFIED_MEMORY workaround, which did not fix batched inference.
* test-backend-sampler: skip top-k subtests on HIP backend
The ROCm backend does not support the TOP_K/ARGSORT op at vocab scale (no CUB; bitonic argsort is capped at ncols <= 1024), so top-k/top-p backend samplers cannot be offloaded. The penalties, set_sampler, mixed, and top_p subtests assert that offload happened, so they fail on HIP. Skip them until TOP_K is supported on the ROCm backend.
* Update tests/test-backend-sampler.cpp
Co-authored-by: Aaron Teo <taronaeo@gmail.com>
* Update tests/test-backend-sampler.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
Co-authored-by: Aaron Teo <aaron.teo1@ibm.com>
Co-authored-by: Jim Wu <ywu@xilinx.com>
Co-authored-by: Aaron Teo <taronaeo@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Resolve -1 to 1024 instead of ctx-len for samplers
Because of backend-sampling we initialize samplers before the complete
llama_context is there. Therefore, we cannot infer the resolved context
length yet at the time we construct the samplers.
* Shared default of 64 for history-based samplers, remove context_size
* common/chat: update DeepSeek V4 templates
Align the DeepSeek V4 templates with the official encoders while keeping parser behavior out of this change.
- Default drop_thinking for DeepSeek V4 history so prior thinking is omitted unless preserve_reasoning is requested or tools are present.
- Add structured output response-format instructions to the V4 templates and pass the schema into template rendering.
- Add a separate Flash 0731 template for the updated high and max reasoning effort mapping.
- Cover reasoning effort, drop_thinking, structured output prompts, preserved reasoning, continuations, and empty tool arguments in template rendering tests.
Official references:
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/encoding/encoding_dsv4.pyhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/blob/main/encoding/encoding_dsv4.py
Assisted-by: Codex
* Fix deepseek v4 0731 template selection
* remove unneeded lower normalization
* Fix DSML parser to consume the tool call separator
* address aldehir requests
* address aldehir comment
* tests: add model resolution test on synthetic repo listings
Include download.cpp and arg.cpp inside a namespace with hf_cache
monkey patched to serve hardcoded listings, so the resolution and the
model handler assembly are tested end-to-end through the real CLI
parsing, without modifying the tested code and without network access.
Covers the primary, shard, mmproj, sidecar and preset resolution on
layouts mimicking real vendor conventions, replays every case on
permutations of the listing to assert determinism, and asserts the
final wired paths for the spec type auto-selection, the -md precedence
and the fallback suppression.
* tests: keep model resolution checks active and let the handler resolve
Replace assert with a REQUIRE macro alive in Release builds, key the
fake hf_cache by repo id so the real handler init resolves every plan
itself, check the exact shard sets, restrict the permutation exception
to the order dependent picks, and cover dflash and eagle3.
* tests: fix model resolution build on fatal warnings CI and Windows
The namespaced copy of the sources leaves many static functions unused
in this TU, exempt it from the unused warnings. Pre-include the
windows headers so arg.cpp does not pull them inside the namespace.
Declare the renamed copies of the download.h functions, verbatim from
the header and renamed in sync by the macros, so missing declaration
and missing prototype warnings are satisfied on every toolchain.
* tests: fix winsock inclusion order for the model resolution test
WIN32_LEAN_AND_MEAN and winsock2.h before windows.h, so http.h does
not redefine the socket types afterwards.
* tests: link cpp-httplib to the model resolution test
The test compiles its own copy of download.cpp, which calls httplib
directly, and the private link of llama-common does not propagate the
symbols under lld-link.
* common_http_client
* common: finish the http client wrapper
Add the virtual Head, Get and Post methods and the passthrough
setters to the common_http_client skeleton, move follow_location
into the constructor, expose the underlying client for the ranged
pull path, and rename the missed common_http_client_init call sites.
* tests: rewrite model resolution on the http client stub
Replace the namespace inclusion of the sources by a plain TU: the
common_http_client factory returns a stub serving hardcoded HF API
responses, so the real hf_cache parsing, resolution and CLI handler
run against synthetic listings in an isolated cache directory.
Failures print the named case, the reordering and the actual versus
expected values, the assembly cases use the full command line as
context, and the empty result cases are checked once to keep the
logs short.
* tests: fix the model resolution on Windows and the builds without TLS
Assert the exact expected paths composed like the cache does instead of
suffix matching on forward slashes, set the environment portably, and
serve the stub through an http endpoint so the builds rejecting the
https scheme still reach it. Pause the log so the negative cases can be
replayed on every reordering.
* tests: make the model resolution failures self explanatory
Resume the paused log before the failure report so the CI shows why
the tested code bailed, and format the stub oids portably.
* common: hold the http client factory behind exported functions
The factory was an inline variable, and the Windows shared builds
export functions but not data, so the executable and the DLL each had
their own instance: the stub installed by the test was invisible to
the library, which reached for the real endpoint and resolved nothing.
Route the creation through functions compiled into the library and
format the stub oids portably.
* common: add the http client factory source missed in the previous commit
* common: typedef the http client factory callback
Address review from @ngxson
* tests: serve the model resolution repos over the loopback
Replace the client stub by a real httplib server bound to the
loopback, so no C++ object crosses the module boundary anymore and
the library exercises its own client and transport end to end. The
debug shared build on Windows crashed inside the stubbed path.
* common: add portable common_get_env and common_set_env helpers
Address review from @ngxson
* common: drop the http client factory left without a caller
The loopback server made the stub substitution unnecessary, the client
init builds the real client directly again.
* common: read the model endpoint through the env helpers
* nit: drop the stub leftovers from the model resolution test
* common: align common_set_env and isolate the test cache per run
The POSIX branch now behaves like _putenv_s, so the helper has a single
contract on every platform, and common_get_env already reads an unset and
an empty variable alike.
The model resolution test keys its cache directory on the loopback port,
where two concurrent runs on the same machine used to share one directory
and the initial cleanup of either wiped the other.
* tests: move the model resolution server into main
* tests: support the DSpark sidecar resolution
* common: revert the http client to the plain httplib client
address review from @ngxson
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* sampling: enhance penalty handling in common_sampler_init
- Set default value for penalty_last_n based on model context if not specified.
- Ensure penalty_last_n and n_prev are non-negative.
- Update llama_sampler_penalties structure to inherit from llama_sampler_backend and add backend input handling for penalties.
- Implement backend initialization and application logic for penalties, including frequency and presence adjustments.
* tests: add backend penalties sampling tests and utility functions
- Introduced `accept_prompt` and `unique_prompt_tokens` functions to handle prompt acceptance and token uniqueness.
- Implemented `compare_penalties_logits` to compare logits from backend and CPU samplers with penalties.
- Added `test_backend_penalties_sampling` to validate backend penalties with various configurations.
- Enhanced the test suite for better coverage of penalty handling in sampling.
* sampling: add support for top-k penalties in backend sampling
* sampling: add fix to ensure stable numerical results. Preserve masked logits as -Inf and no longer generate NaN.
* sampling: enhance penalty comparison tests with masking penalties logic
* add comments on padding
* sampling: add comments on modifications
* add the unit test to cover masked-out token as -INF
* validate repeat penalty to ensure it is finite and greater than 0; add tests for invalid values
* refactor: test functions to share logic and be less verbose
* add test to cover case where previously penalized token is not part of candidates
* remove comments
* remove redundant penalty_last_n initialization and validation in common_sampler_init
* add support for penalties in sampler chain with configurable positions
* add validation for penalty parameters and enhance tests for non-finite values
* add context parameter to common_sampler_init and set default for penalty_last_n
* add llama_n_ctx parameter to common_sampler_init for improved sampler initialization
* replace penalty_last_n x n_candidates comparison matrix with a vocabulary-sized count tensor
* add tests for backend penalties sampling without filler entries , token_count.size() == n_active == n_max == 64
* add test for backend penalties sampling after top-p with large history window
* remove as unused
* add is_disabled method, tensor logits reshape, add rest review suggestions
* clarify comment
- Implement GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and
GGML_OP_DSV4_HC_POST with SIMDgroup register and shuffle optimized kernels.
- Add Metal dispatch and support plumbing and test the production Sinkhorn
iteration count and embedding width.
Assisted-by: Codex
Co-authored-by: Thiago Padilha <thiago@padilha.cc>
* vulkan : add pool1d push constants and pipeline field
Declared data structures needed for POOL1D OP, which are the vk_op_pool1d_push_constants struct and pipeline_pool1d_f32 field.
* vulkan : add pool1d compute shader
Added pool1d.comp for Vulkan backend mirroring the existing pool2d shader.
* vulkan : add full GGML_OP_POOL_1D support
Added pipeline creation and op dispatch for 1D pooling in the Vulkan backend.
* vulkan : fix pool1d shader logic
Registered pool1d_f32 in vulkan-shaders-gen.cpp and fixed tensor dimension indices and avg pool scale.
* vulkan : fix pool1d end boundary crash and expand test coverage
Fixed an issue where the shader crashed when the end boundary was negative when k0 < p0. Also, added more test cases related to this fix.
* SYCL: add oneMKL GEMM flash attention for XMX-accelerated prompt processing
* fattn-mkl: fix interleaved dst layout in normalize kernel
- Fix mkl_fa_normalize_head: use interleaved dst layout
((query * n_q_heads + head) * DV) matching TILE's
flash_attn_combine_results. Previously used dense head-major
layout which wrote head outputs to wrong addresses, corrupting
attention for all models except Qwen3.6-27B (where GQA=6 heads
were sparse enough to avoid visible overlap).
- Remove 7 redundant stream->wait() calls — SYCL in-order queue
already serializes pure SYCL kernel dependencies. Retain only
the 4 MKL GEMM ↔ SYCL handshake barriers (oneMKL GEMM uses its
own internal queue that does not respect SYCL in-order).
- Remove unused dst_row_stride, diagnostic clutter, and dead
K/V hex dump (fa_diag block in fattn-mkl.cpp).
- Add MKL_FA_DISABLE=1 env var for A/B testing.
- Add FA-DISP watchdog (MKL_FA_DEBUG=1) and FA-DIAG output
fingerprint (MKL_FA_DIAG=1) in fattn.cpp.
Tested: Gemma-4-26B, Gemma-4-31B, Qwen3.6-27B, Qwen3.6-35B-A3B
Perf (B70/Battlemage, 32K, q8_0 KV):
Gemma-4-26B: 1473 t/s MKL vs 746 TILE (1.97x)
Qwen3.6-27B: 609 t/s MKL vs 330 TILE (1.85x)
Co-Authored-By: Claude Code on DeepSeek-v4-Pro
* Thank you for the review feedback: rename env vars, use GGML_LOG_INFO, document in SYCL.md
Completed the following:
- Rename MKL_FA_DISABLE → GGML_SYCL_ENABLE_MKL_FA (inverted: 0 to disable)
- Rename MKL_FA_DEBUG → GGML_SYCL_MKL_FA_DEBUG
- Rename MKL_FA_DIAG → GGML_SYCL_MKL_FA_DIAG
- Replace fprintf(stderr, ...) / fflush(stderr) with GGML_LOG_INFO() macro
- Document all three env vars in docs/backend/SYCL.md under Runtime
- Add comment explaining MKL FA activation trigger (flash-attn + quantized
KV cache + batch-size >= 1024 + n_kv >= 1024)
Resolves review feedback from arthw.
Again, thank you!!!
Co-Authored-By: Claude Code on DeepSeek-v4-Pro
* Thank you for the review feedback round 2: use ggml_sycl_get_env, remove dup waits, gate perf macros
- Replace raw getenv() with ggml_sycl_get_env() in all 4 env-var checks
(fattn.cpp: GGML_SYCL_ENABLE_MKL_FA, GGML_SYCL_MKL_FA_DEBUG,
GGML_SYCL_MKL_FA_DIAG; fattn-mkl.cpp: GGML_SYCL_MKL_FA_DEBUG)
- Remove duplicated stream->wait() before ev.wait_and_throw() in GEMM
KQ and GEMM VKQ — ev.wait_and_throw() already waits for completion
- Gate MKL_ACCUM macro behind do_print so timing accumulators are
no-ops in normal operation
- Remove redundant MIT/Intel copyright header from fattn-mkl.cpp
- Remove unused #include <cfloat>
- Expand SYCL.md MKL FA docs with step-by-step activation trigger
and example llama-cli command
Again, thank you!!!
Co-Authored-By: Claude Code on DeepSeek-v4-Pro
* fattn-mkl: enable MKL FA for all KV cache types
Remove the quantized-only restriction on MKL activation — the MKL
kernel converts any non-F16 K/V to F16 via to_fp16_sycl before GEMM,
so F16 (default), BF16, and F32 caches all benefit from XMX hardware
acceleration. The type restriction was an unnecessary gate.
Before (F16/BF16 default cache + FA on at 32K prefill): ~356 t/s (TILE path)
After: ~670 t/s (MKL path, matching quantized-cache baseline)
Minimal change: two conditions removed, one comment updated in fattn.cpp.
No kernel or conversion code changes — the dequant pipeline already
covers all types.
* fattn-mkl: rename mkl_disable -> mkl_enable for clarity
* fattn-mkl: refine MKL FA dispatch gates
Three changes:
1. Remove quantized-only restriction - MKL FA activates for all
KV cache types (F16 default, BF16, F32, quantized). The MKL
kernel converts non-F16 K/V via to_fp16_sycl before GEMM.
2. Rename mkl_disable -> mkl_enable to match env var
(GGML_SYCL_ENABLE_MKL_FA).
3. Replace batch-size threshold with Q->ne[1] >= 32 gate.
Keeps TG (Q=1) and MTP drafts (Q=3-8) on VEC path where
fused kernel beats MKL launch overhead. Routes all
multi-token prefill through XMX-accelerated GEMM.
Production data confirms Q patterns: 1-8 TG, 32-127 cache reuse,
128+ full reprocess. At 32K F16/BF16 FA-on: 356 -> 670 t/s.
* ggml-sycl: fix F16 cache + MKL FA multi-turn corruption; add gate guards
Two changes:
1. Always copy F16 K/V to dense row-major buffers before MKL GEMM.
Previously F16 was read in-place with raw tensor strides. During
multi-turn conversations, the accumulated KV cache had different
stride properties than a fresh prefill, producing corrupted outputs.
Now dense F16 gets a fast memcpy; interleaved (Gemma) gets a strided
copy kernel. This matches what the quantized paths already did through
to_fp16_sycl.
2. Gate MKL FA on unsupported op params (max_bias, logit_softcap, batch
dim mismatch) and pathological F16 strides (nb[1] not a multiple of
ne[0]*2). These conditions would previously crash inside the MKL
kernel. Pathological strides (test-only) and ALiBi/softcap fall
through to TILE/VEC which handle them correctly.
The stride check uses modulo rather than equality, so both dense
(nb1 == ne0*2) and interleaved (nb1 == H * ne0*2) pass — all real
models use these layouts. Only test cases with overlapping rows
(nb1=32 or nb1=75 for ne0=40) are blocked.
Thanks to hmscider for the oneDNN FA PR (#25222) which surfaced the
same insight: always normalize inputs to contiguous F16 before GEMM.
Co-Authored-By: Claude Code using DeepSeek-V4-Pro <noreply@anthropic.com>
* fattn-mkl: fix quant+GQA KV strides, tighten MKL gate, add K>=1024 tests
Adding K>=1024 flash-attn test cases surfaced several MKL bugs:
- Quant K/V with a padded seq-view (real KV cache) used the wrong
strides in the dequant path... only the true Gemma interleave
layout should reconstruct strides. nb[2] vs ne[1]*nb[1]
- Gate was firing on shapes the kernel doesn't handle: head_dim < 64
or not a multiple of 64, MHA, attention sinks, and
bf16 decode... fell through to vec which no bf16 case.
Gate MKL to the validated envelope: gqa>=2, head_dim 64 through 512
(has to be a multiple of 64) with matching K/V head size, mask,
no sinks/alibi/softcap... everything else falls back to tile.
Covers Qwen Dense/MoE and Gemma4 Dense/MoE
Ran test-backend-ops -o FLASH_ATTN_EXT: 3641/3641 pass.
Perplexity unchanged... 6.7267 MKL vs 6.7290 stock using
Qwen 27b q5_k_xl
* Update ggml/src/ggml-sycl/fattn.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* Update ggml/src/ggml-sycl/fattn.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* Update ggml/src/ggml-sycl/fattn.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* fattn-mkl: bound attention scratch so it doesn't grow with batch or context... also dropped the bf16 comment in fattn.cpp per arthw review.
* Update ggml/src/ggml-sycl/fattn-mkl.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* Update ggml/src/ggml-sycl/fattn-mkl.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* apply arthw suggestions: enum for dequant modes, macro for wg_size, env-var one-liners
---------
Co-authored-by: Claude Code using DeepSeek-V4-Pro <noreply@anthropic.com>
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* add bool cwhn = true to conv_2d test cases
* add layout check at graph building time
* extend layout checks for conv2d.cu kernel
* in CPU back-end kernel needs to be stored contiguously to prevent test failures with cwhn=1
* trim white space
* do op support check in vulkan backend
* fix CI failure and vulkan run-time assert failure by introducing new graph build-time check in ggml_backend_vk_device_supports_op
* add additional check in support_op function for Vulkan to fix run-time assert failure
* metal: fix memory leak if model is freed without any GPU operations
* metal: run dummy work only if residency sets are used
* metal: wrap function in #if defined
* metal: measure system-wide wired memory in test
* metal: always build regression test
Co-authored-by: YiChen Lv <63285796+forforever73@users.noreply.github.com>
---------
Co-authored-by: YiChen Lv <63285796+forforever73@users.noreply.github.com>
* Add overlap glu variant to support all archs, fix recurrent-state-rollback test
* format
* Fix all arch overlapped ranges
* format
* diagnose bus error on apple ci
* More testing
* more testing
* more targeted testing
* Fix bug in alignment for > 4gb buffer offsets
* Fix bug in view offsets
* Try avoiding multi_buffers
* not fixed yet, more logging :(
* Handle edge case in set_rows
* Try looking at view source
* Skip deepseek32 for now and clean up trace infrastructure
* simplify skipping
* last cleanup
* actually final cleanup
* update handling of overlap
* format
* try skipping other failing model
* ggml-cuda: add chunked SSD matmul for Mamba-2 prefill acceleration
* cuda: added SSD CICD fixes for CUDA / HIP / MUSA / MSVC.
* ggml-cuda: review comments fixed.
* ggml-cuda: Fuse M matrix materialization into pre_matmul kernel and enabled test.
* ggml-cuda: test updates and fixes
* ggml-cuda: test updates to remove hardcoding of tensor initialise data limits.
* ggml-cuda: ssd minor review comment fixed.
* ggml-cuda: ssd minor CICD fixed.
* CUDA SSD: Fixes correctness by promoting s0_stride_seq to int64_t, improves memory coalescing in ssm_ssd_prepare_dt_kernel, and boosts efficiency by merging B_weighted and C_scaled; also addresses prior review comments.
* cuda: fix sdata read-write race in prepare_dt fallback scan loop
* sycl: fix use-after-return of the SDPA scale in the oneDNN flash-attention path
The scale was uploaded with an async memcpy sourced from a stack local. On the
in-order queue that copy is ordered behind the K/V staging kernels; once n_kv is
large enough (>= ~26k observed on Arc Pro B70) the staging outlives the host
stack frame and the copy reads recycled memory, feeding the SDPA a garbage scale.
Output then collapses to a single repeated token and the KV cache is poisoned
for the rest of the session.
Short contexts win the race by accident, and test-backend-ops caps
FLASH_ATTN_EXT at kv=1024, which is why CI never caught it. The previous
device_count > 1 wait_and_throw() gate (and reverting it, PR #25741) fixes the
symptom only by keeping the frame alive across the copy at the cost of a host
sync on every FA call.
Fix: cache one device scalar per (device, value) -- the scale is constant per
model -- and upload it synchronously once. The single-device fast path (no
per-call host sync) is then safe: every device-side hazard already serializes
on the in-order queue. The multi-GPU conservative wait is kept unchanged.
Also:
- GGML_SYCL_FA_ONEDNN_MAX_KV env (0 = unlimited): optional n_kv ceiling that
routes very long sequences to the native FA kernel.
- test-backend-ops: FLASH_ATTN_EXT F16 cases up to kv=65536 (Qwen3.6-27B
geometry hsk=hsv=256 GQA 6, and hsk=128 GQA 4), closing the kv=1024 blind
spot. Note the race itself needs a live multi-op pipeline to reproduce;
single-op runs pass even on broken builds.
Verified on Arc Pro B70 (bmg_g31), Qwen3.6-27B Q4_K, -c 131072: output
byte-identical at temp 0 to the native FA path through 32k-deep prefill, with
prefill depth-flat at 820-840 t/s (vs 340-350 native at 32k depth).
Assisted-by: Claude Fable 5
* sycl: handle GGML_SYCL_FA_ONEDNN_MAX_KV like the other runtime env vars and document it
Review feedback on #25880:
- read the variable once at backend init into g_ggml_sycl_fa_onednn_max_kv via
ggml_sycl_get_env, and print it in the startup env listing (-lv 4 shows it)
- document GGML_SYCL_FA_ONEDNN and GGML_SYCL_FA_ONEDNN_MAX_KV in the SYCL.md
runtime table
Also trim the added FLASH_ATTN_EXT cases to kv={4096,16384}: the 32768/65536
shapes exceed the legacy NMSE threshold on both the oneDNN and native kernels
(long-sequence fp16 accumulation drift, present before this PR) and would fail
CI for an unrelated reason.
Assisted-by: Claude Fable 5
* sycl: clarify GGML_SYCL_FA_ONEDNN_MAX_KV default is disabled
Assisted-by: Claude Fable 5
* sycl: state default behavior of GGML_SYCL_FA_ONEDNN_MAX_KV explicitly
Assisted-by: Claude Fable 5
* Update ggml/src/ggml-sycl/fattn-onednn.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* sycl: write the SDPA scale from a kernel instead of caching it
The per-(device, value) scale cache was a function-local static
unordered_map with no synchronization, so concurrent backend instances
could access and rehash it at the same time.
Write the scalar with a single_task instead. The value is captured into
the command, so no host memory has to outlive the call -- which is what
the use-after-return fix needed in the first place. That removes the
shared container, the leaked device allocation and the string key, and
it also closes the remaining async-memcpy-from-a-stack-local on the
first flash-attention call.
Ordering does not rely on timing: the queue is created with
sycl::property::queue::in_order and the dnnl stream wraps that same
queue, so the write completes before the SDPA reads the scalar. The
multi-GPU wait_and_throw() branch is unchanged.
Also drop the <cstdlib> include, which is unused.
Assisted-by: Claude Opus 5
---------
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* add common/subproc.h|cpp
* add compile flag LLAMA_SUBPROCESS
* disabled by default on android and ios
* test-jinja: use common subproc
* mtmd: disable video if subproc is not set
* disable subproc on wasm
* make is_created atomic
* migrate server-mcp
* Add preliminary MiniMax-M3 support
Text-only port that re-uses existing components: MiniMax-M2 style GQA with
per-head QK-norm and partial rotary, DeepSeek-V3 style leading-dense and
routed/shared experts, and swigluoai activation. Sparse attention is not
yet supported (dense fallback); vision tower and MTP heads are dropped.
* MiniMax-M3 vision tower (mmproj + clip graph)
* Delete m3_vision_ref.py
* Update clip.cpp
* MSA
* Update constants.py
* Update minimax.py
* Cache creation. Working withotu flash attention
* Added flash attention for sparse layers
* Decomposed slow cpu OP into GPU + CPU ops. Massive speedup over long ctx
* Rewrote indexer op to be cuda native. Modified flash attention to match per group block picking
* Implement sparse attention calc out of stock ops.
* Fix a cache allocation and cont issue
* Fixed -fa auto crash, flagged debug spots
* Delete vocab.json
* Delete model.safetensors.index.json
* Delete generation_config.json
* Delete Minimax directory
* Handled multi stream case to fall back on Dense Attention
* Development scaffolding cleanup. No functional change to the decode or
4-way paths. Full debug harness remains at <8136a9c68ed7a5eb009aa67bba3fda8062f4648f> for reproducing the
selection-parity validation.
* Remove redundant comment from minimax-m3.cpp
* Changed 3 Gelu Ops for vision into Gelu_erf ops
* Assert that n_kv is multiple of 128
* Rename MSA index tensors to indexer convention
Note: All GGUFs generated before this change will need to be regenerated.
* Fix incorrect Assert
* Review driven changes (#3)
* Remove comment from conversion minimax.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Remove whitespaces from constants.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Tighten comment in minimax.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* inherit MiniMax-M3 from MiniMax-M2
* drop dead text_config fallbacks
* Add indexer writer methods
* Reuse LLM_FFN_SWIGLU_OAI_MOE
* Remove duplicate indexer setters, add only block_size/local_blocks, follow value naming convention
* Fix conversion error /gguf_writer.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Update gguf-py/gguf/gguf_writer.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Update gguf-py/gguf/tensor_mapping.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Update conversion/minimax.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Update conversion/minimax.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Remove whitespace in src/llama-kv-cache.cpp
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Remove Whitespace in Update src/llama-model.h
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Remove whitespace in src/llama-hparams.h
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* remove multimodal code upon maintainer request. Will be made as a separate PR
* Whitespace clean in tensor_mapping.py
* Log cache size on launch, block ctx shift, support prompt caching
Log indexer cache size on launch
Disallow ctx shift
Support prompt caching
* Update minimax-m3.cpp
* Optimize implementation, add multi stream support.
Fully rewrote minimax-m3.cpp for speed and buffer size gains:
Unified the 4-way + decode, 1 FA call per layer instead of 4, with the groups mapped onto ne[3]
Custom CPU op now emits block-level mask, expanded on GPU, which causes CPU to GPU transfer to shrinks at prefill
Decode: ~25 nodes/layer vs ~50, no per-group concats/conts
Unified selection semantics, so both regimes rank bs + local bias (position-anchored local force), which means prefill/decode can no longer disagree on selection
can_reuse on the MSA bias input. Graph reuse at decode restored (was rebuilding the full graph every token)
In-place mask adds, shrinking compute buffer ~6.8 to ~4.2 GiB at ub2048/62k
Multi-stream: MSA now runs with -np N when kv_unified=false. Decode stays batched across streams (still 1 FA call), prefill loops per stream. dense fallback only for --kv-unified + multi-seq
Measured effect on expert offload bound setup: decode 6.2(4WAY)–7.15(MSA_decode) -> 7.7~7.8 t/s, flat from 5k to 60k+. prefill around 10% faster. buffer about 20% smaller, multi-user support.
* set default cache type to F32
* Fix potential DSA double indexer cache allocation bug, only allocate in-cache k_idx for archs that opt in
* remove F16 downcasts in MSA attention, force F32 indexer score accum
* Add Minimax eos to llama vocab
* Guard edge case where idx cache can become stale after a tail trim
* Update llama-kv-cache.h
* Update llama-kv-cache.cpp
* Update llama-kv-cache.cpp
* Update llama-kv-cache.h
* Update llama-kv-cache.cpp
* Review driven changes
* style fix
* indexer hparams are required
* fix tests
* fix lint
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* common : extract trie/ac to a separate file
* common : support multiple token sequences in the reasoning budget sampler
* common/trie : return matched word index
* common/trie : rename "word" to "pattern"
* common/reasoning-budget : expose matched end sequence
* common/sampling : replay end sequence when reasoning budget is done
* cont : update to use multiple end sequences
* cont : clean up
