Commit Graph
222 Commits
Author SHA1 Message Date
Oliver Simons 202c1a229f llama : Add FP8 KV-cache support (#28392)
* Add FP8 KV-cache support

* Add tests for scaled, FP8 KV-cache to test-backend-ops

* Fix get_*_scale mapping and dequnatize FP8 also before rope shift

* Normalizes cache-scale names independently of projection layout.

While fused KQV is currently impractival for non-shared scales, this is
an improvement for future use regardless

* Enable Gemma 4 Assistant with FP8 by storing and fetching scale ptrs

* Update comment wording

* Move E2E FP8 KV cache tests to test-llama-archs

They belong there as they are not just a test for ggml's backend ops

* Increase support surface, remove seed

* make scaled kv generic

* Fix CI failings for test-kv-cache

https://github.com/ggml-org/llama.cpp/actions/runs/38718763447/job/116367842578?pr=28392

* CI: execute kv-cache-scale tests

* exclude fp8-kv tests from OpenVINO

* Fix linting errors of test_kv_cache_scales

* Disable kv-ffload=false + multi-GPU tests for FP8

See https://github.com/ggml-org/llama.cpp/actions/runs/38952722084/job/117148353699?pr=28413

* Adjust test label for test-kv-cache.cpp

* Fix conflict-resolution mistakes in test-kv-cache.cpp

https://github.com/ggml-org/llama.cpp/actions/runs/40554674463/job/122367426461?pr=28392

* Move KV tests to test-fp8.cpp

https://github.com/ggml-org/llama.cpp/pull/27512#discussion_r4410518158

* Remove tmp file for test-kv-cache-scales.py on Windows as well

* Fix fp8 kv cache test by population kv_source_impl

* Exclude models that consume KV cache directly

* Remove obsolete comment

* Allow BF16 scales as well for KV cache

* Rename *_cache_scale to *_cache_s

https://github.com/ggml-org/llama.cpp/pull/28392#discussion_r4517194999

* Move `is_type_scaled()` to `llama-kv-cache.cpp`
2026-11-10 18:08:32 +02:00
Georgi Gerganov e24d0b4c34 Merge pull request #32015 from ggml-org/gg/fix-sampling-output-aliasing 2026-11-10 17:12:27 +02:00
Georgi Gerganov 40f4f4ac9d graph: mark every sampling chain as an output to prevent aliasing
The sampling chains are read on the host after the graph compute, and the
graph addresses are reused across builds with different selected rows: the
token-generation reserve graph has one output row per sequence, while the
actual block draft has n_max rows per sequence. The reserve allocates the
unselected chains without keeping them, so their buffers alias, and the
actual graph then reuses those addresses by node index.

set_outputs only marked the tensors stored for the selected rows, so the
unselected per-(sequence, row) chains were freed and reused. Collect every
chain tensor in t_sampling_all and mark all of them as outputs there.

Reproduces with -np 2 (non-unified KV) and 2 concurrent requests for
draft-dspark, draft-dflash and draft-eagle3; with different prompts one
sequence's acceptance collapses (eagle3 0.15 vs 0.40). After the fix:
dspark 0.71/0.78, dflash 0.56/0.44, eagle3 0.35/0.42.

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
2026-11-10 14:44:14 +02:00
Aman Gupta a8fad21e99 mtp: read the target hidden states from a device buffer (#31927) 2026-11-09 20:52:32 +08:00
Alexey Kopytko cf34d31133 moe-cache : ensure CPU part does not wait for the cached experts (#31888) 2026-11-08 22:29:11 +09:00
PascalandGeorgi Gerganov d6be065572 llama, mtmd: follow the reference precision of ggml_prec_set_acc calls (#31851)
* llama, mtmd: follow the reference precision of ggml_prec_set_acc calls

Requests that the reference implementation does not keep in F32 move
to GGML_PREC_BF16: the GLM4 family down and output projections, the
MiMoVL vision matmuls and the Qwen3-TTS speaker encoder conv1d. The
DeepSeek V4/V4.1 routers and the MiniMax-M3 indexer stay in F32 with
a link to the reference line that computes them in F32.

* ggml: handle GGML_PREC_BF16 the same way as GGML_PREC_F32 in backends

Backends that reject GGML_PREC_F32 on the src1 of MUL_MAT_ID now also
reject GGML_PREC_BF16 (CUDA, Hexagon, OpenCL, WebGPU, Vulkan), and
backends that accept GGML_PREC_F32 for flash attention now also accept
GGML_PREC_BF16 with the same F32 accumulation (CPU, SpaceMIT, ET).
Vulkan keeps the F32 transfer between devices for both. The CUDA
cuBLAS path already handles GGML_PREC_BF16 and is unchanged.

* tests: cover GGML_PREC_BF16 in the flash attention tests

The flash attention cases run with GGML_PREC_BF16 next to F32 and
DEFAULT, and var_to_str prints bf16 so these cases keep their own name.

* llama: request GGML_PREC_BF16 for the KQ multiplication

* cont : add TODO in clip.cpp

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2026-11-08 08:37:46 +01:00
e95426f169 llama_moe_cache: run missed experts on CPU, add tuning script (#30966)
* moe-cache : compute part of the misses on the CPU

--moe-cache-cpu-share F computes a share F of the MoE cache misses on the CPU
from the host experts instead of uploading them. The graph builds the expert
FFN twice: on the device from the cache slots, where the misses left to the CPU
read zero slots, and on the CPU from the host experts, where the experts
computed on the device have id -1. The two results are added before the down
bias.

- ggml-cpu: MUL_MAT_ID skips negative ids and writes zero rows
- sched: the copy callback is called for all the inputs of a split, and before
  the host sync when there are no events
- the slot lookup callback reads the inputs of the CPU part together with the
  selected experts, so the CPU part does not wait for the device, and its
  result is copied back asynchronously
- each token uses distinct zero slots, mul_mat_id does not support an expert
  selected twice by a token
- LLAMA_MOE_CACHE_TRACE writes the selected experts of each lookup, for a
  simulator of the cache

* scripts : add moe-cache-tune.py to pick the MoE cache arguments

The script wraps a llama-server command: it starts the server at two cache
sizes and three CPU shares, fits the upload and CPU speeds of the machine, and
replays the routing of the runs through a copy of the cache logic to predict
other sizes and shares. It prints the recommended --moe-cache-mib and
--moe-cache-cpu-share.

docs/moe-cache.md describes the MoE cache arguments and the script.

* add am17an as codeowner for MoE cache

* moe-cache : upload all the misses until the cache is full

With a high --moe-cache-cpu-share, few misses are uploaded, so an empty cache
fills slowly and more misses than needed go to the CPU. The CPU share now
starts once all the slots of a cache group hold an expert. moe-cache-tune.py
replays the same rule.

* add backend sync for async transfer queue

* moe-cache : let the scheduler copy the results of the CPU part

* Fix MSVC compilation

* remove sync change

* llama-graph : extract the CPU share condition into can_cpu_share

* llama-graph : make llm_graph_params::moe_cache const

The graph construction no longer updates the MoE cache directly: the CPU
parts of the layers are accumulated in the graph result and applied to
the cache in llama_context::process_ubatch, after the graph inputs are
set and before the graph is computed.

Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (Mac Studio, M5 Ultra, 256GB)

* scripts : replay the ClassLRU policy in moe-cache-tune.py

* moe-cache : pick CPU misses by CRF instead of random (#31853)

* moe-cache : remove the backend logic of the CPU part from the graph

The CPU part reads the host experts, so the scheduler places it on the
CPU without set_tensor_backend. The cache gives no CPU map for batches
that a device would offload. Graphs with a CPU part look up their own
slot map, so the graph does not register the part with the cache.

ggml-backend : the inputs of the next split that are already computed
are copied to host memory before the current split runs, so a CPU split
does not wait for the device split before it. This replaces the stash
of the cache, and the copy callback is again called for host weights
only.

* llama-graph : return the MoE cache slots and CPU ids in a struct

The scheduler prefetch no longer checks that the next split runs on
another backend: the inputs of a split are never on its own backend,
so the input loop already skips that case.

---------

Co-authored-by: Gaurav Garg <gaugarg@nvidia.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Alexey Kopytko <alexey@kopytko.com>
2026-11-08 00:20:58 +08:00
usrlocalben bcb0152876 model: add Kimi-K3.1-Pro text model (#31779)
* model: add Kimi-K3.1-Pro text model

K3 with the full-attention MLA layers replaced by DSA: on this checkpoint 24
of the 93 layers carry a lightning indexer (32 heads x 128, top_k 2048, no
rope). The KDA (linear-attention) layers and the rest of the K3 skeleton -
cross-layer residuals, latent MoE, situ, MLA output gate, full-rank KDA gate -
are unchanged, so this reuses the K3 graph and converter and adds the indexer.

Indexer:
  - the selection is an additive mask over the whole cache, not a gather of
    the top-k rows, so it reuses the existing sparse flash attention path;
    off that path the mask stays correct but still costs a kq over the cache
  - the indexer key LayerNorm takes its eps from the model's rms_norm_eps,
    because dsa_config declares no eps of its own

Graph and cache:
  - new arch with four indexer tensors per DSA layer, and a hybrid memory
    that keeps a K-only MLA cache for those layers, recurrent state for the
    KDA layers, and the indexer key cache
  - the graph node budget shares the K3 entry

Conversion:
  - KimiK31Model(KimiK3Model), picked by KimiK31ForConditionalGeneration
  - dsa_config becomes attention.indexer.{head_count,key_length,top_k,types}
    plus the indexer LayerNorm eps; the layer list is 1-indexed like
    linear_attn_config's, and the converter asserts that it covers exactly
    the full-attention layers
  - the routed experts stay MXFP4 and are repacked as in K3, not dequantized
  - the vision tower and mm_projector are skipped: text only for now
  - the arch key follows the checkpoint's own model_type ("kimi_k31") and the
    loader registers KimiK31ForConditionalGeneration; the release name carries
    a -Pro suffix that nothing in the config or the tensor names uses

Quantisation: the indexer's q_b, k and proj are kept out of the quant
overrides, since the top-k selection is sensitive to their precision.

Verified on an 8-layer slice against the HF checkpoint: 38 tensors checked
exactly, the indexer metadata against the config, and the selection against
full attention - identical PPL at top_k beyond the context length, divergent
once top_k is forced below it. Arch test fixture and the full arch suite pass.

* converter: avoid clutter in the k3 base

* narrow comment to match model

* update metal fusion expectations

* restore

* sync with PR 31733 refactoring
2026-11-06 23:42:46 +01:00
Xuan-Son Nguyen 6023692b1e llama: fix kv shifting, add tests, add per-model build_rope (#31683)
* llama: fix kv shifting and add tests

* llama: add graph build_rope (#31693)

* llama: add graph build_rope

* improve

* migrate new models

* update review skill

* fix ci

* add --kv-shift to CI
2026-11-05 21:10:36 +01:00
Xuan-Son Nguyen e0f96a1a12 llama: refactoring: move model to class member of llm_graph_context (#31733) 2026-11-05 15:41:03 +01:00
Pascal 040ce199a1 graph: store the MTP conv snapshots of every recurrent model in one set_rows (#31660)
* graph: store the conv state of every rollback arch in one set_rows

The rollback slots of the causal conv history were written with one
copy per slot in qwen4exp, glm5-next, kimi-k3, bailingmoe3, lfm and
mamba2, while delta_net already wrote them with a single set_rows.
build_conv_state_store now does this for all of them: one copy without
rollback, otherwise one set_rows of the overlapping windows into the
min(n, n_rs_seq + 1) slots a rollback can read.

The slot order {n_rs_seq, ..., 0} is computed once per graph and each
conv layer takes a view of its tail. Its size does not depend on the
ubatch, so the reserved graph covers every ubatch shape, and the graph
input that held it is gone.

A model that is not causal never rolls back, so the context clamps
n_rs_seq to 0 for it, like for an arch without rollback support.

* nits: compute the conv slot order in the graph and keep the n_head name of bailingmoe3

* graph: build the conv slot order in each conv layer

The slot order {K-1, ..., 0} of the conv state store was one tensor
shared by every conv layer of the graph, built on the device of the
first one, so each layer on another device copied the view of its
tail. Each conv layer now builds its own order next to its set_rows,
and llm_graph_input_rs no longer holds it.
2026-11-05 09:28:03 +01:00
Aman GuptaandPascal b74298c109 delta_net_base: refactor conv state (#30878)
* delta_net: access conv state via views

* fix sm tensor

* delta_net: fill the conv slot index on the host

The rollback slot index of the conv windows is a constant per graph
shape. It is now an I32 input of the recurrent memory filled on the
host, instead of an arange and an argsort computed on the device, and
the meta backend no longer needs a mirrored split for ARANGE.

---------

Co-authored-by: Pascal <admin@serveurperso.com>
2026-11-03 08:28:35 +01:00
Georgi Gerganov ca8daad199 llama : drop the redundant cnv from llm_attn_mask (#31558)
The dspark block mask is now derived at input-fill time: llm_attn_mask::block_uniform
fills every token of a sequence with the mask row of its first token, so the
graph-level reshape/repeat in dflash is gone and cnv always aliases raw.

Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (Mac Studio, M5 Ultra, 256GB)
2026-11-03 08:11:06 +02:00
Aman Gupta 2cbca389f5 dflash: copy hidden features via device copies (#31506) 2026-11-03 09:02:32 +08:00
Aman GuptaandGeorgi Gerganov 1b6935b227 llama : flash_attn_ext_rows - more efficient unified kv for multi-seq (#29510)
* llama : use llm_attn_mask and kv_rows for unified attention

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* cont : optimize `get_seq_cells`

* llama : build the kv_rows lists from per-sequence cell bitmaps

Keep a bitmap over the cells for each sequence (the transpose of `seq`), updated in seq_pos_inc/dec, and build the cell lists by walking its set bits. The lists stay in cell order, cells shared by several sequences go to every list, and nothing has to be rebuilt when cells are removed.

Also leave out of the lists the cells that the SWA hides from every token of the slice (standard and chunked SWA).

* cont : make get_kv_rows_p_min private

* cont : minor metal

* llama : use kv_rows for every ubatch of a unified cache

A ubatch made of equal-length runs with a single sequence each (f.ex. the drafts of several slots) now splits into per-sequence slices.

Any other ubatch (interleaved, uneven or coupled sequences) becomes a single slice over the cells of all its sequences instead of falling back to the dense KQ mask, so the graph keeps the topology of the reserved graph.

* llama : merge get_kv_rows_seq_ids and get_kv_rows_p_min into get_kv_rows_slices

Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (Mac Studio, M5 Ultra, 256GB)

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2026-11-02 18:13:33 +02:00
Anton IandPascal fb5997274e kv-cache : keep rotation constants on the GPU (#31073)
* kv-cache : keep rotation constants on the GPU

Assisted-by: Codex

* tests : verify resident KV rotation buffers

Assisted-by: Codex

* kv-cache : document remaining multi-device rotation work

Assisted-by: Codex

* kv-cache : support per-layer rotation buffers

Assisted-by: Codex

* kv-cache : fix hybrid SWA rotation inputs and mixed placement check

Assisted-by: Codex

* models : use per-layer rotation in mistral-large4 and dflash

* tests: drop the rotation buffer tests

---------

Co-authored-by: Pascal <admin@serveurperso.com>
2026-10-30 10:15:06 +02:00
Oliver SimonsandSigbjørn Skjæret 68d20b884c Quant: OCP FP8 E4M3 support (#27512)
* WIP OCP FP8 E4M3 support

* Fix python type-checker

* Materialize input_scales also for FP8 when converting

* Fix conversion of https://huggingface.co/RedHatAI/Qwen3.8-27B-NVFP4

1. Per row FP8 wrongly fell to NVFP4 conversion, as NVFP4 checked on
   dims alone. Also check on dtype for FP8
2. Need to reshape FP8 QKV projections scales in the same way that
   weights are reshaped

* Allow remaining FP8 tensors to be dequantized during conversion

* Allow tensor-strategy for FP8-only compressed-tensor ckpts

* Fix and test gguf-py behavior for FP8_E4M3

* Disallow QKV-fusion for FP8 temporarily

* Fix conversion of fused qkvz & FP8 for Qwen3-Next

* Avoid fusing FP8 for --fuse-gate-up-exps as we may have different ws

* Emit warnings when not fusing FP8 during conversion

* More FP8 conversion fixes

* Add supported for strided copies to/from FP8 in CPU

* remove unneeded scratch buffer from FP8 cpu cpy/dup path

* add explanatory comments and TODOs

* Add dedicated encoder and decoder tests for FP8

* Whitelist ggml_op_get_rows in metal instead of blacklisting

required to fix failing CI, and white-listing should be more
future-proof than adding FP8 to the black-list

https://github.com/ggml-org/llama.cpp/actions/runs/38433916498/job/115422880633?pr=27512

* Fix more conversion bugs

* Defer weight materialization during conversion

* Blacklist fp8 in metal backend instead of white-list

* Adjust indents

* Fix lint + execute tests

* Add fast CPU kernel for F8_E4M3

* Conversion fixes

* Fix remaining gaps

* Enable tied weights with shared scales

* Make mandatory-F32 rules apply to FP8 during conversion

* Fix python type checks

https://github.com/ggml-org/llama.cpp/actions/runs/39765670052/job/119780379595?pr=27512

* Fix kaleidai unit-test failures

https://github.com/ggml-org/llama.cpp/actions/runs/39765670125/job/119780383150

* Relax numeric precision constraints for pre-check

* Leave fast-FP8 CPU on RISC-V for follow-up

We already exclude arm from the fast-path

* Allow output scale tensors to be located in any gguf shard

* Move test_conversion_fp8.py test execution to behavior-check.yaml

* Apply batched suggestions from code review

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Move mixed compressed-tensor logic into dequant_fp8 + add unit-tests for it

* Install all requirements also on Windows for behavior-check

Fixes https://github.com/ggml-org/llama.cpp/actions/runs/40535852590/job/122300758223?pr=27512

* Unmap lazy tensors before clean-up in Windows

Should fix https://github.com/ggml-org/llama.cpp/actions/runs/40539797824/job/122314782459?pr=27512

* LF line endings

https://github.com/ggml-org/llama.cpp/actions/runs/40539798168/job/122314791118?pr=27512

* Extract FP8 tests from llama-arch to its own test-fixture

See https://github.com/ggml-org/llama.cpp/pull/27512#discussion_r4400628221

* Update gguf-py/gguf/scripts/gguf_convert_endian.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
2026-10-29 18:14:37 +01:00
Xuan-Son Nguyen 78feec062d model: mistral le chonk (#31376)
* llama: extend DSA and sdpark to new model

* fix ci
2026-10-29 15:43:33 +01:00
Tarek DakhranandXuan Son Nguyen d05706b4f3 model : add LiquidAI/d1-2B and LiquidAI/d1-omni-2B (#31275)
* model : add LiquidAI/d1-2B and LiquidAI/d1-omni-2B

Assisted-by: Claude Opus 5.5

* convert : keep the speech task templates out of the d1 GGUFs

Assisted-by: Claude Opus 5.5

* small refactoring: no_batching

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
2026-10-27 23:36:32 +01:00
Aman GuptaandGeorgi Gerganov 27b7a8a009 llama : build the k-pool sparse selection independently of the kq_mask shape (#31243)
* llama : build the k-pool sparse selection independently of the kq_mask shape

qwen4exp (QSA) and glm5-next pick cache cells by index and no longer read the kq_mask values, but they still shaped the selection like the dense kq_mask. Build the selection as [n_kv, n_tokens] over the cache cells instead, and split it into the streams of the K view right before the attention.

This lets the dense mask change its layout (e.g. per-sequence row lists) without touching the sparse layers.

* improve comment

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* llama : drop the unused build_fused_top_k_mask args, review nits

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2026-10-27 21:57:45 +08:00
Alexey Kopytko d78f40d9d3 glm5-next : QSA without kq_mask (+ DSA) (#31009)
* glm5-next : QSA without kq_mask

* DSA: dsa top-k mask change tested with layers 0,3 of DeepSeek-V3.2
2026-10-27 18:54:36 +09:00
begcdn 961c376a2d llama : enable moe amax rescaling for laguna (#31147)
* llama : enable moe amax rescaling for laguna

Assisted-by: Devin

* llama : declare laguna sm tensor support
2026-10-26 23:04:19 +01:00
Aman GuptaandDaniel Han 74bca39500 model: DSV4.1 flash (#31086)
* model : add DeepSeek-V4.1-Flash (deepseek41)

* deepseek41 : remove MXFP8, convert the Engram table to BF16

* deepseek41 : compute the compressed-KV read layers once in the cache

* deepseek41 : use get_cells() for the shared-history check, drop llama_kv_cache::seq_pos_tok_le

* deepseek41 : gather the Engram rows in-graph like qwen4exp, drop the host gather

* deepseek41 : keep the Engram hash constants in hparams.ple_*, one module per Engram layer

* deepseek41 : drop the LLAMA_DSV41_KV_MIRROR toggle

* deepseek41 : keep il last in build_hc_mixes

* deepseek41 : sort the top-k rows before the sparse gather

* tests : add deepseek41 to the Metal fusion baseline

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
2026-10-25 09:43:51 +08:00
Georgi Gerganov 3032209758 llama : deny -sm tensor for arches with moe amax rescaling (#31081)
The amax POOL_1D reduction spans the whole matmul row, so a device cannot
compute it from an axis-0 split. The Meta split-state case reported
MIRRORED, which gives each device a partial max of its own slice. Drop
that case and deny -sm tensor for the arches which build the amax graph,
see [TAG_MOE_AMAX_SCALE].

Assisted-by: dsh:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
2026-10-24 13:48:32 +03:00
Alexey Kopytko 5c31142682 qwen4exp : build the QSA selection without the kq_mask (#30965)
* qwen4exp : build the QSA selection without the kq_mask

* clean

* lang

* qwen4exp : reshape the fused QSA selection to the kq_mask streams

ggml_top_k_mask with a single keep row returns [n_kv, n_tokens], while
build_attn_qsa expects the kq_mask layout [n_kv, n_tokens/n_stream, 1, n_stream].
With several streams the shape assert failed at graph reserve.

* test the QSA optimization

* reword

* update build_fused_top_k_mask instead of handwiring

* simplify

* keep il at the end for consistency
2026-10-23 19:49:08 +02:00
Hoang Vien Duy 0682a576a7 lora : support prism.hadamard folded weights (#30910)
* lora : support prism.hadamard folded weights

* tests : remove LoRA test from test-llama-archs
2026-10-22 08:12:17 +03:00
alexey-rsiandSigbjørn Skjæret a37cc46104 models : add Flint model (FlintForCausalLM) (#30928)
* Add Flint model (FlintForCausalLM). Assisted by Codex

* cleanup

* one more

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
2026-10-21 23:26:22 +02:00
Aman GuptaandPascal d454a204ef llama: create new fused op for QSA mask (#30640)
* llama: create new fused op for QSA scoring

* enroll glm5.3-next too

* tests : initialize the remaining tensors in test_kpool_mask

The harness sentinels were left uninitialized, so leftover memory could read as NaN and fail the NaN check.

* ggml: generalize the k-pool mask op into ggml_top_k_mask

The k-pool mask op only covered the pooled QSA selection of qwen4exp
and glm5-next, while the DSA models and the DeepSeek V4 CSA build the
same top-k mask by hand with fill, set_rows and add.

ggml_top_k_mask(mask, top_k, score, blocks, extra) keeps the mask
cells of the live picks and the extra cells, and sets the rest to
-inf. Every row tensor is 4D in the layout of top_k, and the mask is
broadcast to those rows, so a selection can also be made per head.
score, blocks and extra are optional: without score every pick is
live, without blocks a pick is a cell, and extra holds the cells a
row keeps without a pick.

llm_graph_context::build_fused_top_k_mask views the row tensors in
the layout of kq_mask and registers the fused node. The k-pool
models, the DSA attention of build_attn and the DeepSeek V4 CSA mask
use it when the probe enables the op, and keep their current path on
backends without it.

---------

Co-authored-by: Pascal <admin@serveurperso.com>
2026-10-22 01:02:22 +08:00
Georgi Gerganov 8baabd53b7 llama : scale moe down projection per expert token (#30861)
* llama : scale moe down projection per expert token

Replace the global amax rescaling for the MoE down projection with a per-expert-token factor computed by a POOL_1D max reduction along the intermediate dimension. Undo the scale on the output with the same exact power-of-two factor. Handle POOL_1D in the Meta split-state logic so a split along axis 0 becomes a replicated per-device partial result, enabling tensor parallelism.

Assisted-by: pi:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

* cont : simplify comments

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* llama : refactor amax src1 scaling result

Return the rescaled activation and the optional scale tensor through a small struct instead of an output pointer.

Assisted-by: pi:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

* cont : add TODO + clean-up
2026-10-21 09:01:00 +03:00
Daniel Han 01fc3aae84 model: add Beam (Reflection) (#30670) 2026-10-17 00:07:55 +02:00
Mikhail Podvitskii 5e03a1ff45 graph : add precision handling for MELLUM architecture (#30584) 2026-10-16 11:32:26 +08:00
Xuan-Son Nguyen 2851b3bcbf mtmd: add mtmd_audio_stream + support voxtral-realtime (#30629)
* mtmd: add mtmd_audio_stream + support voxtral-realtime

* address review
2026-10-16 01:46:42 +02:00
ba3c05f71d model: add Muse Spark Open (text, vision and audio) (#30616)
* model: add Muse Spark Open (text, vision and audio)

Muse Spark Open is a latent space MoE: the residual stream stays at 8192
while the 256 experts live in 4096, with a projection down and up around
them, top-8 sigmoid routing with a correction bias. Attention interleaves
three sliding layers at a 2048 window with rope for one global layer
without rope, with a sigmoid output gate and a weightless QK-norm.
Residuals are norm preserving, their gate is folded at conversion time.

Vision reuses the Muse Glimmer tower. Audio is a causal convolutional
pyramid with no attention and no spectrogram, four stages folding 16, 6,
5 and 4 samples into the channel dim, one embedding per 80 ms of 24 kHz.

The quantizer keeps the latent projections and the attention gate at
Q8_0, a little over 1% of the weights, since the whole residual stream
passes through them and the low-bit types turned the output into a
repeating loop. The model saver writes the post norm epsilon, and the
arch is registered in test-llama-archs.

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* Muse Spark: drop the redundant forward expand

* Muse Spark: drop the redundant forward expand and tighten the comments

* Update src/llama-quant.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update src/llama-graph.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update conversion/muse_spark.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update conversion/muse_spark.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update conversion/muse_spark.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update conversion/muse_spark.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
2026-10-15 23:27:12 +02:00
Pascal 1ddc262558 graph: keep the graph non empty without embeddings (#30522)
An embedding-only graph has no logits, and build_pooling() returns
early when embeddings are off, so since the models no longer expand
t_embd the graph ends up empty and its inputs are never allocated.
Expand t_embd in that case.
2026-10-15 06:38:12 +02:00
Pascal ace20f679e graph: pad the KQ bias of the no-cache attention (#30446)
The no-cache attention pads K, V and the mask to 256, but T5 also adds
a relative position bias of [n_tokens, n_tokens], so the add with KQ
fails unless n_tokens divides the padded width. Pad the bias with
zeros up to the KV dimension of the mask, like K and V.
2026-10-13 16:34:10 +02:00
Pascal cb55ffe6e2 graph: skip the LM head for pooled embeddings (#30445)
* graph: skip the LM head for pooled embeddings

A context that extracts pooled embeddings outputs one vector per
sequence, but every model still expanded its LM head, so the graph
computed n_vocab logits per token and the context kept a host buffer
for them.

The models now only declare t_logits, and set_outputs() expands it
unless the context extracts embeddings with a pooling type other than
NONE. Since the toggle changes the graph shape for a pooled context,
set_embeddings() schedules a re-reserve in that case.

* graph: drop the redundant t_embd expand

Address review from @CISC: with embeddings on, build_pooling() already
expands its output for every pooling type, and a decoder graph without
embeddings does not read t_embd, so set_outputs() only expands
t_logits.

* graph: add build_logits() for the LM head output

Address review from @ggerganov: the LM head is expanded by a dedicated
build_logits(), called from llama_model::build_graph() like
build_pooling(), and set_outputs() only marks the outputs again.

* graph: call build_logits() before build_sampling()

Address review from @ggerganov: build_sampling() reads t_logits, so the
LM head is built first. A pooled context then skips the sampling layers
too, since it has no logits to sample from.

* nit
2026-10-13 16:16:14 +02:00
Georgi Gerganov 9880157881 graph : avoid reallocations in the "no cache" path (#30439) 2026-10-13 11:22:19 +03:00
Pascal 358a420ec4 graph: pad the KV dimension of the no-cache attention to 256 (#30416)
The no-cache path builds a [n_tokens, n_tokens] mask and passes K and V
unpadded, so FLASH_ATTN_EXT sees K->ne[1] == n_tokens. The CUDA kernels
for head size 512 require a multiple of FATTN_KQ_STRIDE, so the op falls
back to the CPU for most lengths, and the -fa auto probe (1 token)
disables flash attention for such models.

Pad the mask like llama_kv_cache::get_n_kv and pad K and V with zeros up
to the KV dimension of the mask. clef builds its own causal mask and
pads it the same way.
2026-10-12 23:00:27 +02:00
bri-prismandGeorgi Gerganov a9d27ac693 model : support for Prism Bonsai 2 27B (#29600)
* Runtime support for Prism Bonsai 2 27B

Assisted-by: Claude Code

* address prism hadamard runtime feedback

* move hadamard tensors into method, define folded weight

* load_* hadamard fixes

* cont : clean-up

* conversion cleanup

Assisted-by: Claude Code

* prism hadamard key and methods for converter

Assisted-by: Claude Code

* address review bot feedback

Assisted-by: Claude Code

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2026-10-11 14:39:16 +02:00
69f201a205 model : support MiniCPM-V 4.7 (#29416)
* mtmd : add MiniCPM-V 4.7 support

Signed-off-by: tc-mb <tianchi_cai@icloud.com>

* model : allow mrope time from an extra position slot

Signed-off-by: tc-mb <tianchi_cai@icloud.com>

* Update conversion/minicpm.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Slim down comments

Signed-off-by: tc-mb <tianchi_cai@icloud.com>

* fix for "do not hand-wrap comments"

Signed-off-by: tc-mb <tianchi_cai@icloud.com>

* fix ci

Signed-off-by: tc-mb <tianchi_cai@icloud.com>

* rm 3d repo for pr one

Signed-off-by: tc-mb <tianchi_cai@icloud.com>

* gguf: add rope.section_order metadata

* fix comments

* handle grid layout

* allow compat

---------

Signed-off-by: tc-mb <tianchi_cai@icloud.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
2026-10-10 20:41:59 +02:00
Xuan-Son Nguyen 781dbc5ac9 spec: properly handle mtmd input for mtp (#30257)
* spec: properly handle mtmd input for mtp

* nits
2026-10-10 11:22:38 +02:00
Georgi Gerganov 8e2d31e0eb graph : reorder get_rows for embeddings (#30160)
* graph : reorder get_rows for embeddings

* cont : fix gemma4 and improve input embedding construction logic

* cont : add TODO for lora

* cont : fix raw embeddings path

* gemma4 : avoid ple cast in embeddings path
2026-10-09 20:43:05 +03:00
bosh 6184e92c57 model : use exact GELU for ModernBERT encoders (#30108)
* model : use exact GELU for ModernBERT encoders

Assisted-by: Codex

* model : keep tanh GELU aliases on ggml_geglu

Assisted-by: Claude Opus 5.5

* model : map gelu_python to ggml_geglu_erf

Assisted-by: Claude Opus 5.5
2026-10-09 23:26:56 +08:00
Pascal a518119d30 llama: keep the backend sampling graph static across ubatches (#30223)
The reserve builds n_outputs_max_per_seq sampling chains per sampler,
while a decode built one per output row, so the graph changed its
topology after the reserve and GGML_SCHED_NO_REALLOC builds aborted on
the next same sized graph. Every sampler now builds
n_outputs_max_per_seq chains, the ones without a row of the ubatch on
the padding row and not selected, and graph_max_nodes counts them.
2026-10-09 17:20:25 +03:00
Aman Gupta c811cb8f0a llama: support MoE cache over multiple GPUs (#30112) 2026-10-08 15:50:49 +05:30
bosh 37ac634566 model : support classifier_activation for rerankers (#29692)
* model : support classifier_activation for rerankers

Assisted-by: Claude Opus 5.5

* model : map classifier gelu to gelu_erf and accept tanh

Assisted-by: Claude Opus 5.5

* model : default act_cls to tanh, ModernBERT falls back to gelu_erf

Assisted-by: Claude Opus 5.5
2026-10-08 09:29:20 +03:00
Aman Gupta d6cf9acb25 llama : add a GPU cache for MoE experts kept in host memory (#29887)
* llama : add a GPU cache for MoE experts kept in host memory

Assisted-by: Claude

* use llama_moe_cache_ptr
2026-10-07 21:07:45 +03:00
Xuan-Son Nguyen 0bb496dbd3 llama: support both embd + raw tokens in batch (#29622)
* llama: support both embd + raw tokens in batch

* add to test-llama-archs

* also check case llm_arch_supports_mixed_batch = false

* constant graph topology

* have dedicated input for mixed case

* rm set_tensor_backend

* is_embd --> type

* consolidate m-rope pos handling into one place

* nits
2026-10-05 01:35:49 +02:00
Pascal 436f6f89e1 graph: gather the recurrent states once so the reserve covers every split (#29856)
build_rs gathered the extra states (n_rs - n_seqs rows) with their own
get_rows. The worst-case reserve has n_rs == n_seqs, so that node was
sized at zero rows, and any ubatch whose cells are not contiguous forced
a graph reallocation at an unchanged node count, which aborts under
GGML_SCHED_NO_REALLOC.

A single get_rows now gathers the n_rs states: the ubatch states and the
extra states are views of it, and its size only depends on n_rs, which
the reserve already sets to the maximum. A custom getter (mamba ssm_scan)
gathers from the second state, so a single sequence ubatch copies no
state. The views are built once per graph in the input to keep the host
overhead of the graph unchanged.
2026-10-03 14:02:05 +02:00
ynankaniandJohannes Gäßler b56f34ab13 CUDA: Handle compute type for NVFP4 on cublass path (#29173)
* CUDA: Handle compute type for NVFP4 on cublass path

Signed-off-by: ynankani <ynankani@nvidia.com>

* Use BF16 compute type for quantized models if HW allows

Signed-off-by: ynankani <ynankani@nvidia.com>

* Set acc prec to bf16 for nvfp4 as it needs atleast bf16 range

Signed-off-by: ynankani <ynankani@nvidia.com>

* Update ggml/src/ggml-cuda/ggml-cuda.cu

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* preserve op_params for per-expert matmul

Signed-off-by: ynankani <ynankani@nvidia.com>

---------

Signed-off-by: ynankani <ynankani@nvidia.com>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2026-10-01 16:53:52 +05:30