* 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>
* 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>
* 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
* moe-cache : LRFU/MQ-inspired hybrid cache
* make the clock advance by N tokens: use the real token count (ne[1] is always 1, ids have n_expert_used entries per token)
* sel is a RESHAPE node, the orig input is in src[0], so sel->src[0]->ne[1] is the count
* c for the class index, last_use for last use time
* explain the roles of these
* compute the key once
* use the key directly instead of re-reading
* show how the score is computed from the age
* remove the branch and clarify exp2 and log2 relationship
* reword the header comment and elaborate on unused slots and sorting
* not a copy, not just skip
* std::array maybe
* Revert "std::array maybe"
This reverts commit 7649b1f172db4941a854ad115bec0c7e062b65b9.
* half life it is
* add key and iteration helpers
* extract decayed_crf
* bump classes to 8, document the formula, display the cache size in token worth count
* assert the computed number of classes matches the half life
* feat : added ricursiveai ripple support
* fix : mandatiry SWA, scale through helper, comments back
* fix : metal
* model : make ricursive RoPE/NoPE explicit and validate the layer pattern
Write attention.rope_pattern and read it as optional, falling back to the
sliding window pattern so existing GGUFs keep loading. Refuse a
sliding_window_pattern that does not cover every layer, a model with no
sliding layer, and a gate or value residual present in only some layers.
The converter derives layer_types when the config omits it, ignores a stale
hybrid_layer_pattern and drops tensors of features the config disables.
Assisted-by: Claude Opus 5.5
* model : make ricursive swiglu_clamp per-layer
Store swiglu_clamp as a per-layer array, read it with get_key_or_arr so the
scalar in existing GGUFs still applies to every layer, and refuse a negative
or non-finite value. Add a test-llama-archs fixture with mixed per-layer
clamps and no rope_pattern, so the saver roundtrip checks both.
Assisted-by: Claude Opus 5.5
* model : use the ricursive per-tensor scales and RoPE parameters
Pass the attn_gate and output weight scales to build_lora_mm, take
ext_factor and attn_factor from the graph instead of hardcoding them, and
copy rope_freq_scale_train to the SWA layers instead of pinning it to 1.
Cite the reference lines for the value residual and the NoPE layers.
Assisted-by: Claude Opus 5.5
* convert : drop the ricursive chat_template.jinja override
The precedence between tokenizer_config.json and chat_template.jinja is to
be changed in SpecialVocab for all models instead. Until then a new
conversion stores the short tokenizer_config.json template; GGUFs that
already carry the tool-aware template are not affected.
Assisted-by: Claude Opus 5.5
* fix : remove NVFP4 checkpoint quantization block
* match #31733 refactor
---------
Co-authored-by: Piotr Wilkin <ilintar@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* 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.
This commit adds nvfp4 scales to qwen2's build_ffn call which are
currently passed as NULL.
The motivation for this is that a model converted to NVFP4 will
currently crash because the scales are loaded but never applied in
the FFN:
```console
llama-completion: llama.cpp/src/llama-sampler.cpp:1218:
void llama_sampler_dist_apply(llama_sampler*, llama_token_data_array*):
Assertion `found' failed.
```
The attention already applies its scales but ffn is missing them.
* llama : fix on-device state copy sizes for quantized KV caches
The on-device sequence state copies (LLAMA_STATE_SEQ_FLAGS_ON_DEVICE) computed
their element count as size / ggml_element_size(). For a quantized type such as
q8_0 or q4_0 that is the number of blocks, not elements, so only part of each
range was copied: the restored state was wrong, or the copy crashed.
Compute the count as size / ggml_type_size() * ggml_blck_size() for the save,
the restore and the copy between differently chunked ranges, and assert that
each range covers whole blocks.
* llama : drop the one-line comments
* llama : drop the redundant cast in the state copy size
* ggml: bound views by the bytes their strides reach
A view was checked against its source with the size of a contiguous
tensor of the same shape, computed before its strides were set. Views
whose rows overlap failed that check although they stay inside their
source, and views with wider strides passed it while reaching past
the end. The view functions now pass their strides to the tensor
constructor, which checks the bytes they actually reach, like
ggml_nbytes. The conv windows of delta_net are created with all their
rows at once instead of being widened after the check.
* ggml: assert that a strided view starts from contiguous elements
The view functions set nb[0] to the type size, so the source must have
contiguous elements along dim 0 for the view to read them.
* Optimizing Gemma4 graph for prompt processing
TODO:
- Currently complete microbatch needs to contain tokens which does not require the logits; need to optimize this further
- There are a few more operations that are performed after the kv cache is calculated, need to safely optimize them.
* Not filtering the rows when they are unmasked
* Cropping out the attention mask too.
* gemma-4: improved pruning, mask cropping and PLE scaling
Prune before the first KV-shared layer with the single-stream, causal, non-embedding safety gates.
Use cropped masks only during shared-layer attention.
Preserve requested layer-input embeddings in the KV prefix.
Remove the PLE contiguous materialization and fold equivalent PLE scaling into RMSNorm parameters/weights.
* spec-mtp: Enable masked h_nextn extraction only when the draft context shares the target KV cache
This lets Gemma 4 prune unused prompt rows before its shared layers so that we keep the full perf P.P gains with MTP.
Avoid extracting target hidden states for output-free prompt chunks in that shared-MTP path.
When a batch has a selected final token, extract only that row unless the whole verification batch is selected.
Keep ordinary MTP and classic speculative decoding behavior unchanged: they still use unmasked h_nextn because the draft
context must catch up on all target rows.
speculative-simple needs an update to select the final prompt row only for shared-KV MTP,
so the optimization works there without changing other speculative modes.
And it looks like the server already selects its final prompt row for first-token sampling.
* spec-mtp: make sure to not drop first draft round
* gemma4: enable pruning based on static cparams.kv_unified/n_seq_max=1
* gemma4-mtp: simplify prunning logic and avoid graph branches/re-allocs
* gemma4: update pruning gates to use only n_seq_max==1, kv_unified can now be used for multi-seq
* spec-simple: revert spec-simple changes for now
* gemma4: remove causal_attn from pruning gate (last source of reallocs)
* cont : adapt to llm_attn_mask changes
* cont : add TODO
---------
Co-authored-by: Meet Amin <meetamin378@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
build_conv_state_at makes a contiguous copy of the conv input tail
before writing it into the recurrent state, so every GDN layer issues
two device copies per generated token where one is enough. On
Qwen3.8-Flash-Next (36 GDN layers) that is 72 copies per token, now 36.
Speculative decoding adds one more per rollback slot. ggml_cpy reads
the strided view directly, as the shared build_conv_state already does,
so the extra copy is dropped.
* meta: support GLM-5.3 next sm tensor
* glm5-next : cont q_absorbed for -sm tensor with non-unified KV
* meta : keep GGML_OP_ARANGE in ggml.h order
* glm5-next : split MLA heads in pairs for -sm tensor
The MLA cache has a single KV head, so the heads on a device are its GQA
ratio. The HIP tile FA kernel for D=512 needs an even GQA ratio; with 3
devices the 16 heads of the test model were split 5/5/6 and it aborted.
* meta : keep the split for views inside one slab of the split axis
With a non-unified KV cache, build_attn_mha splits q into streams with a
view. For a permuted q, e.g. the absorbed query of glm5-next, that view
aborted with "view of permuted tensor not implemented". The split is now
kept if every other dim of the source and the view stays inside one slab
of the split axis. The per-device strides are only rescaled above it.
From tetherto/qvac-fabric-llm.cpp#353.
Co-authored-by: Pratik Narola <pratik.narola@tether.io>
* glm5-next : drop the q_absorbed cont
The meta backend can now split the per-stream view of the permuted q, so
the cont is not needed for -sm tensor with a non-unified KV cache. With
the cont, the q of flash attention was not contiguous, and when FA ran on
the CPU (WebGPU) the meta get_tensor aborted.
---------
Co-authored-by: Pratik Narola <pratik.narola@tether.io>
* 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>
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)
A cache that shares the cells of another one (f.ex. the gemma4 assistant) can build a graph before its sequences have any cells, f.ex. at the server warmup. get_n_kv_rows() then returned 0, which means that kv_rows are not used, so the graph fell back to the dense KQ mask and the scheduler re-planned for it. Later graphs with larger masks then aborted with GGML_SCHED_DEBUG_REALLOC=1.
* 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>
The duplicated output looks up token_embd in its buffer context, so
token_embd must exist first. With the old order a load where input and
output share a buffer type (CPU-only, -ngl 0) held the table twice:
gemma-4-E2B-it Q4_K_M, no mmap, CPU model buffer 3163.73 -> 2947.73 MiB.
* memory: refuse rollbacks past the snapshots of the last ubatch
The recurrent memory and the DSV4 cache accepted any rollback up to
n_rs_seq, whether or not the last ubatch wrote that snapshot. Each now
records the deepest rollback its last ubatch guarantees and seq_rm
refuses anything past it:
- recurrent: n - 1 for a ubatch of n <= n_rs_seq + 1 tokens, 0 for a
longer one, which only guarantees the final state
- DSV4: min(n, n_rs_seq), the snapshots are rebuilt from the state
before the ubatch
The dead rollback block of the rollback test becomes the assertion it
asked for, and rollback_bound checks the exact bound of each memory
after ubatches of 1, 2, n_rs_seq + 1 and n_rs_seq + 2 tokens.
* tests: address review from ggerganov
Detect architectures with llama_arch_from_string in comp_padding and
arch_reserves_single_seq, as in rollback_bound.
* 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>
* gguf-py : add Limite arch
Assisted-by: Claude Opus 5.5
Signed-off-by: Giuseppe Scrivano <giuseppe@scrivano.org>
* convert : add Limite support
Assisted-by: Claude Opus 5.5
Signed-off-by: Giuseppe Scrivano <giuseppe@scrivano.org>
* llama : add Limite support
Assisted-by: Claude Opus 5.5
Signed-off-by: Giuseppe Scrivano <giuseppe@scrivano.org>
---------
Signed-off-by: Giuseppe Scrivano <giuseppe@scrivano.org>
* llama : fix DSV4 compressor padding for simple batches
Count actual tokens per sequence when padding compressor writes. Add a graph reuse regression test for odd-sized DeepSeek V4 and V4.1 batches.
Assisted-by: Codex
* tests : cover multiple sequences in DSV4 padding regression
Exercise one and two sequences at different positions while keeping a fixed ubatch shape.
Assisted-by: Codex
* llama : remove simple batch splitting for DSV4
Use the existing sequence splitting to keep n_seq_tokens consistent with
the compressor padding calculation. Extend the graph reuse regression to
6-token batches.
Assisted-by: Codex