* dflash : move the DSpark markov head and backbone graphs to dspark.cpp
Move-only: build_dspark_markov_head, graph_dsv4, graph_dsv41 and
graph_mistral_large4 move verbatim. build_inp_target_features and
build_dspark_markov_head become static members of llama_model_dflash so
both files can use them.
* dflash : detect the head and the backbone once, split the loaders per file
load_arch_hparams sets llama_model_dflash::variant (DFlash, DFlash2,
DSpark, LiliCorr) and ::backbone (base, DSV4, DSV4.1, Mistral Large 4).
The head, backbone and base layer loaders become member helpers in
dflash.cpp, dflash2.cpp, lilicorr.cpp and dspark.cpp. build_arch_graph
dispatches on the backbone.
The DFlash2 and LiliCorr loaders create the DFlash2 conv tensors through
a helper in dflash2.cpp, right after their head tensors instead of in
the base layer loop; the other tensors keep their creation order. The
DFlash2 loader drops its copy of the conv group size check, which master
never reached.
* dflash : share the injection input, token embeddings and lm head lookups
build_inject_inp (target features + fc + encoder norm), get_tok_embd and
get_lm_head replace the copies in the base, DSV4, DSV4.1 and Mistral
Large 4 graphs. build_inp_target_features is file-static again.
* dflash : split the DFlash decoder into a graph per head
graph_base creates the shared inputs and provides build_inject,
build_inp_tokens, build_layers and build_logits. graph_dflash
(dflash.cpp), graph_dflash2 (dflash2.cpp), graph_dspark (dspark.cpp) and
graph_lilicorr (lilicorr.cpp) assemble them and add their head; the
selector, scorer and markov head builders move to the file of their
graph, and the DFlash2 local conv run by build_layers to dflash2.cpp.
build_arch_graph dispatches on the backbone, then on the head.
The op order and the tensor names are unchanged.
* speculative : split the DFlash driver into a class per head
common_speculative_impl_draft_dflash stays the DFlash implementation and
becomes the base of draft_dspark, draft_dflash2 and draft_lilicorr. Its
protected constructor only reads the draft metadata; init() then sets up
the contexts with the options of the draft (draft length, backend
sampling, device features, masked h_nextn). Each derived class overrides
draft() with the path of its head from the former draft(). DFlash2 also
overrides begin() to reseed its generator per request, and LiliCorr
overrides process() to store the context features of each injected
chunk, which its noise block carries as its state.
The factory keeps the current routing: draft-dspark picks DSpark,
draft-dflash picks DFlash2 or LiliCorr from the draft metadata and
DFlash otherwise. draft-dspark with a DFlash2 or LiliCorr draft now runs
the DSpark reader instead of a mix of both. sample_from_anchor is logged
by DSpark only.
* speculative : add the draft-dflash2 and draft-lilicorr types
draft-dflash2 and draft-lilicorr name the DFlash2 and LiliCorr drafts,
which draft-dflash keeps running too: it picks them from the draft
metadata. draft-dflash2 and draft-lilicorr must match the draft model,
draft-dspark rejects a DFlash2 or LiliCorr draft, a model that is not a
DFlash draft is rejected, and a run uses at most one DFlash family type;
these errors leave speculative decoding uninitialized.
The type detected from the draft GGUF is now the concrete one, and
draft-dflash2 and draft-lilicorr download the dflash sidecar.
* docs : describe the DFlash2 and LiliCorr drafts
Add the DFlash2 and LiliCorr sections and the new types, and explain
that draft-dflash also runs DFlash2 and LiliCorr drafts.
* speculative : drop the unused DFlash is_mrope
Nothing reads it since the migration to common_batch (#29385), which
picks the number of M-RoPE position rows from the rope type of the draft.
* doc: update spec dec
* address review comments
* spec : skip p_min for sampled MTP and DFlash2 drafts
Under rejection sampling the acceptance at a position is 1 - TV(p, q),
which the draft's top probability does not tell, so the p_min cutoff
dropped draft tokens the target would have accepted. Skip it when
MTP or DFlash2 samples the draft. draft-simple keeps it: each extra
draft token costs a full draft model pass.
* spec : replace the sampled-draft p_min skip with a warning
p_min behaves as on master again for every drafter and sampling mode.
Instead, log a warning when --spec-draft-p-min > 0 is used with
--spec-draft-sampling probabilistic, suggesting p_min 0: under
rejection sampling the draft's top probability does not tell how much
of the draft the target accepts. The cut is not skipped because with
MTP at long drafts it can still pay off.
* spec : warn about p_min only for drafters that use it
The n-gram drafters ignore p_min, and DSpark uses it as the cutoff of its confidence head, so the warning is skipped when only these are configured.
---------
Co-authored-by: Pranesh Gonegandla <pgonegandla@nvidia.com>
* llama : remove deprecated `llama_set_warmup`
Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
* llama : remove remaining deprecated API
Drop the APIs marked DEPRECATED in llama.h, the deprecated
llama_context_params::defrag_thold field and the --defrag-thold option.
Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
* fit-params : remove stale `llama-ext.h` include
Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
* llama : document `llama_model_dflash_selector_top_k`
Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
* llama : promote the NextN / speculative API from `llama-ext.h`
Move the NextN, layer input, embd row and state row APIs to `llama.h`
and rename `llama_set_embd_state_src` to `llama_set_ctx_state_src`, with
the internal `llama_context::set_state_src` following the same name.
Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
* llama : polish the NextN API promotion
Move the NextN draft model queries next to the Adapters section, rename
the `src` parameters of `llama_set_ctx_state_src` and
`llama_context::set_ctx_state_src` to `ctx_src`, and drop the now-redundant
`llama-ext.h` includes from `common/speculative.cpp` and `tools/imatrix`.
Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
* 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`
normalize_quotes_to_json copies raw control characters verbatim while
converting pythonic tool arguments to JSON, so calls like
python(code='a<LF>b') produce invalid JSON in the arguments returned to
clients; replaying that history then fails hard. Escape C0 control
characters inside quoted strings via escape_json_string_inner.
* model : allow loading external PLE files
Assisted-by: Codex
* convert : add separate PLE exports
Assisted-by: Codex
* tests : fix PLE export test lint
Assisted-by: Codex
* model : make --ple an alias of --shard, allow multiple shard locations
Reuse the multi-shard loading path for external PLE files instead of a
separate option.
* --ple is an alias of --shard: both take a comma separated list and may
be given more than once, and the shards may live anywhere
* metadata comes from the main file; numbered shards are still
auto-discovered and strictly validated
* explicit files outside the discovered set are appended as extra shards,
exempt from split.no and split.count checks
* -isnm/--ignore-shard-num-mismatch loads only the given files, in order,
with no auto-discovery and no split metadata validation
* ple_path, path_ple and llama_model_loader::add_ple_file are removed;
common_model_load() uses llama_model_load_from_splits for two or more
paths
* common/fit.{h,cpp} take the shard path vector, so --fit and the router
memory measurement load every shard
Assisted-by: GPT-6-Astra
* rm -isnm
* make sure to dedup paths
* rm test_convert_ple
* address comments
* allow ple shard to override tensor
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* 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>
* checkpoints: reduce overhead
Assisted-by: Codex
* rebase
Adapt the slot save/restore checkpoint appendix from #26004 to the
shared_ptr checkpoint list and the common_state_data buffers.
Assisted-by: Claude Opus 5.5
---------
Co-authored-by: cwriter <cwriter@localhost>
* fit : account for skipped NextN layers when setting n_gpu_layers
Without MTP, fit counts n_layer + 1 layers while the model loader counts
n_layer_all + 1 for n_gpu_layers, including the NextN layers which hold no
weights. A fit result that offloads all layers then leaves layer 0 on the CPU
("offloaded 65/66 layers" for Qwen3.8-27B on two GPUs).
Add the skipped NextN layers to n_gpu_layers and to the tensor split of the
device that holds the output layer so the loader assigns the layers that
fit measured.
Fixes#31531
* fit : pass hp_ngl_skip by reference like the other hparams
* fit : add hp_ngl_skip to common_get_device_memory_data
* common : fix -hf tag matching inside longer quant names
":F16" picked a BF16 file when the repo had both, since "BF16"
contains "F16". The same applied to ":Q4_XS" vs "IQ4_XS". The tag
now has to start at a token boundary in the path.
Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (Mac Studio, M5 Ultra, 256GB)
* common : accept underscore as -hf tag boundary
Some repos separate the quant with an underscore, e.g.
stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S ships
step3p5_flash_Q4_K_S-00001-of-00012.gguf, which :Q4_K_S no longer
matched. Add _ to the token boundary and pin the tag and default
lookup cases.
---------
Co-authored-by: Pascal <admin@serveurperso.com>
* 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>
* speculative : save and restore MTP hidden state
Assisted-by: Codex
* tests : remove MTP state test
Assisted-by: Codex
* speculative : simplify MTP checkpoint state
Save only the boundary hidden row; verification rows are rebuilt on the supported restore paths.
Assisted-by: GPT-6
---------
Co-authored-by: zouyongxing <zouyongxing@bytedance.com>
* chat : accept the think block in the qwen3-coder response format with reasoning_format none
Assisted-by: Claude Opus 5.5
* chat : define both qwen3-coder think block variants up front
* common: bound control vector layer indices by the model layer count
A direction tensor layer index is parsed from the tensor name with no
upper bound, and the buffer size and write offset are computed in int.
A large index overflows both, so the accumulation writes before the
start of the buffer.
Pass the model layer count to the loader, reject indices above it like
the existing zero index check, and compute the sizes in size_t.
* mtmd: fill M-RoPE positions by whole merge tiles
The position loops of the Qwen-VL, GLM, PaddleOCR and Qwen2.5-VL
encoders hardcode a 2x2 merge inside a stride of n_merge, so a patch
grid that is not a multiple of the merge writes past the positions
buffer. Iterate over whole merge tiles with n_merge inner steps, which
bounds the writes by the patch count and gives the same positions on
aligned grids.
PaddleOCR also requires a merge aligned image at graph build, as the
Qwen-VL family already does.
* chat : strip trailing whitespace from parsed reasoning and content
Assisted-by: Claude Fable 5.1
* chat : stop qwen3-coder reasoning and content before template separators
Assisted-by: Claude
* tests : remove the Qwen parser case from the chat session tests
Assisted-by: Claude
* common : handle partial prefix until match at end of generation
---------
Co-authored-by: Zeshawn Zahid <zzshawnn@gmail.com>
Co-authored-by: Alde Rojas <hello@alde.dev>
* 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>
* server: expose the reasoning effort levels of the chat template
The capability probe renders a user turn for each candidate level
(minimal, low, medium, high, xhigh, max) and for an unknown level. A
level is honored when it renders without raising and differently from
the unknown level, which shows the template fallback. Levels sharing a
render are aliases and the highest one is kept.
The caps serialize to JSON, and chat_template_caps in /props reports
the list as reasoning_effort_levels, from lowest to highest.
* jinja: keep the caps free of the json dependency
The caps keep reporting their flags as a map, and the effort levels
stay a plain member. common_chat_templates_get_caps() builds the JSON
the server reports from both, so jinja depends on json only for the
probe inputs.
* model : add LiquidAI/LFM3-Audio-2B and LiquidAI/LFM3-Omni-2B
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Assisted-by: Claude Opus 5.5
* convert : read the LFM3-Audio config.json as released
A transformers that knows lfm3_audio rewrites the config into another layout (no lfm/encoder keys).
Assisted-by: Claude Opus 5.5
* mtmd : disable flash attention for the LFM3-Audio depthformer
Assisted-by: Claude Opus 5.5
* tts : add a voice option for LFM3-Audio and LFM3-Omni
The model card offers a female and a male voice; --tts-voice for llama-tts, "voice" for the server /tts.
Assisted-by: Claude Opus 5.5
* clean up template handling
* convert : detect LFM3-Audio and LFM3-Omni by model_type
Assisted-by: Claude Opus 5.5
* move banned tokens --> logits bias --> mtmd_helper
* leave mtmd-cli unchanged
* no more template_asr_system
* left over
* llama : keep the LFM3-Audio encoder convs and codec unquantized
Assisted-by: Claude Opus 5.5
* common, mtmd : shorten comments
Assisted-by: Claude Opus 5.5
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* common : support deferred tools and tool_reference content parts
Assisted-by: Claude Opus 5.5
* common : credit the deferred tools co-author
Co-authored-by: Aldehir Rojas <765227+aldehir@users.noreply.github.com>
Assisted-by: Claude Opus 5.5
* tests : fix the type check of the deferred tools test
Assisted-by: Claude Opus 5.5
* server: add --models-memory-max parameter to allow dynamically unloading models when they exceed a memory size threshold
estimate with to-be-loaded model size included
use no_alloc to get memory requirements for model load
only set model memory_mb if not previously calculated
use memory margin instead of total size limit, apply to each device separately
add server memory debug logging
move llama_context_device_memory function to llama-ext.h
fix model count exceeded check
improve memory_per_device map naming
improve variable naming, fix style
also strip models memory margin from child processes
cont : clean-up
replace device memory map with buft memory map. Use llama_get_memory_breakdown
extract duplicated check into helper function
move model memory estimation to subprocess
precompute name->buft map, map GPU host types to CPU buft
cleanup unused variable
remove duplicated init calls
* adapt to upstream changes
* wip
* common_memory_reports internface
* SERVER_ROLE_MEASURE
* server-models-sched
* fix build
* server : fix memory accounting of measure children and claim spin
A measure child runs no_alloc, so the free memory it reports does not
include its estimate: subtracting self there erased the memory held by
other processes, down to zero with --fit on, and the next load could
overshoot the margin. Subtract self only for instance reports.
A queued load only checked --models-max before claiming, so when the
memory margin blocked it behind a busy model, load() threw and the
waiter retried in a tight loop. Refuse the claim while the limits are
exceeded and no idle model can be unloaded, so the waiter sleeps on
the condition variable until one goes idle.
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Pascal <admin@serveurperso.com>
* model : add MiniCPM-OCR-Jev support
The model is a Qwen3.5-VL trunk with a value head read at the last token: it
scores an OCR text against an image rendered from it, it does not generate.
Add the openbmb-minicpm-ocr-jev decision type, served by POST /v1/systemone.
The class scores are the pooled output of the sequence, read with last pooling,
so the conversion turns score.weight into cls.output and sets
embedding_length_out. The checkpoint has no LM head, so text generation is
rejected instead of using the output weights duplicated from the token
embeddings.
The task is fixed, so its prompt is entirely in the systemone template, the
same way h2o-lightning keeps its default there. The processor of the checkpoint
allows up to 16384 merged tokens per image, more than the Qwen-VL default of
mtmd, so the mmproj keeps the image budget of the preprocessor config. The
checkpoint declares MiniCPMOCRForSequenceClassification in its config.json, as
other models do.
Signed-off-by: tc-mb <tianchi_cai@icloud.com>
* fix code
Signed-off-by: tc-mb <tianchi_cai@icloud.com>
* use gguf meta classifier_output_labels
* remove redudant code changes
---------
Signed-off-by: tc-mb <tianchi_cai@icloud.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>