* mimo2 : always emit h_nextn
the other nextn-capable models set it unconditionally
Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-MOPD
* models : consolidate nextn row cropping into shared helpers
- replace the duplicated crop conditions and the per-model flags (narrow_early,
crop_before_ffn, crop_last_layer, emit_h_nextn) with two helpers on llm_graph_context:
crop_before_nextn() / crop_after_nextn()
- models that only tested embeddings_nextn_masked now share the same condition, so they
crop the last layer before the nextn capture whenever extraction is off
- t_h_nextn is now set unconditionally in mimo2, qwen4exp and deepseek4 (as in the other
nextn-capable models); host-side reads stay gated by cparams.embeddings_nextn
Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-MOPD
The speculative MTP init enables NextN extraction on the target and draft
contexts after both were created and their schedulers reserved. With
unmasked extraction the trunk graph keeps every token through the last
layer instead of cropping to the output rows, so the first decode
reallocates to that batch's shape and the next, wider batch trips
GGML_SCHED_DEBUG_REALLOC. Invalidate the reserve when the flags change so
the next compute re-reserves with the new graph shape.
Assisted-by: Claude
* ggml: refactor selective expert copying to user code
* tests: enroll two models into selective expert copy test
* tests: use deepseek2 as test model
* improve comment in ggml-backend.h
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* cont: fix whitespace
* cont : better comments
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama: re-pool each shared k-pool rep once
With shared cells every pool is re-pooled, and since the pooled keys
are always scattered, the pools a seq_cp shares between sequences
wrote the same rep row from several scatter entries, a data race on
the CPU backend. Mark each rep once: the sharing sequences read the
same row through pool_cells.
* llama: assert whole-sequence seq_cp in the hybrid idx memory
The recurrent state is always copied whole whatever the range, and a
k-pool cell shared by a partial copy could carry two pool groupings
with a single pooled row. Every caller copies whole sequences, so
reject partial ranges instead of supporting them.
* llama: drop the k-pool cache_safe mode
With whole-sequence seq_cp, sequences sharing cells share their pools
too, so the pooled row of a shared rep is valid for all of them. Mark
each rep once in every ubatch instead of re-pooling everything while
cells are shared, which removes the sharing scan and the stale-all
workarounds in seq_rm, state_read and state_drop. seq_cp now only
stales the destination.
* server: support vision input for Clef
* move input_attn_causal to private
* extend old server_batch::embd
* server_batch::token::pos to multi dim
* nits
* fix abort
* fix img tokens cap
* fix yield_to_queue mutate data
* llama : fix unexpected graph reallocation in the k-pool models
Both k-pool models built a graph shape that depends on state the
full-context reserve cannot know:
- qwen4exp branched on inp->cache_safe, which turns false as soon as
llama_memory_seq_cp shares cells (e.g. batched-bench -pps): the QSA
layers swapped scatter+gather for fill+concat and dropped the
new_pool_rep leaf, so the decode graph had 12 fewer nodes than the
reserved one
- glm5-next branched on gather = n_tokens <= 16 && n_kv > n_sel, so the
TG decode built the gather shape (7564 nodes) while the last reserve,
the PP one, had the dense shape (7762 nodes)
Either mismatch forces a decode-time re-reserve that drops the
worst-case sizing and bakes in the current state, so the next state
growth (n_pool, n_kv, n_new) needs more room at an unchanged graph size
and aborts under GGML_SCHED_DEBUG_REALLOC=1. Reproduce with, e.g.:
GGML_SCHED_DEBUG_REALLOC=1 ./bin/llama-batched-bench \
-hf ggml-org/GLM-5.3-Flash-GGUF:Q2_K -npp 2500 -ntg 32 -npl 1,2 \
-c 32768 -pps -kvu
Always scatter+gather the pooled keys, and pick gather from context
constants only: n_ubatch bounds every ubatch, top_k + kpool - 1 bounds
n_sel. Every graph of a context then shares one shape, which the
reserve covers, and the dense path measured faster than the gather path
at 2.5k and 16k context.
Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-MOPD
* llama : drop the unused k-pool cache_safe graph API
The k-pool graphs no longer branch on cache_safe, so nothing reads
get_kpool_cache_safe() or the conditional new_pool_rep any more: both
models always pass the scatter target, which set_input_kpool now
requires instead of merely preferring.
Also drop the cache_safe copy in kpool_build_sizes(), a sizes-only
helper. The layout and state flag itself stays, it still decides which
pools a layout with shared cells must re-pool.
Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-MOPD
* tests : add a shared-seq graph reserve regression test
Decode a prompt into seq 0, share its cells with seq 1 via
llama_memory_seq_cp (what llama-batched-bench does for -pps), then keep
decoding both sequences. For the k-pool models sharing clears
cache_safe, which changes the graph topology while the pools keep
growing, so a scheduler that re-reserves with the current state
instead of the worst-case one aborts under GGML_SCHED_DEBUG_REALLOC=1.
The test registration sets that flag, and the test aborts on both
k-pool models before 2220411ec1.
kimi-linear and minimax-01 are skipped: they reserve the final pp graph
with n_seqs = 1 (see [TAG_RESERVE_DIAG_DECAY] in llama-context.cpp), so
every multi-seq graph has a different layout and re-reserves by design.
Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-MOPD
* cont : add TODOs
* cont : fix comment
* cuda: match the moe weighted reduction on empty ubatches
ggml_cuda_match_moe_weighted_reduction rejected tensors with zero
rows. A ubatch without outputs shrinks the last layer to zero rows
through inp_out_ids, so graph_optimize dropped its alloc dep there and
the scheduler graph lost one node compared to the reserved one. The
scheduler then re-reserved at the size of that ubatch, and the next
ubatch with the same node count but larger tensors aborted under
GGML_SCHED_DEBUG_REALLOC=1.
The compute loop already skips empty nodes before trying any fusion,
so the guard only made the alloc deps depend on the row count.
* tests: build the rollback test only where internal symbols link
The shared-seq case calls llm_arch_from_string, which libllama does
not export through LLAMA_API, so linking test-recurrent-state-rollback
fails on Windows with shared libraries. Its build now sits in the
NOT WIN32 OR NOT BUILD_SHARED_LIBS block, next to test-llama-archs and
the test registration it already lives under.
* tests: skip archs by name in the shared-seq reserve test
The skip of kimi-linear and minimax-01 went through llm_arch_from_string,
which libllama does not export through LLAMA_API, so the test could not
link on Windows with shared libraries. It now compares the
general.architecture string directly, and the test builds on every
platform again.
---------
Co-authored-by: Pascal <admin@serveurperso.com>
* kv-cache: save exact KV rotation metadata, reject restoring mismatched rotation
* tests : move the state rotation test to test-save-load-state
the test is now part of the save/load test matrix and runs against
every model under test, like the rest of the suite
it probes the KV cache type combinations supported by the model and
treats models that do not use attention rotation as passing vacuously
Assisted-by: pi:llama.cpp/Qwen3.8-27B
* cont : skip unsupported KV caches
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* 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
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.
* qwen4exp : halve the indexer score memory
The indexer scored all heads in one product and rectified a copy of it,
so two [n_pool, n_idx_h, n_tokens] f32 tensors were live at once, the
largest buffers of the graph at long context. Each head now gets its
own product, rectified and summed in place into one [n_pool, n_tokens]
score.
* qwen4exp: let the allocator reuse the indexer score buffers
Address review from CISC: use plain ggml_add and ggml_relu in the
indexer head loop. The graph allocator already runs them in place when
their source has no other consumer, so the _inplace variants are not
needed. The compute buffer and the speed are unchanged.
* cuda: support 4 heads in the lightning indexer
Dispatch 4 heads to the vector kernel, too few for a wmma tile, and
accept them in supports_op. test-backend-ops covers 4 heads.
* metal: take the lightning indexer head count as a function constant
The kernel reads the head count from a function constant and zero fills
the last head tile, so any head count runs and 64 heads is unchanged.
* qwen4exp: compute the indexer score with the lightning indexer
Address review from am17an: the unweighted sum of the rectified head
scores scaled by 1/sqrt(head_dim) is the lightning indexer with every
head weight set to that scale, so the indexer calls
ggml_lightning_indexer on the pooled keys with an f16 pool mask. The
keys are read once for all heads and no per head score is
materialized.
* vulkan: tile the lightning indexer over keys and tokens
A workgroup scores 64 keys against 8 tokens: the keys are staged once
in shared memory, the queries one head at a time, and each invocation
owns one key for two tokens, so no dot product needs a cross invocation
reduction. The subgroup variant and the flat dispatch are gone, the grid
is keys x tokens x streams.
* vectorize vulkan loads and use fp16 dot product
---------
Co-authored-by: Ruben Ortlam <rortlam@redhat.com>
* init conversion
* convert: ok
* model loaded
* add server code
* improve conversion script
* support shared prompt prefix
* add docs, imorove UX a bit
* add vision support
* add openjev tiny model for testing
* add dev docs
* support lev & kev
* clean up
* fix lev noul
* fix py lint
* nits docs
* clarify about not supporting date_facts
* tests : simplify function signature
* llama : clamp kpool re-pool bound to existing pools
The n_tokens/kpool + n_seqs_unq bound on n_new_g overshoots when a batch
fills the whole cache: n_ctx tokens complete exactly n_ctx/kpool pools, so
the +1 pads new_pool_idxs/new_pool_rep one entry past n_pool_real. Graph
reserve only covers n_pool_real entries, so the first full-context decode
builds bigger tensors than reserved and ggml-alloc demands a graph
reallocation (abort under GGML_SCHED_DEBUG_REALLOC=1).
Clamp the bound to n_pool_real: a ubatch can never mark more pools than
the cache holds, and reserve's n_pool_max already covers that.
Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-MOPD
* cont : cap to n_pool_max
* 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>
* vocab : honor BOS/EOS settings for PLaMo-2 and PLaMo-3
The original tokenizer configs for PLaMo-2 and PLaMo-3 have
`add_bos_token: true` and `add_eos_token: false` , but
_set_vocab_plamo() did not write the BOS/EOS metadata. The
PLAMO2 tokenizer path also ignored add_bos/add_eos during
tokenization.
Write the settings from tokenizer_config.json and honor them in
the PLAMO2 tokenization path. GGUFs without these keys keep the
previous behavior.
* Update conversion/base.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
#27941 disabled -sm tensor for qwen4exp because test-llama-archs asserted on the
Meta device once the fixture carried a PLE layer:
GGML_ASSERT(ggml_backend_buffer_is_meta(tensor->buffer)) at ggml-backend-meta.cpp:476.
With host-resident embeddings the PLE gather is a CPU node and hc_init (the REPEAT
that fans the embedding out to the hc streams) was first reached through layer 0's
PLE path, after that gather. ggml_backend_sched_split_graph pass 2 expands a device
assignment upwards only until it meets a CPU node, so the REPEAT stayed on the CPU
and the later reshape of hc_init inside the meta split viewed a host-resident node.
Expanding hc_init right after it is built puts the REPEAT directly before the first
device node, where pass 2 assigns it; the embedding reshape stays in the CPU split
and is copied in as a split input, as in deepseek4.
* llama: preserve original batch order for layer inputs
Assisted-by: Codex
* tests: cover layer-input order across KV layouts
Assisted-by: Codex
* tests: exercise layer-input ordering on CUDA devices
Assisted-by: Codex
* llama: make layer input reordering compatible with tensor split
Copy each microbatch tensor from offset zero and restore original row order after synchronization. Extend the layer-input regression to cover tensor split and repeated reads and decodes.
Assisted-by: Codex
* llama: restore token order for unmasked NextN embeddings
Use the original-token mapping for unmasked NextN rows, including when
layer-input capture is disabled. Keep masked NextN rows on the logits
output mapping and preserve offset-zero tensor copies.
Extend the existing regression to cover NextN alone, combined layer
capture, and masked outputs with repeated decodes and getters.
Validation: all 256 CPU/CUDA/tensor configurations pass. Qwen3.8-27B
Q4_K_M MTP completes MT-Bench at concurrency 16 before and after.
Assisted-by: Codex
* ggml: fix WebGPU reservation and OpenVINO hidden-state capture
Reserve WebGPU vector attention scratch across batch sizes and refresh reservations when NextN capture settings change. Preserve requested OpenVINO outputs, dynamic shapes, sequence counts, and current graph bindings.
Extend existing WebGPU regression coverage and enable strict allocation checks.
Assisted-by: Codex
* llama: defer regression test and backend fixes to follow-ups
Keep this PR focused on restoring token order for layer inputs and unmasked NextN embeddings. Remove the added regression test, OpenVINO and WebGPU changes, and the separate NextN reservation change.
Assisted-by: Codex
* llama: keep n_embd declaration in its original position
Assisted-by: Codex
* llama : pass token count to layer input extraction
Assisted-by: Codex
* llama : name original batch indices batch_idxs
Assisted-by: Codex
* llama : name extracted embedding indices embd_batch_idxs
Assisted-by: Codex
* llama : tag target embedding reordering
Assisted-by: Codex
* llama : tag extraction and name the index capture flag
Assisted-by: Codex
The sparse indexer mask is built with a set_rows scatter. Padded pools,
absent sequences and missing tail cells all pointed to the same n_kv
sentinel row, and invisible pools picked by top_k to fill the selection
overlap the tail cells of the token, so several CPU threads wrote the
same element (ThreadSanitizer data race in the sanitize CI).
Allocate the slot mask for both selection paths and route every dead
slot to its own dump row n_kv + slot. Live slots address disjoint cells,
so the scatter indices of a token are unique.
* llama: llama_prefetch_rows
* llama: support row prefetch on Windows
Apply the Windows port contributed by @praneshgo unchanged.
Source: https://github.com/ggml-org/llama.cpp/pull/29599#issuecomment-5887721014
* avoid exposing llama-mmap in model code, route via llama-impl
* add windows check, only prefetch in lazy mode
* cont : clean-up
* cont : fix build
* cont : clarify padding token for gemma4
---------
Co-authored-by: Pranesh Gonegandla <pranesh.iitp@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Rebase GLM-Next support onto master, and migrate to llama-memory-hybrid-idx
* Add initial MTP support
* Merge branch optimizations. Reduce allocated compute buffer size, speed up long context decode, fla, and slight MTP improvements.
* Review driven changes, remove env vars, protect tensors
* Strip MTP for initial PR
* Clean up after mtp strip
* Clean up after mtp strip
* Update speculative.cpp
* Update llama-context.h
* Clean up after mtp strip
* Fix tokenizer ignore merges
* Improve quantization protection selection
* Refactor mhc helpers, graph base
* Lint Fixes
* Apply suggestions from code review
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Skip glm5-next in model saver, fix CRLF
* Skip glm5-next in sweep
* Remove T4 fallback
* Review cleanup
* Review suggestions
* Defer separate MTP gguf handling to MTP PR, drop filter
* Repad n_head_kv
* kpool init apply
* Order by descending score
* Drop guard
* read kpool from hparams, clarify kpool cache flags, remove kpool_build_state(nullptr)
* Add glm5-next support to model saver and add arch test fixture
* Review cleanup
* Kpool pooled caching clarify
* Add multi stream support
* Finish Rebase
* Sparse FA fir DSA prefill
* Const
* Update llama-model.cpp to fix rebase error
* gguf-py : merge tensor map entries for HC tensors
* model : use build_gdn_l2_norm in GLM5_NEXT implementation
* chore : remove trailing whitespace
* model : use new OP precision setting API in GLM5_NEXT implementation
* mtmd : use ggml_swiglu_clamp in GLM5V and apply the image token limit
The two clamps around swiglu_split are what ggml_swiglu_clamp already does,
so the clamp bounds collapse back to one value. GLM5V also never called
set_limit_image_tokens(), so --image-max-tokens had no effect.
Assisted-by: Claude Opus 5
(cherry picked from commit 46d18e12d422be4cc04a70e4a9a9e0168bb3d5b7)
* llama : keep the GLM5-Next k-pool layout across ubatches
The layout was rebuilt from a full cell scan on every ubatch. Pools are fixed
by the positions relative to the sequence's first one, so the layout now lives
on the memory and a ubatch only appends to it.
A sequence edit no longer stales every pooled key either, only the ones at or
after the edited position, which makes a tail seq_rm free. The pooling subgraph
is built unconditionally so the graph shape no longer changes every kpool
tokens, and the pool axis is folded into rows before soft_max, which otherwise
exceeds the CUDA gridDim.y limit past n_kv 262144.
Assisted-by: Claude Opus 5
(cherry picked from commit 5d1c40b93e17fddbf73b785efe43e0d02ccb3977)
* model : write the GLM5-Next recurrent rollback checkpoints
The conv state and the delta net state were only written to the live row, so a
rollback restored whatever the checkpoint rows happened to hold. Take the same
route as kimi-k3: build_recurrent_attn for the state, and write all K_rs conv
groups. That also drops a state view that assumed contiguous rows.
Enroll the arch in test-recurrent-state-rollback, which catches this under its
garbage-filled cache pass.
Assisted-by: Claude Opus 5
(cherry picked from commit 5ace37e86d5d448e83ef5dde5632c748185b18cd)
* llama: fix PR #27773 test-save-load-state restore failure
Clear the attention and indexer cache data after a failed hybrid state restore so restored NaNs cannot affect a later sequence.
Assisted-by: Codex
* llama: fix PR #27773 gpu-rocm graph reallocation
Reserve the full GLM5-Next pool capacity and dirty pool count. The gpu-rocm Test step aborts when n_new grows while the graph node count stays fixed; CUDA, Vulkan, Metal, and WebGPU checks report the same error.
Assisted-by: Codex
* llama : fix GLM5-Next k-pool layout staleness after edits and shared teardown
Two defects in the cross-ubatch k-pool layout added by the k-pool commit:
1. Wrong results. An edited sequence only rebuilt its pool layout when its cell
count changed, so if the first ubatch after an edit added back exactly as many
cells as were removed, the stale position-to-cell list survived. With a unified
cache and more than one sequence, where another sequence takes the freed cells,
the reused layout points at the wrong cells (CPU: large logit drift, CUDA: NaN).
Rebuild whenever the sequence is stale, not only on a size mismatch.
2. Slowdown. "shared" mode was assumed to end only with an edit that forces a
rebuild, but sharing also ends when the other sequence is removed. The survivor
kept shared = true, pinning cache_safe off and re-pooling every pool on every
ubatch (server trigger: n>1 completions with -kvu, via the seq_cp in
copy_state_to). In seq_rm, if the layout has shared cells, stale every sequence
so one rebuild re-derives sharing and cache_safe returns to 1.
Assisted-by: Claude Opus 5
* llama : fix build_attn_mha stream stride for non-contiguous q
build_attn_mha split the batch into streams with a stream stride of
q->nb[3]/n_stream. That only equals one stream's span, (ne[2]/n_stream)*nb[2],
when q is contiguous. GLM5-Next is nope-only, so it does not concat a rope part
and passes the permuted q_absorbed straight in, where nb[3] != ne[2]*nb[2]; the
stride was then n_head times too large and every stream s >= 1 read another
head's queries. Split-KV (-np N without --kv-unified) multi-stream prefill was
wrong for every stream past the first. Unified KV and decode were unaffected
(n_stream == 1, and decode takes the gather path). Other MLA models concat rope
so q is contiguous and the computed value is unchanged for them.
Compute the stride from the token dimension, which is identical for a
contiguous q.
Assisted-by: Claude Opus 5
* llama : re-derive GLM5-Next k-pool sharing on state_read/state_drop
The shared-cell teardown added to seq_rm (stale every sequence when the layout
has shared cells, so a survivor does not keep shared = true and pin cache_safe
off) was missing from the other paths that can free shared cells: state_read
and state_drop staled only the one sequence. Apply the same re-derivation there
and correct the comment that claimed sharing ends only via an edit or seq_rm.
Assisted-by: Claude Opus 5
* quant : drop duplicate GLM5-Next hc_ filter
The hc_ name filter was listed twice in the GLM5_NEXT protection block.
Assisted-by: Claude Opus 5
* glm5-next: scope K-pool cache access to indexed operations
* glm5-next: keep K-pool access in hybrid index memory
* glm5-next: keep mHC graph builders model-local
* glm5-next: mark only touched pools per ubatch
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Piotr Wilkin <ilintar@gmail.com>
* model : support classifier_pooling for ModernBERT rerankers
Assisted-by: Claude Opus 5.5
* model : read classifier pooling type in load_hparams
Write classifier.pooling_type from _try_set_pooling_type whenever the
config has classifier_pooling, and read it in
llama_model_base::load_hparams. ModernBERT falls back to mean when it
is unspecified.
Assisted-by: Claude Opus 5.5
* conversion : only accept cls and mean for classifier_pooling
Assisted-by: Claude Opus 5.5
* model : rename classifier_pooling_type to pooling_type_cls
Assisted-by: Claude Opus 5.5
* vocab : keep </s> NORMAL in PLaMo-2 and PLaMo-3
The PLaMo-2 and PLaMo-3 vocabularies mark </s> as NORMAL. Current
EOG token heuristic matched it by text and added its attribute
to CONTROL.
Skip this heuristic for the PLAMO2 vocab type so </s> stays NORMAL
and is not treated as EOG.
* use <|plamo:eos|> for detection
* models: pad on the left with ggml_pad_ext
The Parakeet, LFM2-Audio, Granite Speech and Gemma 4 audio encoders
build a left padding as a right pad followed by a roll, and DFlash2
concatenates a zero filled block in front of the previous tokens.
ggml_pad_ext does both in one node now that every backend supports a
left padding. The Gemma 4 audio embeddings are bit identical.
* models: skip the DFlash2 taps that only read padding
A tap at or past block_size shifts every row out of the block, so its
term is zero. The loop runs min(kernel_size, block_size) taps.
* context : do not re-reserve the scheduler when toggling causal_attn
`llama_context::set_causal_attn()` marks the scheduler to do a full re-reserve on every change of the flag. For vision inputs, this flag is flipped twice around each non-causal image chunk for Gemma models, resulting in two expensive `sched_reserve()` passes per image. This is especially slow for multi-image or video inputs.
The cost of a re-reserve scales with context and ubatch configurations, so larger settings pay more per image (see table below).
The re-reserve is unnecessary in this case because `causal_attn` only changes the values written to KQ mask, not tensor shapes or any other buffer sizes.
Note: `causal_attn` is a graph reuse key (`llm_graph_params` via `cparams`), so a new graph is built regardless of `sched_need_reserve`, so this doesn't change the graph rebuilding behaviour.
llama-server with gemma-4-26B-A4B Q4_0 + BF16 mmproj, 130-token images,
cache_prompt=false, prompt_ms median of 3 (before -> after):
| images | config | H200 before -> after | RTX 4090 before -> after |
|-|-|-|-|
| 1 | `-c 8192 -ub 512` | 134 -> 105 ms (1.27×) | 201 -> 119 ms (1.69×) |
| 24 | `-c 8192 -ub 512` | 2278 -> 1562 ms (1.46×) | 3559 -> 1748 ms (2.04×) |
| 24 | `-c 32768 -ub 2048` | 5379 -> 1584 ms (3.40×) | 13377 -> 1759 ms (7.61×) |
Generated output remains identical before and after.
* qwen4exp : make the indexer bias shape independent of causal_attn
The block/cell bias path was selected on cparams.causal_attn, so the
causal and non-causal graphs differed in tensor shapes and ops. With the
re-reserve removed (previous commit), a runtime flip resulted in
reallocating the compute buffers, which would fail under
GGML_SCHED_NO_REALLOC.
This commit selects the block path from the mask shape only, independent
of causal_attn. causal_attn is instead passed to set_input_qsa.
causal_attn is fixed per graph as it's part of the reuse key. Causal
values are unchanged. Non-causal values now follow the reference rule,
where every visible block competes on score and only unpooled cells are
always selected.
* context : state the causal_attn shape rule in the comment
* cont : add TODOs
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* server : allow splitting RANK pooling for causal LLM rerankers
Rerank models fall into two categories: bidirectional cross-encoders
(BERT, etc.) that require all tokens in a single physical batch, and
causal LLMs repurposed as rerankers (Qwen3, Qwen3-VL) that can use
chunked prefill like any other decoder.
Previously the server rejected all RANK-pooling inputs larger than
n_ubatch, and the graph builder hardcoded QWEN3/QWEN3VL arch checks to
determine last-token pooling. This broke long-document and multimodal
reranking for causal models.
Fix: expose llama_get_causal_attn(ctx) so the server can check the
effective runtime attention type (reflecting any --attention override
or set_causal_attn call). Also expose llama_model_is_causal(model)
for querying the static architectural property from GGUF metadata.
can_split() now permits chunked prefill for RANK pooling when the
context is causal. The graph builder's inline arch check is replaced
with the same cparams.causal_attn predicate, removing the duplication.
Assisted-by: Opencode/Qwen3.8-27B
* remove unused llama_model_is_causal, fix whitespace
Assisted-by: opencode
---------
Co-authored-by: timothywang21 <timothywang21@users.noreply.github.com>
* llama : add discard for deferred state writes
* llama : add tensor zeroing helper for backends without tensor memset
* llama : clear K/V data after failed sequence restore
* llama : clear recurrent state data after failed sequence restore
* llama : simplify discard and restore cleanup
* llama : report error when abnormal cell count is found in state_read_meta
* llama : clear attention state on hybrid restore failure
* tests : cover failed state restore cleanup
* llama : clear MLA state on dsa restore failure
* tests : update test for rebased test suite
* llama : clarify comment in llama_memory_recurrent::state_read
* model : fold Ling 3.0 VL into the BailingMoeV3 architecture
Assisted-by: Scout
* model : keep shared NORM rope list intact when gating bailingmoe3 on mrope sections
---------
Co-authored-by: aetherbird <aetherbird@users.noreply.github.com>
* dspark: add Gemma 4 draft support
Add GGUF conversion and runtime support for full-attention and SWA Gemma 4
DSpark drafts, including tied output weights and boolean backbone metadata.
Assisted-by: Codex
* dflash: infer Gemma draft features from metadata
* metal : gate mul_mm_id src1 rescale behind ggml_prec
Assisted-by: Claude Fable 5.1
* ggml-webgpu: reject MUL_MAT_ID when src1 precision is F32
* cuda/vulkan: reject MUL_MAT_ID in supports_op when src1 prec is F32
fix `supports_op` to return false for failing backends when the specified src1 precision is f32
Assisted-by: Claude Fable 5.1
---------
Co-authored-by: yomaytk <yoshimura.masashi.frbs@gmail.com>
* model : add DFlash layer-input taps for HunyuanVL
DFlash speculative decoding needs the target graph to expose the residual
stream entering each layer (res->t_layer_inp[il]) - the draft model reads
those tensors to build its cross-context. Qwen3 and the other DFlash-capable
targets register them, but the Hunyuan graphs do not, so serving a DFlash
draft against a HunyuanOCR target aborts during the first graph build:
GGML_ASSERT(t_layer_inp[il] != nullptr && "layer input tensor is null")
Register the tensor at the top of the layer loop, mirroring qwen3. The
layer input is the residual stream entering layer il, i.e. the output of
layer il-1, which is what the draft's target_layers metadata refers to
(the converter writes target_layer_ids+1). hunyuan-dense.cpp reuses this
graph, so it is covered as well; hunyuan-moe has a separate graph and is
untouched.
The vector is only read when a speculative implementation enables those
layer ids, so there is no behaviour change without a draft model.
Tested with tencent/HunyuanOCR 1.5 and its DFlash draft: image requests now
run, draft acceptance is ~0.5 and the OCR output is byte-identical to the
non-speculative run.
Co-authored-by: wendadawen <wendadawen@qq.com>
* convert : fix DFlash draft conversion against HunYuan targets
Converting a DFlash draft with a HunYuan target failed in two ways.
1. DFlashModel.set_vocab() reuses the target class' vocab handling by
calling it unbound with the draft instance, but HunYuanModel.set_vocab()
called self._fix_special_tokens(), a method that only exists on
HunYuanModel, so the conversion always aborted with
AttributeError: 'DFlashModel' object has no attribute '_fix_special_tokens'
Make the vocab helpers module-level functions taking the model
explicitly, so they do not depend on the instance being a HunYuanModel.
They have no other callers, so the two id lookups are folded into
_fix_special_tokens().
2. The delegated call runs with self.dir_model pointed at the target but
keeps the draft's self.hparams, so config lookups inside the target's
vocab code (the pad_token_id < 0 guard, eod_token_id) read the draft's
config instead of the target's. That aborts on targets with
pad_token_id = -1 (e.g. the HunyuanOCR v1.0 checkpoint) and otherwise
writes special token ids that disagree with the target.
Add _vocab_hparams(): it returns the target's config (with text_config
merged to the root, as TextModel does) when the model is a draft
converted with --target-model-dir, and the model's own hparams
otherwise, so a normal conversion is unaffected.
Tested: converting tencent/HunyuanOCR/dflash succeeds with both the 1.5 and
the v1.0 target; converting the base model without --target-model-dir
produces a byte-identical GGUF to before.
Co-authored-by: wendadawen <wendadawen@qq.com>
* convert : fix DFlash draft vocab against HunYuan targets
Switch hparams to the target config for the duration of the borrowed
set_vocab(), matching the existing dir_model swap, instead of teaching
HunYuanModel::set_vocab about draft models.
* convert : fix HunYuan special token ids for DFlash drafts
* convert : use load_hparams for HunYuan special token ids
Xuan-Son Nguyen
Georgi Gerganov
Piotr Wilkin (ilintar)
Aman GuptaandGeorgi Gerganov
Pascal
Evan HuusandGeorgi Gerganov
Adrien Gallouët
Johannes Gäßler
ynankaniandJohannes Gäßler
Pranesh GonegandlaandPranesh Gonegandla
Toki NasinandSigbjørn Skjæret
Kevin Hopper
Hrishith Thadicherla
bosh
Sihan YuandGeorgi Gerganov
Tim Wangandtimothywang21
Chipmunk
Daniel Kuts
Tobyandaetherbird
Max Krasnyansky
Michael de Gansandyomaytk
wendadawen