* GLM-DSA: improve TG performance even more
* Fix crash when using MTP
* GLM-DSA: much better PP performance (CPU-only)
* GLM-DSA: allow for quantized indexer cache
* Add GLM-5.2/DeepSeek-V3.2 DSA lightning indexer (batch-local, single-seq prefill)
Implements the sparse top-k "lightning indexer" attention for LLM_ARCH_GLM_DSA
in build_deepseek2_layer_attention (ik's deepseek2 graph).
What it does (per layer, gated on model.arch==GLM_DSA && indexer_attn_q_b):
- indexer_q = indexer_attn_q_b(q_lora latent), split rope(64)/nope(64), NEOX-rope
the pe part, concat. indexer_k = indexer_attn_k(attn_norm out), LayerNorm w/ bias,
same rope/concat (single key head, MQA).
- scores = relu(indexer_k . indexer_q), scaled per-head weights (indexer_proj),
summed over heads, + base causal mask, then ggml_top_k(min(top_k, n_tokens)).
- sparse mask: ggml_fill(-inf) -> ggml_set_rows(0) at top_k positions -> + causal,
used in the soft_max_ext attention path (-mla 1 -fa 0) instead of KQ_mask.
Simplifications (intentional, proven sound):
- Batch-local: no indexer KV-cache. Indexer keys are the current batch tokens.
- Walsh-Hadamard transform omitted: orthonormal rotation, (Hq).(Hk)==q.k, no score change.
Validation (GLM-5.2-UD-IQ2_M, 3x P100, -mla 1 -fa 0):
- Compiles clean (CUDA sm_60); loads and runs.
- c512 -b512 (n_seq=1) PPL = 2.7760, byte-identical to dense baseline (indexer
disabled) = 2.7760, all 8 chunks match -> indexer is an exact no-op when
top_k>=n_tokens. Proves correctness-preservation.
- 3105-token prompt completion (top_k=2048 < 3105 -> indexer ACTIVELY masks):
prompt-eval produces coherent, accurate continuation, identical to dense for the
prompt+early-gen tokens. No NaN/crash. Confirms the masking path works in prefill.
Known limitations (documented follow-ups, NOT handled):
- Single-sequence prefill only. Multi-sequence batches (n_seq>1, e.g. perplexity
default n_batch>n_ctx) and kv_head>0 (decode) break the batch-local key->slot
mapping. n_seq>1 -> NaN (use n_batch==n_ctx). Decode (kv_head>0): each generated
token sees only itself as an indexer key, so generation degenerates into repetition
after the prompt (dense A/B stays coherent) -- this is the decode-cache stub, the
documented next step.
- Flash-attn path (-fa 1, F16 mask) still uses dense KQ_mask (soft_max path only).
- Decode indexer KV-cache + Hadamard cached-K storage not implemented.
Runtime gate: DSA_INDEXER_DISABLE=1 falls back to dense attention (for A/B).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-5.2 DSA indexer: decode-correct via persistent indexer-K cache
Make the lightning-indexer correct for DECODE (not just prefill). Previously the
indexer was batch-local, so a generated token only scored against itself and
generation degenerated. Now the indexer keys are cached across the full context.
Changes
- llama_kv_cache: add per-layer indexer-key cache `kr_l` [indexer_head_size, kv_size]
(F16, MQA single head), allocated alongside the MLA latent cache for GLM_DSA.
- build_deepseek2_dsa_indexer: write the batch's (Hadamard-rotated) indexer keys to
kr_l at kv_head, read back the full [128, n_kv] cached keys, and score the indexer
queries against ALL past keys. Returns the full descending argsort of the scores.
- Walsh-Hadamard rotation of indexer q/k (cparams.dsa_indexer_hadamard, default on;
filled in llama_set_inputs). Score-preserving; improves cached-K F16 precision.
- build_deepseek2_dsa_sparse_mask: rank-based full-coverage scatter (write a 0/-BIG
penalty into EVERY key slot keyed by rank) instead of partial set_rows into a -inf
fill — the CUDA in-place set_rows does not preserve an un-written base, which had
corrupted decode when n_kv > top_k.
- Attention-sink force-inclusion (DSA_SINK, default 1): boost the first key(s) so the
sink always survives top-k. The IQ2_M-quantized indexer under-ranks the sink, and
masking it collapsed decode; with the boost, top_k=2048 over n_kv>2048 stays coherent.
ggml backend fixes (needed by the indexer)
- CUDA argsort: report unsupported when padded ncols > 1024 (one-thread-per-column
bitonic launch limit) so the scheduler falls back to the CPU argsort. Fixes
"invalid configuration argument" for top_k over a large n_kv.
- CUDA cpy/dup: support I32 -> I32 (top_k index copies / cross-backend moves).
Validation (GLM-5.2-UD-IQ2_M, 3xP100 + --cpu-moe, -mla 1 -fa 0)
- c512 PPL = 2.0743, byte-identical to dense (all 8 chunks): no-op path exact.
- Short-context decode (300 tok): coherent, identical to dense.
- Long-context decode (2521-tok prompt, n_kv>top_k, real masking of ~474 keys,
120+ tok generated): coherent with the sink boost; dense A/B also coherent.
Gated behind arch==GLM_DSA + indexer tensors + kr_l cache; DSA_INDEXER_DISABLE=1
forces dense. Remaining: FA path still uses the dense KQ_mask; multi-sequence
(n_seq>1) batches; deepseek32 arch wiring.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-5.2 DSA indexer: wire sparse mask into the flash-attention path (-fa 1)
The DSA sparse top-k mask is now applied on the -fa 1 path (our serving config),
not just -fa 0 soft_max. c512 PPL on -fa 1 = 2.0743, byte-identical to dense
(no regression, indexer no-op exact at n_kv <= top_k). Gated arch==GLM_DSA with
DSA_INDEXER_DISABLE escape; -fa 0 path unchanged.
Long-context -fa 1 decode coherence (n_kv > top_k, mask actually biting) validation
is still running at commit time; the FA mask reuses the same full-coverage scatter
proven coherent on the -fa 0 decode path, so it should hold, but confirm before
relying on long-context -fa 1.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-5.2 DSA indexer: UPDATE 4 — MLA-FA fix merged, FA path validated, multi-seq characterized
Document the re-validation after cherry-picking the MLA-FA vec-decode fix (5f18dcc0):
- FA path is ALIVE. Long-ctx -fa 1 decode (2521-tok prompt > top_k, mask actively
biting) is now COHERENT at -mla 1 and -mla 3, vs the pre-fix degeneration into
"0.0.0.0..." repetition. Matches dense (DSA_INDEXER_DISABLE) and -fa 0 controls.
- c512 -fa 1 PPL: indexer-ON == dense == 2.0854, byte-identical all 8 chunks (exact
no-op when n_kv <= top_k; no regression). The 2.0743->2.0854 shift is the MLA-FA
fix changing V accumulation, not an indexer artifact (ON==dense proves it).
- Indexer is feature-complete + validated for single-seq prefill+decode on both
-fa 0 and -fa 1, at -mla 1 and -mla 3 (the R740 serving target).
Remaining PR gaps, characterized honestly:
- Multi-seq (n_seq>1) with active mask is BROKEN (n_seq=2 c4096 PPL 62.6 vs dense
multi-seq 2.54 and single-seq indexer 3.05). No NaN/crash anymore. Root cause:
the indexer uses a single scalar kv_head/n_kv for the whole ubatch; multi-seq
needs per-sequence cache writes + per-sequence top-k. Fix deferred (structural).
- deepseek32 arch: N/A in this fork. DSA lives entirely under LLM_ARCH_GLM_DSA;
there is no LLM_ARCH_DEEPSEEK32 enum. Documented the steps to add one if a real
deepseek32 GGUF is ever served.
Also commit DSA_REFERENCE.md (verbatim mainline deepseek32/glm-dsa source, the port
reference), trimmed of a stray agent-handoff footer.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-5.2 DSA indexer: per-sequence attention sink — fix multi-seq (n_seq>1)
UPDATE 5. The DSA lightning indexer was numerically broken for multi-sequence
batches once the top-k mask bites (n_kv > top_k): c4096 n_seq=2 PPL 62.6 vs
dense 2.54, while single-seq was fine. Root cause: the attention-sink
force-include boosted the GLOBAL key range [0, n_sink) by +1e20, which only
protects sequence 0's sink. With several sequences packed contiguously into one
ubatch (seq 0 at cells [0,n0), seq 1 at [n0,n1), ...), every non-first
sequence's sink lives at cell n0.. (not cell 0), got no boost, and was dropped
from top-k once the mask bites — collapsing that sequence (chunk[2]=61.2 while
chunk[1]=2.33).
The cache write and score/argsort were already per-sequence correct: tokens are
placed contiguously like the main K cache, and the base KQ_mask (filled from
kv_self.cells[i].has_seq_id) already drives cross-seq keys to -inf before
argsort. Only the sink was anchored at the wrong (global) cell.
Fix: replace the global arange sink boost with a per-graph input tensor
inp_dsa_sink {n_kv, n_tokens} (F32), filled on the CPU in llama_set_inputs from
kv_self.cells exactly like the KQ_mask:
inp_dsa_sink[j,i] = 1e20 iff cell[i].pos in [0,n_sink) AND
cell[i].has_seq_id(seq_of_query_j), else 0
so each query force-includes only its OWN sequence's sink. For a single
contiguous sequence from pos 0 this is exactly the old "cell index < n_sink"
set with the same magnitude, so n_seq==1 is byte-identical.
Validation (3x P100, -ngl 99 --cpu-moe -mla 3 -fa 1, wikitext-2):
- c4096 n_seq=2 indexer chunk[2]: 61.2 -> 3.07 (== single-seq 3.05).
- c2048 topk=1024 (mask bites): n_seq=4 == n_seq=1 chunk-for-chunk
(2.5005/2.6080/2.7759/3.1137 vs .../3.1138) -> multi-seq is numerically
identical to processing each sequence alone.
- c512 n_seq=1 indexer ON == dense, all 4 chunks byte-identical (no regression).
n_seq=4 at full c4096 (n_kv=16384) OOMs the P100 compute buffer (capacity, not
correctness; n_seq=4 proven correct at c2048/n_kv=8192).
GLM-5.2 DSA indexer is now sequence-correct for n_seq>=1, prefill+decode,
soft_max+FA, -mla 1/-mla 3. Fully general and PR-ready.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-5.2 DSA indexer: UPDATE 6 — serving-correctness (kr_l maintained across shift/defrag/seq-ops; per-seq sink on first-present pos)
An adversarial review found the indexer was proven on the perplexity path but
not the serving path: the persistent indexer-K cache kr_l was written/read but
never *maintained* by the KV-cache mutators, and the attention sink anchored on
absolute pos<n_sink (wrong after multi-turn seq_rm). This closes those gaps and
pins down what is actually reachable on the MLA model.
kr_l maintenance:
- build_k_shift (llama-build-context.cpp): rotate the indexer keys by the same
per-cell delta as the main K. The cached key is H*concat(RoPE(k_pe,pos),k_nope),
so un-Hadamard (H sym/orthonormal => H*H=I) -> RoPE-delta the pe sub-block ->
re-Hadamard. Exact because GLM-DSA has no rope-scaling metadata (ext_factor=0,
attn_factor=1, freq_scale=1), so NEOX RoPE is pure/composable. Params mirror the
forward indexer RoPE exactly (rope_factors=nullptr); no DEEPSEEK2 yarn-shift leak.
Non-in-place (cont->rope->concat->re-Had->cpy), no aliasing. K-shift Hadamard
input filled in llama_set_k_shift with the identical Sylvester construction.
- build_defrag: kr_l row-move mirrors the k_l move (defrag never changes pos, so
no re-RoPE). max_moves divisor 6->9 *n_layer when the indexer cache is present.
- seq_rm/seq_cp/seq_keep are metadata-only (verified) so kr_l rows stay matched to
cells; seq_add/seq_div set has_shift and route through K-shift. No seq-op change.
Per-seq sink (llama.cpp llama_set_inputs): anchor on each sequence's FIRST PRESENT
pos (min present pos over the scored n_kv span), not absolute pos<n_sink. After
multi-turn seq_rm drops a sequence's early tokens its earliest survivor has
pos>=n_sink; the absolute test would protect nothing. Fresh seq at pos 0 => min=0
=> byte-identical to the old behaviour.
Serving-shift finding (the whole point): a RoPE context-shift on this model is
REFUSED BY THE ENGINE. get_can_shift() returns false for all MLA models
(is_mla_model() includes GLM_DSA); llama_kv_cache_update returns 1 ->
"main : failed to eval". Reproduced AND isolated with a dense control
(DSA_INDEXER_DISABLE=1): dense fails identically at the same token. The failure is
pre-existing MLA engine behaviour, independent of the indexer. On the MLA path the
shift never happens, so the indexer's kr_l can never desync via K-shift; the
build_k_shift kr_l block is correct-and-dormant (documented loudly in code).
Validation (3x P100, -ngl 99 --cpu-moe -mla 3 -fa 1, GGML_CUDA_NO_PINNED=1,
numactl --interleave=all, wikitext-2):
- No regression: c512 n_seq=1 indexer ON == dense == 2.1957 +/- 0.12031,
byte-identical all 4 chunks (2.2770/2.8741/2.3956/2.1957).
- Multi-seq: c4096 n_seq=2 chunk[1]=2.33 chunk[2]=3.07 healthy (== UPDATE 5;
per-seq sink change did not regress).
- Serving shift: engine-refused for MLA, dense control fails identically.
- Independent adversarial review: GO, no correctness defect in the diff.
- Build clean (llama-cli, llama-perplexity, sm_60).
Comments updated (build_deepseek2.cpp): multi-seq+FA no longer limitations; sink
description matches per-seq min-pos anchoring; BIG=1e30 masks on both soft_max and
FA paths.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-5.2 DSA indexer: UPDATE 7 — FIX latent graph-reuse cache-fixup omission for the kr_l indexer cache
update_cache_copies() re-points the K/V cache writes to the current kv_head whenever a
compute graph is REUSED (can_reuse_graph reuses iff kv_self.n == prev->n_kv). The persistent
indexer-key cache write (kr_l) is a separate ggml_cpy whose destination view bakes kv_head at
graph-build time, and it was NEVER registered for that fixup. Under FA the cache pads to 256,
so consecutive single-token decode ubatches share the same padded n_kv and the graph IS reused;
without the fixup the kr_l write keeps landing in the first ubatch's slot and later ubatches
never populate their own recent index-key cells (those cells stay at the alloc-zeroed 0.0).
Structurally identical to the MiniMax MSA bug (fork commit 133d14c9).
Fix (mirrors the K/V cache_copies fixup, same shape as MSA 133d14c9):
- llama-context.h: new std::vector<CacheCopy> dsa_cache_copies.
- llama.cpp ctor: resize dsa_cache_copies to n_layer (null entries -> no-op when DSA off).
- build_deepseek2.cpp: register the kr_l ggml_cpy as dsa_cache_copies[il] = {kr_cpy, kr->nb[1]}.
- llama.cpp update_cache_copies(): re-point each registered cpy view_offs = kv_head*step and
patch src[1]->data/data, exactly like K/V, with the c.cpy->view_src == kv_self.kr_l[il]
(+ null/op) guard the MSA fix omitted. soft_max / non-DSA paths byte-identical.
Validation (GLM-5.2-UD-IQ2_M, 3x P100 -ngl 99 --cpu-moe -t 32, NO_PINNED, P2P-disable patch
re-applied to get a working multi-GPU baseline — see UPDATE 7.3; that patch was lost in the
upstream rebase and is required separately):
- c512 -fa1 -mla3 indexer ON: 2.1983 (== prior baseline; build healthy).
- Long-ctx FA decode, 2735-tok recall prompt, -mla3 -fa1 temp0, reuse ON (default): coherent,
correct deep-context recall ("Dr. Mariana Velasquez ... Daniel Okonkwo") on BOTH the fixed and
the unfixed binary.
- ub128 PPL -fa1 -mla3 reuse ON, unfixed: 1.7239/1.8211/2.1888/2.4517, healthy (no inflation).
Honest scope: the bug is real in code but LATENT for GLM-DSA at its configured top_k=2048
(permissive selection keeps the genuinely-attended recent blocks even when reuse leaves some
recent index-key cells stale), unlike MSA's tighter top-k where it inflated PPL ~2x. The fix is
correct and prevents the latent corruption from biting at any tighter top_k / longer ctx /
future serving config. The pre-P2P-patch "nan" seen at ub128 was P2P corruption, not this bug.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-DSA: convert sparse-attention control from env vars to CLI args (off by default)
Implements ikawrakow's direction from discussion #2040: the DSA sparse
indexer must be controllable via command-line argument (not environment
variables), and must be OFF by default for now.
Control surface, before -> after:
DSA_INDEXER_DISABLE (env, inverted: on-by-default) -> --dsa / -dsa
(cparams.dsa, default false; opt-in, dense-by-default)
DSA_TOPK_OVERRIDE (env) -> --dsa-top-k N / -dsatk N
(cparams.dsa_top_k, default -1 == model's configured indexer_top_k)
DSA_HADAMARD_DISABLE, DSA_SINK (env) -> kept as DEBUG-ONLY env
knobs (clearly commented; no CLI surface, not system on/off controls)
Plumbing mirrors existing boolean/int feature flags (-mla, -khad):
include/llama.h llama_context_params {bool dsa; int dsa_top_k;}
src/llama.cpp default_params (false / -1); cparams assignment
src/llama-cparams.h llama_cparams {bool dsa=false; int dsa_top_k=-1;}
common/common.h gpt_params {bool dsa=false; int dsa_top_k=-1;}
common/common.cpp arg parse + help text + cparams copy
src/graphs/build_deepseek2.cpp gate now checks cparams.dsa instead of
getenv; top-k override reads cparams.dsa_top_k. Stays arch-gated to
LLM_ARCH_GLM_DSA. When --dsa is off (default) the indexer function is
never called -> existing dense MLA path, byte-identical to no-feature.
Validation (GLM-5.2-UD-IQ2_M, 3x P100, -ngl 99 --cpu-moe -mla 3 -fa 1,
wikitext-2, 4 chunks @ c2560):
--dsa OFF (default, dense): PPL 2.4151 (graph nodes 4166)
--dsa ON, default top_k=2048: PPL 2.4697 (graph nodes 8846)
--dsa ON, --dsa-top-k 1024: PPL 3.5107
Off-by-default runs the dense path; ON activates the indexer (node count
jumps, PPL shifts as the top-k mask bites once n_kv > top_k). No env var
is consulted for the primary on/off or the top-k knob.
Graph-parallel (-sm graph) interaction (the item ikawrakow flagged):
Under -sm graph the MLA layers are TP-split (wo->extra) and route to
build_deepseek2_tp_attention(), which contains NO indexer code. So --dsa
is silently a NO-OP under -sm graph: it does not error or crash, it runs
dense. Empirically, --dsa --dsa-top-k 1024 under -sm graph gives
PPL 2.4308 (chunks 1.6967/1.7906/2.1664/2.4308) -- the dense baseline
(2.4151), NOT the DSA top_k=1024 numbers (3.5107). The 0.016 delta is
f16 TP-reduce numerics, not DSA. Conclusion: DSA "works under deepseek2"
only on the non-TP (layer) path; serving DSA with -sm graph would require
wiring the indexer into the TP attention path (or a dedicated DSA arch).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-DSA: warn that --dsa is inactive under -sm graph/attn (TP path runs dense MLA)
The DSA lightning indexer is built only in the layer-mode (non-TP) attention
path. Under -sm graph / -sm attn the tensor-parallel attention path has no
indexer, so --dsa would silently run dense MLA. Emit a clear one-time
LLAMA_LOG_WARN at context creation instead of degrading silently.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-DSA: drop in-tree dev reference docs from the PR branch
DSA_REFERENCE.md and the R740 progress note are development scratch, not
part of the submission. Remove them so the PR diff is code-only.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* GLM-DSA: fix CPU-only crashes in the sparse-attention path
PR #2045 adds GLM-DSA sparse attention but was validated on CUDA (--cpu-moe).
A CPU-only build (-ngl 0 --dsa) crashes in four spots where the CUDA backend
tolerates something the CPU backend does not. These make GLM-5.2 --dsa run
coherently on CPU; with --dsa off they are no-ops (DSA CPU path only).
1. set_rows into an F32 dest segfaults (ggml.c set_rows_f32):
type_traits[F32].from_float is NULL, so the DSA sparse-mask scatter calls a
NULL fn (segfault at ip=0). memcpy when the dest is F32. CUDA has a real F32
set_rows path, so this only bit the CPU build.
2. ggml_add(F32 score, F16 mask) aborts on CPU (build_deepseek2_dsa_indexer and
build_deepseek2_dsa_sparse_mask): under -fa 1 the dense KQ_mask is F16 and CPU
add only accepts F32+F16 when src0 is F16. Cast the causal mask view to F32.
CUDA's add accepts the mixed types.
3. dsa_fa_mask dim-1 concat must be F32 on CPU (build_deepseek2_dsa_fa_mask):
CPU ggml_concat only supports F16 along dim 0; do the row (dim-1) concat in
F32 then cast the result to F16. CUDA supports the F16 dim-1 concat.
4. indexer k_norm epsilon is 0 -> ggml_norm aborts (llama-hparams.cpp): the
lightning-indexer k_norm is a non-RMS LayerNorm using f_norm_eps, but the
GLM-DSA GGUF only carries the RMS eps so f_norm_eps stays 0
(GGML_ASSERT(eps > 0)). Mirror the RMS eps. CUDA's norm doesn't assert on eps=0.
Validated: GLM-5.2 UD-Q4_K_M, single-socket Xeon w7-2475X, CPU-only (-ngl 0 --dsa)
- coherent at 49K+ ctx, correct 30K needle retrieval, prefill flat with length
(~32 tok/s, the O(L) DSA signature) vs the dense build's O(L^2) decline.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* DSA: loop over attention heads + use builtin Hadamard
* DSA: ggml_blend
* DSA: remove a bunch of unnecessary ggml_cont
* DSA: fix CUDA blend - but something is still wrong
* DSA: use ggml_top_k instead of ggml_argsort when FA is ON
* CUDA: add CUB based argsort
* DSA: avoid graph leaves
* Various
* GLM-5.2 DSA: IndexShare (shared layers reuse full-layer top-k)
GLM-5.2's indexer_types marks 21 'full' layers that compute their own
lightning-indexer top-k and 57 'shared' layers that reuse the previous
full layer's top-k. This port computed an independent top-k on every
layer, which mis-selects keys on the 57 shared layers (the transformers
reference sets indexer=None on shared layers and reuses prev_topk).
Shared layers now reuse the most-recent full layer's selection. Full/
shared map derived from the config rule (full iff il<=1 or il%4==2),
which reproduces indexer_types exactly; loader can later override from
GGUF metadata. Built on #2063's tree; head-loop/ggml_hadamard/ggml_blend/
argsort/FA-mask unchanged.
4K PPL (unsloth IQ2_M, top_k 2048, CPU): DSA-on 3.1922 -> 2.7111, dense
2.6972 (~97% of the gap). top_k>=n_kv reproduces dense exactly. Single-
seq and 4x8 parallel decode coherent.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Apply suggestion from @ikawrakow
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: mgkwill <168222+mgkwill@users.noreply.github.com>
Co-authored-by: Kawrakow <iwankawrakow@gmail.com>
examples/llava has been replaced by mtmd since late 2025, and has
been out-of-build in ik_llama.cpp since examples/CMakeLists.txt removed it
in #798.
Repointed descriptions from llava to mtmd where they remained.
* Refactor speculative decoding: move logic outside of server
* remove duplicated tokens in mtp kv cache
* narrow to only discard draft cells in MTP
* revert mtp_speculative_gen_draft
* (qwen3vl) Correct calculation for injection point of deepstack image embeddings
INjection point for deepstack embeddings used Hyperparameter n_embd_inp(), which caused the hidden state to be double accounted for, causing an OOB array access. The correct accessor is n_embd()
* Fix m-rope when pipeline parallelism is enabled
* MLA tensor parallelism under -sm graph (DEEPSEEK2/GLM_DSA/MISTRAL4)
Extends -sm graph (split-mode graph) to MLA-style attention across the
DEEPSEEK2, GLM_DSA, and MISTRAL4 architectures. Previously these archs
fell back to -sm layer regardless of the user's flag.
Implementation:
- Per-rank attention build in build_deepseek2_tp_attention with
view-sliced FlashAttention, split-buffer output projection, and
ggml_reduce across devices
- wk_b / wv_b absorbed weights replicated per device via materialize()
in llm_prepare_mla (these can't live in a split buffer)
- KV cache replication path (replicated_k_l) for graph-mode TP
- distribute_mla_tensors_for_split_mode_graph routes attention/norm
tensors into ctx_split; expert tensors stay per-layer
- Implements ggml_backend_cuda_split_buffer_get_tensor for the
replicated / row-split / col-split inverse paths
- Early-reject guard in src/llama.cpp that auto-downgrades -sm graph
to -sm layer (with a warning) when incompatible loader flags are set:
-ncmoe, -cmoe, -ot, -rtr, -muge
New CLI flag:
- -gap | --graph-attn-precision <f16|f32> (default f16)
See the PR description for the full validation matrix (3 archs x 2/4/8
GPU counts), perf numbers, VRAM accounting, and known limitations.
* Some tweaks
* materialize lambda: per-head split for graph-mode tp_replicate
7dd19e19 changed wk_b/wv_b distribution from mirror to per-head split
(split_dim=2) via prepare_split_tensors. That path only fires when
wk_b/wv_b are loaded from GGUF.
Models that store only wkv_b in GGUF derive wk_b/wv_b at load via
llm_prepare_mla, going through the materialize lambda, which was
untouched and still produced mirror replicas (split_dim=-1, full n_head
per device).
build_deepseek2_tp_attention now does mul_mat(wk_b_local, q_nope_perm)
without the prior view_3d slice, so a mirror replica passes an n_head
tensor where the kernel expects n_head_local. Result: silent SIGSEGV
right after model load.
Mirror logic in materialize is replaced with the same per-head split as
prepare_split_tensors: head_offsets derived from wo split, each rank
gets a tensor with ne[2]=n_head_local, data copied from the appropriate
source byte slice. Singular `computed` tensor keeps full metadata for
tensors_by_name lookups.
Tested: 8x3090, -sm graph -mla 3 -fa on now boots cleanly and
sweep-benches without crash. Log confirms new path: "Computed
blk.X.attn_k_b.weight ... split across N devices on dim=2".
* cleanup: indent fix + remove dead view_3d slicing and debug printf
- build_deepseek2.cpp: re-indent the self_attention block in
build_deepseek2_layer_attention (lines 253-670). Block was at column 0
inside a function body; now at the expected 4/8-space indent.
- build_deepseek2.cpp: drop the commented-out view_3d slicing and debug
printfs left over after 7dd19e19's switch to direct mul_mat on
per-rank wk_b_local / wv_b_local. Update the stale 'wk_b is
replicated (split_dim=-1)' comment to match the new split_dim=2
reality.
- ggml-cuda.cu: remove the leftover debug printf in
ggml_backend_cuda_split_buffer_get_tensor.
No behavior change. Verified with a clean rebuild and DSV2.5 +
GLM-4.7-Flash sweep-bench runs.
* llm_load_tensors: gate incompatible-flag warning to MLA archs
The -ncmoe / -rtr / -muge / -ot warning under -sm graph currently fires
for all archs that support graph mode. That's an over-reach: the
incompatibility is specific to the MLA TP paths (DEEPSEEK2, GLM_DSA,
MISTRAL4) — Gemma4 graph mode existed pre-PR and works with those flags.
Gate the warning to MLA archs only.
Also refreshes two stale comments left over from the wk_b/wv_b
mirror -> per-head-split rewrite:
- src/llama.cpp llm_prepare_mla: "Replicate wk_b/wv_b ..." now reads
"Per-head split wk_b/wv_b ..." to match what the materialize lambda
actually does post-823a39e2.
- src/llama-load-tensors.cpp distribute_mla_tensors_for_split_mode_graph:
drop the wkv_b row-split mention (wkv_b is no longer created under
graph mode after 7dd19e19) and correct the wk_b/wv_b distribution
description (per-head split, not per-device replicated).
---------
Co-authored-by: Kawrakow <iwankawrakow@gmail.com>
Add `-tm` / `--threads-mtmd` to control CPU thread count used during
multimodal image/audio processing (mmproj encoding), separate from the
main LLM thread count.
This allows running the LLM on GPU with minimal CPU threads (e.g. `-t 1`)
to reduce sync overhead, while using many threads (e.g. `-tm 16`) for
CPU-bound mmproj encoding with `--no-mmproj-offload`.
Fallback chain when `-tm` is not specified:
1. `--threads-batch` (-tb) — multimodal encoding is a batch/prefill-like
operation, so it makes sense to track with batch thread count
2. `--threads` (-t) — final default
Works with both mtmd-cli and llama-server.
AI: ubergarm/Qwen3.6-27B-GGUF MTP IQ4_KS 15.113 GiB (4.752 BPW) + pi.dev
The get_batch_ubatch() function unconditionally inflated n_batch and
n_ubatch whenever --mmproj was specified, regardless of whether the
mmproj model actually ran on the GPU. This boosted batch size applies
to both the main context and the MTP draft context, since
params_base.speculative.cparams_dft is derived from
common_context_params_to_llama(params_base).
When mmproj runs on CPU (--no-mmproj-offload), this batch inflation
is unnecessary for mmproj itself (CPU compute is sized by image
dimensions independently), but it still inflates the MTP compute buffer
proportionally. For large images (e.g. --image-max-tokens 4096), the
MTP compute buffer ballooned to ~2020 MiB and triggered an OOM even
though the mmproj model was fully on CPU and should have saved VRAM.
Restrict the batch inflation to !params.mmproj.path.empty() &&
params.mmproj_use_gpu so it only triggers when mmproj actually occupies
GPU memory. When mmproj runs on CPU, the existing per-chunk decode
splitting in mtmd_helper_decode_image_chunk_impl handles large images
correctly with the default batch size.
AI: ubergarm/Qwen3.6-27B-GGUF MTP IQ4_KS 15.113 GiB (4.752 BPW) + pi.dev
* server: spec checkpoints for recurrent models
* fix: save/restore sampler state during speculative checkpoint
When speculative decoding rejects draft tokens and restores the
recurrent state checkpoint, the sampler (RNG, grammar, prev tokens)
must also be restored to maintain consistency. Without this, the
sampler state reflects the rejected draft tokens, leading to
potential divergence.
Uses common_sampler_clone() to snapshot the sampler before the
speculative batch decode, and restores it on rejection.
* server: snapshot recurrent state in tensor
* reset ngram mod state for rejected tokens
* server: refactor checkpoint state logic
* speculative: fix sampler for checkpoints
* recurrent model: implement recurrent kernel checkpoint
* recurrent model: refactor api
* spec: free rbudget before overwriting
* Autoparser - complete refactoring of parser architecture
Autoparser: add optional argument reshuffle capability
Autoparser: True streaming (#20177)
* Relax atomicity constraint for nicer, more pleasent, True Streaming parsing
* Whitespace
* Remove redundant atomics
Revert to OAI-compatible args (#20213)
* Revert to OAI-compatible args
* Apply workaround::func_args_not_string
Fix structured outputs (#20223)
* Fix structured outputs
* Update common/chat-auto-parser-generator.cpp
Co-authored-by: Aldehir Rojas <hello@alde.dev>
---------
Co-authored-by: Aldehir Rojas <hello@alde.dev>
Fix compile bug (#20203)
* Fix compile bug
* Update common/chat-auto-parser-helpers.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
# Conflicts:
# common/chat-auto-parser-helpers.cpp
common : gracefully handle incomplete output (#20191)
* common : handle incomplete UTF-8 at end of input in PEG parser
* cont : if reached end prematurely, emit needs_more_input to propagate partial output
* cont: refactor peg parse context to add lenient flag
* cont : remove partial flag, keep lenient flag
PEG parser for LFM2 (#20251)
* PEG parser for LFM2
* Simplify using python_value()
common: map developer role to system (#20215)
* Map developer role to system
* Simplify
common: consolidate PEG string parsers (#20263)
* common : consolidate PEG string parsers
* cont : fix json_string_content()
examples : fix empty items in json_schema_to_grammar.py [no ci] (#19968)
* Fix logic for retrieving schema items in `json_schema_to_grammar.py`
If `schema['items']` is `{}` and `prefixItems not in schema', as `{}` is Falsy, the original code here will raise an error.
I think if `schema['items']` is `{}`, them items should just be `{}`
* Apply suggestion from @CISC
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Add tests for arrays with empty items
Add two unit tests to `tests/test-json-schema-to-grammar.cpp` that validate handling of arrays when 'items' is an empty schema and when 'prefixItems' is present alongside an empty 'items'. Both tests expect the same generated grammar, ensuring the JSON Schema->grammar conversion treats an empty 'items' schema (and the presence of 'prefixItems') correctly and covering this edge case.
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Reduce level of content parser warning message to avoid log spam on non-debug verbosity (#20347)
do not return if template parse failed
add arg to enable parallel tool call
common : fix incorrect uses of stoul (#20313)
# Conflicts:
# common/arg.cpp
# src/llama-grammar.cpp
examples : fix empty items in json_schema_to_grammar.py [no ci] (#19968)
* Fix logic for retrieving schema items in `json_schema_to_grammar.py`
If `schema['items']` is `{}` and `prefixItems not in schema', as `{}` is Falsy, the original code here will raise an error.
I think if `schema['items']` is `{}`, them items should just be `{}`
* Apply suggestion from @CISC
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Add tests for arrays with empty items
Add two unit tests to `tests/test-json-schema-to-grammar.cpp` that validate handling of arrays when 'items' is an empty schema and when 'prefixItems' is present alongside an empty 'items'. Both tests expect the same generated grammar, ensuring the JSON Schema->grammar conversion treats an empty 'items' schema (and the presence of 'prefixItems') correctly and covering this edge case.
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Add support for MiroThinker with new jinja template
common/parser: handle reasoning budget (#20297)
* v1
* Finished!
* Handlie cli
* Reasoning sampler
* Apply suggestions from code review
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Less explosive terminology :)
* Add utf-8 case and tests
* common : migrate reasoning budget sampler to common
* cont : clean up
* cont : expose state and allow passing as initial state
* cont : remove unused imports
* cont : update state machine doc string
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Alde Rojas <hello@alde.dev>
common/parser: use nlohmann::ordered_json to preserve parameter order (#20385)
common/parser: add GigaChatV3/3.1 models support (#19931)
Co-authored-by: Mishusha <pmv26021975@gmail.com>
common/parser: gracefully handle undetected tool parser, print error message. (#20286)
fix: prevent nullptr dereference (#20552)
common : fix iterator::end() dereference (#20445)
# Conflicts:
# common/regex-partial.cpp
jinja : add capability check for object args (#20612)
common/parser: add `--skip-chat-parsing` to force a pure content parser. (#20289)
* Add `--force-pure-content` to force a pure content parser.
* Update common/arg.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
common : rework gpt-oss parser (#20393)
* common : rework gpt-oss parser
* cont : fix gpt-oss tests
* cont : add structured output test
* cont : rename final to final_msg
common : fix gpt-oss content removal (#20745)
common/parser: add proper reasoning tag prefill reading (#20424)
* Implement proper prefill extraction
* Refactor cli parameters, update docs, move reasoning budget sampler part to common/reasoning-budget.cpp
* Update tools/server/server-task.cpp
* refactor: move grammars to variant, remove grammar_external, handle exception internally
* Make code less C++y
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
chat : handle tool calls with no required args in TAG_WITH_TAGGED format (#20764)
* chat : handle tool calls with no required args in TAG_WITH_TAGGED format
* Update tests/test-chat.cpp [no ci]
Co-authored-by: Aldehir Rojas <hello@alde.dev>
---------
Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
Co-authored-by: Aldehir Rojas <hello@alde.dev>
common/parser : fix out_of_range crash in throw path (#20424 regression) (#20777)
* chat : fix out_of_range crash in throw path (#20424 regression)
#20424 introduced effective_input = generation_prompt + input, but the
throw path uses input.substr(result.end) where result.end is a position
within effective_input. Every thinking model with a non-empty
generation_prompt crashes with std::out_of_range instead of the intended
error message.
Test crashes on unpatched master, passes with fix:
cmake -B build -DLLAMA_BUILD_TESTS=ON -DLLAMA_BUILD_TOOLS=OFF
cmake --build build --target test-chat
./build/bin/test-chat
* Update test-chat.cpp
* Update test-chat.cpp
* Update test-chat.cpp
---------
Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
jinja : fix heap OOB read in value equality comparison (#20782)
Address GHSA-q9j6-4hhc-rq9p and GHSA-2q4c-9gq5-5vfp.
The three-iterator overload of std::equal in value_array_t::equivalent()
and value_object_t::equivalent() reads past the end of the shorter
container when comparing arrays or objects of different lengths.
Use the four-iterator overload (C++14) which checks both range lengths.
Found-by: Pwno
common : fix typo in debug log ('extracft' -> 'extract') (#20807)
common/parser: fix nasty bug causing subtle corruption of generation prompt (#20825)
jinja : refactor token advancement (#20864)
* refactor token advancement
* exercise sub-expressions
common/autoparser : detect reasoning markers when enable_thinking changes system prompt (#20859)
common : replace wrap_for_generation with a prefix convenience function and fix gpt-oss (#20912)
jinja: fix macro with kwargs (#20960)
* jinja: fix macro with kwargs
* Apply suggestions from code review
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* fix newline problem
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
common : inhibit lazy grammar sampler while reasoning is active (#20970)
* common : inhibit grammar while reasoning budget is active
* cont : update force_pos in accept
* cont : fix tests
* cont : tweak should apply logic
* cont : return early not using grammar sampler
* Add tests
* cont : prevent backend sampling when reasoning budget enabled
* cont : fix typo
---------
Co-authored-by: Piotr Wilkin <piotr.wilkin@syndatis.com>
# Conflicts:
# common/reasoning-budget.h
# common/sampling.cpp
# tools/cli/cli.cpp
# tools/server/server-common.cpp
# tools/server/server-task.cpp
common/parser: fix reasoning whitespace bugs + extra parser tests (#21085)
* fix whitespace reasoning issues + add reconstruction tests
* Proper fix
* fix Nemotron autoparser test expectations to include newline in marker
common : add reasoning_format = none support to gpt-oss (#21094)
common/json-schema: fix: handle non-capturing groups (?:...) in JSON schema pattern converter (#21124)
The regex-to-grammar converter in _visit_pattern() crashes with SIGSEGV
when a JSON schema "pattern" field contains a non-capturing group (?:...).
Root cause: when the parser sees '(' followed by '?', it pushes a warning
but does not advance past '?:'. The recursive transform() call then
interprets '?' as a quantifier and calls seq.back() on an empty vector,
causing undefined behavior.
This commonly occurs when serving OpenAI-compatible tool calls from
clients that include complex regex patterns in their JSON schemas (e.g.,
date validation patterns like ^(?:(?:\d\d[2468][048]|...)-02-29|...)$).
The fix:
- Skip '?:' after '(' to treat non-capturing groups as regular groups
- For unsupported syntax (?=, ?!, etc.), skip to matching ')' safely,
handling escaped characters to avoid miscounting parenthesis depth
- Adjust the ')' unbalanced-parentheses check using direct char
comparisons instead of substr
- Add test cases for non-capturing groups (C++ only, as the JS/Python
implementations do not yet support this syntax)
common/parser: fix handling of tool definition with missing properties key (#21128)
jinja : handle empty expressions correctly (#20913)
* Reject empty computed member expressions before returning slices[0] from parse_member_expression_arguments().
* Treat empty computed member expressions with Jinja2 undefined semantics
Treat empty computed member expressions like `a[]` as undefined instead of
raising a parser error, to match Jinja2 behavior.
- return a noop expression for empty computed member arguments
- return undefined when a computed member key evaluates to undefined
- add Jinja tests covering `a[]|default('fallback')` and `a[] is undefined`
* Handle undefined computed member properties
Move undefined-property handling to the common member access path, and add a test covering `a[undefined] is undefined`.
* Use default undefined value in member access
Initialize val and then return it when property is undefined.
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* empty statement parses to blank_expression instead of noop_statement
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
common : gpt-oss handle builtin and unsolicited tool calls (#21213)
fix: tool call parsing for LFM2 and LFM2.5 models (#21242)
* fix: tool call parsing for LFM2 and LFM2.5 models'
* refactor: add test / break out lfm2 and lfm2.5 parsing logic
# Conflicts:
# common/chat.cpp
Relax prefill parser to allow space. (#21240)
* Relax prefill parser to allow space.
* Move changes from prefix() to parser generation
* Only allow spaces if we're not having a pure content parser next
common : add commentary rules for gpt-oss-20b (#21286)
add reasoning budget
model, mtmd: fix gguf conversion for audio/vision mmproj (#21309)
* fix gguf conversion for audio/vision mmproj
* fix test
# Conflicts:
# convert_hf_to_gguf.py
# examples/eval-callback/eval-callback.cpp
# examples/mtmd/CMakeLists.txt
# examples/mtmd/clip-impl.h
# examples/mtmd/mtmd.cpp
# gguf-py/gguf/constants.py
# gguf-py/gguf/gguf_writer.py
# gguf-py/gguf/tensor_mapping.py
# src/CMakeLists.txt
# src/llama-arch.cpp
# src/llama-arch.h
# src/llama-model.cpp
# src/llama-model.h
# src/llama-vocab.cpp
# src/models/models.h
# tests/test-llama-archs.cpp
# tools/mtmd/clip-graph.h
# tools/mtmd/clip-model.h
# tools/mtmd/clip.cpp
# tools/mtmd/models/models.h
fix: gemma 4 template (#21326)
chat : avoid including json in chat.h (#21306)
jinja: coerce input for string-specific filters (#21370)
common : fix tool call type detection for nullable and enum schemas (#21327)
* common : fix tool call type detection for nullable and enum schemas
* common, tests : fix grammar delegation for nullable/enum schemas and add tests
Fix enum type inference to scan all enum values (not just index 0) so
schemas like {"enum": [0, "celsius"]} correctly detect string type.
Fix schema_delegates in peg-parser to handle nullable type arrays
(["string", "null"]) and typeless enum schemas in raw mode, allowing
the tagged parser to use raw text instead of JSON-formatted strings.
Add test cases for Qwen3-Coder (TAG_WITH_TAGGED format):
- nullable string ["string", "null"]
- nullable string with null first ["null", "string"]
- nullable integer ["integer", "null"]
- enum without explicit type key
common/parser: fix call ID detection (Mistral parser mostly) + atomicity for tag-json parsers (#21230)
* Fix call ID detection (Mistral parser mostly) + atomicity for tag-json parsers
* Rename
* Update common/chat-auto-parser-generator.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
common : add gemma 4 specialized parser (#21418)
* common : add gemma4 dedicated parser
* cont : add '<|tool_response>' as eog
* cont : emit JSON from Gemma4 tool call AST
* cont : more fixes
* cont : refactor convert function
* cont : refine rules and mapping
* cont : add more tests
* cont : clean up
* cont : remove autoparser gemma4 implementation
* cont : more cleanup
* cont : rename gemma4.jinja to match the others
* cont : add custom template to support interleaved thinking
* cont : preserve reasoning in model turns
* cont : fix initializer error
* cont : fix unused vars
* cont : fix accidental static
* cont : fix specialized_template signature
* fix extra semicolon
* remove debug line and extra space [no ci]
fix reasoning budget
parser: fix MiniMax handling (#21573)
jinja : support ensure_ascii=true, string repetition and int/float self-filtering (#21623)
* feat: jinja engine improvements for reka-edge
Port three Jinja engine improvements needed for the reka-edge model:
1. Python-style string repetition ("ab" * 3 → "ababab")
2. ensure_ascii=true support for tojson filter (escapes non-ASCII to \uXXXX)
3. int() builtin on value_int_t (identity, needed for Reka Edge template)
* fix: escape invalid utf8 bytes when ensure_ascii=true
The json_ensure_ascii_preserving_format function does not correctly
handle an edge case where if UTF-8 parsing fails, it adds the non-ascii
character back to the output as a raw byte.
This commit fixes that by adding the unicode standard replacement
character \\ufffd to the output instead. This is the standard behavior
for various programming languages like Python, Rust, Go, etc.
* chore: address PR comments
1. Add todo comment for supporting string repetition for array/tuples
2. Add support for float identity operation
3. Move invalid ascii test case to test_fuzzing
* chore: accept suggestion for common/jinja/value.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
common : simplify autoparser tagged parser rules (#21216)
* common : simplify autoparser tagged parser rules
* cont : remove upper limit on optional args
* cont : revert changes to parsing at the end
* cont : undo arbitrary ordering of optional args
* cont : fix uninitialized required parameters
* revert to simplify merge
* re-apply patches
* restore flexible optional arg ordering tests
common : fix ambiguous grammar rule in gemma4 (#21661)
* common : fix ambiguous grammar rule in gemma4
* cont : fix missing comma...
common : enable reasoning budget sampler for gemma4 (#21697)
* fix: enable reasoning budget sampler for gemma4
Add thinking_start_tag and thinking_end_tag to
common_chat_params_init_gemma4(). Without these, the reasoning
budget sampler never activates for gemma4.
Make the newline after "thought" optional in the PEG parser to
handle budget=0 (sampler forces end tag before the newline).
Add test case for empty thinking block.
Fixes#21487
* use p.space() instead of p.optional(p.literal("\n")) in gemma4 thought parser
common : better align to the updated official gemma4 template (#21704)
fix: Fix broken structured output when using $refs in json_schema (#21699)
chat: dedicated DeepSeek v3.2 parser + "official" template (#21785)
Hide render_message_to_json warning
common/gemma4 : handle parsing edge cases (#21760)
common: skip reasoning budget sampler when no budget is requested (#21870)
* common: skip reasoning budget sampler when no budget is requested
After I added thinking_start_tag / thinking_end_tag for gemma4 in #21697, the reasoning budget sampler gets unconditionally created even when no budget is configured (the default -1). The same applies to kimi_k2, lfm2, lfm2_5, and ministral_3 which also set these tags. The budget gets converted to INT_MAX, so the sampler never actually forces any tokens but still runs per-token checks (start tag matching in IDLE state, token-to-piece conversion + UTF-8 checks in COUNTING state).
More importantly, the mere existence of the sampler (non-null rbudget) disables backend sampling. Backend sampling lets the GPU select tokens directly, avoiding a full logits transfer from GPU to CPU every token. This could explain the 30% speed regression reported in #21784 (98 t/s to 70 t/s on Vulkan).
So I added a reasoning_budget_tokens >= 0 check to the sampler creation condition. When the budget is unlimited, the sampler is not created, backend sampling stays enabled, and no per-token overhead is added. When a budget is explicitly set (0, 128, 1024, etc.), the sampler is created and works as before.
* common: preserve rbudget when grammar is lazy
Following up on the review feedback on #21870: keep the reasoning budget sampler when grammar_lazy is true, so the thinking-block grammar suppression from #20970 still works when tools are in use. This way, we only skip the sampler when both no budget is set AND grammar is not lazy.
autoparser: support case of JSON_NATIVE with per-call markers (test case: Reka-Edge) (#21892)
* fix grammar
* fix add sampled token
---------
Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
Co-authored-by: firecoperana <firecoperana>
* wip: build spec tuner for spefic args
* wip: test different reward system
* spec-tune: fix the reward to find best params given a good TPS
* spec-tune: refactor logic for its own file
* minor clean for comments and modules
Purpose:
Add --minilog flag to llama-sweep-bench that filters log output to show only essential GPU/layer distribution information while suppressing verbose model metadata and per-layer device assignment messages.
Changes:
- Add llama_selective_log_callback with blacklist approach (sweep-bench.cpp)
Blacklisted patterns (hidden):
- Per-layer device assignments ('Setting default device in layer')
- KV metadata dump header and entries
- Tensor type counts
- Model validation messages
- EOG/special token cache info
- Metadata printout (llm_load_print_meta, print_info)
- Layer sizes table
- Tensor loading info (llm_load_tensors)
- Separator lines
- Most common cases of incomplete/continuation lines are also hidden
All other log output is shown, including:
- GPU VRAM info
- Split/buffer distribution per device
- Graph split estimates
- Final benchmark table and timings
Two changes:
1. Add four missing environment variable bindings to
gpt_params_parse_from_env():
- LLAMA_ARG_CACHE_TYPE_K (string, e.g. "q8_0")
- LLAMA_ARG_CACHE_TYPE_V (string, e.g. "q8_0")
- LLAMA_ARG_MLOCK (bool, "1"/"true")
- LLAMA_ARG_K_CACHE_HADAMARD (bool, "1"/"true")
2. Call gpt_params_parse_from_env() from gpt_params_parse() so that
ALL tools (llama-cli, llama-bench, etc.) respect env vars, not
just llama-server. Env vars act as defaults; CLI flags override.
Follows the existing get_env() pattern and uses the same
LLAMA_ARG_ prefix convention as the other env vars.
Co-authored-by: Pipboyguy <>
* server: enable checkpoint for recurrent models
create checkpoint after cancel
fix ban string and rm context during rewind
add checkpoint interval
only save recurrent cache
* save checkpoint during pp
---------
Co-authored-by: firecoperana <firecoperana>
* Revive fused delta-net
* Add command line argument for fused delta net
* Simplify/improve CUDA delta-net
* Add -fdn to llama-bench
* More CUDA fused delta net optimizations
* CPU optimizations
* Much faster fused delta-net on the CPU
It seems it is faster than the chunked implementation!
* Change meaning of fdn from bool flag to threshold value
* Use eps = 1e-6
* Give some nodes a name
