Mirror fork commit d9b9be0be as patch 0050 and record the Phase 4 W4A16 shared-memory padding gates, benchmarks, and mirror verification.
Assisted-by: Codex:gpt-5
Record the Phase 3 scale-broadcast experiment, its md5 and MUL_MAT_ID gates, the prefill regression, and the decision to ship no 0050 patch.
Assisted-by: Codex:gpt-5
Mirror fork commit 7dfa0e175 as patch 0049 and record the Phase 2 GB10 W4A16 shape sweep, md5 gates, MUL_MAT_ID checks, and mirror verification.
Assisted-by: Codex:gpt-5
Mirror the fork-first W4A16 packed tile metadata commit into the LocalAI paged patch series, record the Phase 1 benchmark result, and keep the implementation plan checked off.
Assisted-by: Codex:gpt-5
Record the clean forced W4A16 baseline, default comparison, selected metadata target, and completed plan checkpoint for the GB10 parity reopen.
Assisted-by: Codex:gpt-5
Record the clean DGX build retry, binary provenance, canonical greedy md5 gates, and completed plan steps for the GB10 parity reopen.
Assisted-by: Codex:gpt-5
Add the read-only DGX artifact review for the Phase 0 parity reopen, including supported paged measurements and missing vLLM difference-method evidence.
Assisted-by: Codex:gpt-5
Add the clean llama.cpp fork state, base merge point, patch count, and tree-match result for the Phase 0 parity reopen workflow.
Assisted-by: Codex:gpt-5
Add the Superpowers implementation plan for the GB10 parity reopen, including Phase 0 provenance, decode repro, W4A16 kill gates, and later kernel workstream entry criteria.
Assisted-by: Codex:gpt-5
Add a phased follow-up spec for challenging the GB10 vLLM-parity closure, including provenance gates, W4A16/GDN/MoE workstreams, and subagent ownership boundaries.
Assisted-by: Codex:gpt-5
Reconcile the paged backend pin prose with the current Makefile pin, mark the 0044 patch tracking note as resolved, and add DGX Docker worker idleness to the benchmark preflight.
Assisted-by: Codex:gpt-5
The fork mudler/llama.cpp branch localai-paged is the canonical source of
truth for all paged-backend kernel/patch work. Always update it FIRST: commit
the change on the fork branch and push it, then regenerate the LocalAI patch
series (backend/cpp/llama-cpp-localai-paged/patches/paged/) from the fork via
git format-patch so the series is a 1:1 drift-free mirror of the branch. Never
edit the LocalAI patch files directly, and never add a patch with no
corresponding fork-branch commit. The series is a derivative; the fork is the
source. The fork branch is also where the build and the per-path bit-exact md5
gate actually run, so it is the only place a change is truly validated.
Codified in two places:
- .agents/llama-cpp-localai-paged-backend.md: new "Fork-first workflow
(MANDATORY)" section at the top of the patch/pin-sync material, plus the
"Encapsulating your work" bullet now points at it.
- backend/cpp/llama-cpp-localai-paged/docs/PARITY_HANDOFF.md: strengthened the
hard-gate (section 2.5) into "Fork-first is MANDATORY", and corrected a stale
numbering example (fork 51168c5ee "patch 0044" maps to worktree 0044, not the
f32-only M5 which is worktree 0047).
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The fork mudler/llama.cpp branch localai-paged is the canonical source of
truth for the paged-backend patch series. This file is the git format-patch
of fork commit 51168c5ee ("feat(paged): fused gated RMSNorm + SiLU gate-mul
CUDA op (patch 0044)"), verified byte-identical to that commit's format-patch
output. The full on-disk series applies clean in numeric order on the pinned
base and the resulting tree is byte-identical to the fork commit tree (tree
hash a73d759350277532a14e853e1fe78f08bbb74ce8), so the LocalAI series is a
drift-free 1:1 mirror of the fork branch.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Ground-check follow-up to 2431090ff. Two factual corrections:
- Section 7 worktree line had the ahead/behind counts swapped ("25 ahead,
197 behind"); the branch is actually ~199 ahead / 25 behind origin/master.
- Discrepancy item 5 flagged only the MoE CDEF PAGED_GATE_MD5 (0921716...);
the dense run is symmetric (COMBINED_DEFINITIVE.txt records ecfe924d... for
dense, which likewise differs from the canonical dense gate 5951a5b4). Both
CDEF values come from combined_definitive.sh's own gate command, not the
canonical bit-exact gate in section 3.3, so neither is sanctioned and both
must be KL-validated.
Everything else in the handoff verified accurate: fork branch localai-paged
HEAD 51168c5ee (patch 0044) on dgx:~/llama-paged-fork, dev-tree HEAD a7d439e,
all md5/KL numbers, the 86%/1078/924 decode record, bench env, and all
referenced file/artifact paths.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Adds docs/PARITY_HANDOFF.md: the operational how-to for an agent with zero
context picking up the GB10 vLLM-parity work. Complements VLLM_PARITY_FINAL.md
(the why/record) with TL;DR state, the hard gates (per-path bit-exact md5,
KL-gate, no LLAMA_MAX_BATCH_TOKENS, fork-is-canonical), a copy-pasteable
operational quickstart (ssh/lock/build/bench + the --cuda-graph-trace=node
decode-profiling rule that caused 4 wrong analyses), the complete tested-and-
rejected lever map, methodology lessons, the three forward directions, and a
key file/artifact index with the open discrepancies to reconcile.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The decode-serving section characterized the high-N gap as "BW-floored, vLLM
pays equally / 56-68%". A clean uncontended graph-node-traced profile
(dgx ~/highN_prof2 + ~/highN_vllm, 2026-06-30) shows that was a profiling
artifact: decode runs as a replayed CUDA graph, and nsys without
--cuda-graph-trace=node collapses each replay into one opaque launch, so every
prior decode decomposition (159 us/tok, "host-bound", "5.4x more efficient")
was wrong. Corrected via --cuda-graph-trace=node + the ntg=64-minus-ntg=16
difference method.
Real picture (paged npl=256): 99% GPU-busy (idle 1.4%), NOT host-bound. GDN
recurrent scan 553 us/tok (51%, linear in batch, dominant), NVFP4 expert GEMM
254 (23%), bf16 proj 73 (7%), elementwise 57, SSM conv 31. Gap reconciled:
vLLM-server 1177 -> vLLM true GPU-steady 1078 (chunked-prefill overlap inflates
its window ~8pt) -> llama GPU-steady 924 (= 86% of 1078) -> llama-server 718
(61%, the ~17pt S3-recoverable serving graph-reuse overhead). So vs vLLM's true
GPU-steady decode we are ~86%, not 56%. GDN is a shared BW floor where paged
leads (83% vs 79% of 273 GB/s peak; both 1.17-1.18x for 2x batch).
The residual ~14pt is vLLM's mature fused kernels (Marlin MoE +11ms, Triton
elementwise +10ms); both ggml fusions rejected: act-quant-into-MMQ -79.4%
(ggml MMQ re-quantizes y per row-tile x stream-k split, no single-pass tiling),
norm+quant+silu infeasible via ggml_cuda_can_fuse. Added rejected levers:
Q8_0/FP8 projection (regime error, closes <=6%; vLLM FP8-proj confirmed from
hf_quant_config.json MIXED_PRECISION), the two decode fusions; refined BV-block
GDN occupancy to -1.04% (wave-hidden).
Revised verdict: PREFILL genuinely capped (36-43%, not graph-replayed so real);
DECODE-SERVING near-parity ~86% of vLLM true GPU-steady (headline 56% was a
measurement/operating-point artifact). GB10-vs-datacenter framing kept.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Add docs/VLLM_PARITY_FINAL.md: the standing, never-re-litigate record of the
exhaustive GB10 (sm_121) vLLM-parity investigation for the Qwen3.6 NVFP4 hybrid
models. Captures the definitive same-session both-engine benchmark (prefill
S_PP, decode/serving per-seq + aggregate, TTFT, PEAK_GB, paged-as-%-of-vLLM for
both the MoE 35B-A3B and dense 27B models), the complete lever map (every
prefill-GEMM, prefill-GDN, decode and serving/engine attempt with its verdict
and key number), the structural floors (LPDDR5x bandwidth, FP4-MMQ optimality,
GDN O(C^2) intra-chunk + serial recurrence, vLLM's HBM-tuned FLA/Marlin), the
shipped bit-exact wins, and the parity verdict: parity is a hardware ceiling on
GB10, not missing optimizations; the path to parity is datacenter Blackwell.
Every number cites its artifact (dgx:~/bench/COMBINED_DEFINITIVE.txt, the
marlin_gate / gdn_p1_ab A/Bs, PREFILL_GEMM_RESULTS, VLLM_PARITY_LEVER_MAP,
DECODE_SERVING_SCOPE, the patch headers); figures not pinned to an artifact are
marked estimated. Add a section-9 summary + link in the backend README.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The 0044/0045 patches were exported from the old bf16/hybrid dev tree and no
longer apply on the f32-only series (0026 ssm_bf16_tau is dropped), so the
build broke at `git apply`. Re-sync the vendored series to the now
feature-complete fork branch mudler/llama.cpp:localai-paged, which is the
canonical source (pin 0ed235ea + the paged patch commits in order).
- git rm the dev-tree-based 0044 (GDN M5, bf16-machinery base) and 0045
(Marlin W4A16 offline-repack, not part of the fork branch).
- Add the fork branch's newest commit (2c32ab8b7, "GDN M5 tensor-core
chunked-scan prefill, f32-only re-port") as 0047, generated with a single
git format-patch off that branch. It sequences after 0046 (its parent on
the branch) and recovers the prefill win 0044 encoded (+3.5% S_PP @npp512,
+17.7% @npp2048), bit-exact per-path (test-backend-ops GATED_DELTA_NET
46/46 default and force-M5; greedy md5 default-on == M5-forced == canonical).
- Track patch 0046 (dense-prefill geometry gate), which was on disk but never
committed, so the series is complete in git.
- README: patch-table header 0001-0046 -> 0001-0047, replace the 0044 row with
the f32-only 0047 row, fix the dangling 0044 prose references, note the
bf16 M6/M7/M8 variants are not part of this f32-only series, and add a
maintenance bullet that the series is now generated from the fork branch so
there is no more patch-export drift.
Verified: on a pristine llama.cpp at pin 0ed235ea the full series 0001-0043,
0046, 0047 applies clean in sorted order with the Makefile's exact
`git apply --verbose` method (37/37 OK), and the resulting tree is
byte-identical to the fork branch tip 2c32ab8b7.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
nsys cross-engine decomposition: the MoE prefill 64% gap vs vLLM is engine plumbing, not the kernel (GPU 97% busy, 443 vs 197 us/tok). Three buckets: per-expert W4A4 M-fragmentation (58%), GDN scan (24%), f32<->bf16 casts (15%). Offline-repack (0045) and verbatim vLLM-marlin port both trail FP4-MMQ via wrapper overhead, kept default-off as recorded negatives.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Land the tensor-core forms of the chunked gated-DeltaNet prefill scan (0031)
as a single GDN_TC-selected build and ship the M5 variant (full TC form-T
solve + state-update mma) default-ON when LLAMA_KV_PAGED is set.
The dispatch defaults GDN_TC=5 and GDN_CHUNK_MIN=64 under paged KV (both
env-overridable; OFF/INT_MAX when not paged, so stock/non-paged stays
regression-free). GDN_CHUNK_MIN is the per-call engage threshold and stays > 1
so decode (1 tok/call) keeps the sequential recurrence; 64 was tuned from a
{1,32,64,128,256} sweep (32/64/128 all win on prefill, 256 barely fires because
the MoE-prefill per-call count is < 256, 1 collapses decode S_TG ~25%).
Measured GB10, q36-35b-a3b-nvfp4, LLAMA_KV_PAGED=1 LLAMA_MOE_FORCE_GRAPHS=1,
llama-batched-bench -ngl 99 -fa on -ntg 4 -npl 32:
-npp 512 S_PP 2208.96 -> 2286.5 t/s (+3.5%, mean of 3 interleaved A/B)
-npp 2048 S_PP 2021.5 -> 2379.8 t/s (+17.7%)
Decode S_TG unchanged (~399 vs ~397 t/s, within noise).
Bit-exactness (per-path greedy md5, n=48 --temp 0 --seed 1, paged): default-on
== M5-forced == canonical on the gate prompt - MoE 8cb0ce23, dense 5951a5b4.
test-backend-ops GATED_DELTA_NET 94/94 vs CPU with M5 forced (incl. multi-chunk
up to n_tokens=256). On a long MoE prompt the default (M5 fires at >=64 tokens)
and the sequential path agree word-for-word until one benign greedy token-flip;
dense is byte-identical. The chunked scan is a NEW per-path result (different FP
reduction order), NMSE-validated benign.
CUDA-only, gencode arch=compute_121a,code=sm_121a (GB10 / sm_121a). README
sections 3 (0044 row, 0031 superseded note) and 5 (dev-notes verdict) updated.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Lever map records the full prefill/decode gap decomposition vs vLLM, the ranked levers, and the rejected dead ends. GDN build plan is the per-product mma mapping + A-inverse + occupancy design.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
0042 fuses the pre-norm residual add into RMSNorm (+0.5% prefill, bit-exact). 0043 makes the full-step MoE decode CUDA graph default-on (+2-4% decode, bit-exact; removes ~18x per-step host kernel re-issue, A/B-confirmed). 0034 (native FP4-MMA W4A4) and 0035 (Marlin-style W4A16 grouped MoE GEMM) are correct + bit-exact but regress vs the int8 FP4-MMQ in-backend on GB10 (bf16 MMA is ~half the int8 rate); shipped default-off as validated mechanisms and recorded negatives per the parity methodology.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Patch 0041 (LLAMA_PAGED_DECODE_STABLE) was made default-on-when-paged, but a
measured end-to-end A/B proved that is a serving mistake. S3 defers prefill
admission on the period-8 cadence, which delays prompt admission: 2.5x worse
TTFT (60s vs 24s at N=256) and 20-29% lower end-to-end throughput, with no
end-to-end win at any concurrency. Its apparent decode_agg gain was a metric
artifact (faster per-step decode bought by starving prefill).
Flip the s3_enabled default so an unset LLAMA_PAGED_DECODE_STABLE means OFF; the
mechanism stays available as an explicit opt-in (LLAMA_PAGED_DECODE_STABLE=1) for
decode-dominated, low-arrival traffic where TTFT is not a concern. The default now
prefers prompt prefill admission for good TTFT. S1 (patch 0040) keeps shipping
default-on; only S3's default changes.
Re-exports patch 0041 (change folded into its source commit) and updates the
README 0041 row plus the decode-serving narrative to record the A/B finding.
Greedy md5 gate unchanged (single-sequence llama-completion path, not
update_slots): paged MoE 8cb0ce23, dense 5951a5b4.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The S1 section-(a) padded/fixed-slot decode shape (the scoped follow-up to push
serving graph reuse from ~72% toward ~100%) was implemented in an isolated
worktree off the committed S1/S3/tail base, built CUDA-only, and benched on GB10.
Verdict: REJECTED. It is bit-exact and provably inert, but it regresses serving
throughput at every concurrency and does not close the vLLM gap.
Implementation (default-off, LLAMA_PAGED_PAD_DECODE): on a pure-decode step
(n_prompt_budgeted == 0) emit a masked-inert dummy decode for every idle slot so
n_tokens / n_seqs / n_seqs_unq / n_outputs and the seq-id set stay constant; a
release()-side guard keeps a finished slot warm under padding. Each dummy is its
own sequence (private recurrent state, per-stream paged attention, logits
discarded), so it cannot perturb a real stream.
Gates: single-seq greedy md5 bit-exact (dense 5951a5b4, paged-MoE 8cb0ce23). The
literal per-stream ON-vs-OFF identity gate is unachievable - concurrent cuBLAS/FA
decode is not bit-reproducible run-to-run even with padding off (OFF-vs-OFF
diverging streams: dense 3/16, MoE 8/16). The achievable inertness gate passed:
ON-vs-OFF per-stream prefix-agreement equals the OFF-vs-OFF noise floor exactly
(MoE 0.940/0.940, dense 0.812/0.812), so the dummy slots leak nothing.
Bench (MoE Qwen3.6-35B-A3B-NVFP4, GB10), burst decode tok/s/seq: n=8 S1+S3 28.16
/ PAD 6.05 / vLLM 44.8; n=128 S1+S3 4.53 / PAD 4.32 / vLLM 6.87. Staggered
aggregate tok/s: baseline (reuse 0%) 757.6, S1+S3 (reuse 72%) 763.3, PAD
(reuse 38%) 558.0.
Why it fails: (1) serving decode here is GPU-compute-bound, not host-rebuild-bound
- baseline reuse 0% ~= S1+S3 reuse 72% on aggregate tok/s, so closing reuse buys
~nothing (the earlier 542->762 host-bound delta did not reproduce); (2) padding
adds dummy-row compute proportional to pad_width - real_load, catastrophic at low
load; (3) in continuous serving padding cannot hold a constant width (perpetual
prefill churn) so reuse drops 72% -> 38%; (4) the completion-driven batch shrink
padding prevents is itself a throughput win in a compute-bound regime. The
residual burst gap is GPU-compute, which a host-side reuse lever cannot close.
Patch series unchanged: this rejected lever is NOT added to patches/paged/.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
FIX A (patch 0031 compose break): the chunked GDN prefill patch carried
'#include <cuda_bf16.h>' and '#include <type_traits>' as CONTEXT lines, but
those were introduced by the dropped bf16-tau patch 0026, so on the
bf16-tau-free 0001-0030 base only '#include <cstdlib>' is present and 'git
apply' failed. The same 0026 drop also shifted 0031's later hunks off their
context (the ', hyb' kernel-launch arg, the 'STATE_BF16, HYBRID' template
params, and the GDN_LAUNCH_ARGS list). Regenerated 0031 against a fresh
pin(0ed235ea) + 0001-0030 tree: the chunked kernel now SELF-PROVIDES the
cuda_bf16.h / type_traits includes (adds them, plus the climits it needs for
INT_MAX) and the dispatch guard is the 2-param 'if constexpr (!KDA &&
!keep_rs_t)' form. Behaviour is unchanged: 0031 stays opt-in, default OFF
(GDN_CHUNK_MIN), a recorded negative. The full 0001-0042 series now applies
clean on 0ed235ea ('git apply --check' green for every patch).
FIX B (patch 0041 S3 default): the decode-shape-stable scheduler defaulted OFF.
Make it default ON whenever paged KV is active (LLAMA_KV_PAGED set), still
overridable to off via LLAMA_PAGED_DECODE_STABLE=0. Minimal host-side change in
update_slots(); re-exported from the dev tree, README 0041 row updated to match.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Add the two decode-serving graph-reuse levers (validated on GB10) that close the
host-bound serving gap (paged dropped to ~3.7 vs vLLM ~5.9 tok/s/seq in real
continuous serving while tying it in static batched-bench).
- 0040 S1 paged decode-graph reuse: the paged decode inputs never overrode
llm_graph_input_i::can_reuse (defaults false), so the host rebuilt the ggml
graph on EVERY decode step (layer-A reuse 0%). Add a 256-bucketed-shape
can_reuse + a live-mctx refresh from the owning attn input. Bit-exact (md5
byte-identical reuse on/off). Static batched-bench: paged reuse 0% -> 95.5%.
- 0041 S3 decode-shape-stable scheduling: keep co-batched prefill out of decode
steps so the scheduler emits the reuse-stable pure-decode shape S1 can reuse.
Default-off policy on top of 0016; bit-exact (per-stream independent).
S1+S3 together (128-client staggered serving, MoE Qwen3.6-35B-A3B-NVFP4): graph
reuse 0% -> 72.2%, hostproc 15.98 -> 6.31 ms/step, decode 4.05 -> 5.52 tok/s/seq
median (4.24 -> 5.96 mean, at vLLM's ~5.9). S1 alone is insufficient (13.8%);
S3 is the multiplier. S2 (double-buffer set_inputs) dropped: Phase-0 put
set_inputs at ~0.05 ms/step, so it has nothing to recover. README patch table +
DECODE_SERVING_SCOPE.md updated with results and the padded/fixed-slot follow-up.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Option (a) of PREFILL_GEMM_SCOPE.md: route large-M (prefill) NVFP4 dense weight
GEMMs off the decode-tuned FP4-MMQ kernel onto the dequant->bf16 cuBLAS (nvjet)
tensor-core path, wired via an M-threshold in ggml_cuda_should_use_mmq. Lands the
validated, bit-exact-gated mechanism and records the honest GB10 result: it is a
regression, so it ships default-off (== stock), mirroring the patch-0017
default-off discipline.
Three-edit scaffold (no new kernel): should_use_mmq routes NVFP4+Blackwell+dense
M>LLAMA_FP4_PREFILL_M to cuBLAS; op_mul_mat_cublas gains an NVFP4 branch that
dequants the FP4 weights to a transient bf16 pool buffer (not cached - stays
FP4-resident) and runs cublasGemmEx CUDA_R_16BF/COMPUTE_32F; ggml_get_to_bf16_cuda
gains the NVFP4 case.
Bit-exact gate PASS (benign): test-backend-ops MUL_MAT 1146/1146 + MUL_MAT_ID
806/806; the forced path (LLAMA_FP4_PREFILL_M=64) is green CUDA-vs-CPU at NVFP4
large-M shapes; greedy md5 on q36-27b is byte-identical to FP4-MMQ both for
short prefill (5951a5b4, decode untouched) and for a >threshold prefill that
exercises the bf16 path (5f3967df - no greedy argmax flips).
Performance REGRESSES on GB10 (S_PP, q36-27b dense, A/B via env): M=512 958.99
-> 486.65 (-49%), M=1024 1013.65 -> 587.27 (-42%), M=2048 918.46 -> 649.42
(-29%). The scope premise (FP4-MMQ ~3% of FP4 peak at large M) is false here:
FP4-MMQ beats bf16-cuBLAS because bf16 peak is ~half FP4 peak and the per-step
weight dequant + 4x bf16 weight traffic (~8x total vs the FP4 read) dominate,
only partially amortizing as M grows. Default-off keeps stock S_PP (966.98).
Phase 2 (MoE grouped large-M) not implemented: it inherits the same
bf16-peak<FP4-peak ceiling plus a per-expert dequant, so grouped bf16-cuBLAS
would regress for the same reason; a real prefill GEMM win needs option (b), a
native FP4-MMA large-M kernel. Full A/B in docs/PREFILL_GEMM_RESULTS.md.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Scopes the follow-up recorded by patch 0031 + README section 5: replace the
serial per-thread reductions of the chunked gated-DeltaNet prefill scan with
mma.sync tensor-core matmuls and lift the 1-block/SM occupancy ceiling, the
path that would beat the tuned sequential scan and close the GDN prefill
bucket toward vLLM's ~2.5x-cheaper chunked scan.
Confirmed (not assumed) the GB10/sm_121a tensor-core reality: consumer
Blackwell (SM12x) has NO wgmma (Hopper-only) and NO tcgen05/TMEM (sm_100a
data-center only); the usable path is the extended mma.sync family. So the
kernel is a warp-synchronous mma.sync + cp.async design (reusing ggml's
mma.cuh tiles), not a wgmma/TMA/tcgen05 design - patch 0031's 'mma/wgmma'
shorthand reads as mma only on this part.
Design: register-resident state frees the 64KB that forced C=16, admitting
C=64 under the 99KB shared opt-in; tf32 inputs / f32 accumulate with a 3xtf32
precision ladder; decays/gamma/beta stay f32 outside the mma to preserve the
bounded de-gating; A-inverse via blocked forward substitution (FLA UT
transform) with mma off-diagonal coupling. Mechanism: chunking cuts state-BW
~Cx, mma absorbs the O(C^2) intra-chunk flops the serial 0031 could not.
Honest: multi-week, high risk, no vendor kernel to route to on sm_121; gains
beat the sequential scan and close most of the bucket but not full sm_100-class
parity. KL-gate binding (NMSE likely fails at reduced precision). Phased:
re-profile -> two-product PoC -> full intra-chunk + C=64 + reg-state ->
occupancy/cp.async; opt-in default-OFF until A/B-proven.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Add DECODE_SERVING_SCOPE.md: the decode KERNEL is at parity in static
batched-bench (~6.1 tok/s/seq ~ vLLM ~5.9 at npl128) but continuous serving
through llama-server update_slots() drops to ~3.7 (-39%) while vLLM sustains
~5.9. Scope shows the gap is the scheduler/host loop, not the kernel.
Root-cause hypothesis from source: continuous batching's batch-shape + seq-set
churn breaks BOTH graph-reuse layers every step - llama-context can_reuse/
allow_reuse (n_tokens + seq-set must match) and the CUDA ggml_cuda_graph
update_required memcmp (ne/nb/data ptrs) - so the GPU idles while the host
rebuilds + re-captures the graph and runs un-graphed set_inputs. vLLM avoids
this with padded/bucketed decode shapes + piecewise CUDA graphs. Documents that
the shipped scheduler patches (0008/0013/0016/0024/0025/0029) target prefill
freezing + burst collapse, NOT decode-step graph reuse, which is why the serving
gap survives them; notes the README s.5 'lever 2 graph coverage FLAT' verdict was
static-regime and is reopened here for serving only.
Ranks host-side, bit-exact-safe levers: S1 bucketed/padded decode-step shape for
graph reuse, S2 double-buffer/overlap per-step host work, S3 graph-shape-stable
scheduling (extend 0016). Specifies a Phase-0 profile to confirm host-bound
before any build, reusing the in-tree [L5INSTR] hostproc/set_inputs/
get_block_table timers, the 'graphs reused' perf counter, LLAMA_GRAPH_REUSE_DISABLE
and nsys GPU-busy%, with vLLM ground-truthed at the same concurrency. No kernel
code; no GPU run in this pass.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Adds patch 0031 to the paged llama.cpp series: an FLA-style chunked
parallel-scan prefill kernel for gated DeltaNet (the upstream
gated_delta_net.cu "Add chunked kernel for even faster pre-fill" TODO).
Scope: non-KDA scalar gate, f32 state, final-state-only, homogeneous.
Bit-exact-benign (NEW per-path): test-backend-ops GATED_DELTA_NET 91/91 within
the 1e-7 NMSE gate vs the CPU reference (patch adds 8 S_v=128 prefill cases:
exact-multiple / tail / multi-seq / GQA / permuted); numpy prototype confirms
f32 chunked-vs-sequential NMSE ~1e-13.
OPT-IN, default OFF: GB10's 99KB dynamic-smem opt-in forces C=16 (the 128x128
f32 state is 64KB of the all-shared layout), pinning the kernel to 1 block/SM
with serial dk-reductions. Measured ~761 t/s chunked vs ~971 t/s sequential
(~22%% slower) on q36-27b-nvfp4 prefill, so it defaults OFF (enable with
GDN_CHUNK_MIN=<n>); the backend default is regression-free. Beating the
84.7%-of-peak sequential scan needs tensor-core matmuls / register-resident
state with larger chunks (recorded in README section 5).
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Design + plan for the #1 prefill lever: NVFP4 weight GEMM at large M, where
MMQ (decode/M<=128-tuned, 1 CTA/SM, 128-col tile cap) is ~3.4x slower than
vLLM's marlin/cutlass large-M path (~51% of the prefill gap).
Recommends (a) dequant->bf16 cuBLAS routed by an M-threshold (dense first,
MoE grouped-cuBLAS second); rejects (b) a from-scratch Marlin/FP4 kernel as a
multi-week project. Key enabling finding: NVFP4->bf16 dequant kernels already
exist, and NVFP4 is currently force-excluded from the tensor-core cuBLAS path
(falls to f32 Sgemm) - relaxing that one guard is the pivot. Honest: bf16-cuBLAS
banks ~60-75% of the GEMM gap, not full 68us/tok parity (bf16 TC peak ~half FP4).
Design only - no kernel, no GPU run.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
The opt-in hybrid per-head bf16 SSM-state lever (ssm_bf16_tau, patch 0026) is
removed from the llama-cpp-localai-paged patch series. Clean re-measurement after
the decode fusions (0028 recurrent-state gather-fusion + 0029 block-table cache)
landed shows it buys nothing: forcing ALL gated-DeltaNet heads to bf16
(tau=100000, the most aggressive setting) gives flat decode throughput, 780.6 vs
780.0 t/s. The mode engages but adds zero speed because it is subsumed by the
fusions. The earlier "+12%" was measured before the fusions completed. bf16-tau
was a precision trade (not bit-exact, ~91% same-top-p) plus extra bug surface and
extra CUDA template-instantiation compile cost with no offsetting benefit.
Dependency check: no later patch (0028/0029/0030) depends on 0026. 0030's only
mention is a description comment; its code keys off fused_gdn_ar/ch/auto_fgdn,
which originate in 0018/0019/0021 (before 0026). The remaining series (0001-0025,
0028-0030) applies clean with git apply --check against the pin
0ed235ea2c17a19fc8238668653946721ed136fd. The Makefile applies the series by glob
(patches/paged/0*.patch); the resulting gap at 0026 is tolerated (0005/0027 are
already absent).
Removed:
- patches/paged/0026-qwen35-hybrid-perhead-ssm-state.patch
- the dead ssm_bf16_tau / ssm_hybrid_tau option handler in the shared
grpc-server.cpp (it only set LLAMA_SSM_BF16_TAU, now a no-op the library no
longer reads)
- the patched+bf16-tau benchmark columns and llama-patched-bf16tau rows
(README + final_benchmark.csv), the ssm_bf16_tau option text in backend
index.yaml, the gallery NOTE block, and the docs/features/backends.md mention.
The rejected-lever lesson is kept (why it was dropped: subsumed, tau=100000 flat)
in the backend README section 5, the paged-backend agent guide, and the
vLLM-parity methodology, so it is not re-tried.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
llama.cpp renamed the RPC tool target (tools/rpc/CMakeLists.txt: set(TARGET
ggml-rpc-server)) at the 0ed235ea pin. master already updated the stock
llama-cpp Makefile to match (--target ggml-rpc-server, cp bin/ggml-rpc-server);
the paged backend's separate Makefile copy was left stale and its -grpc (RPC)
variant failed with 'No rule to make target rpc-server' (grpc-server itself
built to 100%). Mirror the stock rename in the paged Makefile.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Add ACCELERATOR_PORTING_SCOPE.md, the umbrella scope for taking the paged
backend's accelerator-portable wins off the CUDA family. It builds on (does
not duplicate) UPSTREAM_LAYER2_SCOPE.md, which stays the GDN/SSM-fusion
detail (benefit #1), and adds:
- Benefit #2 (paged KV in-kernel block-table flash-attn read, 0009-0011):
new per-backend feasibility from source analysis of the Metal/SYCL/Vulkan
flash-attn kernels. SYCL EASY (near line-for-line CUDA mirror), Metal
EASY-MEDIUM (decode already routes to the vec kernel), Vulkan MEDIUM (the
fast coopmat2 NVIDIA decode path cannot do the indexed read; push-constants
are full). Universal constraint: only the vec/scalar decode kernel admits
the per-cell indexed read, so route block-table ops onto vec (as CUDA's
0009-0010 dispatch guard already does) and leave the fast MM/coopmat2 path
contiguous-only. This is the lever that flips paged KV from
neutral-to-slightly-negative to non-negative off CUDA.
- Benefit #3 (decode-first scheduler, 0013/0016): confirmed a free portable
win - host-side update_slots() policy, zero kernel work, runs on any
accelerator as-is.
- Benefit #4 (NVFP4 FP4-MMA, 0017/0023/0025): out of scope (Blackwell only);
flags the backend-agnostic analogues of the act-quant dedup and the
graph-coverage lever without over-claiming a port.
- A ROCm note: ROCm rides the CUDA/HIP path (validate, don't re-port);
FP4-MMA stays Blackwell-only.
Benefits #1 and #2 share the port shape and rank Metal->SYCL->Vulkan, so they
bundle into one per-backend PR behind a shared ops-first PR. Cross-link added
from UPSTREAM_LAYER2_SCOPE.md. All gates are test-backend-ops on-target (no
Metal/SYCL/Vulkan/ROCm hardware here).
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
master auto-bumped the stock llama-cpp pin 9d5d882d -> 0ed235ea and updated the
shared grpc-server.cpp. The paged backend's pin must track the stock pin (the
grpc-server.cpp is shared), so bump its LLAMA_VERSION to match. All 28 paged
patches apply clean on 0ed235ea (verified against a fresh upstream clone). The
bf16-tau state-serialization fix (patch 0026) is included. Bit-exact gate + full
grpc-server build verify on GPU/CI to follow.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The opt-in ssm_bf16_tau hybrid mode splits a gated-DeltaNet layer's
recurrent SSM state into an f32 partition (s_l) and a bf16 partition
(s_l_bf16). The recurrent state serialization paths (state_write_data /
state_read_data) were never updated for the split: they read/wrote s_l
using the FULL hparams.n_embd_s() (S_v*S_v*H) row width, but a split
layer's s_l only holds S_v*S_v*n_f32, so the access overruns the smaller
tensor (a ggml_backend tensor read out of bounds), and the bf16
fast-head partition was never persisted at all.
This is what broke high-concurrency serving with --ssm-bf16-tau: the
server's context-checkpoint feature serializes per-sequence state via
state_seq_get_data. With a checkpoint enabled, even a single request
triggered the out-of-bounds read; at higher concurrency the cell range
starts at a higher base slot so the overrun reaches further (hard abort
in a debug build, silent state corruption then 1-token-then-EOS on
restore in a release build). The static batched-bench never exercises
save/restore so it did not catch it; the GDN decode kernel and per-head
partition offsets were already correct (decode with checkpoints disabled
is fine at N=8/16/32).
Fix: serialize the f32 partition and, when the layer is split, the bf16
partition right after it, each with its OWN row width (tensor ne[0]).
head_slot is rebuilt deterministically at load (same model + tau), so it
is not serialized. Non-split layers have ne[0] == n_embd_s() and no bf16
partition, so their on-disk format and behavior are byte-identical (the
default f32 path and the bit-exact gate are unaffected).
Verified on GB10/DGX with Qwen3.6-35B-A3B-NVFP4 + --ssm-bf16-tau 64 via a
continuous-batching llama-server: with context checkpoints enabled, N=8,
N=16 and N=32 (slot reuse + restore) all now produce full coherent
128-token output and the server stays up; pre-fix the same config
aborted on the first checkpoint.
Assisted-by: Claude:claude-opus-4-8[1m] [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(voice-detect): add Go purego backend for voice-detect.cpp
Add backend/go/voice-detect implementing the Backend gRPC voice subset
(VoiceEmbed/VoiceVerify/VoiceAnalyze) over libvoicedetect.so via purego,
mirroring the parakeet-cpp / omnivoice-cpp backends.
The flat voicedetect_capi C ABI is dlopen'd cgo-less; malloc'd string and
float-vector returns are owned by Go and released through the matching capi
free functions, with the per-ctx last error surfaced into Go errors. Calls are
serialized via base.SingleThread since the C context is not reentrant.
Proto field mapping:
- VoiceEmbed: VoiceEmbedRequest.audio (path) -> embed_path -> Embedding+Model.
- VoiceVerify: audio1/audio2 + threshold (<=0 falls back to the
verify_threshold option, default 0.25) -> verify_paths -> verified/distance/
threshold/confidence/model/processing_time_ms.
- VoiceAnalyze: audio (path) -> analyze_path_json; the JSON age/gender/emotion
document maps to a single VoiceAnalysis segment (start/end 0; gender "label"
-> dominant_gender with the remaining float scores as the gender map; emotion
label/scores -> dominant_emotion/emotion).
The Makefile pins voice-detect.cpp to 47546430, clones+builds libvoicedetect.so
with ggml static-linked (PIC, GGML_NATIVE off) so dlopen needs no external
libggml/libvoicedetect; ldd on the artifact shows only system libs. Ginkgo
tests cover option parsing and analyze-JSON mapping; embed/verify smoke specs
gate on VOICEDETECT_BACKEND_TEST_MODEL + VOICEDETECT_BACKEND_TEST_WAV.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(voice-detect): wire backend into index, gallery and build
Register the voice-detect.cpp speaker-recognition + voice-analysis
backend (added in Voice-INT-A) into LocalAI's distribution surfaces,
mirroring the ced backend (the closest mudler C++/ggml audio analogue):
- backend/index.yaml: add the &voicedetect meta-backend (capabilities
platform map, no top-level uri) plus the full set of concrete per-arch
image entries (cpu/cuda12/cuda13/metal/rocm/sycl/vulkan/l4t and the
-development variants). Referential integrity audited - every alias
target resolves.
- gallery/index.yaml: add 5 model entries on backend voice-detect -
ECAPA-TDNN, WeSpeaker ResNet34, 3D-Speaker ERes2Net, CAM++ and the
wav2vec2 age/gender/emotion analyze model. The engine architecture is
read from GGUF metadata (voicedetect.arch) at load. GGUF artifacts are
not yet published: each files: entry points at the intended
mudler/voice-detect-gguf location with a TODO to fill sha256 after
upload (no fabricated hashes).
- .github/backend-matrix.yml: add the linux build matrix block + the
darwin metal entry mirroring ced.
- .github/workflows/bump_deps.yaml: track mudler/voice-detect.cpp via
VOICEDETECT_VERSION (pin 47546430, = 4754643).
- core/config/backend_capabilities.go: register voice-detect in the
backend capability map (VoiceVerify/VoiceEmbed/VoiceAnalyze ->
speaker_recognition), mirroring speaker-recognition.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(face-detect): add purego Go backend for face-detect.cpp
Add the LocalAI Go backend that dlopens libfacedetect.so (the flat
facedetect_capi_* C-ABI) via purego, mirroring the sibling voice-detect
backend. Implements the Face subset of the Backend gRPC service:
- Embeddings(PredictOptions): Images[0] base64 -> temp file -> embed_path
-> L2-normalized ArcFace embedding.
- Detect(DetectOptions): src -> detect_path_json -> Detection boxes
(class_name "face", [x1,y1,x2,y2] -> x/y/w/h).
- FaceVerify(FaceVerifyRequest): two images + threshold + anti_spoof ->
verify_paths; best-effort img areas via detect.
- FaceAnalyze(FaceAnalyzeRequest): img -> analyze_path_json -> per-face
age + gender ("M"/"F" normalized to "Man"/"Woman").
The Makefile pins face-detect.cpp to 636a1963 and builds the shared lib
with ggml + vendored libjpeg-turbo static (PIC), so the .so is
ldd-clean (no libggml) and exports only facedetect_capi_* (no jpeg_
symbols). Gated Ginkgo e2e mirrors voice-detect.
Note for the gallery-wiring task: backend registration (index.yaml,
gallery, core/config/backend_capabilities.go) is intentionally not
touched here.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* fix(voice-detect): replace em dashes in net-new descriptions
Project style forbids em/en dashes. Replace the three U+2014 chars
introduced by the voice-detect gallery/index wiring with `-`/`:`.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(face-detect): wire backend into index, gallery and build
Register the face-detect.cpp face detection / embedding / verification /
analysis backend (added in Face-INT-A) into LocalAI's distribution
surfaces, mirroring the voice-detect wiring (the closest mudler C++/ggml
recognition analogue):
- backend/index.yaml: add the &facedetect meta-backend (capabilities
platform map, no top-level uri to avoid the meta-backend gotcha) plus
the full set of concrete per-arch image entries (cpu/cuda12/cuda13/
metal/rocm/sycl-f16/sycl-f32/vulkan/l4t and the -development variants),
22 entries. Referential integrity audited: every alias target resolves.
- gallery/index.yaml: add 4 model entries on backend face-detect -
face-detect-buffalo-l/m/s (insightface SCRFD + ArcFace/MBF, NON-COMMERCIAL)
and face-detect-yunet-sface (OpenCV-Zoo YuNet + SFace, APACHE-2.0, the
commercial-friendly alternative). The detector/embedder architecture is
read from GGUF metadata (facedetect.arch) at load; only the real
verify_threshold option is set (0.35 buffalo, 0.363 sface). GGUF
artifacts are not yet published: each files: entry points at the
intended mudler/face-detect-gguf location with a TODO to fill sha256
after upload (no fabricated hashes).
- core/config/backend_capabilities.go: register face-detect in the
backend capability map (Embedding/Detect/FaceVerify/FaceAnalyze ->
face_recognition), mirroring insightface.
- .github/backend-matrix.yml: add the linux build matrix block + the
darwin metal entry mirroring voice-detect.
- .github/workflows/bump_deps.yaml: track mudler/face-detect.cpp via
FACEDETECT_VERSION (pin 636a1963).
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* fix(recon): voice-detect metal build branch + face-detect gallery usecases
Add the missing metal BUILD_TYPE branch to the voice-detect Makefile
forwarding -DVOICEDETECT_GGML_METAL=ON, mirroring face-detect, so the
darwin metal CI artifact is built with the Metal backend instead of
CPU-only.
Expand the 4 face-detect gallery models' known_usecases to
[face_recognition, detection, embeddings] to match the backend
capabilities map and the mirrored insightface-buffalo entries, so
auto-selection for /v1/detect and /embeddings works.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* docs(recon): document voice-detect and face-detect ggml backends
Document the new standalone C++/ggml biometric backends as the
recommended/default option for face and voice recognition, keeping the
existing Python insightface / speaker-recognition backends framed as the
legacy path.
- features/face-recognition.md: add a face-detect (ggml) backend section
with the gallery entries (buffalo-l/m/s non-commercial, yunet-sface
Apache-2.0), licensing, and verify/detect/analyze quickstart.
- features/voice-recognition.md: add a voice-detect (ggml) backend
section with the gallery entries (ecapa-tdnn, wespeaker-resnet34,
eres2net, campplus speaker recognizers; emotion-wav2vec2 non-commercial
analyze head) and quickstart.
- reference/compatibility-table.md: add face-detect.cpp and
voice-detect.cpp rows to the Vision, Detection & Recognition table.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(gallery): publish recon backend GGUF uris + sha256
Fill in the published HuggingFace GGUF uris and verified sha256 for the
9 recon gallery entries (voice-detect-* and face-detect-*), and remove
the TODO publish markers. Correct the eres2net, campplus, and
emotion-wav2vec2 uris to the actual published filenames.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(gallery): re-embed buffalo anti-spoof + add audeering age/gender voice model
Update the 3 buffalo face-detect GGUF sha256 (anti-spoof ensemble now
embedded and re-uploaded under the same filenames/uris) and note the
FaceVerify anti_spoof request flag in each description. Add a new
voice-detect-age-gender-wav2vec2 gallery entry mirroring the emotion
model.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(gallery): add face-detect-buffalo-sc and antelopev2 packs
Add gallery entries for two newly-published insightface face packs on
the face-detect backend: buffalo_sc (smallest pack, SCRFD-500M + small
ArcFace) and antelopev2 (higher-accuracy, SCRFD-10G + ArcFace glint360k
R100, 512-d). Both are non-commercial research-only.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(recon): honor LocalAI per-model threads in voice/face-detect backends
LocalAI spawns one backend process per model and serves requests
concurrently, so the engines' own min(hardware_concurrency, 8) default
can oversubscribe cores. Forward the per-model Threads value from the
gRPC LoadModel options into the engine via VOICEDETECT_THREADS /
FACEDETECT_THREADS (read at backend construction) before the capi load.
A non-positive Threads is treated as unset, leaving the engine default.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump backend pins to CPU-optimized engine commits
voice-detect.cpp -> 0d9c1b3 (radix-2 FFT FBank, threads, flash attn + cached
pos-conv); face-detect.cpp -> 523aee1 (thread-gated direct conv, threads).
Brings the CPU optimizations into the LocalAI backend builds. GGUF format and
parity unchanged, so the published HF GGUFs remain valid.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump backend pins to round-2 CPU-optimized engines
voice-detect.cpp -> fe7e6a3 (ERes2Net 1x1->mul_mat, CAM++ layout+context,
wav2vec2 conv-LN, ECAPA capture-drop, AVX512 dispatch opt-in); face-detect.cpp
-> 9c8adb7 (AVX2 Winograd F(2x2,3x3) for SCRFD/ArcFace 3x3 convs, ArcFace
BN-fold). Parity unchanged (cosine=1.0); GGUF format unchanged, HF GGUFs valid.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump backend pins to round-3 Winograd engines
voice-detect.cpp -> 45122ec (Winograd F(2x2,3x3) for WeSpeaker/ERes2Net 3x3
convs, -22%/-20% @8t); face-detect.cpp -> cd5c962 (Winograd F(4x4,3x3) for
SCRFD large maps, -22% @1t on top of F(2x2), more load-stable). Parity held
(cosine=1.0); GGUF format unchanged, HF GGUFs valid.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump backend pins to round-4 Winograd engines (CPU opt complete)
voice-detect.cpp -> d2839ca (CAM++ FCM 2D convs through Winograd, -15.5%/-10.3%);
face-detect.cpp -> c1db23d (AVX2-vectorized Winograd tile transforms, SCRFD
detect -14%/-9.6%). Final CPU optimization round; the conv-kernel lever class is
now exhausted (parity held cosine=1.0; GGUF/parity unchanged, HF GGUFs valid).
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump face-detect pin to deep-kernel engine (7ae5c4d)
face-detect.cpp -> 7ae5c4d: register-blocked winograd-domain GEMM microkernel
(2.8x isolated GFLOP/s), AVX-512 zmm evolution behind runtime CPUID dispatch
(ship-safe, AVX2 fallback bit-identical), bias/relu fused into the winograd
output transform, and SFace Conv+BN fold + bias/PReLU fusion. SCRFD detect
~1.4x faster end-to-end vs the round-4 baseline; parity bit-exact; portable
single binary (function-multiversioned, no global -mavx512f).
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump voice-detect pin to ECAPA operand-order win (e9c56ae)
voice-detect.cpp -> e9c56ae: weight-as-src0 mul_mat order in ECAPA's F32
conv1d_same (routes through tinyBLAS sgemm); ECAPA embed 1.67x @1t / ~1.3x @8t,
parity cosine=1.0. Isolated to encoder.cpp (ECAPA-only); ERes2Net/CAM++/WeSpeaker
do not call conv1d_same so are provably unaffected.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump pins to FMA-throughput engines (voice f7b9f89, face 2d2d5f0)
face -> 2d2d5f0: route ArcFace 3x3 body convs through the AVX-512 winograd
microkernel (kWinoMinSize 80->14); ArcFace 1.62x @1t, SCRFD detect to 0.966 of
MLAS @1t, no regression. voice -> f7b9f89: runtime-CPUID-dispatched AVX-512
winograd-GEMM microkernel (ship-safe, AVX2 fallback bit-identical); WeSpeaker
1.90x @1t. Parity cosine=1.0 throughout; portable single binaries.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump pins to MLAS-class direct-conv engines (voice 7ecfd07, face be22d67)
Hand-tuned nChw16c AVX-512 register-tiled direct-conv microkernel (~263 GFLOP/s,
within 6-7% of MLAS per-op efficiency), runtime-CPUID-dispatched + AVX2 fallback,
fused bias/relu. voice 7ecfd07: default 3x3-s1 kernel for WeSpeaker (+37%/+32%)
+ ERes2Net, CAM++ pinned to Winograd. face be22d67: shape-gated to the ArcFace
recognizer body (+25-27% @8t); SCRFD detector stays on Winograd (no regression).
Parity cosine=1.0 / detect <=1px on AVX-512 + AVX2 paths. Portable single binaries.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump voice pin to Phase-A blocked backbone (f4e7eef)
WeSpeaker ResNet34 runs as one nChw16c blocked island (2 reorders/forward vs
~60) on AVX-512, default; per-conv directconv fallback on AVX2. +2.9% @1t /
+17-19% @8t vs per-conv directconv, parity cosine=1.0. The conv microkernel is
already FMA-bound near peak (~0.86-0.98x MLAS-implied); residual to MLAS is
sub-peak edge + non-conv tail, documented in docs/cpu-optimization.md.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump pins to breadth blocked-backbone (voice 7f66871, face d80092b)
voice 7f66871: AVX2-vectorized (ymm) blocked island - AVX2-only hosts now run
the blocked backbone for WeSpeaker (2.3x over per-conv-AVX2, cosine=1.0);
ERes2Net stays per-conv (blocked regresses, opt-in only); CAM++ Winograd-pinned.
face d80092b: ArcFace recognizer blocked island, AVX-512 default (-13% @8t, ~0.90x
MLAS, the closest conv result), auto per-conv on AVX2; SCRFD untouched on Winograd
(0 island invocations during detect). Parity cosine=1.0 / detect <=1px throughout.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump pins to small-spatial + stem conv kernels (voice 99b1804, face 47fdab6)
Measured-gap-driven conv kernels: small-spatial (fill the register tile when
output width <= tile width) + small-IC stem + strided-1x1/downsample recovery.
ArcFace recognizer 0.57 -> 0.70x MLAS @1t (the closest conv model), WeSpeaker
0.65 -> 0.79x @1t. Parity cosine=1.0 / detect <=1px. The OC-block-sharing lever
was a measured dead-end (deep stride-1 is L3-weight-bandwidth bound, not
read-port bound) and was NOT shipped. Kernel ceiling reached; further gap needs
an algorithm-class change (cache-blocked weight-stationary GEMM, or q8 weights).
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump pins to GPU persistent-graph + multi-model-safe cache (voice 45d2e6b, face 0a4799a)
GPU wins (CUDA/ggml backend, no CPU-path change): persistent per-shape graph+context
cache in Backend::compute() eliminates the per-call cudaGraph re-instantiation churn
-> wav2vec2 emotion+age-gender now AT GPU parity with torch-cuDNN on GB10 (0.97-0.98x),
CAM++ -5.7ms; bit-identical parity. Cache hardened multi-model-safe (invalidate-on-free
keyed by the ModelLoader weights buffer) so LocalAI multi-model hosting cannot stale-hit.
Conv models still trail cuDNN (im2col-materialization-bound) - cuDNN implicit-GEMM lever next.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump pins to cuDNN-conv-capable engines (voice b6e4356, face 6107a24)
Adds the opt-in cuDNN implicit-GEMM conv path (VOICEDETECT_GGML_CUDNN /
FACEDETECT_GGML_CUDNN, DEFAULT OFF -> zero build/runtime dep until enabled).
On GPU it kills the im2col-materialization bottleneck and reaches torch-cuDNN
parity on the spill-bound convs: SCRFD detect 14.8->6.4ms (2.3x, ~parity),
WeSpeaker ~parity, ERes2Net beats torch (1.10x); ArcFace/CAM++ neutral (no
spill). Parity exact (SCRFD <=1px, cosine=1.0). To USE it in LocalAI, the CUDA
backend build must enable the flag AND bundle libcudnn - deferred until a
cuDNN-bundled GPU image; flag stays OFF here.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(recon): enable cuDNN conv path on arm64+CUDA13 recon backends
The voice-detect.cpp / face-detect.cpp engines have an opt-in cuDNN
implicit-GEMM conv path behind VOICEDETECT_GGML_CUDNN / FACEDETECT_GGML_CUDNN
(default OFF) that kills im2col on the GPU and reaches torch-cuDNN parity
(SCRFD 2.3x, WeSpeaker/ERes2Net parity), measured on the GB10
(arm64, CUDA 13, sm_121a).
Enable it for the CUDA build, but only where cuDNN actually ships: the
arm64 + CUDA 13 image (GB10/Jetson/L4T). x86 CUDA images carry no cuDNN,
so flipping it on globally for BUILD_TYPE=cublas would be a link failure.
The Makefiles gate on CUDA_MAJOR_VERSION=13 + arch (TARGETARCH from the
matrix/Docker build, uname -m fallback for local builds).
backend/Dockerfile.golang already installs the runtime libcudnn9-cuda-13
in the arm64+CUDA13 apt block; add the matching libcudnn9-dev-cuda-13 so
the build-time link resolves.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): bump voice-detect pin to ERes2Net blocked-default (30beecd)
Defaults VD_ERES2NET_BLOCKED ON: routes the ERes2Net Res2Net body through the
blocked nChw16c AVX-512 directconv island instead of the 1x1 mul_mat fast path
(CONT-transpose + skinny low-K GEMM). On the shipped GGML_NATIVE=OFF build (ggml
mul_mat is AVX2-only) this wins ~2x at every thread count (2.07x@1t, 2.2x@4t,
2.05x@8t); pure-AVX2 fallback still 1.3-1.62x. Parity exact (cosine=1.000000 vs
golden), so registered voices + verify/identify thresholds are unaffected. The
prior default-OFF rested on a stale comment whose 23pct regression only held on
the non-shipping GGML_NATIVE=ON build.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* docs(readme): announce native voice-detect + face-detect backends in Latest News
Add a Latest News entry for the new from-scratch C++/ggml biometric backends
(voice-detect.cpp + face-detect.cpp) that replace the Python insightface and
speaker-recognition backends: no Python/onnxruntime at inference, self-contained
GGUF, bit-exact parity, GPU cuDNN parity. Mirrors the parakeet.cpp /
locate-anything.cpp native-backend news entries. Refs PR #10441.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* chore(recon): re-pin to the squashed engine release commits
The voice-detect.cpp and face-detect.cpp histories were squashed to a single
release commit, which orphaned the previous pins (voice 30beecd, face 6107a24).
Re-pin to the new single-commit SHAs (voice 3d51077, face 06914b0); the tree is
identical, so the backend build is unchanged.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* fix(ik-llama): port multimodal path to mtmd API and bump to f96eaddb (#10534)
The IK_LLAMA_VERSION bump to f96eaddba8bed6a9a5e628bbf6a566775c70b49c pulls in
upstream commit "Prune examples/llava", which deletes examples/llava (clip.* /
llava.*). The ik-llama backend's grpc-server.cpp built a local `myclip` library
from those files and called the removed clip/llava C API, so the bump no longer
builds.
ik_llama keeps its multimodal stack in the surviving `mtmd` library
(examples/mtmd/, public headers mtmd.h + mtmd-helper.h). This ports the backend's
multimodal path onto the high-level mtmd_* / mtmd_helper_* API in place, leaving
the text path (which still uses ik_llama's retained old common API) untouched:
- Makefile: bump IK_LLAMA_VERSION to f96eaddb.
- prepare.sh: drop the clip/llava source copy + sed block; mtmd is a library
target, no source copy needed.
- CMakeLists.txt: remove the `myclip` target; link `mtmd` and add its include
dir; build grpc-server as C++17 (mtmd headers require it).
- patches: drop 0002 (targeted the deleted examples/llava/clip.cpp; the mtmd
clip.cpp never calls ggml_quantize_chunk, so the fix is unneeded). Keep 0001
(verified still applies).
- grpc-server.cpp / utils.hpp: replace clip_model_load + clip_image_load_from_bytes
+ llava_image_embed_make_with_clip_img + the manual [img-N] prefix splitting and
per-image llava_embd_batch decode loop with mtmd_init_from_file (moved after the
model load, which it requires), mtmd_helper_bitmap_init_from_buf, mtmd_tokenize
and mtmd_helper_eval_chunks. Legacy [img-N] tags are translated, in order, into
mtmd media markers (mtmd_default_marker()); the post-image suffix text stays on
the normal token path so the sampling loop is unchanged.
Supersedes #10534.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* fix(ik-llama): align json alias to ordered_json to resolve mtmd.h conflict (#10534)
mtmd.h declares `using json = nlohmann::ordered_json` at global scope (and its
mtmd.cpp depends on it), while ik_llama's whole server/common stack also uses
ordered_json. Our grpc-server.cpp/utils.hpp kept a plain `nlohmann::json` alias,
which now collides with mtmd.h once it is included for the multimodal port:
"conflicting declaration 'using json = ...'". Switch our two aliases to
ordered_json to match; it is API-compatible (utils.hpp already used ordered_json
for its log helper) and our json never crosses into an unordered-json API.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* ⬆️ Update ggml-org/llama.cpp
Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* fix(llama-cpp): link server-stream.cpp TU into grpc-server for upstream 0ed235ea (#10536)
Upstream llama.cpp 0ed235ea added an SSE stream-resumption layer in a new
translation unit tools/server/server-stream.cpp, which defines
stream_session, stream_pipe_producer and the g_stream_sessions manager.
server-context.cpp (already #included into grpc-server.cpp) now calls into
it via spipe->cleanup(), stream_aware_should_stop() and
stream_session_attach_pipe(), so without the new TU the grpc-server link
fails on every arch with:
undefined reference to `stream_pipe_producer::cleanup()'
prepare.sh already copies every tools/server/* file into tools/grpc-server/,
so the source is present; the only missing piece was including its
definitions. Add an __has_include-guarded #include "server-stream.cpp"
before server-context.cpp, mirroring the existing server-chat.cpp and
server-schema.cpp guards, keeping the source compatible with older
pins/forks that predate the split. The file is self-contained (its only
external symbols come from server-common, already in the TU) so it adds no
new undefined references; the http route-handler factories it also defines
are unused in the grpc path but harmless.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* fix(llama-cpp): build renamed ggml-rpc-server target for upstream 0ed235ea (#10536)
Upstream renamed the RPC server CMake target and binary from `rpc-server`
to `ggml-rpc-server` (tools/rpc/CMakeLists.txt: `set(TARGET ggml-rpc-server)`),
so the RPC-enabled grpc build failed with "No rule to make target 'rpc-server'".
The grpc-server itself links fine after the server-stream.cpp fix; this only
updates the RPC target name and the binary path copied to llama-cpp-rpc-server.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
---------
Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: mudler <2420543+mudler@users.noreply.github.com>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Vendored paged headers used size_t / uintN_t without including <cstddef> /
<cstdint>. The arm64 DGX toolchain provides them transitively so the build
passed there, but amd64/older toolchains do not, failing the CI amd64 build one
header at a time ('size_t' does not name a type -> cascade).
paged-kv-manager.h was already fixed. This adds the missing includes to the
remaining vendored headers at the point each is created/rewritten in the patch
series so every src/paged*.h self-includes both:
* paged-attn.h (0003): add <cstddef> (had <cstdint>)
* paged-alloc.h (0007): add <cstddef> (had <cstdint>)
* paged-prefix-api.h (0007): add <cstddef> + <cstdint> (had only llama.h)
The .cpp units include their own paged header, so they inherit the includes
transitively. Whole series still applies clean on the pinned llama.cpp.
Compile-only change: no runtime behavior change, bit-exactness unaffected.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Fixes cuda-13 amd64 / non-arm64 build where size_t was used without the
header (arm64 cuda-13 pulled it in transitively; amd64/cuda-12 toolchains
do not). Compile-only change, bit-exactness unaffected.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The paged backend targets Blackwell sm_121a, which CUDA 12.0 cannot target
at all, so the CUDA-12 variants were pointless. They were also broken: the
cublas-12 / nvidia-l4t / arm64 build failed to compile paged-kv-manager.cpp
("no declaration matches ...", a ~10-function mismatch the older
cuda-12-base gcc rejects). CUDA-13 compiles it fine (confirmed on GB10).
Removed (config-only, scoped to the paged backend):
- backend-matrix.yml: the two CUDA-12 paged rows
(-gpu-nvidia-cuda-12-llama-cpp-localai-paged,
-nvidia-l4t-arm64-llama-cpp-localai-paged)
- backend/index.yaml: CUDA-12 capability keys (nvidia-cuda-12,
nvidia-l4t-cuda-12, nvidia-l4t) on both meta-backends, repointed
default/nvidia to the cuda13 amd64 variant, and dropped the orphaned
cuda12-* / nvidia-l4t-arm64-* variant definitions (latest + -development).
Kept CUDA-13 only: cuda13-llama-cpp-localai-paged (amd64) and
cuda13-nvidia-l4t-arm64-llama-cpp-localai-paged (l4t arm64). Matrix
tag-suffixes <-> index variant URIs form a clean 2:2 bijection.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The background=true resumable-stream path had two latent issues.
1. Unbounded resume buffer. AppendEvent grew StreamEvents without limit, so
a long-running or abandoned background generation could consume process
memory without bound. The store now caps the buffer (event count and total
bytes, mirroring llama.cpp's byte-capped slot ring), evicting oldest events
from the front and advancing a droppedThrough watermark. GetEventsAfter
returns ErrOffsetLost when the requested starting_after is below the
watermark, and handleStreamResume surfaces that as HTTP 409 before
committing to the SSE response, so a resuming client gets a clear error
instead of a silently truncated stream.
2. Missing ownership check (IDOR). GET /responses/:id, its stream resume, and
/cancel looked up responses purely by ID, letting any caller who knows or
guesses an ID read or cancel another caller's response. Responses now carry
the creating caller's identity (auth.GetUser), stamped at creation and
compared on read/cancel/resume; a mismatch returns 404 (not 403) so
existence is not leaked. Backward compatible: responses with no owner
(single-key / no-auth deployments) remain accessible.
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
In distributed mode, even when the frontend and workers share the same
models directory via a shared volume mount, starting a model on a worker
re-staged (re-downloaded) it: stageModelFiles always uploads model files
into a tracking-key-namespaced subdir on the worker, and the staging probe
only checks that staged location, so a file already present on the shared
volume at the canonical path was never reused.
Add a config switch LOCALAI_DISTRIBUTED_SHARED_MODELS (default false). When
enabled, the operator asserts that all nodes mount the SAME models directory
at the SAME path, so staging is unnecessary: the frontend's absolute model
paths are already valid on the worker. In that mode stageModelFiles returns
the cloned opts unchanged without uploading, leaving the path fields pointing
at their canonical absolute paths so the worker loads them directly from the
shared volume.
The value is plumbed from DistributedConfig through SmartRouterOptions into
the SmartRouter. Docs and docker-compose.distributed.yaml updated.
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
