mirror of
https://github.com/mudler/LocalAI.git
synced 2026-09-12 22:33:54 -04:00
5effa4752789f2bbe4e17029ee00ff731d8afffa
14
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 mudler's LocalAI [bot]andEttore Di Giacinto
|
7b9167eaad |
feat(llama-cpp): serve Qwen3-TTS through the llama.cpp backend (#11392)
* fix(config): do not read a TTS speaker-encoder mmproj as vision support Qwen3-TTS on llama-cpp ships an mmproj holding the speaker encoder and code predictor. VisionSupported() treated any non-empty MMProj as proof of image input, so every such model would be advertised as vision-capable. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * feat(llama-cpp): add TTS request option parsing helper Validates text and speaker reference presence and strictly parses the top_k / top_p per-request params, in a header with no llama.cpp or gRPC dependencies so the standalone C++ unit test gate picks it up. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(llama-cpp): range-check the TTS top_k and top_p request params Format validation alone let NaN, infinity and out-of-range values through. The consumer copies both values into the audio generation input unconditionally and only guards its separate sampler assignment with "> 0", a test NaN also fails, so a NaN reached llama.cpp with the guard never firing. top_k must now be >= 0 and top_p must fall within 0.0 to 1.0 inclusive, with the bound written as a negated in-range test so NaN is rejected rather than silently accepted. Also cover the two checks the suite could not previously kill: the whole-string check in the float parser and the int32 range check. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * chore(llama-cpp): bump pin to f9e832c10 and carry the TTS server task Picks up ggml-org/llama.cpp#26254 (Qwen3-TTS via mtmd) and #26536 (the short-input audio chunk fix). Adds 0002-add-server-task-type-tts.patch, the server-side half of the still-draft #26603, so TTS runs through the slot scheduler instead of racing it. Remove that patch when #26603 merges. The patch is rebased on top of the score patch: its tokenize-switch hunk collided with the SERVER_TASK_TYPE_SCORE case, and its lone SRV_WRN call passes no variadic argument, which the macro cannot expand. The score patch itself needed no refresh. Also fixes fallout from the bump in grpc-server.cpp: upstream dropped the per-slot n_ctx argument from server_schema::eval_llama_cmpl_schema. Only the schema branch loses it, since forks predating the server-schema split still expect the old argument list. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * feat(llama-cpp): implement the TTS and TTSStream RPCs Both were declared in backend.proto but unimplemented. They now submit a SERVER_TASK_TYPE_TTS task and drain the response reader, the same shape PredictStream uses. The streaming path emits a leading sample_rate message and then raw PCM, because ModelTTSStream builds the WAV header itself; the non-streaming path emits a complete WAV to the requested dst. The streamed samples are converted from the pipeline's float32 to signed 16-bit first. MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM hands back floats, while the header ModelTTSStream writes announces 16-bit samples, so shipping the floats verbatim would decode as noise. prepare.sh and CMakeLists.txt now stage tts_request_options.h alongside the other grpc-server helpers, and register its standalone test with ctest the way passthrough_options_test is registered. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(llama-cpp): mask non-codec tokens for Qwen3-TTS generation The Qwen3-TTS gen-audio pipeline maps a sampled backbone token to a codebook row with an unchecked subtraction, in mtmd-helper-gen.cpp: inp.code0 = sampled - codec_0; For ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF the vocab is 155008 tokens, <|codec_0|> is 151936 and the codec codes end at 153983. The model's own tokenizer.ggml.suppress_tokens holds 1023 ids covering 153984..155007, every special above the codec range except <|codec_eos_token|> (154086) which stays reachable as the stop token. Nothing masks the text range 0..151935, so the backbone can sample a text token at any step, the subtraction goes negative, and ggml_compute_forward_get_rows aborts the whole backend process on GGML_ASSERT(i01 >= 0 && i01 < ne01). Complete the mask upstream started: bias every token below <|codec_0|> to -INFINITY for TTS tasks so only codec codes and the codec EOS remain reachable. The biases are appended to task.params.sampling.logit_bias, which common_sampler_init already merges with the model's suppress tokens into one llama_sampler_init_logit_bias, so no sampler is added to the chain. Measured cost is 0.082 ms per sampled token and 1.16 MB, set against a forward pass in the multi-millisecond range. It lands in launch_slot_with_task rather than in a route handler so that llama.cpp's own POST /tts and LocalAI's TTS/TTSStream RPCs are both covered, and <|codec_0|> is resolved from the vocab rather than hardcoded so a model without it is left alone. This is reproducible with upstream's own llama-tts and no LocalAI code loaded, aborting at frame 55 on Q4_K_M and frame 71 on Q8_0, so it is neither a quantization artifact nor an artifact of the gRPC adapter. Two further defects in the same draft pipeline still prevent end-to-end audio; they are independent of this one and are recorded in the task report for an upstream bug report. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * chore(llama-cpp): bump pin to 9de0fcf2b and drop the TTS codec mask Upstream fixed the Qwen3-TTS abort in ggml-org/llama.cpp c8e03ce81 ("mtmd/ggml: add ggml_build_forward_order", #26649), landed one hour after the previous pin. ggml_build_forward_expand marks a tensor and all its ancestors for compute, so using it as a pure ordering hint defeated ggml_build_forward_select and made GEN_WAV calls execute the GEN_CODE branch against a stale inp_code0, hitting the get_rows bound assert in ggml_compute_forward_get_rows. That single defect accounts for every abort seen on this model, so 0003-mask-non-codec-tokens-for-tts.patch is removed rather than rebased. The mask changed the observed behavior, but it was perturbing a graph ordering bug rather than fixing a sampling one: at the new pin the whole path works without it. Keeping it would have meant carrying a 152k-entry logit bias, and rebasing it on every pin bump, for no benefit. Verified at 9de0fcf2b with only 0001 and 0002 applied, which both apply clean with no fuzz and needed no rebase: non-streaming HTTP 200, 410924 bytes, 8.56 s RIFF (little-endian) data, WAVE audio, Microsoft PCM, 16 bit, mono 24000 Hz streaming HTTP 200, 560684 bytes, 11.68 s, exactly one RIFF at byte 0, same format, which also exercises the float32-to-s16 conversion at runtime for the first time Pristine unpatched llama-tts at the same pin now also completes, 130 frames to a valid WAV, where it aborted at frame 55 before. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(llama-cpp): clear the TTS slot sequence between requests Only the first TTS request in a backend process succeeded. Every later one failed instantly, in about 0.13 s, with "TTS prompt processing failed" from step_prompt, regardless of streaming or non-streaming and regardless of the text. With LOCALAI_SINGLE_ACTIVE_BACKEND=true the process is kept alive between requests, so a deployment would have served exactly one utterance per backend start. The cause is missing KV hygiene, not anything in the gRPC adapter. TTS slots never enter the shared batch: pre_decode() returns early for them and process_tts_slots() drives them instead, so they skip the prompt-cache bookkeeping that clears a slot's sequence between requests. Nothing in the gen-audio path makes up for it: mtmd_helper_gen_audio_reset only clears host-side buffers, and the pipeline always decodes from position 0 into the sequence identified by slot.id. So the second task on a slot writes positions 0..N over the first task's tokens and llama_decode fails. Fix is one call to slot.prompt_clear(), the same helper the normal path uses, in the SERVER_TASK_TYPE_TTS branch of launch_slot_with_task before set_input. It goes into 0002 rather than a new patch file because it is a defect in the code that patch introduces, and the header now records it as ours so we know whether it still needs carrying if #26603 merges without it. Verified in one backend process, different text on every request: three consecutive non-streaming requests, three consecutive streaming requests, and an interleaved non-streaming, streaming, non-streaming, streaming run. All ten returned HTTP 200 with RIFF ... WAVE audio, Microsoft PCM, 16 bit, mono 24000 Hz, the streamed ones carrying exactly one RIFF header at byte 0, and every output measured as real speech rather than silence or a truncated fragment. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * feat(llama-cpp): expose max_frames for TTS requests The Qwen3-TTS backbone does not always emit <|codec_eos_token|>, and when it does not, generation runs to upstream's 512-frame n_predict default. At the model's 12.5 Hz frame rate that is 40.96 s of audio, which a short input can trigger: one request in this session produced 40.96 s for a ten-word sentence. prepareTTSTask hardcoded n_predict to -1, so callers had no way to bound it. Add a max_frames key alongside top_k and top_p, parsed with the same strict whole-string parsing so a typo is an error rather than a silently truncated value, and rejected with a field-naming message when negative. 0 keeps the existing sentinel convention and means unset, so a request that omits it behaves exactly as before. Named max_frames rather than n_predict because frames are what the parameter means at a TTS endpoint: one frame is 0.08 s of audio. The 512-frame default is deliberately unchanged. Lowering it would truncate legitimately long inputs, which is a worse failure than an occasionally overlong one. Verified end to end on one text of thirty words: max_frames=25 HTTP 200, 96044 bytes, 2.00 s, exactly 25 frames max_frames=50 HTTP 200, 192044 bytes, 4.00 s, exactly 50 frames no max_frames HTTP 200, 572204 bytes, 11.92 s, stopped at its own codec EOS after 149 frames, unchanged behavior max_frames=-1 InvalidArgument "max_frames must be >= 0, got \"-1\"" max_frames=many InvalidArgument "max_frames must be an integer, got \"many\"" Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(llama-cpp): send the TTS sample rate up front, and tidy three review items Four items from the Task 4 review. Streaming first-byte latency. TTSStream sent the sample-rate reply only once the first audio result arrived, and a chunk needs a whole 72-frame window, roughly 5.8 s of audio and far longer in wall time on CPU. The Go side blocks on that reply before it can emit the WAV header, so a streaming client sat at zero bytes for the whole stretch. The rate is a property of the loaded model and is available synchronously from mtmd_gen_audio_get_info, so it now goes out immediately after post_task and the rate_sent bookkeeping is gone. Measured on a warm model, first byte drops from 30.48 s to 0.014 s, and the output is still a valid WAV with exactly one RIFF header at byte 0. Unchecked close. The non-streaming path ignored ofstream::close(), so a failure that only surfaces on flush was reported as success while leaving a truncated file at dst. It now returns INTERNAL like the other write failures. Wrong comment on set_lang. gen_audio::inp::get() already maps a stored blank to nullptr, so our guard is behavior-preserving, not behavior-fixing. The comment claimed otherwise; the code was right. Repetition penalty. penalty_last_n = -1 is inert at this pin, because llama_sampler_init_penalties clamps it with std::max(penalty_last_n, 0) and then builds a disabled sampler, so the 1.05 penalty never applies. Upstream's README attributes looping to a missing repeat_penalty, so it was worth testing as a root-cause fix for the model running to the frame cap. Dropping the line lets the sampling default of 64 apply, which was confirmed in the sampler chain trace as penalty_last_n = 64 with repeat_penalty = 1.050. Over 15 uncapped short requests each way it did not help: 0 of 15 ran to the cap with the penalty inert, 1 of 15 with it active. Both lines are therefore kept for parity with upstream's draft, and a comment now records that the pair is inert and why, so the next reader does not believe a penalty is applied. max_frames remains the way to bound output. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * build(llama-cpp): let unpatched forks opt out of the TTS task turboquant and bonsai copy grpc-server.cpp into llama.cpp forks that do not carry our patches. disable-tts-task.sh injects the same kind of preprocessor switch disable-score-task.sh already uses, so those builds answer UNIMPLEMENTED rather than failing to compile. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(config): keep a TTS speaker-encoder projector out of vision detection Task 1 exempted a declared-TTS model's mmproj from VisionSupported, but the first real gallery entry with an mmproj still came back vision-capable through two paths the earlier fix did not close. GuessUsecases has no FLAG_VISION branch, so it falls through to true for any chat-ish model. That is not just a wrong answer at the call site: syncKnownUsecasesFromString rewrites KnownUsecaseStrings from HasUsecases, and the loader calls it more than once per config file, so the guessed FLAG_VISION is written out and parsed back into KnownUsecases as if the operator had declared it. Give GuessUsecases a FLAG_VISION branch that defers to the same explicit signals VisionSupported uses. Second, llama.cpp builds an mtmd context for the speaker-encoder projector and reports its media marker on the first chat probe, which resurrected vision after the model had been used once. Apply the same declared-TTS exemption to MediaMarker that the mmproj check already had. Verified against the qwen3-tts-llamacpp-q4 gallery entry: no vision capability and no image input modality, before load, after a TTS request, and after a chat probe. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * feat(gallery): add Qwen3-TTS entries for the llama-cpp backend Two entries over upstream's own GGUF conversion, Q8_0 and Q4_K_M, each pairing a backbone with the Q8_0 projector. Named to sit alongside the existing qwen3-tts-cpp entries rather than replace them. Also tags the llama-cpp backend text-to-speech / TTS so the backend browser surfaces the capability. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * docs: cover Qwen3-TTS on the llama-cpp backend Adds the gallery variants, the two-file mmproj configuration, the required voice reference, and the language and sampling knobs. Also corrects the streaming-support list, which named only voxcpm. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(config): register llama-cpp as a TTS and voice-cloning backend The branch taught the llama-cpp backend to serve Qwen3-TTS and shipped two gallery entries for it, but never told the capability table. llama-cpp still declared only the text RPCs and usecases, so: - VoiceCloningForModel returned nil at the capability check, before it ever reached the model's own tts.voice_cloning override, and /tts answered 400 "selected model does not support reference-audio voice cloning" for any localai://voice-profiles/... voice. No model YAML could opt back in. - GET /api/backends/usecases did not list tts for llama-cpp, so the gallery greyed out the TTS filter for the entries this branch adds. - The React TTS page saw voice_cloning: null and kept both models out of the Voice Library. Add the TTS RPCs and usecase, and the reference-audio contract. The contract needs narrowing, because the per-backend switch in VoiceCloningForModel ends in a permissive default: an unnarrowed entry would have advertised reference-audio cloning on every GGUF chat model in the gallery. Narrow on the declared TTS usecase rather than the model name. The TTS checkpoints are the only llama-cpp models carrying known_usecases: [tts]; name matching would have to guess at third-party repacks, and "base", the substring the neighbouring Qwen and vLLM cases key on, is a routine word in text-model names. The check reads the declared bit directly instead of going through HasUsecases, which falls through to GuessUsecases and would hand the decision to a heuristic that never had a llama.cpp TTS model in mind. DefaultUsecases stays [chat]: a bare GGUF served by llama.cpp is a chat model, and both the gallery filter and the importer read that field. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(gallery): declare what nemotron-3-nano-omni actually accepts The entry is backend: vllm-omni with known_usecases: [chat, completion], no mmproj and no media marker, so it used to report vision only through the blanket GuessUsecases fallthrough that the vision branch in this branch removed. Nemotron 3 Nano Omni is a multimodal understanding model: image, video and audio in, text out. Declaring that is what the sibling vllm-omni-qwen3-omni-30b already does. known_usecases gains vision only. FLAG_VIDEO is video GENERATION, an output modality, and this model generates none; video and audio input belong in known_input_modalities, which is where AudioInputSupported and VideoInputSupported read them from. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(importers): import a Qwen3-TTS GGUF repo as TTS, not chat The llama-cpp importer hardcodes known_usecases: [chat] and assigns any mmproj-matching file as a vision projector, so ggml-org/Qwen3-TTS-12Hz-1.7B- Base-GGUF imported as a chat model with vision. Both fields were wrong, and the model was unreachable from /tts and from the Voice Library. Filenames cannot fix this. A Qwen3-TTS repo has the exact shape of a vision repo, one backbone GGUF plus one mmproj-*.gguf, so the projector's own header is the only honest signal: mtmd writes clip.has_gen_audio_encoder for the projectors it can drive as a speech pipeline and refuses to build one without it. Probe the selected mmproj for that flag, reusing the range-fetch the MTP detection already does, and declare tts when it is set. The mmproj assignment then stops reading as vision on its own, since a declared-TTS model already exempts its projector from vision detection. The probe is best-effort like the MTP one: a network blip leaves the chat default in place rather than failing the import. Verified against the real artifacts on disk: the Qwen3-TTS projector reports gen-audio, its backbone does not. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(llama-cpp): stop non-TTS models crashing on the new pin Two regressions, both hit every ordinary llama-cpp model and neither was caught locally because every test on this branch loaded a TTS model. The first is a null dereference. server_slot::tts_ctx::reset() called mtmd_helper_gen_audio_reset() unconditionally, but the gen-audio pipeline is only allocated for models carrying a gen-audio mmproj, and upstream's implementation reads ctx->pipeline before null-checking anything. Since server_slot::reset() runs during slot initialization for every model, any non-TTS model segfaulted the backend the moment it loaded. Guard the call on the is_supported() predicate already defined beside it, and keep the plain field resets unconditional. The second is unrelated to TTS and came in with the pin bump. PredictOptions.Penalty is a bare proto float, so a caller that names no repetition penalty sends 0 rather than omitting the field. Since 9de0fcf2b, common_sampler_init() rejects a non-positive penalty_repeat outright because it would divide logits by zero, turning every such request into "Failed to initialize samplers". Treat 0 as unset and leave llama.cpp's own neutral default in place. Verified with the same suite CI runs, which is what caught both: tests/e2e-backends passes 6 of 6 including the load and predict specs that were red. Qwen3-TTS still synthesises on both paths, 24 kHz mono 16-bit WAV with exactly one RIFF header on the streamed output. Assisted-by: Claude:claude-fable-5 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> --------- Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Co-authored-by: Ettore Di Giacinto <mudler@localai.io> |
||
|
mudler's LocalAI [bot]andmudler
|
7e4a60c701 |
chore: ⬆️ Update TheTom/llama-cpp-turboquant to 8a891f4b566efdbd3cea92fafee3227a0a267683 (#11258)
⬆️ Update TheTom/llama-cpp-turboquant Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: mudler <2420543+mudler@users.noreply.github.com> |
||
|
Richard Palethorpe
|
49ef40a187 |
feat(classifier/VAD): support voice control on low power devices (#10804)
* feat(llama-cpp): route Score through the slot loop Score previously bypassed the slot loop with a direct llama_decode: a conflict guard aborted the whole process if scoring raced generation, the config validator had to reject score alongside chat/completion/embeddings, and every candidate re-decoded the full shared prompt. Add SERVER_TASK_TYPE_SCORE to the (patched) upstream server so score tasks are scheduled like any other slot work: generation and scoring serialize naturally, the shared prompt is decoded once per call, and the slot's prompt cache carries the conversation prefix across calls. Context checkpoints at the score boundary and at the cache-divergence point keep SWA/hybrid/recurrent models (e.g. LFM2.5) from re-prefilling the whole prompt per candidate: warm-turn scoring on a 6-option set drops from ~8s to ~0.5s on a desktop CPU. The conflict guard and the validation split are removed; declaring score with generation usecases on one config is now supported and shares the slot cache. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * feat(realtime): classifier wire types and pipeline config Wire types and YAML config for realtime classifier mode: sessions carry a localai_classifier extension (options with canned replies/tool calls, softmax threshold, normalization, history trimming, fallback modes, and a deterministic wake-word address gate), mirrored by pipeline.classifier in the model YAML and surfaced in the config-meta registry. The localai.classifier.result server event reports the full score distribution per turn. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * feat(realtime): classifier response flow Classifier-mode responses: instead of autoregressive generation, each user turn is prefill-scored against the option list (router.ScoreClassifier prompt/candidate shapes over the Score primitive) and the winning option's canned reply and tool call are emitted through the existing response machinery. Below-threshold turns take the configured fallback (none / canned reply / generate); empty transcripts and unaddressed turns (wake word not mentioned) skip scoring entirely. The scoring probe defaults to the latest user message only — small scorers echo canned replies from prior turns back as the top option otherwise. Built for hardware that can afford prompt processing but not decode: with slot-based Score the option list stays KV-cached across turns, so a turn costs roughly one forward pass over the new words. session_update_error events now carry the validation cause instead of a generic message. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(realtime): bound the VAD tick's scan window and buffer retention The VAD tick loop re-scanned the entire input buffer every 300ms and only trimmed it on zero-segment ticks or commits. Audio that keeps producing segments without a committing pause (steady noise a mic pipeline lets through, music, continuous speech) grew the buffer toward the 100MB cap with each tick rescanning all of it — O(n^2), measured at ~3.3ms of silero per buffered second: past ~90s retained, ticks run back to back and pin ~4 cores until the stream stops. Silero's recurrent state only carries a few hundred ms of context, so rescanning old audio buys nothing. Clip the slice handed to the VAD to the largest silence the commit test can need to measure (server_vad silence window or the semantic eagerness fallback) plus a warm-up margin, and rebase the returned segment times so every downstream consumer keeps whole-buffer coordinates. An open turn whose clipped window is all silence now commits (the silence outran the window) instead of being discarded as no-speech. Independently, retain at most 90s of raw buffer, rebasing the live-feed and EOU cursors on trim — this also bounds the previously unbounded VAD-error path. Turn boundaries are otherwise unchanged: no forced commits, no new coordinator states. pipeline.turn_detection.vad_window_sec can widen the scan window; values below the automatic floor are ignored. The tick body is extracted into vadTick so specs can drive turn detection synchronously (same shape as classifySoundWindow); the babble reproduction that pinned 4 cores now plateaus under 10% of one core. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(backend): let per-model threads override the global default ModelOptions overrode a set per-model threads value with the app-level --threads whenever the latter was non-zero — and WithThreads defaults it to the physical core count, so it always was. The YAML threads: knob has been dead config: a tiny VAD model could never opt down from the global pool size. SetDefaults already fills an unset per-model value from the app config, which is the intended precedence; resolve threads through a helper that honors it (explicit threads: 0 still means unset). Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * chore(gallery): single-thread the silero VAD Silero is a ~2MB recurrent model with no exploitable graph parallelism: measured per-call latency is identical at 1 and 10 ORT threads, while every extra pool thread just spin-waits between the realtime loop's frequent tiny inferences. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * docs(realtime): classifier mode, VAD scan window, threads precedence Document the realtime classifier mode (options, threshold guidance, wake-word address gate, empty-transcript handling), the VAD scan window and 90s buffer retention (pipeline.turn_detection.vad_window_sec), the per-model threads precedence, and the M3 classifier note in the realtime state-machine design doc. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * perf(llama-cpp): score all candidates in one batched decode One scoring call is now a single SERVER_TASK_TYPE_SCORE task: the slot decodes the shared prefix (prompt + longest common candidate token prefix) once, then forks one sequence per candidate off it (metadata-only for the unified KV cache, copy-on-write for recurrent state) and decodes every candidate's unique tail in one llama_decode. Previously each candidate was its own task that restored the boundary checkpoint and re-decoded its full tail sequentially, paying per-candidate task and decode overhead. The context reserves SERVER_SCORE_FORK_SEQS extra sequence ids (and recurrent-state cells) beyond the parallel slots via the new common_params::n_seq_score_forks. Forking requires the unified KV cache (already this backend's default) since per-sequence streams would shrink n_ctx_seq; an explicit kv_unified:false disables forking and Score calls that need it fail cleanly. Candidates beyond the fork/output budget decode in successive chunks. Wire contract and scores are unchanged: per-token logprobs are stitched from the shared region and the forked tails. Verified bitwise deterministic call-to-call and independent of candidate order (no cross-fork leakage via equal-length candidate swap); ranking matches the per-candidate implementation on the drone battery (winner softmax 0.99996 vs 0.99997), and >16-candidate chunking, prefix-of-another and empty candidates all pass. Measured on a desktop CPU: warm /api/score calls 0.52s -> 0.23s; warm realtime classifier turns 196-303ms. The 9-candidate drone turn decodes ~17 unique tail tokens in one batch instead of nine sequential ~220ms checkpoint-restore tasks. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(realtime): gate scoring capacity by model usecase Reserve llama.cpp scoring slots only for models that explicitly declare the score usecase, while allowing score to coexist with chat and completion. Reject incompatible unified-KV settings and classifier activation on models without scoring capacity. Propagate application defaults when resolving realtime and preload pipeline stages so unset thread counts are resolved consistently without overriding explicit model settings. Assisted-by: Codex:gpt-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(ci): honor APT mirrors in the prebuilt llama-cpp compile step The builder-prebuilt path installs gcc-14 with apt directly and ignored the APT_MIRROR/APT_PORTS_MIRROR build args the from-source path already honors, so an ubuntu mirror outage broke every arm64 backend build. Pass the args into the stage and run apt-mirror.sh (already in the build context via COPY . /LocalAI) before the apt step. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * feat(realtime): classifier argument slots via constrained completion Hybrid classify-then-complete: a classifier option's canned tool call can declare typed argument slots (number | enum | string, with defaults and prompt hints) referenced as "{{name}}" in the arguments template. When the option wins, the slots are filled by a short grammar-constrained completion that continues the exact scoring prompt — rendered by the same cached ScoreClassifier, so the llama.cpp prompt cache is already warm — with the chosen route JSON re-opened at the first slot field. A GBNF grammar pins the field skeleton and frees only the values; temperature 0, a couple dozen tokens at most (~300ms on a desktop CPU for two slots). Slot declarations and hints ride the option descriptions in the shared system prompt, informing scoring and the fill alike at no per-turn token cost. The localai.classifier.result event carries the final arguments and a fill_latency_ms. On inference failure the slots' defaults apply; a slot without a default fails the response (or falls through with fallback.mode: generate). Slot filling requires completion alongside score in the scoring model's known_usecases. Verified end-to-end on the Pi drone demo: "fly forward three meters" in distance mode classifies forward and infers {"distance": 3, "units": "meters"} in ~310ms, and the drone flies exactly 3 units. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * feat(realtime): splice filled slot values into classifier replies A classifier option's spoken reply can now reference its tool's argument slots ("Going forward {{distance}} {{units}}."): the values inferred by the slot-fill completion — or the recovery defaults — are spliced into the reply as plain text before it is emitted, so what the assistant says confirms what it actually inferred. Placeholders without a value stay literal, and options without slots are untouched. FillToolArguments now returns the raw slot values alongside the spliced arguments JSON to make the reply templating possible. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(realtime): harden classifier slot completion Reserve context for constrained slot filling, size completions from their encoded output, and encode enum grammar literals as valid JSON. Reject empty enum values and cover the failure modes with regression tests. Assisted-by: Codex:gpt-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * feat(realtime): prewarm the classifier scoring prompt on registration Swapping a session's classifier option list (a voice-switched command mode, for instance) made the next turns pay a full re-prefill of the new option-list prompt — measured 2.4s vs 0.3s warm on a desktop CPU, and worse: on hybrid-memory models like LFM2.5, whose state cannot be partially rewound (llama.cpp can only restore checkpoints), *every* probe change re-prefilled from scratch whenever the last checkpoint missed the probe boundary, so even same-list turns intermittently cost full prefills. Registering an option list (pipeline seed or session.update) now fires a best-effort background prewarm: two throwaway scores with distinct probes. The first prefills the new option-list prompt; the second, diverging exactly where per-turn probe text starts, plants the backend's rewind point (KV checkpoint) at the stable-prefix boundary that every real turn reuses. The prewarm hides behind the canned mode-switch reply — by the time it finishes speaking, the cache is warm. Idempotent per option set, detached from the registering request's lifetime. Measured on the drone demo (LFM2.5-1.2B, desktop CPU): first turn after a mode switch 2374ms -> 340ms; intermittent same-list full prefills (1.3-2.1s) all -> under 0.5s. For clients that swap lists frequently, options: [parallel:2] on the scoring model additionally keeps one slot per list via prefix-similarity routing (+26MB RSS, unified KV). Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * perf(llama-cpp): checkpoint scoring at the caller-declared stable prefix Hybrid-memory models (LFM2.5 shortconv, Qwen3.5 deltanet — where new small models are headed) cannot rewind their state, so any prompt-cache reuse that needs a rewind falls back to a full re-prefill. For classifier scoring that meant every probe change re-processed the whole option-list prompt: the server's checkpoints were placed reactively (at wherever the previous task happened to diverge), so a checkpoint past the next divergence was erased rather than restored — measured as intermittent 2-10s turns on prompts with a 95%+ common prefix. The classifier now computes the probe-invariant prompt prefix once (the byte-wise common prefix of two synthetic probe renders) and declares its length with every Score request; the server maps it to a token boundary and forces a KV checkpoint exactly there on each score prefill. That checkpoint sits at or before every future divergence under the same option list, so it always survives and always restores — repeat scoring costs probe+candidates regardless of how the probe changes. Also: - prewarm reruns on every option-list registration instead of memoizing per list: with boundary checkpoints a redundant rewarm costs two probe-sized decodes, while skipping one after a slot eviction (three lists sharing fewer slots evict in LRU cascades) silently moves a full re-prefill onto the user's next turn - new llama.cpp backend option rs_seq:N exposes bounded recurrent-state rollback outside speculative decoding; measured impractical for deltanet-scale states (65GB for 64 snapshots on Qwen3.5-4B) but cheap insurance for small-state models - docs: the multi-list recipe (parallel:N + sps:0.5 — the default slot similarity threshold funnels distinct lists onto one slot) Measured on the drone demo (LFM2.5-1.2B scorer, desktop CPU), steady state: every turn 285-421ms including mode switches, vs 2.4s post-switch and intermittent 1.3-2.9s re-prefills before. Assisted-by: Claude:claude-fable-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(realtime): align classifier cache guidance Document the single-score prewarm behavior and clean the vendored score patch formatting. Assisted-by: Codex:gpt-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(llama-cpp): guard score task for fork backends TurboQuant and Bonsai reuse the primary gRPC server against llama.cpp forks that do not carry LocalAI's slot-based Score patches. Compile the Score integration only for the patched primary backend and return UNIMPLEMENTED from fork builds instead of referencing absent task types and common_params fields. Assisted-by: Codex:gpt-5 [gh] Signed-off-by: Richard Palethorpe <io@richiejp.com> * fix(dev): generate gRPC code before commit lint The coverage phase regenerates ignored protobuf bindings, but lint runs first and can fail against missing or stale output. Generate the pinned bindings before lint so the gate always type-checks the current schema. Assisted-by: Codex:gpt-5 Signed-off-by: Richard Palethorpe <io@richiejp.com> --------- Signed-off-by: Richard Palethorpe <io@richiejp.com> |
||
|
|
9fbb8e89cf |
fix(turboquant): supersede stale dependency bump (#11064)
* ⬆️ Update TheTom/llama-cpp-turboquant Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(turboquant): refresh HIP compatibility patch The updated fork now carries its own HIP-safe peer-copy path, so the old hunk no longer applies. Keep only the event-creation compatibility change that the fork still needs. Assisted-by: Codex:gpt-5 [Codex] --------- Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: mudler <2420543+mudler@users.noreply.github.com> Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com> |
||
|
LocalAI [bot]andEttore Di Giacinto
|
1f53dff436 |
fix(turboquant,bonsai): do not apply vendored llama.cpp patches to fork trees (#10866)
The turboquant and bonsai backends copy backend/cpp/llama-cpp/ wholesale into their build directories and reuse its Makefile/prepare.sh against their own llama.cpp forks. When PR #10837 added backend/cpp/llama-cpp/patches/0001-add-minimax-m3-support.patch, the copied patches/ directory was mis-applied to the fork checkouts: the fork trees diverge from upstream, hunks rejected, and because the patch-apply loop in prepare.sh ran before set -e took effect the build kept going and died much later with a confusing compile error ("'LLM_ARCH_MINIMAX_M3' was not declared in this scope"). This broke tests-turboquant-grpc on that PR. Two hardening changes: - turboquant/bonsai Makefiles: delete the copied patches/ directory right after the cp -rf of backend/cpp/llama-cpp/. Patches vendored for upstream llama.cpp must never be applied to the forks; each fork carries its own patch series under backend/cpp/<backend>/patches/, applied by its apply-patches.sh. - llama-cpp prepare.sh: run the patch-apply loop under set -e so a rejecting patch fails fast and loudly at apply time instead of surfacing as a downstream compile error. A missing or empty patches/ directory remains a no-op success, so all existing callers (the llama-cpp Makefile targets and the turboquant/bonsai copies) are unaffected when no patches ship. Exposed by PR #10837. 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> |
||
|
LocalAI [bot]andEttore Di Giacinto
|
4ac67d255d |
feat: single-build ggml CPU_ALL_VARIANTS for llama-cpp + turboquant (x86/arm64/apple) (#10497)
* feat(llama-cpp): single x86 CPU build via ggml CPU_ALL_VARIANTS
Replace the per-microarch avx/avx2/avx512/fallback multi-binary build on
x86 with a single grpc-server plus the dlopen-able libggml-cpu-*.so set
that ggml's backend registry selects at runtime by probing host CPU
features. One build instead of four, broader microarch coverage (adds
alderlake AVX-VNNI, zen4 AVX512-BF16, sapphirerapids AMX), and the
shell-side /proc/cpuinfo probing in run.sh goes away.
Build/link notes:
- CPU_ALL_VARIANTS requires GGML_BACKEND_DL + BUILD_SHARED_LIBS=ON, so
ggml/llama become shared objects. SHARED_LIBS is now a make variable
(default OFF) so the override survives the recursive sub-make into the
VARIANT build dir instead of being re-clobbered by the base flags.
- The cpu-all target also builds "--target ggml": the per-microarch
backends are runtime-dlopened, not link deps, so they only compile via
ggml's add_dependencies().
- hw_grpc_proto is pinned STATIC. Under BUILD_SHARED_LIBS=ON it would
otherwise become a DSO referencing hidden-visibility symbols in the
static libprotobuf.a, which fails to link ("hidden symbol ... is
referenced by DSO"). Keeping it static links gRPC/protobuf into the
executable while only ggml/llama stay shared, so no PIC or base-image
change is required.
- package.sh bundles the libggml-*.so set into package/lib; ggml finds
them by scanning the bundled ld.so directory (/proc/self/exe), which
run.sh launches from.
Scope: x86 only. arm64/darwin keep the single fallback build. The
ik-llama-cpp / turboquant forks and the other ggml C++ backends are
unchanged; the same recipe applies but is out of scope here.
Validated with a full docker build plus a live inference smoke test:
the model loads, ggml selects the AVX512_BF16 variant on a Zen-class
host, and tokens generate correctly.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(llama-cpp,turboquant): extend CPU_ALL_VARIANTS to arm64 + turboquant
- llama-cpp: x86 AND arm64 now use the single llama-cpp-cpu-all build
(only hipblas keeps the fallback build). ggml's arm64 variant table
(armv8.x / armv9.x, plus apple_m* on darwin) is selected at runtime.
- turboquant: same recipe via a turboquant-cpu-all target. turboquant
copies backend/cpp/llama-cpp's CMakeLists.txt + Makefile per flavor, so
the hw_grpc_proto STATIC fix and the SHARED_LIBS / EXTRA_CMAKE_ARGS
make-vars are inherited; the target just passes SHARED_LIBS=ON, the DL
flags and --target ggml through, then collects the .so set. run.sh and
package.sh updated to ship/select turboquant-cpu-all.
- Makefile lib-collection find now also matches *.dylib (for the darwin
build, which emits dylibs rather than .so).
ik-llama-cpp is intentionally left unchanged: its pinned ggml has no
CPU_ALL_VARIANTS support and its IQK kernels require AVX2, so the
per-microarch dynamic backend set does not apply.
Scope still excludes the darwin packaging wiring (separate change).
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* feat(llama-cpp,turboquant): arm64 gcc-14 for SME variants + darwin cpu-all packaging
- arm64: ggml CPU_ALL_VARIANTS builds armv9.2 SME variants whose -march=...+sme
is rejected by the Ubuntu 24.04 default gcc-13. Build the arm64 variants with
gcc-14 (installed in the compile step). The host only selects a variant it
actually supports at runtime, but every variant must still compile.
- darwin: scripts/build/llama-cpp-darwin.sh builds llama-cpp-cpu-all instead of
the fallback binary, keeps Metal (GGML_METAL stays ON; --target ggml also builds
ggml-metal). The per-microarch libggml-cpu-*.dylib are placed in the package
root next to the binary (darwin has no bundled ld.so, so ggml's executable-dir
scan looks there), while the other shared dylibs go in lib/ for DYLD_LIBRARY_PATH.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* fix(llama-cpp-darwin): distribute ggml backends by suffix (.so root, .dylib lib)
ggml emits its loadable backends (per-microarch CPU variants, metal, blas) with a
.so suffix even on darwin, while the core libraries (ggml-base/ggml/llama/
llama-common/mtmd) use .dylib. Split the distribution by suffix: .so DL backends
go in the package root for ggml's executable-directory scan, .dylib core libs go
in lib/ for DYLD_LIBRARY_PATH. The previous .dylib name-pattern matched none of the
variants.
Verified on an M4: ggml loads the apple_m4 CPU variant (SME=1) and Metal, model
loads and generates correct tokens.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* fix(llama-cpp,turboquant): only CPU_ALL_VARIANTS for pure-CPU builds, GPU uses fallback
The previous gate sent every non-hipblas build through llama-cpp-cpu-all, so the
GPU image builds (cublas, sycl_f16/f32, vulkan, nvidia l4t) compiled the whole CPU
microarch variant matrix on top of their already-huge GPU backend - blowing the
build time (the sycl job was only 59% done after 2h11m) - and the arm64 l4t build
failed at `apt-get install gcc-14` (exit 100) on the Jetson base.
Gate on an empty BUILD_TYPE instead: only the pure CPU image (build-type: '' in
.github/backend-matrix.yml) builds the CPU_ALL_VARIANTS set; every GPU build gets a
single fallback CPU grpc-server, since the accelerator does the compute. This also
confines the arm64 gcc-14 step (needed for the armv9.2 SME variants) to the CPU
build, away from the GPU base images.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
* docs(llama-cpp): correct run.sh comment for arm64/darwin cpu-all
arm64 and darwin CPU images now also ship llama-cpp-cpu-all (not fallback-only);
only GPU images ship fallback-only. Fix the stale comment to match.
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>
|
||
|
LocalAI [bot]andEttore Di Giacinto
|
7402d1fd20 |
chore(turboquant): bump to 7d9715f1 + fix compilation against rebased fork (#10205)
* chore(turboquant): bump TheTom/llama-cpp-turboquant to 7d9715f1 Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Assisted-by: Claude:claude-opus-4-8 [Claude Code] * fix(turboquant): drop obsolete legacy-spec shim after fork rebased The TheTom/llama-cpp-turboquant fork (pin c9aa86a) rebased past the upstream common_params_speculative refactor (ggml-org/llama.cpp #22397/#22838/#22964), the model_tgt rename (#22838) and get_media_marker (#21962). The old fork-compat shim forced now-wrong legacy code paths, breaking the build with errors like 'struct common_params_speculative has no member named mparams_dft / type' and 'server_context_impl has no member named model'. Remove the obsolete LOCALAI_LEGACY_LLAMA_CPP_SPEC branches from the shared grpc-server.cpp (stock llama-cpp and the modern fork both take the modern path now), and narrow the one remaining gap (the fork still lacks common_params::checkpoint_min_step) to a dedicated LOCALAI_TURBOQUANT_NO_CHECKPOINT_MIN_STEP guard injected by patch-grpc-server.sh. The patch script now only adds the turbo2/3/4 KV-cache types and injects that one macro. Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Assisted-by: Claude:claude-opus-4-8 [Claude Code] * fix(turboquant): HIP-port the fork's CUDA additions (copy2d 3D-peer + cudaEventCreate) The turboquant fork adds/modifies a few ggml-cuda.cu spots with CUDA APIs that ggml's HIP/MUSA shim does not provide, breaking the -gpu-rocm-hipblas-turboquant build. patches/0001-hip-guard-copy2d-peer-fastpath.patch (applied by apply-patches.sh) ports them: - Guard ggml_cuda_copy2d_across_devices's 3D-peer copy fast path with #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) so HIP/MUSA fall through to the existing cudaMemcpyAsync staging fallback (HIP genuinely lacks cudaMemcpy3DPeerAsync, per the fork's own comment). - Create the device event in ggml_backend_cuda_device_event_new with the HIP-aliased cudaEventCreateWithFlags(.., cudaEventDisableTiming) instead of the un-aliased plain cudaEventCreate, matching this file's own usage elsewhere. CUDA builds are unaffected. Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Assisted-by: Claude:claude-opus-4-8 [Claude Code] * ci(turboquant): drop the ROCm/hipblas build flavor The TheTom/llama-cpp-turboquant fork is not ROCm-clean at the current pin: beyond the CUDA-API gaps already patched (3D-peer copy, cudaEventCreate), its llama.cpp base fails to compile the flash-attention MMA f16 kernels for head-dim 640 under HIP (cols_per_warp evaluates to 0 -> division-by-zero / non-constant static asserts in fattn-mma-f16.cuh). That is a deep ggml-on-ROCm kernel issue, not something a small fork patch can paper over. Drop -gpu-rocm-hipblas-turboquant from the build matrix so turboquant still ships for cpu / cublas / vulkan / sycl. Re-add it once the fork's HIP path compiles (or upstream ggml fixes the large-head-dim MMA kernels for ROCm). 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> |
||
|
LocalAI [bot]andmudler
|
ddbbdf45b9 |
chore: ⬆️ Update TheTom/llama-cpp-turboquant to 5aeb2fdbe26cd4c534c6fa15de73cb5749bd0403 (#9740)
⬆️ Update TheTom/llama-cpp-turboquant Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: mudler <2420543+mudler@users.noreply.github.com> |
||
|
LocalAI [bot]andmudler
|
a315c321c1 |
chore: ⬆️ Update TheTom/llama-cpp-turboquant to 69d8e4be47243e83b3d0d71e932bc7aa61c644dc (#9638)
⬆️ Update TheTom/llama-cpp-turboquant Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: mudler <2420543+mudler@users.noreply.github.com> |
||
|
LocalAI [bot]andmudler
|
f5c268deac |
chore: ⬆️ Update TheTom/llama-cpp-turboquant to 11a241d0db78a68e0a5b99fe6f36de6683100f6a (#9571)
⬆️ Update TheTom/llama-cpp-turboquant Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: mudler <2420543+mudler@users.noreply.github.com> |
||
|
LocalAI [bot]andmudler
|
806ea24ff4 |
chore: ⬆️ Update TheTom/llama-cpp-turboquant to 67559e580b10e4e47e9a6fd6218873997976886d (#9497)
⬆️ Update TheTom/llama-cpp-turboquant Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: mudler <2420543+mudler@users.noreply.github.com> |
||
|
Ettore Di Giacinto
|
369c50a41c |
fix(turboquant): drop ignore-eos patch, bump fork to b8967-627ebbc (#9423)
* fix(turboquant): drop ignore-eos patch, bump fork to b8967-627ebbc
The upstream PR #21203 (server: respect the ignore_eos flag) has been
merged into the TheTom/llama-cpp-turboquant feature/turboquant-kv-cache
branch. With the fix now in-tree, 0001-server-respect-the-ignore-eos-flag.patch
no longer applies (git apply sees its additions already present) and the
nightly turboquant bump fails.
Retire the patch and bump the pin to the first fork revision that carries
the merged fix (tag feature-turboquant-kv-cache-b8967-627ebbc). This matches
the contract in apply-patches.sh: drop patches once the fork catches up.
* fix(turboquant): patch out get_media_marker() call in grpc-server copy
CI turboquant docker build was failing with:
grpc-server.cpp:2825:40: error: use of undeclared identifier
'get_media_marker'
The call was added by
|
||
|
Ettore Di Giacinto
|
5837b14888 |
chore: ⬆️ Update TheTom/llama-cpp-turboquant to `45f8a066ed5f5bb38c695cec532f6cef9f4efa9d' (#9385)
chore: ⬆️ Update TheTom/llama-cpp-turboquant to `45f8a066ed5f5bb38c695cec532f6cef9f4efa9d`
Drop 0002-ggml-rpc-bump-op-count-to-97.patch; the fork now has
GGML_OP_COUNT == 97 and RPC_PROTO_PATCH_VERSION 2 upstream.
Fetch all tags in backend/cpp/llama-cpp/Makefile so tag-only commits
(the new turboquant pin is reachable only through the tag
feature-turboquant-kv-cache-b8821-45f8a06) can be checked out.
|
||
|
Ettore Di Giacinto
|
95efb8a562 |
feat(backend): add turboquant llama.cpp-fork backend (#9355)
* feat(backend): add turboquant llama.cpp-fork backend
turboquant is a llama.cpp fork (TheTom/llama-cpp-turboquant, branch
feature/turboquant-kv-cache) that adds a TurboQuant KV-cache scheme.
It ships as a first-class backend reusing backend/cpp/llama-cpp sources
via a thin wrapper Makefile: each variant target copies ../llama-cpp
into a sibling build dir and invokes llama-cpp's build-llama-cpp-grpc-server
with LLAMA_REPO/LLAMA_VERSION overridden to point at the fork. No
duplication of grpc-server.cpp — upstream fixes flow through automatically.
Wires up the full matrix (CPU, CUDA 12/13, L4T, L4T-CUDA13, ROCm, SYCL
f32/f16, Vulkan) in backend.yml and the gallery entries in index.yaml,
adds a tests-turboquant-grpc e2e job driven by BACKEND_TEST_CACHE_TYPE_K/V=q8_0
to exercise the KV-cache config path (backend_test.go gains dedicated env
vars wired into ModelOptions.CacheTypeKey/Value — a generic improvement
usable by any llama.cpp-family backend), and registers a nightly auto-bump
PR in bump_deps.yaml tracking feature/turboquant-kv-cache.
scripts/changed-backends.js gets a special-case so edits to
backend/cpp/llama-cpp/ also retrigger the turboquant CI pipeline, since
the wrapper reuses those sources.
* feat(turboquant): carry upstream patches against fork API drift
turboquant branched from llama.cpp before upstream commit 66060008
("server: respect the ignore eos flag", #21203) which added the
`logit_bias_eog` field to `server_context_meta` and a matching
parameter to `server_task::params_from_json_cmpl`. The shared
backend/cpp/llama-cpp/grpc-server.cpp depends on that field, so
building it against the fork unmodified fails.
Cherry-pick that commit as a patch file under
backend/cpp/turboquant/patches/ and apply it to the cloned fork
sources via a new apply-patches.sh hook called from the wrapper
Makefile. Simplifies the build flow too: instead of hopping through
llama-cpp's build-llama-cpp-grpc-server indirection, the wrapper now
drives the copied Makefile directly (clone -> patch -> build).
Drop the corresponding patch whenever the fork catches up with
upstream — the build fails fast if a patch stops applying, which
is the signal to retire it.
* docs: add turboquant backend section + clarify cache_type_k/v
Document the new turboquant (llama.cpp fork with TurboQuant KV-cache)
backend alongside the existing llama-cpp / ik-llama-cpp sections in
features/text-generation.md: when to pick it, how to install it from
the gallery, and a YAML example showing backend: turboquant together
with cache_type_k / cache_type_v.
Also expand the cache_type_k / cache_type_v table rows in
advanced/model-configuration.md to spell out the accepted llama.cpp
quantization values and note that these fields apply to all
llama.cpp-family backends, not just vLLM.
* feat(turboquant): patch ggml-rpc GGML_OP_COUNT assertion
The fork adds new GGML ops bringing GGML_OP_COUNT to 97, but
ggml/include/ggml-rpc.h static-asserts it equals 96, breaking
the GGML_RPC=ON build paths (turboquant-grpc / turboquant-rpc-server).
Carry a one-line patch that updates the expected count so the
assertion holds. Drop this patch whenever the fork fixes it upstream.
* feat(turboquant): allow turbo* KV-cache types and exercise them in e2e
The shared backend/cpp/llama-cpp/grpc-server.cpp carries its own
allow-list of accepted KV-cache types (kv_cache_types[]) and rejects
anything outside it before the value reaches llama.cpp's parser. That
list only contains the standard llama.cpp types — turbo2/turbo3/turbo4
would throw "Unsupported cache type" at LoadModel time, meaning
nothing the LocalAI gRPC layer accepted was actually fork-specific.
Add a build-time augmentation step (patch-grpc-server.sh, called from
the turboquant wrapper Makefile) that inserts GGML_TYPE_TURBO2_0/3_0/4_0
into the allow-list of the *copied* grpc-server.cpp under
turboquant-<flavor>-build/. The original file under backend/cpp/llama-cpp/
is never touched, so the stock llama-cpp build keeps compiling against
vanilla upstream which has no notion of those enum values.
Switch test-extra-backend-turboquant to set
BACKEND_TEST_CACHE_TYPE_K=turbo3 / _V=turbo3 so the e2e gRPC suite
actually runs the fork's TurboQuant KV-cache code paths (turbo3 also
auto-enables flash_attention in the fork). Picking q8_0 here would
only re-test the standard llama.cpp path that the upstream llama-cpp
backend already covers.
Refresh the docs (text-generation.md + model-configuration.md) to
list turbo2/turbo3/turbo4 explicitly and call out that you only get
the TurboQuant code path with this backend + a turbo* cache type.
* fix(turboquant): rewrite patch-grpc-server.sh in awk, not python3
The builder image (ubuntu:24.04 stage-2 in Dockerfile.turboquant)
does not install python3, so the python-based augmentation step
errored with `python3: command not found` at make time. Switch to
awk, which ships in coreutils and is already available everywhere
the rest of the wrapper Makefile runs.
* Apply suggestion from @mudler
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
---------
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
|