anodyine/vllm-omni
1
0 Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity
2,928 Commits 235 Branches 31 Tags
82955b3a225c6c0a7d6f736cdbaeb09da30e67c4
Commit Graph
507 Commits
Author SHA1 Message Date
tzhouamandClaude Fable 5.1 82955b3a22 [Edge][CPU] Default the code predictor's CPU fast path to fp32 + compiled static step; int8 opt-in 
In-situ parity against the bf16 model (VLLM_OMNI_PREDICTOR_PARITY_CHECK=1):
fp32 control keeps 93.8 % top-1 / 95 % top-5 / KL 0.0015; torch.ao dynamic
int8 falls to 22 % top-1 (max logit diff 7), so per-tensor activation
quantization is not acceptable as a default. Adds the ao-nohead variant
(int8 transformer body, fp32 lm heads) for measurement.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
 2026-09-10 02:54:53 +08:00
tzhouamandClaude Fable 5.1 94485152d2 [Edge][CPU] Predictor fast-path modes: torch.ao int8 (default), torchao weight-only + compiled step, fp32 control; parity metrics 
torchao's dynamic-activation int8 is numerically wrong under torch.compile on
CPU (cosine 0.02 vs fp32; caused runaway generations), so it is not offered;
weight-only + compile is faithful but slower (78 ms vs 36 ms per frame for the
15 sub-steps), so torch.ao dynamic int8 stays the default. Parity mode now also
records top-5 overlap and KL vs the bf16 model; compiled-step equivalence test.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
 2026-09-10 02:44:29 +08:00
tzhouamandClaude Fable 5.1 6ce8de1b28 [Edge][CPU] KV-cached code predictor, int8-dynamic CPU fast path, model-side CPU sampler 
Measured on a Xeon 8480C (12+4 threads, 0.6B): the 15 residual-codebook sub-steps
re-ran the 5-layer predictor over the whole prefix each time (97 ms per frame of
the 134 ms talker step) and vLLM's stock sampler spent 9 ms per step in its
torch.compile'd random-sample kernel on CPU.

- code predictor: forward_cached (KV cache per layer, 2-token prefill then one
  token per sub-step) and a shape-static forward_static_step (tensor position +
  boolean mask) that torch.compile handles as one graph; incremental decoding
  is the default on CPU (VLLM_OMNI_PREDICTOR_INCREMENTAL), optional compile
  (VLLM_OMNI_PREDICTOR_COMPILE=step), optional thread cap.
- CPU fast path: dynamic int8 Linear layers (fp32 activations) for the predictor
  and its lm heads (VLLM_OMNI_PREDICTOR_INT8, default on; micro-benchmark 70 ->
  36 ms per frame for the 15 sub-steps); VLLM_OMNI_PREDICTOR_PARITY_CHECK=1
  records top-1 agreement / max logit diff against the bf16 model in situ.
- Qwen3-TTS talker: prefer_model_sampler on CPU with a sample() that mirrors
  vLLM's Sampler (logits processors + penalties, temperature, top-k/top-p,
  exponential-noise sampling with per-request generators) without the compiled
  kernel (VLLM_OMNI_CPU_FAST_SAMPLER, default on).
- step_stats: sampler phase timers and sample-metadata timers.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
 2026-09-10 01:56:00 +08:00
tzhouamandClaude Fable 5.1 abf81f4114 [Edge][WP8] Model-level step timers: talker forward, sampling, talker-MTP, code-predictor sub-steps 
Counters model.forward_ms / model.sample_ms / model.talker_mtp_ms and
predictor.substep_ms / predictor.loop_ms (VLLM_OMNI_STEP_STATS_DIR); with
VLLM_OMNI_STEP_STATS_SYNC=1 the accelerator is synchronized around the model
timers so async kernels are attributed (attribution runs only).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
 2026-09-09 21:32:15 +08:00
tzhouamandClaude Fable 5.1 49a839ecf9 [Edge][WP5] Add codec-token streaming mode with on-device decoding 
* qwen3_tts_talker_only pipeline (single terminal LLM_AR stage,
  final_output_type "latent") + vllm_omni/deploy/qwen3_tts_talker_only.yaml;
  a terminal talker publishes its per-step codes.audio rows to the client
  (omni_client_multimodal_output_keys).
* /v1/audio/speech/stream: session.config.output_mode="codec_tokens" streams
  codec.start (JSON) + binary frames (<IHB seq/n_frames/flags + int16 rows,
  vllm_omni/entrypoints/openai/codec_stream.py) + codec.done; new input.cancel
  message aborts an in-flight generation for both codec and PCM sessions
  (single reader task + session inbox replaces direct socket reads); optional
  session.config.seed; audio.done gains "cancelled" only when it happened.
* serving_speech._generate_codec_chunks aligns the cumulative code tensor to
  output token ids (drops prefill placeholder/EOS rows) and emits deltas.
* TTS adapters: supports_codec_stream / codec_stream_spec (Qwen3-TTS: 16
  codebooks, 2048 ids, 12.5 Hz, 24 kHz, decoder qwen3_tts_tokenizer_v2_12hz).
* examples/online_serving/text_to_speech/qwen3_tts/edge_decoder_client.py
  decodes locally with Qwen3TTSTokenizerV2Decoder (ramp schedule, 72-frame
  context, cancel/reset), codec_parity_check.py and
  benchmarks/tts/codec_stream_e2e.py compare against the two-stage server.
* docs/serving/codec_token_streaming.md; CPU unit tests for the wire format,
  alignment, session config and handler (frames, unsupported adapter, cancel);
  L3 e2e tests/e2e/online_serving/test_qwen3_tts_codec_stream.py.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
 2026-09-09 11:10:13 +08:00
tzhouamandClaude Fable 5.1 35e43b9983 [Edge][WP3] Add CPU OmniPlatform (AR + generation stages on vLLM's CPU backend) 
* vllm_omni/platforms/cpu/platform.py: CPUOmniPlatform(OmniPlatform, CpuPlatform)
  selected when VLLM_TARGET_DEVICE=cpu or a +cpu vLLM wheel is installed
  (builtin plugin "cpu"; explicit CPU target short-circuits accelerator probes
  like vLLM does). NUMA-based memory queries, one logical device, no graph
  wrapper (sentinel type for isinstance checks), inductor opt-in via
  VLLM_OMNI_CPU_INDUCTOR=1, diffusion stages unsupported.
* vllm_omni/platforms/cpu/worker: CPUARWorker / CPUGenerationWorker compose
  vLLM's CPUWorker with the Omni mixin (XPU pattern); CPUARModelRunner /
  CPUGenerationModelRunner compose the Omni runners with CPUModelRunner
  (no graphs, no async omni output; sleep/wake return error ACKs).
* patch.py skips the CuMem patch on CPU-only torch; Qwen3-TTS speaker encoder
  and prompt-embeds dtype follow the model dtype (float32/float16 on CPU).
* vllm_omni/deploy/qwen3_tts.yaml: `platforms: cpu:` overlay (eager, no async
  scheduling, absolute kv_cache_memory_bytes, disjoint OMP thread ranges).
* requirements/cpu.txt (+psutil, intel-openmp), docs/getting_started/installation/cpu.md.

Measured (this x86 Xeon 8480C host, vllm 0.28.0+cpu, bf16, 12+4 threads):
Qwen3-TTS 0.6B CustomVoice end-to-end on CPU, 12 prompts: TTFA p50 292 ms,
RTF ~1.7 (25-frame chunk ~4.6 s), init 185 s. Result JSON + init timeline
committed under benchmarks/tts/edge_results/.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
 2026-09-09 11:02:14 +08:00
tzhouamandClaude Fable 5.1 073d31b9cd [Edge][WP2] Add initialization timeline and cold-start profiler 
* vllm_omni/engine/init_timeline.py: InitTimeline phase collector; engine
  phases (stages_init, g3_device_lock, g2_spawn/ready) and worker phases
  (proc_entry, engine_core_init, init_device, load_model/load_weights,
  talker_mtp_init, attn_prewarm, cudagraph_capture, talker_mtp_capture,
  code2wav_cudagraph, predictor_setup_compile/warmup/cudagraph) recorded via
  worker_phase/timed_phase; workers append JSONL under
  VLLM_OMNI_INIT_TIMELINE_DIR (inherited at spawn) and the engine merges them.
  Enabled by VLLM_OMNI_INIT_TIMELINE=<json> or log_stats; table logged at INFO,
  partial table on orchestrator startup timeout, suggested timeouts logged.
* omni_base.py: named deploy profiles default to stage_init_timeout=900 /
  init_timeout=1500 unless given explicitly.
* gpu_model_runner._init_talker_mtp tolerates max_cudagraph_capture_size=None;
  gpu_ar_model_runner.shutdown skips CUDA graph teardown on CPU.
* benchmarks/tts/init_profile.py: runs stream_latency_bench --init-only per
  configuration (default, cold_cache, eager_stage1, eager_both,
  parallel_stage_init, edge, ...) and tabulates phase durations.
* stream_latency_bench: --parallel-stage-init / --deploy-profile flags; the
  resource poller no longer shadows Thread._stop and skips GPU sampling on CPU.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
 2026-09-09 11:01:39 +08:00
BruceLoveDecimal 5cd925631c [Feature][LingBot World]stepwise execution (#6844) 
Signed-off-by: BruceLoveDecimal <liuqihao970610@gmail.com>
Signed-off-by: liuqihao <liuqihao970610@gmail.com>
 2026-09-07 13:38:23 +08:00
NATUREandamy-why-3459 ff6e906a25 [Model] Preserve MiniCPM-o Talker full attention until capacity (#6799) 
Signed-off-by: natureofnature <wzliu@connect.hku.hk>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>
 2026-09-06 22:25:35 +08:00
Hongsheng Liu 715b8b8746 [Model] MiniMax-H3: validate versioned text-conditioning handoff (#6720) 
Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>
 2026-09-06 21:45:10 +08:00
SYLARandhsliuustc0106 14d2d00536 [Bugfix] Fix Hunyuan Image-3 expert mapping contract (#7021) 
Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>
Co-authored-by: hsliuustc0106 <liuhongsheng4@huawei.com>
 2026-09-06 21:43:37 +08:00
chickeytonClaude Fable 5Gao Han
 		5d7fdf9350 [Core] Graduate MiniCPM-o 4.5 and PersonaPlex full-duplex serving out of experimental (#6196) 
Signed-off-by: chickeyton <ngton2014@gmail.com>
Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
 2026-09-05 08:20:21 +00:00
boatman 66bbebd9d3 [Bugfix] Fix fallback loading for IndexTTS external models (#6137) 
Signed-off-by: boatman <109857087+sphinxkkkbc@users.noreply.github.com>
 2026-09-04 19:58:33 -07:00
BruceLoveDecimalandClaude Opus 5 1e74807c35 [Feature][Moss-TTS] Reference Audio Preprocessing (#4982) 
Signed-off-by: BruceLoveDecimal <39156883+BruceLoveDecimal@users.noreply.github.com>
Signed-off-by: liuqihao <liuqihao970610@gmail.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
 2026-09-04 16:36:26 -07:00
LOGO127 b6b992ac7b [Bugfix] Fix Higgs Audio v3 voice-clone token validation (#7065) 
Signed-off-by: luozijian <luozijian0924@gmail.com>
 2026-09-04 16:06:47 -07:00
Inesh Reddy ChappidiandGao Han 66d938100b [Bugfix] Give exact-shape codec models an unpadded input_ids view (#6775) 
Signed-off-by: Inesh Reddy <ineshreddy249@gmail.com>
Signed-off-by: IneshReddy249 <ineshreddy249@gmail.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
 2026-09-04 15:28:49 -07:00
boatman 348fd0fc07 [Feat][dots.tts] Add Online Serving Support (#6235) 
Signed-off-by: boatman <109857087+sphinxkkkbc@users.noreply.github.com>
 2026-09-04 14:44:10 -07:00
Yuekai ZhangandClaude Fable 5 4179b0cae1 [Perf][NemotronVoiceChat] Fast defaults for all three deploy profiles: 2.8x offline, realtime full-duplex audio (#6354) 
Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-09-03 18:49:19 -07:00
Jim Ban 711daacabf [Bugfix][TTS] Make CosyVoice3 TRT plan publication concurrency-safe (#6955) 
Signed-off-by: BANANASJIM <bananasjim1@gmail.com>
 2026-09-03 06:44:42 +00:00
Tai An d1a5a7c178 [Bugfix][Qwen2.5-Omni] Fix two stale call sites in the speech helpers (#6886) 
Signed-off-by: Anai-Guo <antai12232931@outlook.com>
 2026-09-02 21:41:05 -07:00
Joshna-MedisettyandCursor 24c49cf392 [Bugfix][XPU] Add MoTRMSNorm.forward_xpu delegating to forward_native (#5568) 
Signed-off-by: Joshna Medisetty <joshna.medisetty@intel.com>
Signed-off-by: Joshna-Medisetty <joshna.medisetty@intel.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-02 23:06:17 -05:00
ZhanshenzhuiandSy03 48ae412a2f [model]Add experimental Mage-VL full-duplex adapter and offline integration (#6537) 
Signed-off-by: Zhanshenzhui <99891002+zyforsure@users.noreply.github.com>
Co-authored-by: Sy03 <1370724210@qq.com>
 2026-09-02 20:25:32 -07:00
DanaerLee ca6b4e2112 [Perf][Bugfix][OmniVoice] Restore float16 serving, and fuse the generator hot loop (#6317) 
Signed-off-by: MrlixiangWE <mrdanaer@gmail.com>
 2026-09-02 19:54:55 +00:00
Jim Ban b9a88bf51e [Bugfix][TTS] Isolate concurrent StepAudio2 TRT builds (#6957) 
Signed-off-by: BANANASJIM <bananasjim1@gmail.com>
 2026-09-02 19:53:02 +00:00
GXINandHongsheng Liu 3ab582c993 [Bugfix][Qwen3-TTS] Reject task mismatches before EngineCore dispatch (#6113) 
Signed-off-by: XIN GAO <1037396230@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
 2026-09-02 11:16:55 -07:00
jingchengtian 67ef0b64b9 [Hardware][Ascend] Use npu_rotary_mul fused kernel for MOSS-TTS codec RoPE (#6908) 
Signed-off-by: jingchengtian <tjc1995@126.com>
 2026-09-02 08:58:23 +00:00
Rakesh Kariya b864374089 [Bugfix][MiMo-Audio] Restore make_empty_intermediate_tensors on the talker for vLLM 0.28 (#6803) 
Signed-off-by: rk9595 <rakesh.kariya@somaiya.edu>
 2026-08-31 15:26:35 +08:00
akshatvishu f321cdbe16 [Bugfix] Use Seed-TTS reference audio with MiniCPM-o 4.5 (#6628) 
Signed-off-by: akshatvishu <akshatnayak197@gmail.com>
 2026-08-31 10:27:34 +08:00
akshatvishu c41925689d [BugFix][Higgs-Audio-V3] Fix transcript and concurrent sampling tests (#6422) 
Signed-off-by: akshatvishu <akshatnayak197@gmail.com>
 2026-08-30 08:52:15 +00:00
Yueqian Lin a6b559cf54 [Bugfix][MiniMax-Music3] Resolve stage subdirs against a real snapshot (#6640) 
Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
 2026-08-30 06:51:24 +00:00
NATURE 87736a3cc5 [Bugfix][MiniCPM-o] Bound full-duplex Talker context with sliding recompute (#6626) 
Signed-off-by: natureofnature <wzliu@connect.hku.hk>
 2026-08-30 00:49:42 +08:00
psv666andGao Han 5c47fd0bc8 [Bugfix][Qwen3-Omni] Handle missing packed modules mapping (#6748) 
Signed-off-by: psv666 <2693925048@qq.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
 2026-08-29 15:24:49 +08:00
abrahamzewoudieHongsheng Liuamy-why-3459
 		143ccc5a2f Fix Qwen3-Omni audio encoder TP when heads are not divisible by TP size (#4322) 
Signed-off-by: abrahamzewoudie <azewoudi@amd.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>
 2026-08-29 14:52:02 +08:00
boatman bfe8967ca2 [Bugfix][StepAudio2] Fix async-chunk metadata mismatch and boundary glitches (#5917) 
Signed-off-by: boatman <109857087+sphinxkkkbc@users.noreply.github.com>
 2026-08-28 21:45:24 -07:00
WallbreazzzCanlin GuoCanlin Guo
 		be335a86f9 fix(npu): extend bf16 autocast to NPU in MossAudioTokenizer decode (#6664) 
Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com>
Signed-off-by: Canlin Guo <961750412@qq.com>
Co-authored-by: Canlin Guo <961750412@qq.com>
Co-authored-by: Canlin Guo <canlinguosdu@gmail.com>
 2026-08-28 04:01:49 -07:00
Sy03 009b992c5b [Bugfix][MiniCPM-o] Isolate duplex handoff failures and barge-in (#6170) 
Signed-off-by: Sy03 <1370724210@qq.com>
 2026-08-28 15:46:41 +08:00
Canlin Guolishunyang12yuanwu
 		b6a4baa659 [Feature] Support text encoder Disaggregation for H3 (#5885) 
Signed-off-by: Canlin Guo <canlinguosdu@gmail.com>
Signed-off-by: lishunyang12 <lishunyang12@163.com>
Signed-off-by: yuanwu <yuan.wu@intel.com>
Co-authored-by: lishunyang12 <lishunyang12@163.com>
Co-authored-by: yuanwu <yuan.wu@intel.com>
 2026-08-28 14:27:05 +08:00
BruceLoveDecimalandamy-why-3459 ab8d45d3cc [Bugfix][MiniCPM-o] Retire CFM DiT CUDA graphs a generation at a time (#6587) 
Signed-off-by: BruceLoveDecimal <liuqihao970610@gmail.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>
 2026-08-28 14:07:57 +08:00
Hongsheng Liuandhsliuustc0106 4245fd6955 [Fix] Restore MammothModa2 support (#6694) 
Signed-off-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>
Co-authored-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>
 2026-08-28 08:32:32 +08:00
Zhou TaichangandClaude Opus 5 651b03225b [Rebase] Rebase to vllm 0.28.0 (#6606) 
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
 2026-08-27 06:12:55 +00:00
Sy03andsy03 432b2c6ca3 [Model] Add native full-duplex Nemotron VoiceChat serving (#6089) 
Signed-off-by: Sysy <1370724210@qq.com>
Signed-off-by: sy03 <137072421@qq.com>
Co-authored-by: sy03 <137072421@qq.com>
 2026-08-26 13:19:20 -07:00
Allen WuTRAE CLISy03TRAE CLI
 		f90e992dc1 [Model] Optimize CosyVoice3 TensorRT stream handoff (#5673) 
Signed-off-by: Allen Wu <allenwu2795@gmail.com>
Co-authored-by: TRAE CLI <noreply@bytedance.com>
Co-authored-by: Sy03 <1370724210@qq.com>
Co-authored-by: TRAE CLI <traecli@bytedance.com>
 2026-08-25 22:51:46 -07:00
wy17003andHongsheng Liu b454e34966 [Refactor][2/N Scheduler]Replace _FULL_PAYLOAD_INPUT_STAGES with a resolved stage transport capability. (#6149) 
Signed-off-by: wy17003 <yaow17003@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
 2026-08-25 15:24:59 +08:00
NATURE 702a409664 [Feature][MiniCPM-o] Support configurable concurrent duplex sessions (#6021) 
Signed-off-by: natureofnature <wzliu@connect.hku.hk>
 2026-08-25 02:27:30 +00:00
Pranav ThakurandZhou Taichang 53b60bd9bc [Bugfix] Invalidate stale ref_audio cache on local file modification (#4873) (#5670) 
Signed-off-by: pranavthakur0-0 <pranavthakurdew@gmail.com>
Co-authored-by: Zhou Taichang <tzhouam@connect.ust.hk>
 2026-08-25 10:06:32 +08:00
bjf-frzandHongsheng Liu 072bfc02dd [FEAT]Add FastVideo VSA backend for Wan2.2 (#4820) 
Signed-off-by: bjf-frz <frz123db@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
 2026-08-24 14:48:31 +08:00
SYLAR 1c917b2751 [Bugfix] Fix Qwen3-Omni AWQ quantization name mapping (#5687) 
Signed-off-by: lishunyang12 <lishunyang12@163.com>
 2026-08-23 15:13:36 +00:00
wangyu 1e2a6446f7 [CI] Remove AudioX and MammothModa2 support (#6353) 
Signed-off-by: wangyu <410167048@qq.com>
 2026-08-23 04:22:05 +00:00
akshatvishu 39c16d75a9 [tts][bugfix] Fix CosyVoice3 sampling and stage handoff (#6424) 
Signed-off-by: akshatvishu <akshatnayak197@gmail.com>
 2026-08-22 21:20:57 -07:00
wangyu 2e096788a6 [CI]Remove SoulX-Singer support (#6362) 
Signed-off-by: wangyu <410167048@qq.com>
 2026-08-23 07:16:32 +08:00