modelscope/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR.git synced 2026-10-12 02:19:18 +00:00
Code Issues Packages Projects Releases Wiki Activity
5,910 Commits 468 Branches 66 Tags
85516946f7df4cff17af8544aa6a8776ae80ccda
Commit Graph
54 Commits
Author SHA1 Message Date
zhifu gao 564c4d91a3 docs: align English vLLM installation and native validation contracts 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-07 11:02:41 +00:00
zhifu gao 2dfe391936 docs: finish reference navigation and correct model and vLLM contracts 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-06 17:43:37 +00:00
zhifu gaoandLauraGPT 649b5e5218 fix(realtime): reduce default partial decode window (#3632) 
Signed-off-by: LauraGPT <[email protected]>
Co-authored-by: LauraGPT <[email protected]>
 2026-09-03 16:32:40 +08:00
qiulangandLauraGPT fddf9b486b docs(vllm): clarify native transcription boundary (#3601) 
Signed-off-by: LauraGPT <[email protected]>
Co-authored-by: LauraGPT <[email protected]>
 2026-09-02 17:25:03 +08:00
LauraGPT 331369fd64 docs: clarify Fun-ASR-Nano vLLM precision boundary 2026-09-02 07:21:20 +00:00
zhifu gao ff7f76a684 Merge pull request #3598 from modelscope/codex/docs-l20-realtime-tuning 
docs(vllm): add measured L20 tuning profile
 2026-09-02 14:06:24 +08:00
LauraGPT 4604a0bc64 docs: clarify Fun-ASR-Nano HF CTC boundary 
Signed-off-by: LauraGPT <[email protected]>
 2026-09-02 00:02:26 +00:00
LauraGPT 02ff040a1e docs(vllm): add measured L20 tuning profile 
Signed-off-by: LauraGPT <[email protected]>
 2026-08-31 14:17:38 +00:00
LauraGPT 606837e23d docs(realtime): clarify websocket receive buffering 
Signed-off-by: LauraGPT <[email protected]>
 2026-08-28 14:27:56 +00:00
LauraGPT 88a553e709 fix(realtime): avoid false keepalive timeouts under load 
Signed-off-by: LauraGPT <[email protected]>
 2026-08-28 14:26:58 +00:00
zhifu gao 55daa5a193 fix(realtime): batch concurrent websocket decoding (#3529) 
Fix the realtime WebSocket concurrency regression by removing the process-wide session lock and batching compatible vLLM decode requests. Isolate per-session VAD runtime state while sharing encoder weights, keep batch failures contained, and document reproducible concurrency benchmarks.

Fixes #3528

Signed-off-by: LauraGPT <[email protected]>
 2026-08-26 23:30:20 +08:00
zhifu gao 680b1b3f10 docs(vllm): clarify official model paths (#3476) 
Clarify official and native vLLM model paths. Closes #3475.
 2026-08-05 16:23:20 +08:00
zhifu gao ca3ce33ff3 docs: require FunASR 1.3.26 for Nano guides 
Align Nano docs with the 1.3.26 server hub fix release.
 2026-07-23 00:46:26 +08:00
zhifu gao 876641e10d docs: require FunASR 1.3.25 for Nano guides 2026-07-22 23:29:02 +08:00
zhifu gaoandLauraGPT 8ca8e03cd9 docs: require FunASR 1.3.23 for Nano guides (#3310) 
Co-authored-by: LauraGPT <[email protected]>
 2026-07-22 21:26:16 +08:00
zhifu gao f53fe68b1c docs: quote current FunASR install commands 
Quote FunASR install specifiers and align Fun-ASR-Nano vLLM/finetune docs with the current 1.3.19 release floor.
 2026-07-19 13:29:31 +08:00
zhifu gao 3d6e811175 docs: add realtime session stats guidance 
Document realtime WebSocket session-stats diagnostics for long sessions with speaker diarization.
 2026-07-19 09:50:24 +08:00
zhifu gao ac84c4bf67 feat: expose realtime websocket keepalive options (#3121) 
Expose ping interval, ping timeout, close timeout, and max message size options for long-lived realtime WebSocket deployments.
 2026-07-07 17:21:32 +08:00
zhifu gao d8df61a17a docs: add realtime websocket benchmark (#3119) 
Add a reusable benchmark client and documentation for measuring serve_realtime_ws.py under realtime WebSocket workloads.
 2026-07-07 17:11:31 +08:00
zhifu gao a6ebdb7fed docs: add vLLM repeated punctuation troubleshooting (#3118) 
Add English and Chinese FAQ guidance for repeated punctuation outputs in external Fun-ASR-Nano vLLM integrations.
 2026-07-07 16:56:26 +08:00
zhifu gao 1b7c1b4222 docs: add benchmark RTF reproducibility notes (#3116) 
Add benchmark RTF / RTFx reproducibility notes and link them from README, vLLM guides, and docs index.

Documents the public vLLM benchmark scope, the separate website H100 benchmark scope, and required fields for comparable reports.
 2026-07-07 16:43:08 +08:00
zhifu gao 4a412f8c22 Document realtime WS partial rendering and concurrency (#3092) 2026-06-30 19:08:33 +08:00
zhifu gao 0573e013d3 docs: add vLLM API repeat troubleshooting 
Add production API troubleshooting guidance for AutoModelVLLM repeated-request output issues.
 2026-06-30 02:30:49 +08:00
zhifu gao e483164677 Improve realtime websocket multi-client defaults 
Disable streaming SPK by default, keep bounded partial previews testable, and document multi-client tuning for the realtime WebSocket service.
 2026-06-29 22:53:12 +08:00
qiulang b201dfa94b docs(vllm_guide): explain EmbedsPrompt/streaming partial/SPK/concurrency setting (#3011) 2026-06-21 23:39:28 +08:00
qiulang e192f0ca9e docs(vllm_guide): let vLLM pin torch/torchaudio in the installation steps (#3009) 
* docs: installation tips

let vLLM pin torch/torchaudio in the installation steps

* Revise vLLM installation instructions in Chinese guide

Updated installation instructions for vLLM, including specific pip install commands based on CUDA version and clarifying dependencies.

* Update vLLM installation guide for clarity

Clarified installation instructions for vLLM and dependencies, emphasizing the importance of matching CUDA versions and avoiding manual installation of PyTorch components.

* Fix formatting and wording in vllm_guide_zh.md

* Clarify installation instructions for PyTorch dependencies
 2026-06-21 12:10:22 +08:00
zhifu gao fa695e057a docs(vllm_guide): drop stale repetition_penalty hardcode note (#3007) 
* docs(vllm_guide): drop stale repetition_penalty=1.3 hardcode note (removed in #2974)

* docs(vllm_guide): drop stale repetition_penalty=1.3 hardcode note (removed in #2974)

* docs(vllm_guide): drop stale repetition_penalty=1.3 hardcode note (removed in #2974)
 2026-06-20 23:40:52 +08:00
游雁 5a37a31462 docs: avoid unbounded vLLM install guidance 2026-06-11 15:54:58 +00:00
游雁 ef46055056 fix: respect explicit VAD silence threshold 2026-06-11 15:16:24 +00:00
zhifu gao 768750587d docs: add torch+torchaudio to vllm_guide install section 2026-05-27 21:17:20 +08:00
游雁 8325d5eabf docs: rewrite English vLLM guide with high-quality manual translation 2026-05-26 12:08:07 +08:00
游雁 0f780ce592 docs: rename to '离线 SDK 推理' + add design principles, sync zh/en/v2 2026-05-26 12:05:07 +08:00
游雁 bb92a779fd docs: restore vLLM guide — add English version and Chinese (vllm_guide_zh.md) 
The original docs/vllm_guide.md was accidentally emptied. Restored the full
content as docs/vllm_guide_zh.md (Chinese) and created docs/vllm_guide.md
(English translation) as the primary guide.
 2026-05-26 02:25:58 +08:00
游雁 7d3c2161db feat: WebSocket SPK as client-controlled param (SPK:true/false, default off) 2026-05-26 02:12:31 +08:00
游雁 a67e7d7356 feat: add SPK support to all protocols (OpenAI API + WebSocket) 
- OpenAI API: added spk param (FunASR extension)
- WebSocket: SPK clustering on STOP (full re-cluster all sentences)
- All 3 protocols tested: HTTP REST ✓, OpenAI API ✓, WebSocket ✓
- docs: merge VAD tables into one, remove "no SPK" notes
 2026-05-26 02:07:40 +08:00
游雁 374afebb39 docs: note WebSocket also no SPK support 2026-05-26 02:00:52 +08:00
游雁 800bd2cd30 docs: fix offline service - split by protocol with per-protocol clients, add WebSocket, note OpenAI no SPK 2026-05-26 01:57:56 +08:00
游雁 ba2a3ac4f0 docs: final vllm_guide - benchmark first, architecture section, clear service separation 2026-05-26 01:47:06 +08:00
游雁 4cc2b75446 docs: rewrite vllm_guide - separate offline/streaming services, full protocol + client examples 2026-05-26 01:40:10 +08:00
游雁 e46abebc88 docs: rewrite vllm_guide.md - unified structure, dynamic VAD only, stream/offline separation 2026-05-26 01:33:57 +08:00
游雁 6a18244602 feat: finalize vLLM service + benchmark matrix 
Updates:
- Non-streaming dynamic VAD: (10000,2000)→(60000,200)→(inf,100)
  Fun-ASR-Nano CER 8.90%→8.20%; GLM-ASR not supported (long audio)
- SPK is opt-in (default off), configurable via API param spk=true
- serve_vllm.py: unified HTTP/OpenAI/WebSocket server
- benchmark_vllm.py: unified benchmark for all models
- docs/vllm_guide.md: full benchmark matrix (8 configs)

Benchmark (184 files, 11541s):
  Fun-ASR-Nano vLLM batch (fixed): RTFx=371, CER=8.88%
  Fun-ASR-Nano vLLM batch (dynamic): RTFx=340, CER=8.20%
  Fun-ASR-Nano Service (no SPK): RTFx=102, CER=8.14%
  Fun-ASR-Nano Service (with SPK): RTFx=46, CER=8.19%
  GLM-ASR vLLM (fixed): RTFx=265, CER=12.93%
 2026-05-26 01:26:09 +08:00
游雁 005e9d8458 merge vllm: new dynamic VAD + unified server + updated benchmark 2026-05-25 23:22:45 +08:00
游雁 3bc9018854 feat: add unified vLLM inference server (serve_vllm.py) 
Three interfaces:
- HTTP REST: POST /asr (file upload, VAD + ASR + SPK + timestamps)
- OpenAI API: POST /v1/audio/transcriptions (Whisper-compatible)
- WebSocket: ws://host:port/ws (streaming with dynamic VAD)

All share: vLLM engine + dynamic VAD + speaker diarization + CTC timestamps.
Performance: 227s audio processed in 3.4s (RTF=0.015), CER=9.30%.
 2026-05-25 22:20:14 +08:00
游雁 6a89af3847 docs: add CTC timestamp output example to vllm_guide.md 2026-05-25 22:03:41 +08:00
游雁 06558dfdd2 docs: merge benchmark tables into single unified table 2026-05-25 20:29:12 +08:00
游雁 653d787c21 refactor: unify benchmark_vllm.py at project root 
- Single script supports all models (Fun-ASR-Nano, GLM-ASR, AutoModelVLLM)
- Remove old model-specific benchmark from examples/
- Update docs with unified benchmark commands
 2026-05-25 20:22:12 +08:00
游雁 674ba5e53e docs: add GLM-ASR vLLM benchmark results to vllm_guide.md 2026-05-25 20:09:24 +08:00
游雁 4484557424 docs: update vllm_guide.md with benchmark results (RTFx=393.9, CER=8.91%) 2026-05-25 19:28:13 +08:00
xiaoyunchong.xyc 7be505f00e refactor: move DynamicStreamingVAD to funasr/models/fsmn_vad_streaming/ 2026-05-24 18:26:08 +08:00
xiaoyunchong.xyc ca352d992f feat: extract DynamicStreamingVAD as reusable component 
- New funasr/utils/dynamic_vad.py: generic dynamic-threshold streaming VAD
- Supports both streaming (feed/finalize) and non-streaming (process) APIs
- Configurable silence_schedule for custom threshold curves
- Refactor serve_realtime_ws.py to use DynamicStreamingVAD
- Add documentation in vllm_guide.md appendix
 2026-05-24 18:22:57 +08:00