Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
Record vLLM guide GPU validation (#3372)
zhifu gao5c83517d4a65
diff --git a/docs/community_growth_20k.md b/docs/community_growth_20k.mdindex 91c7db423..f583bb982 100644--- a/docs/community_growth_20k.md+++ b/docs/community_growth_20k.md@@ -352,7 +352,7 @@ Paste the Markdown output into launch notes, weekly community updates, or releas- [x] Publish API server curl examples. On 2026-07-23, the OpenAI-compatible API docs already publish copy-paste curl paths in `examples/openai_api/README.md` and `examples/openai_api/README_zh.md`, plus scripted smoke tests in `examples/openai_api/smoke_test.sh` and `examples/openai_api/smoke_test.py`. The manual flow downloads `sample.wav`, checks `/health`, and posts to `/v1/audio/transcriptions` with `-F [email protected]`, `-F model=sensevoice`, and `response_format=verbose_json`; no-code and gateway users can use `examples/openai_api/POSTMAN.md` and `examples/openai_api/OPENAPI.md`.- [x] Publish WebSocket streaming examples. On 2026-07-23, realtime WebSocket examples are published in `docs/vllm_guide.md`, `docs/vllm_guide_zh.md`, and `examples/industrial_data_pretraining/fun_asr_nano/docs/realtime_demo.md`. They show how to start `serve_realtime_ws.py`, connect to `ws://localhost:10095`, run `client_python.py --server ws://localhost:10095 --mic` or `client_python.py --server ws://localhost:10095 --file audio.wav`, interpret `partial`, `partial_start_ms`, and `is_final`, and run `realtime_ws_benchmark.py` for multi-client load checks.- [ ] Validate Docker CPU and GPU images. On 2026-07-23, `examples/openai_api/validate_docker.sh` was prepared so a Docker-capable host can run `bash validate_docker.sh` for the portable CPU image and `bash validate_docker.sh --gpu` for the NVIDIA Container Toolkit path. Current ops-host evidence: `ind-gpu8` has H100 GPUs but no `docker` CLI, and `funasr-web` also has no `docker` CLI, so do not mark this item complete until the script passes on a Docker host.-- [ ] Validate vLLM guide on one GPU server.+- [x] Validate vLLM guide on one GPU server. On 2026-07-23, `ind-gpu8` validated the Fun-ASR-Nano vLLM guide on one H100 with the existing `funasr-3239-vllm0191-py312-20260717` environment (`torch 2.10.0+cu128`, `vllm 0.19.1`, `funasr 1.3.26`, `modelscope 1.38.1`, CUDA available with 8 H100s). The first attempt with the HF `Fun-ASR-Nano-2512-hf` snapshot correctly failed before vLLM init because that Transformers-format snapshot lacks the guide-required `Qwen3-0.6B` directory. The verified run used the ModelScope snapshot with `Qwen3-0.6B` and `model.pt`, then ran `CUDA_VISIBLE_DEVICES=0` `examples/industrial_data_pretraining/fun_asr_nano/demo_vllm.py` with `tensor_parallel_size=1`, `gpu_memory_utilization=0.35`, `max_model_len=1024`, and the bundled `example/zh.mp3`; `model.pt` was converted to `Qwen3-0.6B-vllm/model.safetensors`, vLLM loaded on GPU 0, Model loaded in 82.6s, and inference processed 1 files in 1.95s with output `开饭时间早上九点至下午五点。` plus token timestamps. Logs are kept under `/cpfs_speech/user/zhifu.gzf/.cache/funasr-ops/vllm-guide-smoke-ms-20260723.log` and the JSONL result under `/cpfs_speech/user/zhifu.gzf/.cache/funasr-ops/vllm-guide-smoke-ms-20260723.jsonl`.- [ ] Record concise terminal outputs for release notes.### Week 3: Launch packagediff --git a/tests/test_growth_plan_doc.py b/tests/test_growth_plan_doc.pyindex 1c1cbab3f..470438762 100644--- a/tests/test_growth_plan_doc.py+++ b/tests/test_growth_plan_doc.py@@ -101,3 +101,29 @@ def test_growth_plan_records_websocket_streaming_examples_completion():assert marker in textassert "[ ] Publish WebSocket streaming examples" not in text+++def test_growth_plan_records_vllm_guide_gpu_validation():+ text = PLAN.read_text()++ required_markers = [+ "[x] Validate vLLM guide on one GPU server",+ "funasr-3239-vllm0191-py312-20260717",+ "`torch 2.10.0+cu128`",+ "`vllm 0.19.1`",+ "`funasr 1.3.26`",+ "ModelScope snapshot",+ "`Qwen3-0.6B`",+ "`model.pt`",+ "`CUDA_VISIBLE_DEVICES=0`",+ "`examples/industrial_data_pretraining/fun_asr_nano/demo_vllm.py`",+ "`gpu_memory_utilization=0.35`",+ "Model loaded in 82.6s",+ "1 files in 1.95s",+ "开饭时间早上九点至下午五点。",+ "HF `Fun-ASR-Nano-2512-hf` snapshot",+ ]+ for marker in required_markers:+ assert marker in text++ assert "[ ] Validate vLLM guide on one GPU server" not in text