siliconflow/qwen3_asr_agent

siliconflow/qwen3_asr_agent

Language: Python

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-09-16T11:51:38Z

Pushed: 2026-09-16T11:51:42Z

Default branch: main

Fork: no

Archived: no

README:

Qwen3-ASR Agent

OpenAI-compatible ASR proxy sidecar for Qwen3-ASR served by vLLM.

This project extracts the proxy sidecar program out of the Qwen3-ASR GPU Function template (siliconflow/ready-to-use/qwen3-asr) into a standalone, version-controlled repository, mirroring the siliconflow/vllm_agent / siliconflow/sglang_agent pattern. The template consumes the prebuilt image published from this repository; the sidecar source no longer lives inline inside the template's command: block.

Components

| File | Purpose | |------|---------| | app.py | FastAPI sidecar (port 80 by default, overridable via PORT) proxying /v1/audio/transcriptions to vLLM (127.0.0.1:8080). Cleans the Qwen3-ASR output format (language X... → plain text), repairs repetition hallucinations, applies the multi-language strategy, converts audio to 16 kHz mono WAV, exposes Prometheus metrics, and drains in-flight requests on SIGTERM. | | Dockerfile | python:3.10-slim image; no build-time GPU/torch dependency. | | requirements.txt | Runtime dependencies (fastapi / uvicorn / httpx / prometheus-client / av / python-multipart). | | VERSION | Semantic version used as the image tag suffix. | | test_proxy.py | Local smoke test against a running proxy + vLLM. | | build-image.sh | Local build + push helper (personal namespace by default). |

Endpoints

| Path | Purpose | |------|---------| | POST /v1/audio/transcriptions | OpenAI-compatible transcription (multipart file upload). | | GET /health | Liveness — proxy process only. | | GET /ready | Readiness — cascades to vLLM /ping with a cooldown window. | | GET /metrics | Prometheus metrics. | | ANY /{path} | Catch-all pass-through to vLLM. |

Environment variables

| Variable | Default | Purpose | |----------|---------|---------| | VLLM_HOST | 127.0.0.1 | vLLM host (pod-shared network namespace). | | VLLM_PORT | 8080 | vLLM port. | | PORT | 80 | Sidecar listen port; must match the template's entrypointPort. | | CLEAN_OUTPUT | true | Strip language X prefixes from the transcript. | | ENABLE_REPETITION_FIX | true | Enable repetition repair. | | REPETITION_THRESHOLD | 20 | Repetition count that triggers repair. | | MULTILANG_STRATEGY | major | first \| major \| mixed — how language is picked for multi-language audio. | | MAX_FILE_SIZE_MB | 500 | Max request body size. | | READY_INFERENCE_COOLDOWN | 60 | Seconds between readiness cascade probes to vLLM. | | GRACEFUL_SHUTDOWN_TIMEOUT | 30 | SIGTERM drain timeout; keep below the pod's terminationGracePeriodSeconds. |

Quick start

pip install -r requirements.txt
python app.py # proxy on :80, expects vLLM on 127.0.0.1:8080
python test_proxy.py # smoke test (needs a reachable proxy + vLLM)

Docker

docker build -t qwen3-asr-proxy .

# Local build + push (bumps nothing; fails if the tag already exists)
./build-image.sh

# Build for the shared registry namespace used by the templates (CI does this)

Image publishing

.github/workflows/docker-build.yml builds and pushes on push to main / feat/sf-gpu-function-adaptation and on manual dispatch:

hub.6scloud.com/d1r7umcsfi9c73b4drdg/qwen3-asr-proxy:{YYYYMMDD}-v{VERSION}

Tag rule: {YYYYMMDD}-v{SEMVER} from VERSION (no GPU suffix — the image is pure Python and architecture-independent). The tag format and image name are intentionally kept identical to the image previously built from Qwen3-ASR/proxy/, so already-published tags such as 20260914-v1.0.0 remain valid and the template needs no image rename.

Template wiring

siliconflow/ready-to-use/qwen3-asr/ in the templates repository carries this repository as a git submodule at qwen3_asr_agent/, and runs it as a sidecar container next to the vLLM main container:

Gateway :80 -> [qwen3-asr-proxy :80] -> [vLLM :8080]

The submodule gitlink records the exact sidecar revision the pinned image tag was built from.

Relationship to the Qwen3-ASR fork

app.py is the same program as Qwen3-ASR/proxy/app.py (branch feat/sf-gpu-function-adaptation). Once this repository is the published source of the sidecar image, the copy inside the fork should be retired (or turned into a pointer README) to avoid two drifting sources.