vLLM split-engine inference

Benchmark

Test set: 184 files, 11,541 seconds total. Models: Fun-ASR-Nano / GLM-ASR-Nano. See Benchmark RTF and Reproducibility Notes for the RTFx definition, timing scope checklist, and fields required for comparable reports.

vLLM matches PyTorch CER exactly (delta < 0.2%) while achieving 16–340x speedup.

Table of Contents

1. Installation & Environment
2. vLLM Engine Architecture
3. Offline SDK Inference
4. Streaming SDK Inference
5. Offline Speech Recognition Service
6. Streaming Speech Recognition Service
7. Dynamic VAD
8. API Reference
9. FAQ

1. Installation & Environment

Install vLLM first, choosing a version compatible with your NVIDIA driver's CUDA. vLLM pins and installs a matching torch / torchaudio / torchvision trio automatically, so do not install torch/torchaudio yourself — the three are ABI-locked, i.e. they must be the matching set built against each other (e.g. torch 2.10.0 ↔ torchaudio 2.10.0 ↔ torchvision 0.25.0).

# 1) Install vLLM first. Pick the version by the CUDA version shown in `nvidia-smi`
#    (the driver's max CUDA), NOT the runtime CUDA. vLLM brings a matching torch/torchaudio/torchvision.
#    driver CUDA 12.x  -> pip install vllm==0.19.1   (ships torch 2.10 / cu128)
#    driver CUDA >= 13 -> pip install vllm           (latest; ships torch 2.11 / cu130)
pip install "vllm==0.19.1"   # adjust to your driver CUDA; see note below

# 2) Then FunASR and the rest.
pip install "funasr>=1.3.26"

cd /path/to/FunASR && pip install -e .

Choose the model path before you start

There are two different Fun-ASR-Nano vLLM integrations. Their checkpoints and APIs are not interchangeable:

A. FunASR split-engine integration (this guide)

Use the official FunAudioLLM/Fun-ASR-Nano-2512 checkpoint from ModelScope or Hugging Face. The current Hugging Face model.pt has incomplete CTC weights, so it remains usable for transcription but FunASR disables the affected CTC path rather than returning unreliable timestamps or speaker diarization. Use the ModelScope checkpoint when your deployment requires timestamps or speaker diarization; the publication repair is tracked in #3496. The Qwen3-0.6B/ subdirectory intentionally contains only the LLM config and tokenizer; it is not a standalone model download.

from funasr.auto.auto_model_vllm import AutoModelVLLM

# Choose one official hub. ModelScope is the default.
model = AutoModelVLLM(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    hub="ms",
)
# model = AutoModelVLLM(
#     model="FunAudioLLM/Fun-ASR-Nano-2512",
#     hub="hf",
# )

On first load, FunASR calls prepare_vllm_model_dir() automatically. It copies the config/tokenizer from Qwen3-0.6B/, extracts the llm.* tensors from the root model.pt, and writes Qwen3-0.6B-vllm/model.safetensors. Do not point model at Qwen3-0.6B/ and do not try to serve that config-only directory directly with vLLM.

B. Native vLLM transcription integration

The official native checkpoint is FunAudioLLM/Fun-ASR-Nano-2512-vllm. The vLLM supported-model table also lists allendou/Fun-ASR-Nano-2512-vllm, a community-converted full checkpoint hosted outside the official FunAudioLLM organization, for its native FunASRForConditionalGeneration architecture. Use either native checkpoint only when you intentionally choose vLLM's native transcription API. Do not substitute either one for the official checkpoint in the FunASR AutoModelVLLM examples below.

Both native paths use vllm serve for non-realtime request/response transcription at /v1/audio/transcriptions; they do not register /v1/realtime. vLLM registers that WebSocket endpoint only for models that declare the realtime task, and FunASRForConditionalGeneration is not in its Realtime Transcription table. For realtime streaming, use the FunASR streaming SDK inference or streaming ASR service. The AutoModelVLLM examples in path A are offline inference as well.

Hardware: GPU ≥ 8 GB VRAM, CUDA ≥ 11.8. 16 GB+ recommended.

Why not pip install torch torchaudio? The torch/torchaudio/torchvision versions are determined by the vLLM release — each major vLLM version bumps them together (see vLLM's requirements/cuda.txt). Installing them by hand pulls the newest wheel, which may be built for a newer CUDA runtime than your driver supports; PyTorch then fails during CUDA initialization with The NVIDIA driver on your system is too old before FunASR even starts. Letting vLLM own the trio avoids this. If you still hit a driver-too-old error, install a vLLM version whose CUDA build matches the CUDA reported by nvidia-smi (e.g. vllm==0.19.1 for CUDA 12.x), or update the NVIDIA driver first.

2. vLLM Engine Architecture

Overall Architecture

FunASR's vLLM integration splits the ASR model into two independently running components:

┌──────────────────────────────────────────────────────────────┐
│                  FunASR + vLLM Inference Architecture        │
├──────────────────────────────────────────────────────────────┤
│                                                              │
│  ┌─────────────── PyTorch (single GPU) ───────────┐          │
│  │                                                │          │
│  │  Audio ──→ Frontend ──→ Audio Encoder ──→ Adaptor         │
│  │            (fbank)      (SenseVoice/     (Transformer/    │
│  │                          Whisper)         MLP)            │
│  │                              │                            │
│  │                              ▼                            │
│  │                     Audio Embeddings                      │
│  │                              │                            │
│  │  Text Prompt ──→ Tokenize ──→ Embed                       │
│  │  (system/user/                  │                         │
│  │   hotwords/language)            │                         │
│  │                                 ▼                         │
│  │                          [Concat Embeddings]              │
│  └─────────────────────────────────┼─────────────┘           │
│                                    │                         │
│                                    ▼ EmbedsPrompt            │
│  ┌─────────────── vLLM Engine ────────────────────┐          │
│  │                                                │          │
│  │   PagedAttention + Continuous Batching         │          │
│  │   KV Cache management + CUDA Graph             │          │
│  │   Tensor Parallel (multi-GPU)                  │          │
│  │                                                │          │
│  │   Qwen3-0.6B / Llama-2B (LLM decoding)         │          │
│  │                                                │          │
│  └────────────────────┬───────────────────────────┘          │
│                       │                                      │
│                       ▼                                      │
│                Generated Text                                │
│                       │                                      │
│  ┌────────────────────┼──────────────────────────┐           │
│  │  (Optional) CTC Decoder ──→ Forced Alignment  │           │
│  │           ──→ Character-level timestamps      │           │
│  └───────────────────────────────────────────────┘           │
└──────────────────────────────────────────────────────────────┘

Why vLLM?

Supported Models

Key Implementation Details

1. Weight separation: LLM weights are extracted from model.pt and converted to HuggingFace format for vLLM loading
2. EmbedsPrompt: a vLLM input mode that feeds precomputed embedding vectors (rather than the usual token IDs) directly as the prompt (enabled via enable_prompt_embeds=True). Fun-ASR-Nano requires it because the audio, after the adaptor, is a sequence of continuous vectors — not tokens — so the audio embeddings and text embeddings are concatenated along the sequence dimension and submitted to vLLM as a whole
3. use_low_frame_rate: Fun-ASR-Nano's adaptor output must be truncated to the correct token count via a formula (critical for consistency)
4. Batch encode: Multiple audio files pass through extract_fbank → audio_encoder → audio_adaptor in a single forward pass
5. CTC timestamps: Encoder output is retained; after text generation, forced alignment yields character-level timing

3. Offline SDK Inference

Best suited for large-scale audio transcription and offline batch processing. vLLM's batching capability provides the greatest advantage in this scenario.

Design Principles

Offline SDK inference splits the ASR pipeline into two stages executed independently:

┌─────────────────────────────────────────────────────────────────────┐
│            Stage 1: Audio Encoding (PyTorch, single GPU)            │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  Audio file list ──→ Group (batch of 8) ──→ Frontend (Fbank)        │
│       │                                          │                  │
│       │                                          ▼                  │
│       │                                 SenseVoice Encoder          │
│       │                                          │                  │
│       │                                          ▼                  │
│       │                                 Audio Adaptor               │
│       │                                 (dim transform + LFR trunc) │
│       │                                          │                  │
│       └─── Shared text prompt encoding ────┐     ▼                  │
│            (system/hotwords/language)      │  audio_embeds          │
│                     │                      │     │                  │
│                     ▼                      │     ▼                  │
│                prefix_emb ──→ [concat: prefix | audio | suffix]     │
│                                                  │                  │
│                                                  ▼                  │
│                                        EmbedsPrompt (N samples)     │
└──────────────────────────────────────────────────┼─────────────────┘
                                                   │
                                                   ▼
┌─────────────────────────────────────────────────────────────────────┐
│        Stage 2: LLM Decoding (vLLM, multi-GPU Tensor Parallel)      │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  EmbedsPrompt × N ──→ vLLM Continuous Batching                      │
│                        (PagedAttention + CUDA Graph)                │
│                              │                                      │
│                              ▼                                      │
│                     Generated token_ids × N                         │
│                              │                                      │
│                              ▼                                      │
│                     Decode + post-processing (strip special tokens) │
│                              │                                      │
│                              ▼                                      │
│                    (Optional) CTC Forced Alignment → char timestamps│
└─────────────────────────────────────────────────────────────────────┘

Key design decisions:

1. Weight separation: On first run, weights with the llm.* prefix are extracted from model.pt and saved in HuggingFace safetensors format for vLLM (cached in the Qwen3-0.6B-vllm/ directory)
2. Embedding concatenation: The text prompt is encoded through the LLM's embed_tokens layer into embeddings, then concatenated with the audio adaptor output along the sequence dimension: [prefix_emb | audio_emb | suffix_emb], and submitted to vLLM as an EmbedsPrompt
3. Low Frame Rate truncation: Adaptor output must be truncated to the correct length using: fake_token_len = ((((fbank_len - 3 + 2) // 2 - 3 + 2) // 2) - 1) // 2 + 1, ensuring consistency with the PyTorch training pipeline
4. Batch audio encoding: Multiple audio files are grouped in batches of 8 through the encoder + adaptor forward pass, reducing GPU kernel launch overhead
5. Shared text prompt: When hotwords and language are identical within a batch, prefix_emb and suffix_emb are computed only once
6. CTC timestamps: Encoder output is preserved; after LLM text generation, forced alignment produces character-level timestamps

Why faster than PyTorch generate()?

Universal Interface (Recommended)

from funasr.auto.auto_model_vllm import AutoModelVLLM

model = AutoModelVLLM(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    hub="ms",                    # or "hf"
    tensor_parallel_size=2,      # multi-GPU parallel
    gpu_memory_utilization=0.8,
)

results = model.generate(
    ["audio1.wav", "audio2.wav"],
    language="中文",
    hotwords=["张三", "北京"],
)
for r in results:
    print(f"[{r['key']}] {r['text']}")

Direct Interface

from funasr.models.fun_asr_nano.inference_vllm import FunASRNanoVLLM

engine = FunASRNanoVLLM.from_pretrained(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    tensor_parallel_size=4,
)

results = engine.generate(
    inputs="wav.scp",  # supports scp/jsonl/file lists
    hotwords=["开放时间"],
    language="中文",
    max_new_tokens=512,
)

Command Line

cd examples/industrial_data_pretraining/fun_asr_nano

# Single file
python demo_vllm.py --input audio.wav --language 中文

# Batch + multi-GPU
python demo_vllm.py --input wav.scp --tensor-parallel-size 4 --batch-size 32

# With hotwords + save results
python demo_vllm.py --input audio.wav --hotwords 张三 北京 --output results.jsonl

4. Streaming SDK Inference

Processes audio in 720 ms chunks incrementally, outputting progressively stable recognition results. Suited for SDK-integrated real-time subtitle scenarios.

Design Principles

Audio stream (720 ms chunks)
    │ Cumulative re-encoding (each chunk covers all audio from the start)
    ▼
┌──────────────────────────┐
│ Stage 1: First 10 chunks │  ← No prev_text; batch generation
│ Identify stable output   │
└──────────┬───────────────┘
           ▼
┌──────────────────────────┐
│ Stage 2: Subsequent      │  ← Use stable output as prev_text
└──────────┬───────────────┘
           ▼
Each chunk: [fixed region (confirmed)] + [8-char unfixed (may change)]

Usage

from funasr.models.fun_asr_nano.inference_vllm_streaming import FunASRNanoStreamingVLLM

engine = FunASRNanoStreamingVLLM.from_pretrained(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    chunk_ms=720,
    rollback_chars=8,
)

for result in engine.streaming_generate("audio.wav", language="中文"):
    if result["is_final"]:
        print(f"Final: {result['text']}")
    else:
        print(f"[{result['audio_duration_ms']:.0f}ms] Confirmed: {result['fixed_text']}")

Output Characteristics

Note: repetition_penalty cannot be used with EmbedsPrompt. Here the prompt is a sequence of embedding vectors with no corresponding token IDs, whereas repetition_penalty needs the prompt's token IDs to down-weight already-seen tokens in the logits; applying it under EmbedsPrompt indexes out of bounds and triggers a CUDA device-side assert.

Production API Stability Checklist

When wrapping AutoModelVLLM in a long-running API service, keep request state isolated and pin safe decoding defaults:

common = dict(
    language="auto",
    temperature=0.0,
    repetition_penalty=1.0,
    max_new_tokens=200,
)

for _ in range(2):
    results = model.generate(["vad_segment_01.wav", "vad_segment_02.wav"], **common)
    print([r["text"] for r in results])

If the same audio is normal on the first request but repeats on the second request:

1. Run the minimal script above outside the API layer with the same VAD segments.
2. If the script is stable, check whether the API wrapper reuses per-request variables, previous VAD segment lists, previous results, or accumulated text across requests.
3. If the script also repeats, capture the exact funasr, vllm, and torch versions, plus the first and second outputs, before tuning any decoding parameter.

Do not increase repetition_penalty to suppress repeats on Fun-ASR-Nano vLLM. The prompt-embeds path should stay at the neutral value 1.0.

5. Offline Speech Recognition Service

5.1 Service Architecture

Client                                  serve_vllm.py
  │                                        │
  │── HTTP / OpenAI / WebSocket ─────────→│
  │                                        │
  │                                   ┌────┴────────────────────────┐
  │                                   │ 1. Receive complete audio   │
  │                                   │ 2. Dynamic VAD (≤60 s/seg)  │
  │                                   │ 3. vLLM batch all segments  │
  │                                   │ 4. CTC timestamps (per-char)│
  │                                   │ 5. Speaker diarization (opt)│
  │                                   └────┬────────────────────────┘
  │                                        │
  │←── JSON result ───────────────────────│

Characteristics: - Processes audio only after it arrives in full — ideal for file transcription - Dynamic VAD preserves long segments (≤60 s), reducing boundary-cut losses - Batch inference over all VAD segments maximizes throughput - Automatically outputs character-level timestamps - Speaker diarization is off by default; clients can enable it

5.2 Starting the Service

CUDA_VISIBLE_DEVICES=0 python examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py \
    --port 8899 \
    --model FunAudioLLM/Fun-ASR-Nano-2512 \
    --gpu-memory-utilization 0.5

About CUDA_VISIBLE_DEVICES: the =0 in the examples is just an illustrative value ("use GPU 0"), not a fixed requirement. It selects which GPUs are visible to this process (indexed as in nvidia-smi), a single GPU machine does not need to set it.

• Single GPU: small models like 0.6B / 1.7B can run several instances on one card — point multiple processes at the same GPU (e.g. all =0) sharing it via MPS, or split across cards with process A =0, B =1 (see §6.7).

5.3 Protocol 1: HTTP REST — POST /asr

The most feature-complete interface, supporting speaker diarization, timestamps, and hotwords.

Request: multipart/form-data

Response:

{
    "text": "Full transcription text",
    "segments": [
        {
            "text": "Segment text",
            "start": 1.7,
            "end": 14.8,
            "speaker": "SPK0",
            "words": [
                {"word": "砸", "start": 2.02, "end": 2.08},
                {"word": "了", "start": 2.26, "end": 2.32}
            ]
        }
    ],
    "duration": 227.4,
    "processing_time": 3.422,
    "rtf": 0.015
}

Client examples:

# cURL
curl -X POST http://localhost:8899/asr \
    -F "file=@meeting.wav" -F "language=中文" -F "spk=true"

# Python requests
import requests
resp = requests.post("http://localhost:8899/asr",
    files={"file": open("audio.wav", "rb")},
    data={"language": "中文", "spk": "true"})
result = resp.json()

// JavaScript fetch
const form = new FormData();
form.append("file", audioBlob, "audio.wav");
form.append("language", "中文");
form.append("spk", "true");
const resp = await fetch("http://localhost:8899/asr", { method: "POST", body: form });
const result = await resp.json();

5.4 Protocol 2: OpenAI Whisper Compatible — POST /v1/audio/transcriptions

Compatible with the OpenAI Whisper API standard; works directly with the OpenAI SDK.

Request: multipart/form-data

Response (verbose_json):

{
    "task": "transcribe",
    "language": "zh",
    "duration": 5.17,
    "text": "我一直没有照顾孩子，但是我想要抚养权。",
    "segments": [
        {
            "id": 0, "start": 0.0, "end": 5.15,
            "text": "我一直没有照顾孩子，但是我想要抚养权。",
            "words": [{"word": "我", "start": 0.42, "end": 0.48}, ...]
        }
    ]
}

Client examples:

# OpenAI SDK (recommended)
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8899/v1", api_key="none")
result = client.audio.transcriptions.create(
    model="fun-asr-nano",
    file=open("audio.wav", "rb"),
    response_format="verbose_json",
)
print(result.text)

# cURL
curl -X POST http://localhost:8899/v1/audio/transcriptions \
    -F "file=@audio.wav" -F "model=fun-asr-nano" -F "response_format=verbose_json"

5.5 Protocol 3: WebSocket — ws://host:port/ws

WebSocket interface for the offline service. Send complete audio, then receive results. Speaker clustering is performed automatically on STOP, and results include the spk field.

Client → Server:

Server → Client:

{"event": "started"}
{"event": "language_set", "language": "中文"}
{"sentences": [{"text":"...","start":..,"end":..}], "is_final": true, "duration_ms": 5170}
{"event": "stopped"}

Client example:

import asyncio, websockets, json, numpy as np, soundfile as sf

async def offline_ws(audio_path):
    audio, sr = sf.read(audio_path)
    pcm = (audio * 32768).astype(np.int16)

    async with websockets.connect("ws://localhost:8899/ws") as ws:
        await ws.send("START")
        await ws.recv()
        await ws.send("LANGUAGE:中文")
        await ws.recv()

        # Send complete audio
        await ws.send(pcm.tobytes())
        await ws.send("STOP")

        # Receive result
        async for msg in ws:
            data = json.loads(msg)
            if data.get("is_final"):
                for s in data["sentences"]:
                    print(f"[{s['start']/1000:.1f}s] {s['text']}")
                break

asyncio.run(offline_ws("audio.wav"))

6. Streaming Speech Recognition Service

6.1 Service Architecture

Client (microphone / audio stream)     serve_realtime_ws.py
  │                                      │
  │── WebSocket PCM16 16 kHz ───────────→│
  │   (~100 ms per frame, continuous)    │
  │                                      │
  │                                 ┌────┴──────────────────────────┐
  │                                 │ Real-time loop:               │
  │                                 │  ├─ Dynamic VAD (60 ms chunk) │
  │                                 │  ├─ Endpoint → vLLM decode    │
  │                                 │  ├─ No endpoint → partial     │
  │                                 │  └─ Streaming SPK assignment  │
  │                                 └────┬──────────────────────────┘
  │                                      │
  │←── JSON real-time push ──────────────│

Characteristics: - Audio arrives frame by frame; processing starts immediately - Natural sentence segmentation based on VAD endpoints - Confirmed segment text is locked and never changes; partial text updates in real time - Optional streaming speaker assignment (--enable-spk) + global re-clustering on STOP - First-word latency ~480 ms

6.2 Starting the Service

CUDA_VISIBLE_DEVICES=0 python examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py \
    --port 10095 --language 中文 --hotword-file hotword_list

For multi-client or long continuous-speech workloads, start by bounding partial previews and lowering the refresh rate:

CUDA_VISIBLE_DEVICES=0 python examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py \
    --port 10095 --language 中文 \
    --partial-window-sec 8 --decode-interval 0.8

Speaker diarization is disabled by default; add --enable-spk only when the spk field is required.

The server sends WebSocket pings every 20 seconds by default, while ping-timeout closure is disabled. Under concurrent long-audio load, model and VAD work can delay control-frame handling even though the connection is healthy. A fixed timeout can therefore close valid sessions during compute or queue backpressure.

CUDA_VISIBLE_DEVICES=0 python examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py \
    --port 10095 --language 中文

Set a positive --ws-ping-timeout only after measuring the worst-case decode and queue delay for the production traffic shape; keep it above that delay and coordinate it with the gateway idle-timeout policy. The websockets library's max_queue setting bounds receive buffering for incoming messages; it doesn't change ping/pong timeout semantics, so increasing it doesn't fix keepalive timeouts. Set --ws-ping-interval 0 only when an external gateway already owns keepalive/reconnect policy.

For long-session debugging, especially with --enable-spk, enable periodic session-state logs:

CUDA_VISIBLE_DEVICES=0 python examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py \
    --port 10095 --language 中文 --enable-spk \
    --log-session-stats-interval 30

This prints a Session stats: line every 30 seconds. Include the last few lines in issue reports together with tail RTF, process RSS, GPU memory, and the final disconnect log.

6.3 WebSocket Protocol

Connection: ws://host:10095

Client → Server:

Server → Client:

{"event": "started"}
{"sentences": [{"text":"你好","start":300,"end":1200}], "partial": "世界", "is_final": false}
{"sentences": [...], "is_final": true}
{"event": "stopped"}

Fields: sentences[] = locked segments, partial = text being spoken (may change), partial_start_ms = where the current provisional partial begins, is_final = true after STOP. When --enable-spk is enabled, sentences[] also includes spk.

Sequence diagram:

Client              Server
  │── START ───────→│
  │←─ started ──────│
  │── [audio] ─────→│
  │←─ {partial} ────│  #refer to 6.5
  │── [audio] ─────→│
  │←─ {sentences+partial} ─│  (VAD cut a sentence)
  │── STOP ────────→│
  │←─ {is_final:true} ────│
  │←─ stopped ─────│

6.4 Client Usage

Python CLI:

python client_python.py --server ws://localhost:10095 --mic
python client_python.py --server ws://localhost:10095 --file audio.wav

Realtime benchmark:

python examples/industrial_data_pretraining/fun_asr_nano/realtime_ws_benchmark.py \
    audio_16k_mono_pcm16.wav --server ws://localhost:10095 --clients 4 \
    --output-jsonl realtime_ws_4c.jsonl

For metric definitions and reporting fields, see Realtime WebSocket Benchmark.

Browser: Open client_mic.html

Custom Python:

import asyncio, websockets, numpy as np, json

async def stream(audio_path):
    import soundfile as sf
    audio, sr = sf.read(audio_path)
    pcm = (audio * 32768).astype(np.int16)

    async with websockets.connect("ws://localhost:10095") as ws:
        await ws.send("START")
        await ws.recv()

        for i in range(0, len(pcm), 1600):
            await ws.send(pcm[i:i+1600].tobytes())
            await asyncio.sleep(0.05)

        await ws.send("STOP")
        async for msg in ws:
            data = json.loads(msg)
            if data.get("is_final"):
                for s in data["sentences"]:
                    print(f"[{s['start']/1000:.1f}s] {s['text']}")
                break

asyncio.run(stream("audio.wav"))

6.5 Partial preview mechanism and long-sentence behavior

What partial is and how it's produced While the user is speaking, the streaming service periodically (default decode_interval≈0.48s in serve_realtime_ws.py) decodes "the current sentence from its start up to now," emitting provisional text (the partial field in the protocol, which may be overwritten by later refreshes), until VAD detects the sentence end and locks it into sentences. This lets the user see text as they speak.

Note: serve_vllm.py's /ws (§5) has no partial and only returns at sentence end; use serve_realtime_ws.py for live preview.

Frontend rendering rule Treat partial as a replaceable preview, not as text to append. A good UI keeps locked text and preview text separate:

const committed = data.sentences.map((s) => s.text).join("");
const preview = data.partial || "";
render(committed + preview);

If partial_start_ms moves forward because --partial-window-sec is active, the preview only describes the current bounded decode window. Replace the preview area on each message; append only VAD-locked sentences or the final is_final=true result.

Principle: why each partial re-encodes the whole segment from the start Fun-ASR-Nano's acoustic encoder (SenseVoice) is a full-context, non-streaming encoder — each frame's representation depends on the context of the entire segment. When the sentence continues and the audio grows, the context of the earlier frames changes, so the previously computed encoding no longer holds. It therefore cannot cache history and encode only the new frames the way a streaming / causal encoder would; it must run the whole "start → now" segment through the encoder again.

Resulting behavior: partial gets slower on long sentences (O(L²)) Because each refresh re-encodes from the sentence start, the longer a sentence, the longer each partial's audio and the more refreshes occur — so total encoding work grows quadratically with sentence length. In practice a ~29 s continuous utterance is fully re-encoded a dozen-plus times, with single-pass encoder time climbing from tens to hundreds of milliseconds. (The §4 SDK streaming "each chunk contains all audio from the start to now" is the same mechanism; long files behave the same way.)

Usage guidance - Normal conversational speech has natural pauses, so VAD splits it into relatively short utterances and each partial's cost is naturally bounded — usually nothing to worry about. - Only very long, pauseless continuous speech (e.g. reading aloud) makes a single utterance keep growing and the partial preview progressively slower. serve_realtime_ws.py bounds provisional previews with --partial-window-sec 8 by default; raise the window only after measuring headroom for multi-client continuous-monologue load. This only affects provisional partial; VAD-locked sentences and STOP final output still run on the full audio. See the measured L20 starting point in §6.7.

6.6 Cost of speaker diarization (SPK) and how to enable it

serve_realtime_ws.py does not load the SPK model by default. It loads --spk-model (default iic/speech_eres2netv2_sv_zh-cn_16k-common) only when started with --enable-spk, then runs speaker assignment for each VAD-completed sentence during streaming. Note:

• SPK is of limited effectiveness on Fun-ASR-Nano (see #2944); most real-time ASR scenarios do not need speaker separation.
• Streaming SPK is expensive and grows with the session: each sentence re-clusters all historical embeddings (O(N²), more expensive per sentence as the session grows) in that session's worker; the session also re-clusters everything again at the end, so the per-sentence clustering during streaming is overwritten by the final result — redundant as far as the final output is concerned. This is especially pronounced under long sessions + high concurrency.
• Recommendation: keep the default off for multi-client live ASR; if diarization is required, add --enable-spk and treat the final STOP-time labels as authoritative.
• Long-session diagnostics: when a session still slows down or disconnects, rerun with --log-session-stats-interval 30 and check whether audio_buffer_samples, locked_sentences, speaker_history_chunks, speaker_history_embeddings, and speaker_centers stay bounded. If those counters stay near their limits while RTF keeps rising, the remaining bottleneck is more likely model inference, response payload size, or environment scheduling rather than retained session state.

6.7 Production concurrency and multi-process deployment

serve_realtime_ws.py keeps network I/O on one asyncio loop, but runs each connection's blocking session work in worker threads. Compatible ASR requests that arrive together are collected for up to --decode-batch-wait-ms (10 ms by default), flattened into one AutoModelVLLM.generate() call, and capped by --decode-max-batch-size (16 audio segments by default). The shared vLLM engine still has one controlled caller, while audio ingestion and session state are no longer held behind a process-wide lock.

• Streaming VAD defaults to CPU. Each connection gets an isolated FSMN-VAD instance using --vad-device cpu --vad-ncpu 1, avoiding shared mutable cache state and per-frame CUDA allocator synchronization. Override the device only after benchmarking the full WebSocket path; multiplying CPU threads by the number of active sessions can also oversubscribe a host.
• Tune batching for the traffic profile. The 10 ms default is small relative to the 480 ms partial-decode interval and gives simultaneous speakers a chance to share encoder and vLLM batches. Lower it toward 0 for the lowest single-stream queue latency; raise it cautiously when throughput matters more than a few milliseconds of added queue time. Increase the maximum batch only after measuring GPU memory and tail latency.
• One process is the first scaling unit. Benchmark a single process with the built-in batching path before adding replicas. Use multiple processes or one instance per GPU only after a single process reaches its measured GPU, CPU, or tail-latency limit; each extra process duplicates model memory and may reduce batching opportunities.
• vLLM benefits depend on requests arriving together. Turn-taking traffic may have many connected clients but only a few simultaneous decodes, while replaying the same continuous monologue across every client creates deliberately synchronized batches. Report both traffic shape and batching flags with every result.
• Sustainable concurrency has no universal "supports N connections" number. It depends mainly on simultaneous speakers, silence ratio, utterance length, partial refresh interval, speaker diarization, batch wait, and GPU/CPU capacity. Long pauseless speech still costs more because provisional windows are repeatedly encoded (see §6.5). Benchmark your own workload instead of adopting another deployment's connection count as a specification.
• Measured L20 starting point, not a global default. In #3528, one L20 running 16 synchronized clients on a 47-second continuous utterance, with SPK and client ping disabled, performed best at --partial-window-sec 8 --decode-interval 2.0: 408 decode requests, 3,072.1 seconds of encoded audio, 51.18-second p50 completion, 4.5-second output lag, 14.2x aggregate realtime, and 1.31-second first text. The default 15-second window did not complete that exact 16-client workload. Use 8 / 2.0 as an initial L20 high-concurrency profile only, then tune against your own first-text latency, output lag, final completion, request count, and encoded-audio total.

CUDA_VISIBLE_DEVICES=0 python examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py \
  --partial-window-sec 8 --decode-interval 2.0 --log-decode-profile

7. Dynamic VAD

fsmn-vad enables dynamic silence thresholds by default. Offline and streaming modes use different configurations.

Offline mode favors longer segments to reduce boundary-cut losses; streaming mode tightens faster to reduce latency.

Customization

model.generate(input="audio.wav", silence_schedule=[(5000,1500), (20000,800), (float('inf'),300)])

GLM-ASR does not support long-segment inference; pass dynamic_silence=False when using it.

8. API Reference

9. FAQ

Q: Offline or streaming? Complete files → offline (high throughput). Microphone / live stream → streaming (low latency).

Q: Can GLM-ASR use dynamic VAD? It does not support long-segment inference. Use dynamic_silence=False.

Q: Performance impact of SPK? RTFx drops from 102 to 46. CER is unchanged. Disabled by default.

Q: Entry points for custom development? Offline: serve_vllm.process_audio() / FunASRNanoVLLM.generate() Streaming: serve_realtime_ws.RealtimeASRSession

Q: Slow first startup? vLLM initialization takes 60–90 s (KV Cache + CUDA Graph warmup). Subsequent inferences are instant.

Q: What happens when Fun-ASR-Nano vLLM uses dtype="fp16"? The audio frontend and adaptor remain float16, but FunASR runs the Qwen3 decoder in bfloat16 because vLLM float16 decoding can produce degraded repeated output. This is automatic and keeps the two-byte decoder weight footprint. On hardware without BF16 support, use dtype="fp32"; the vLLM path does not claim an end-to-end FP16 decoder mode.

Q: vLLM returns repeated punctuation such as !!!!!!!! but PyTorch/HF generate is normal. What should I check? This usually means the audio frontend and checkpoint can work, but the vLLM prompt-embedding path or decoding parameters differ from the upstream runner. Check these items before changing the model:

• Pass prompt embeddings to vLLM as float32: EmbedsPrompt(prompt_embeds=input_embeds.float()).
• Use ASR-style deterministic decoding. The Fun-ASR-Nano vLLM path defaults to temperature=0.0, top_p=1.0, and skip_special_tokens=True. In prompt-embeds mode, keep repetition_penalty at the neutral 1.0 unless you are using a token-prompt path; other values are normalized by FunASR's vLLM helpers to avoid vLLM CUDA scatter errors.
• Verify that model_dir and vllm_model_dir are the matching Fun-ASR-Nano pair. If clearing vllm_model_dir makes the same audio work through HF generate, keep debugging the vLLM path rather than the audio file.
• Log vLLM finish_reason, generated token ids, prompt embedding dtype, and prompt embedding shape for one failing sample. Repeated punctuation with finish_reason="length" usually points to decode/prompt mismatch rather than VAD or audio loading.