Whisper model support & /v1/audio/transcriptions endpoint & benchmark (#16983)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: MahmoudAshraf97 <hassouna97.ma@gmail.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

benchmark/asr/README.md

@@ -0,0 +1,166 @@
+# ASR Benchmark
+
+This benchmark evaluates the performance and accuracy (Word Error Rate - WER) of Automatic Speech Recognition (ASR) models served via SGLang.
+
+## Supported Models
+
+- `openai/whisper-large-v3`
+- `openai/whisper-large-v3-turbo`
+
+## Setup
+
+Install the required dependencies:
+
+```bash
+apt install ffmpeg
+pip install librosa soundfile datasets evaluate jiwer transformers openai torchcodec torch
+```
+
+## Running the Benchmark
+
+### 1. Start SGLang Server
+
+Launch the SGLang server with a Whisper model:
+
+```bash
+python -m sglang.launch_server --model-path openai/whisper-large-v3 --port 30000
+```
+
+### 2. Run the Benchmark Script
+
+Basic usage (using chat completions API):
+
+```bash
+python bench_sglang.py --base-url http://localhost:30000 --model openai/whisper-large-v3 --n-examples 10
+```
+
+Using the OpenAI-compatible transcription API:
+
+```bash
+python bench_sglang.py \
+    --base-url http://localhost:30000 \
+    --model openai/whisper-large-v3 \
+    --api-type transcription \
+    --language English \
+    --n-examples 10
+```
+
+Run with streaming and show real-time output:
+
+```bash
+python bench_sglang.py \
+    --base-url http://localhost:30000 \
+    --model openai/whisper-large-v3 \
+    --api-type transcription \
+    --stream \
+    --show-predictions \
+    --concurrency 1
+```
+
+Run with higher concurrency and save results:
+
+```bash
+python bench_sglang.py \
+    --base-url http://localhost:30000 \
+    --model openai/whisper-large-v3 \
+    --concurrency 8 \
+    --n-examples 100 \
+    --output results.json \
+    --show-predictions
+```
+
+## Arguments
+
+| Argument | Description | Default |
+|----------|-------------|---------|
+| `--base-url` | SGLang server URL | `http://localhost:30000` |
+| `--model` | Model name on the server | `openai/whisper-large-v3` |
+| `--dataset` | HuggingFace dataset for evaluation | `D4nt3/esb-datasets-earnings22-validation-tiny-filtered` |
+| `--split` | Dataset split to use | `validation` |
+| `--concurrency` | Number of concurrent requests | `4` |
+| `--n-examples` | Number of examples to process (`-1` for all) | `-1` |
+| `--output` | Path to save results as JSON | `None` |
+| `--show-predictions` | Display sample predictions | `False` |
+| `--print-n` | Number of samples to display | `5` |
+| `--api-type` | API to use: `chat` (chat completions) or `transcription` (audio transcriptions) | `chat` |
+| `--language` | Language for transcription API (e.g., `English`, `en`) | `None` |
+| `--stream` | Enable streaming mode for transcription API | `False` |
+
+## Metrics
+
+The benchmark outputs:
+
+| Metric | Description |
+|--------|-------------|
+| **Total Requests** | Number of successful ASR requests processed |
+| **WER** | Word Error Rate (lower is better), computed using the `evaluate` library |
+| **Average Latency** | Mean time per request (seconds) |
+| **Median Latency** | 50th percentile latency (seconds) |
+| **95th Latency** | 95th percentile latency (seconds) |
+| **Throughput** | Requests processed per second |
+| **Token Throughput** | Output tokens per second |
+
+## Example Output
+
+```bash
+python bench_sglang.py --api-type transcription --concurrency 128 --model openai/whisper-large-v3 --show-predictions
+
+Loading dataset: D4nt3/esb-datasets-earnings22-validation-tiny-filtered...
+Using API type: transcription
+Repo card metadata block was not found. Setting CardData to empty.
+WARNING:huggingface_hub.repocard:Repo card metadata block was not found. Setting CardData to empty.
+Performing warmup...
+Processing 511 samples...
+------------------------------
+Results for openai/whisper-large-v3:
+Total Requests: 511
+WER: 12.7690
+Average Latency: 1.3602s
+Median Latency: 1.2090s
+95th Latency: 2.9986s
+Throughput: 19.02 req/s
+Token Throughput: 354.19 tok/s
+Total Test Time: 26.8726s
+------------------------------
+
+==================== Sample Predictions ====================
+Sample 1:
+  REF: on the use of taxonomy i you know i think it is it is early days for us to to make any clear indications to the market about the proportion that would fall under that requirement
+  PRED: on the eu taxonomy i think it is early days for us to make any clear indications to the market about the proportion that would fall under that requirement
+----------------------------------------
+Sample 2:
+  REF: so within fiscal year 2021 say 120 a 100 depending on what the micro will do and next year it is not necessarily payable in q one is we will look at what the cash flows for 2022 look like
+  PRED: so within fiscal year 2021 say $120000 $100000 depending on what the macro will do and next year it is not necessarily payable in q one is we will look at what the cash flows for 2022 look like
+----------------------------------------
+Sample 3:
+  REF: we talked about 4.7 gigawatts
+  PRED: we talked about 4.7 gigawatts
+----------------------------------------
+Sample 4:
+  REF: and you know depending on that working capital build we will we will see what that yields
+  PRED: and depending on that working capital build we will see what that yields what
+----------------------------------------
+Sample 5:
+  REF: so on on sinopec what we have agreed with sinopec way back then is that free cash flows after paying all capexs are distributed out 30 70%
+  PRED: so on sinopec what we have agreed with sinopec way back then is that free cash flows after paying all capexes are distributed out 30% 70%
+----------------------------------------
+============================================================
+```
+
+## Notes
+
+- Audio samples longer than 30 seconds are automatically filtered out (Whisper limitation)
+- The benchmark performs a warmup request before measuring performance
+- Results are normalized using the model's tokenizer when available
+- When using `--stream` with `--show-predictions`, use `--concurrency 1` for clean sequential output
+- The `--language` option accepts both full names (e.g., `English`) and ISO 639-1 codes (e.g., `en`)
+
+## Troubleshooting
+
+**Server connection refused**
+- Ensure the SGLang server is running and accessible at the specified `--base-url`
+- Check that the port is not blocked by a firewall
+
+**Out of memory errors**
+- Reduce `--concurrency` to lower GPU memory usage
+- Use a smaller Whisper model variant

benchmark/asr/bench_sglang.py

@@ -0,0 +1,404 @@
+import argparse
+import asyncio
+import base64
+import io
+import json
+import time
+from statistics import mean, median
+
+import httpx
+import librosa
+import numpy as np
+import soundfile
+from datasets import load_dataset
+from evaluate import load
+from openai import AsyncOpenAI, OpenAI
+from transformers import AutoTokenizer
+
+
+def to_bytes(y, sr):
+    buffer = io.BytesIO()
+    soundfile.write(buffer, y, sr, format="WAV")
+    buffer.seek(0)
+    return buffer
+
+
+async def run_asr_chat(client, model_name, y, sr):
+    """Use chat completions API with audio_url for ASR."""
+    with to_bytes(y, sr) as f:
+        audio_bytes = f.read()
+        audio_base64 = base64.b64encode(audio_bytes).decode("utf-8")
+
+    start_time = time.perf_counter()
+    response = await client.chat.completions.create(
+        model=model_name,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "audio_url",
+                        "audio_url": {"url": f"data:audio/wav;base64,{audio_base64}"},
+                    }
+                ],
+            }
+        ],
+        temperature=0.0,
+    )
+    end_time = time.perf_counter()
+
+    asr_text = response.choices[0].message.content
+    latency = end_time - start_time
+    return latency, asr_text
+
+
+def run_asr_transcription_sync(client, model_name, y, sr, language=None):
+    """Use audio transcriptions API for ASR (sync version)."""
+    audio_buffer = to_bytes(y, sr)
+    audio_buffer.name = "audio.wav"  # OpenAI client needs a name attribute
+
+    start_time = time.perf_counter()
+    kwargs = {
+        "model": model_name,
+        "file": audio_buffer,
+    }
+    if language:
+        kwargs["language"] = language
+
+    transcription = client.audio.transcriptions.create(**kwargs)
+    end_time = time.perf_counter()
+
+    latency = end_time - start_time
+    return latency, transcription.text
+
+
+def run_asr_transcription_stream_sync(
+    base_url, model_name, y, sr, language=None, show_stream=False
+):
+    """Use audio transcriptions API with streaming for ASR."""
+    audio_buffer = to_bytes(y, sr)
+    audio_bytes = audio_buffer.read()
+
+    data = {
+        "model": model_name,
+        "response_format": "json",
+        "stream": "true",
+    }
+    if language:
+        data["language"] = language
+
+    start_time = time.perf_counter()
+    text_chunks = []
+
+    if show_stream:
+        print("[STREAM] ", end="", flush=True)
+
+    with httpx.stream(
+        "POST",
+        f"{base_url}/v1/audio/transcriptions",
+        data=data,
+        files={"file": ("audio.wav", audio_bytes, "audio/wav")},
+        timeout=60.0,
+    ) as response:
+        for line in response.iter_lines():
+            if line.startswith("data: ") and not line.startswith("data: [DONE]"):
+                try:
+                    chunk = json.loads(line[6:])
+                    if "choices" in chunk and chunk["choices"]:
+                        delta = chunk["choices"][0].get("delta", {})
+                        content = delta.get("content", "")
+                        if content:
+                            text_chunks.append(content)
+                            if show_stream:
+                                print(content, end="", flush=True)
+                except json.JSONDecodeError:
+                    pass
+
+    if show_stream:
+        print()  # newline after stream
+
+    end_time = time.perf_counter()
+    latency = end_time - start_time
+    return latency, "".join(text_chunks)
+
+
+async def run_asr_transcription(
+    client,
+    model_name,
+    y,
+    sr,
+    language=None,
+    stream=False,
+    base_url=None,
+    show_stream=False,
+):
+    """Async wrapper for transcription API (runs sync call in executor)."""
+    loop = asyncio.get_event_loop()
+    if stream:
+        return await loop.run_in_executor(
+            None,
+            run_asr_transcription_stream_sync,
+            base_url,
+            model_name,
+            y,
+            sr,
+            language,
+            show_stream,
+        )
+    return await loop.run_in_executor(
+        None, run_asr_transcription_sync, client, model_name, y, sr, language
+    )
+
+
+async def bound_asr(
+    sem,
+    client,
+    model_name,
+    tokenizer,
+    audio,
+    reference,
+    api_type="chat",
+    language=None,
+    stream=False,
+    base_url=None,
+    show_stream=False,
+):
+    async with sem:
+        try:
+            if api_type == "transcription":
+                latency, text = await run_asr_transcription(
+                    client,
+                    model_name,
+                    *audio,
+                    language=language,
+                    stream=stream,
+                    base_url=base_url,
+                    show_stream=show_stream,
+                )
+            else:
+                latency, text = await run_asr_chat(client, model_name, *audio)
+
+            # Calculate tokens for throughput metrics
+            num_output_tokens = len(tokenizer(text, add_special_tokens=False).input_ids)
+
+            # Normalize for WER evaluation
+            # Whisper tokenizer has a normalize method
+            if hasattr(tokenizer, "normalize"):
+                out = tokenizer.normalize(text)
+                ref = tokenizer.normalize(reference)
+            else:
+                out = text.lower().strip()
+                ref = reference.lower().strip()
+
+            return latency, num_output_tokens, out, ref
+        except Exception as e:
+            print(f"Error during ASR: {e}")
+            return None
+
+
+async def process_dataset(
+    model_name,
+    client,
+    data,
+    concurrent_request,
+    api_type="chat",
+    language=None,
+    stream=False,
+    base_url=None,
+    show_predictions=False,
+):
+    sem = asyncio.Semaphore(concurrent_request)
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+    # Warmup
+    print("Performing warmup...")
+    audio_warmup, sr_warmup = (
+        data[0]["audio"]["array"],
+        data[0]["audio"]["sampling_rate"],
+    )
+    await bound_asr(
+        sem,
+        client,
+        model_name,
+        tokenizer,
+        (audio_warmup, sr_warmup),
+        "",
+        api_type=api_type,
+        language=language,
+        stream=stream,
+        base_url=base_url,
+        show_stream=False,  # Don't show stream during warmup
+    )
+
+    tasks = []
+    print(f"Processing {len(data)} samples...")
+    for sample in data:
+        audio, sr = sample["audio"]["array"], sample["audio"]["sampling_rate"]
+        tasks.append(
+            asyncio.create_task(
+                bound_asr(
+                    sem,
+                    client,
+                    model_name,
+                    tokenizer,
+                    (audio, sr),
+                    sample["text"],
+                    api_type=api_type,
+                    language=language,
+                    stream=stream,
+                    base_url=base_url,
+                    show_stream=show_predictions and stream,
+                )
+            )
+        )
+
+    results = await asyncio.gather(*tasks)
+    return [r for r in results if r is not None]
+
+
+def run_evaluation(args):
+    # Use sync client for transcription API, async for chat API
+    if args.api_type == "transcription":
+        client = OpenAI(base_url=f"{args.base_url}/v1", api_key="None")
+    else:
+        client = AsyncOpenAI(base_url=f"{args.base_url}/v1", api_key="None")
+
+    print(f"Loading dataset: {args.dataset}...")
+    print(f"Using API type: {args.api_type}" + (f" (streaming)" if args.stream else ""))
+    dataset = load_dataset(args.dataset, split=args.split)
+
+    # Filter by duration if needed (Whisper max is 30s)
+    def add_duration(sample):
+        y, sr = sample["audio"]["array"], sample["audio"]["sampling_rate"]
+        sample["duration_ms"] = librosa.get_duration(y=y, sr=sr) * 1000
+        return sample
+
+    if "duration_ms" not in dataset.column_names:
+        dataset = dataset.map(add_duration)
+
+    dataset = dataset.filter(lambda x: x["duration_ms"] < 30000)
+
+    if args.n_examples > 0:
+        dataset = dataset.select(range(min(args.n_examples, len(dataset))))
+
+    start = time.perf_counter()
+    results = asyncio.run(
+        process_dataset(
+            args.model,
+            client,
+            dataset,
+            args.concurrency,
+            api_type=args.api_type,
+            language=args.language,
+            stream=args.stream,
+            base_url=args.base_url,
+            show_predictions=args.show_predictions,
+        )
+    )
+    total_test_time = time.perf_counter() - start
+
+    if not results:
+        print("No successful results to evaluate.")
+        return
+
+    # Metrics
+    latencies = [res[0] for res in results]
+    total_tokens = sum([res[1] for res in results])
+    predictions = [res[2] for res in results]
+    references = [res[3] for res in results]
+
+    wer_metric = load("wer")
+    wer_score = 100 * wer_metric.compute(references=references, predictions=predictions)
+
+    print("-" * 30)
+    print(f"Results for {args.model}:")
+    print(f"Total Requests: {len(results)}")
+    print(f"WER: {wer_score:.4f}")
+    print(f"Average Latency: {mean(latencies):.4f}s")
+    print(f"Median Latency: {median(latencies):.4f}s")
+    print(f"95th Latency: {np.percentile(latencies, 95):.4f}s")
+    print(f"Throughput: {len(results) / total_test_time:.2f} req/s")
+    print(f"Token Throughput: {total_tokens / total_test_time:.2f} tok/s")
+    print(f"Total Test Time: {total_test_time:.4f}s")
+    print("-" * 30)
+
+    if args.output:
+        with open(args.output, "w") as f:
+            import json
+
+            json.dump(
+                {
+                    "model": args.model,
+                    "dataset": args.dataset,
+                    "wer": wer_score,
+                    "avg_latency": mean(latencies),
+                    "throughput": len(results) / total_test_time,
+                    "token_throughput": total_tokens / total_test_time,
+                },
+                f,
+                indent=2,
+            )
+
+    if args.show_predictions:
+        print("\n" + "=" * 20 + " Sample Predictions " + "=" * 20)
+        num_to_show = min(args.print_n, len(results))
+        for i in range(num_to_show):
+            print(f"Sample {i+1}:")
+            print(f"  REF: {references[i]}")
+            print(f"  PRED: {predictions[i]}")
+            print("-" * 40)
+        print("=" * 60)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Benchmark sGLang ASR performance.")
+    parser.add_argument(
+        "--base-url", default="http://localhost:30000", help="sGLang server base URL"
+    )
+    parser.add_argument(
+        "--model", default="openai/whisper-large-v3", help="Model name on the server"
+    )
+    parser.add_argument(
+        "--dataset",
+        default="D4nt3/esb-datasets-earnings22-validation-tiny-filtered",
+        help="HF dataset repo",
+    )
+    parser.add_argument("--split", default="validation", help="Dataset split")
+    parser.add_argument(
+        "--concurrency", type=int, default=4, help="Number of concurrent requests"
+    )
+    parser.add_argument(
+        "--n-examples",
+        "-n",
+        type=int,
+        default=-1,
+        help="Number of examples to test (-1 for all)",
+    )
+    parser.add_argument("--output", help="Path to save results in JSON")
+    parser.add_argument(
+        "--show-predictions",
+        action="store_true",
+        help="Print sample predictions and references",
+    )
+    parser.add_argument(
+        "--print-n", type=int, default=5, help="Number of sample predictions to print"
+    )
+    parser.add_argument(
+        "--api-type",
+        choices=["chat", "transcription"],
+        default="chat",
+        help="API type to use: 'chat' for chat completions with audio_url, 'transcription' for audio.transcriptions API",
+    )
+    parser.add_argument(
+        "--language",
+        default=None,
+        help="Language code for transcription API (e.g., 'en')",
+    )
+    parser.add_argument(
+        "--stream",
+        action="store_true",
+        help="Use streaming mode for transcription API",
+    )
+    args = parser.parse_args()
+
+    run_evaluation(args)

python/sglang/srt/configs/model_config.py

@@ -545,6 +545,17 @@ class ModelConfig:
         if "IQuestLoopCoderForCausalLM" in self.hf_config.architectures:
             loop_num = getattr(self.hf_text_config, "loop_num", 1)
             self.num_attention_layers = int(self.num_hidden_layers * int(loop_num))
+        if "WhisperForConditionalGeneration" in self.hf_config.architectures:
+            # Whisper has unique layer ID scheme:
+            # - Encoder self-attention: 0 to encoder_layers-1 (no KV cache)
+            # - Decoder self-attention: encoder_layers to encoder_layers+decoder_layers-1 (uses KV cache)
+            # - Decoder cross-attention: encoder_layers+decoder_layers to encoder_layers+2*decoder_layers-1
+            # Even though cross-attention doesn't save KV cache, attention backend needs buffer to exist
+            encoder_layers = getattr(self.hf_text_config, "encoder_layers", 0)
+            decoder_layers = getattr(
+                self.hf_text_config, "decoder_layers", self.num_hidden_layers
+            )
+            self.num_attention_layers = encoder_layers + 2 * decoder_layers
         self.num_nextn_predict_layers = getattr(
             self.hf_text_config, "num_nextn_predict_layers", None
         )
@@ -1247,6 +1258,7 @@ multimodal_model_archs = [
     "InternS1ForConditionalGeneration",
     "InternS1ProForConditionalGeneration",
     "Phi4MMForCausalLM",
+    "WhisperForConditionalGeneration",
     "Step3VLForConditionalGeneration",
     "POINTSV15ChatModel",
     "DotsVLMForCausalLM",
@@ -1285,11 +1297,17 @@ def is_image_gen_model(model_architectures: List[str]):
 
 
 def is_audio_model(model_architectures: List[str]):
-    return False
+    models = [
+        "WhisperForConditionalGeneration",
+    ]
+    return any(model in model_architectures for model in models)
 
 
 def is_encoder_decoder_model(model_architectures: List[str]):
-    return "MllamaForConditionalGeneration" in model_architectures
+    models = [
+        "WhisperForConditionalGeneration",
+    ]
+    return any(model in model_architectures for model in models)
 
 
 def is_local_attention_model(model_architectures: List[str]):

python/sglang/srt/entrypoints/http_server.py

@@ -46,7 +46,7 @@ import orjson
 import requests
 import uvicorn
 import uvloop
-from fastapi import Depends, FastAPI, HTTPException, Request
+from fastapi import Depends, FastAPI, File, Form, HTTPException, Request, UploadFile
 from fastapi.exceptions import RequestValidationError
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import ORJSONResponse, Response, StreamingResponse
@@ -93,6 +93,9 @@ from sglang.srt.entrypoints.openai.serving_tokenize import (
     OpenAIServingDetokenize,
     OpenAIServingTokenize,
 )
+from sglang.srt.entrypoints.openai.serving_transcription import (
+    OpenAIServingTranscription,
+)
 from sglang.srt.entrypoints.warmup import execute_warmups
 from sglang.srt.environ import envs
 from sglang.srt.function_call.function_call_parser import FunctionCallParser
@@ -298,6 +301,9 @@ async def lifespan(fast_api_app: FastAPI):
     fast_api_app.state.openai_serving_detokenize = OpenAIServingDetokenize(
         _global_state.tokenizer_manager
     )
+    fast_api_app.state.openai_serving_transcription = OpenAIServingTranscription(
+        _global_state.tokenizer_manager
+    )
 
     # Initialize Ollama-compatible serving handler
     fast_api_app.state.ollama_serving = OllamaServing(_global_state.tokenizer_manager)
@@ -1418,6 +1424,38 @@ async def openai_v1_detokenize(request: DetokenizeRequest, raw_request: Request)
     )
 
 
+@app.post("/v1/audio/transcriptions")
+async def openai_v1_audio_transcriptions(
+    raw_request: Request,
+    file: UploadFile = File(...),
+    model: str = Form(default="default"),
+    language: Optional[str] = Form(default=None),
+    response_format: str = Form(default="json"),
+    temperature: float = Form(default=0.0),
+    stream: bool = Form(default=False),
+):
+    """OpenAI-compatible audio transcription endpoint."""
+    if response_format not in ["json", "text"]:
+        return ORJSONResponse(
+            content={"error": {"message": "Only 'json' and 'text' formats supported"}},
+            status_code=400,
+        )
+
+    audio_data = await file.read()
+
+    return (
+        await raw_request.app.state.openai_serving_transcription.create_transcription(
+            audio_data=audio_data,
+            model=model,
+            language=language,
+            response_format=response_format,
+            temperature=temperature,
+            stream=stream,
+            raw_request=raw_request,
+        )
+    )
+
+
 @app.get("/v1/models", response_class=ORJSONResponse)
 async def available_models():
     """Show available models. OpenAI-compatible endpoint."""

python/sglang/srt/entrypoints/openai/protocol.py

@@ -1389,3 +1389,51 @@ class ResponseReasoningTextContent(BaseModel):
 ResponseInputOutputItem: TypeAlias = Union[
     ResponseInputItemParam, "ResponseReasoningItem", ResponseFunctionToolCall
 ]
+
+
+# ================== Transcription API Protocol Definitions ==================
+
+
+class TranscriptionRequest(BaseModel):
+    """Request model for audio transcription (OpenAI-compatible)."""
+
+    model: str = DEFAULT_MODEL_NAME
+    language: Optional[str] = None
+    response_format: str = "json"
+    temperature: float = 0.0
+    stream: bool = False
+    # Internal fields (not from API)
+    audio_data: Optional[bytes] = None
+    audio_duration_s: float = 0.0
+
+
+class TranscriptionUsage(BaseModel):
+    """Usage info for transcription response (duration-based)."""
+
+    type: Literal["duration"] = "duration"
+    seconds: int  # Audio duration in seconds (rounded up)
+
+
+class TranscriptionResponse(BaseModel):
+    """Non-streaming transcription response (OpenAI-compatible)."""
+
+    text: str
+    usage: Optional[TranscriptionUsage] = None
+
+
+class TranscriptionStreamChoice(BaseModel):
+    """Delta content for streaming transcription."""
+
+    delta: DeltaMessage
+    finish_reason: Optional[str] = None
+
+
+class TranscriptionStreamResponse(BaseModel):
+    """Streaming transcription chunk (OpenAI-compatible)."""
+
+    id: str = Field(default_factory=lambda: f"trsc-{uuid.uuid4().hex}")
+    object: Literal["transcription.chunk"] = "transcription.chunk"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+    choices: List[TranscriptionStreamChoice]
+    usage: Optional[UsageInfo] = None

python/sglang/srt/entrypoints/openai/serving_transcription.py

@@ -0,0 +1,220 @@
+# Copyright 2025 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""
+OpenAI-compatible transcription endpoint handler for Whisper models.
+"""
+from __future__ import annotations
+
+import io
+import logging
+import math
+import time
+import uuid
+from typing import TYPE_CHECKING, AsyncGenerator, Optional, Union
+
+from fastapi import Request
+from fastapi.responses import ORJSONResponse, Response, StreamingResponse
+
+from sglang.srt.entrypoints.openai.protocol import (
+    DeltaMessage,
+    ErrorResponse,
+    TranscriptionRequest,
+    TranscriptionResponse,
+    TranscriptionStreamChoice,
+    TranscriptionStreamResponse,
+    TranscriptionUsage,
+)
+from sglang.srt.entrypoints.openai.serving_base import OpenAIServingBase
+from sglang.srt.managers.io_struct import GenerateReqInput
+
+if TYPE_CHECKING:
+    from sglang.srt.managers.tokenizer_manager import TokenizerManager
+
+logger = logging.getLogger(__name__)
+
+
+class OpenAIServingTranscription(OpenAIServingBase):
+    """Handler for /v1/audio/transcriptions requests"""
+
+    def __init__(self, tokenizer_manager: TokenizerManager):
+        super().__init__(tokenizer_manager)
+
+    def _request_id_prefix(self) -> str:
+        return "trsc-"
+
+    def _validate_request(self, request: TranscriptionRequest) -> Optional[str]:
+        """Validate transcription request."""
+        # Validation is done in the route handler for form data
+        return None
+
+    def _convert_to_internal_request(
+        self,
+        request: TranscriptionRequest,
+        raw_request: Request = None,
+    ) -> tuple[GenerateReqInput, TranscriptionRequest]:
+        """Convert transcription request to internal format."""
+        # Build sampling params - include language for WhisperProcessor
+        sampling_params = {
+            "temperature": request.temperature,
+            "max_new_tokens": 448,  # Whisper default max tokens
+            "language": request.language,  # Pass to WhisperProcessor for language-specific decoding
+        }
+
+        # For Whisper, we pass audio_data and let the processor handle it
+        adapted_request = GenerateReqInput(
+            text="",  # Empty text - Whisper processor will set proper decoder tokens
+            audio_data=request.audio_data,
+            sampling_params=sampling_params,
+            stream=request.stream,
+            modalities=["audio"],
+            routing_key=self.extract_routing_key(raw_request),
+        )
+
+        return adapted_request, request
+
+    def _get_audio_duration(self, audio_data: bytes) -> float:
+        """Calculate audio duration in seconds."""
+        try:
+            import soundfile as sf
+
+            audio_array, sr = sf.read(io.BytesIO(audio_data))
+            duration = len(audio_array) / sr
+            return duration
+        except Exception as e:
+            logger.warning(f"Could not calculate audio duration: {e}")
+            return 0.0
+
+    async def create_transcription(
+        self,
+        audio_data: bytes,
+        model: str,
+        language: Optional[str],
+        response_format: str,
+        temperature: float,
+        stream: bool,
+        raw_request: Request,
+    ) -> Union[TranscriptionResponse, StreamingResponse, Response, ORJSONResponse]:
+        """Main entry point for transcription requests."""
+        # Calculate audio duration for usage reporting
+        audio_duration_s = self._get_audio_duration(audio_data)
+
+        # Build request
+        request = TranscriptionRequest(
+            audio_data=audio_data,
+            model=model,
+            language=language,
+            response_format=response_format,
+            temperature=temperature,
+            stream=stream,
+            audio_duration_s=audio_duration_s,
+        )
+
+        # Use the base class handle_request pattern
+        return await self.handle_request(request, raw_request)
+
+    async def _handle_non_streaming_request(
+        self,
+        adapted_request: GenerateReqInput,
+        request: TranscriptionRequest,
+        raw_request: Request,
+    ) -> Union[TranscriptionResponse, ErrorResponse, ORJSONResponse, Response]:
+        """Handle non-streaming transcription request."""
+        try:
+            ret = await self.tokenizer_manager.generate_request(
+                adapted_request, raw_request
+            ).__anext__()
+        except ValueError as e:
+            return self.create_error_response(str(e))
+
+        text = ret.get("text", "")
+
+        # Build response based on format
+        if request.response_format == "text":
+            return Response(content=text, media_type="text/plain")
+
+        # JSON format
+        usage = TranscriptionUsage(seconds=int(math.ceil(request.audio_duration_s)))
+
+        return TranscriptionResponse(text=text, usage=usage)
+
+    async def _handle_streaming_request(
+        self,
+        adapted_request: GenerateReqInput,
+        request: TranscriptionRequest,
+        raw_request: Request,
+    ) -> StreamingResponse:
+        """Handle streaming transcription request."""
+        return StreamingResponse(
+            self._generate_transcription_stream(adapted_request, request, raw_request),
+            media_type="text/event-stream",
+            background=self.tokenizer_manager.create_abort_task(adapted_request),
+        )
+
+    async def _generate_transcription_stream(
+        self,
+        adapted_request: GenerateReqInput,
+        request: TranscriptionRequest,
+        raw_request: Request,
+    ) -> AsyncGenerator[str, None]:
+        """Generate streaming transcription response."""
+        created_time = int(time.time())
+        request_id = f"{self._request_id_prefix()}{uuid.uuid4().hex}"
+        model = request.model
+        stream_buffer = ""
+
+        try:
+            async for content in self.tokenizer_manager.generate_request(
+                adapted_request, raw_request
+            ):
+                finish_reason = content["meta_info"]["finish_reason"]
+                finish_reason_type = finish_reason["type"] if finish_reason else None
+
+                # Calculate delta (new text since last chunk)
+                current_text = content.get("text", "")
+                delta = current_text[len(stream_buffer) :]
+                stream_buffer = current_text
+
+                # Send content delta if there's new text
+                if delta:
+                    choice_data = TranscriptionStreamChoice(
+                        delta=DeltaMessage(content=delta),
+                        finish_reason=None,
+                    )
+                    chunk = TranscriptionStreamResponse(
+                        id=request_id,
+                        created=created_time,
+                        model=model,
+                        choices=[choice_data],
+                    )
+                    yield f"data: {chunk.model_dump_json()}\n\n"
+
+                # Send finish reason when done
+                if finish_reason_type:
+                    choice_data = TranscriptionStreamChoice(
+                        delta=DeltaMessage(),
+                        finish_reason=finish_reason_type,
+                    )
+                    chunk = TranscriptionStreamResponse(
+                        id=request_id,
+                        created=created_time,
+                        model=model,
+                        choices=[choice_data],
+                    )
+                    yield f"data: {chunk.model_dump_json()}\n\n"
+
+        except ValueError as e:
+            error = self.create_streaming_error_response(str(e))
+            yield f"data: {error}\n\n"
+
+        yield "data: [DONE]\n\n"

python/sglang/srt/managers/tokenizer_manager.py

@@ -694,9 +694,15 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
                     "the engine with skip_tokenizer_init=False."
                 )
 
-            input_ids, token_type_ids = await self._tokenize_texts(
+            # For audio-only requests (e.g., Whisper), text may be empty.
-                input_text, is_cross_encoder_request
+            # The multimodal processor will provide input_ids later.
-            )
+            if not input_text and self.mm_processor and obj.contains_mm_input():
+                # Use empty placeholder - multimodal processor will override
+                input_ids = []
+            else:
+                input_ids, token_type_ids = await self._tokenize_texts(
+                    input_text, is_cross_encoder_request
+                )
 
         if self.mm_processor and obj.contains_mm_input():
             if obj.image_data is not None and not isinstance(obj.image_data, list):

python/sglang/srt/models/whisper.py

@@ -0,0 +1,543 @@
+from typing import Any, Iterable, List, Optional, Tuple
+
+import torch
+from transformers import WhisperConfig
+
+from sglang.srt.distributed import get_tensor_model_parallel_world_size
+from sglang.srt.layers.activation import get_act_fn
+from sglang.srt.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.quantization import QuantizationConfig
+from sglang.srt.layers.radix_attention import AttentionType, RadixAttention
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+from sglang.srt.managers.schedule_batch import MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+
+
+class WhisperAttention(torch.nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        bias: bool = True,
+        layer_id: Optional[int] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        is_cross_attention: bool = False,
+        is_encoder=False,
+    ):
+        super().__init__()
+        self.total_num_heads = num_heads
+        head_dim = embed_dim // num_heads
+        self.is_cross_attention = is_cross_attention
+        self.is_encoder = is_encoder
+
+        tp_size = get_tensor_model_parallel_world_size()
+        assert (
+            num_heads % tp_size == 0
+        ), f"num_heads ({num_heads}) must be divisible by tp_size ({tp_size})"
+        self.num_heads = num_heads // tp_size
+
+        if (head_dim * num_heads) != embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = head_dim**-0.5
+        self.head_dim = head_dim
+        self.kv_size = self.num_heads * head_dim
+
+        if is_cross_attention:
+            self.q_proj = ColumnParallelLinear(
+                embed_dim, embed_dim, quant_config=quant_config
+            )
+            self.kv_proj = QKVParallelLinear(
+                hidden_size=embed_dim,
+                head_size=head_dim,
+                total_num_heads=0,
+                total_num_kv_heads=num_heads,
+                bias=bias,
+                quant_config=quant_config,
+            )
+        else:
+            self.qkv_proj = QKVParallelLinear(
+                embed_dim, head_dim, num_heads, quant_config=quant_config
+            )
+        self.out_proj = RowParallelLinear(
+            embed_dim, embed_dim, bias=bias, quant_config=quant_config
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            head_dim,
+            scaling=1.0,
+            num_kv_heads=self.num_heads,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            is_cross_attention=is_cross_attention,
+            attn_type=(
+                AttentionType.ENCODER_ONLY if is_encoder else AttentionType.DECODER
+            ),
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        cross_hidden_states: Optional[torch.Tensor] = None,
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Input shape: Batch x Time x Channel"""
+
+        if self.is_cross_attention:
+            q, _ = self.q_proj(hidden_states)
+            if cross_hidden_states is not None:
+                kv, _ = self.kv_proj(cross_hidden_states)
+                k, v = kv.split([self.kv_size, self.kv_size], dim=-1)
+            else:
+                k = torch.zeros_like(q)
+                v = torch.zeros_like(q)
+
+            q = q * self.scaling
+            num_heads = self.attn.tp_q_head_num
+            head_dim = self.attn.head_dim
+
+            q = q.view(-1, num_heads, head_dim)
+            k = k.view(-1, num_heads, head_dim)
+            v = v.view(-1, num_heads, head_dim)
+
+            q_len = q.shape[0]
+            kv_len = k.shape[0]
+
+            q = q.transpose(0, 1)
+            k = k.transpose(0, 1)
+            v = v.transpose(0, 1)
+
+            attn_weights = torch.bmm(q, k.transpose(1, 2))
+
+            # Apply block-diagonal mask for batched cross-attention
+            batch_size = forward_batch.batch_size if forward_batch else 1
+            if batch_size > 1 and kv_len > 0:
+                encoder_len_per_request = kv_len // batch_size
+                if encoder_len_per_request * batch_size == kv_len:
+                    is_decode = forward_batch.forward_mode.is_decode()
+                    if is_decode:
+                        mask = torch.zeros(
+                            (q_len, kv_len), device=q.device, dtype=torch.bool
+                        )
+                        for i in range(batch_size):
+                            enc_start = i * encoder_len_per_request
+                            enc_end = (i + 1) * encoder_len_per_request
+                            mask[i, enc_start:enc_end] = True
+                        attn_weights = attn_weights.masked_fill(
+                            ~mask.unsqueeze(0), float("-inf")
+                        )
+                    else:
+                        seq_lens = forward_batch.seq_lens
+                        if seq_lens is not None and len(seq_lens) == batch_size:
+                            seq_lens_list = seq_lens.tolist()
+                            mask = torch.zeros(
+                                (q_len, kv_len), device=q.device, dtype=torch.bool
+                            )
+                            q_start = 0
+                            for i, dec_len in enumerate(seq_lens_list):
+                                enc_start = i * encoder_len_per_request
+                                enc_end = (i + 1) * encoder_len_per_request
+                                q_end = q_start + dec_len
+                                mask[q_start:q_end, enc_start:enc_end] = True
+                                q_start = q_end
+                            attn_weights = attn_weights.masked_fill(
+                                ~mask.unsqueeze(0), float("-inf")
+                            )
+
+            attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
+            attn_output = torch.bmm(attn_weights, v)
+            attn_output = attn_output.transpose(0, 1)
+            attn_output = attn_output.reshape(q_len, num_heads * head_dim)
+        else:
+            qkv, _ = self.qkv_proj(hidden_states)
+            q, k, v = qkv.chunk(chunks=3, dim=-1)
+            q = q * self.scaling
+
+            if self.is_encoder:
+                num_heads = self.attn.tp_q_head_num
+                head_dim = self.attn.head_dim
+                batch_size, seq_len, _ = hidden_states.shape
+
+                q = q.view(batch_size, seq_len, num_heads, head_dim).permute(0, 2, 1, 3)
+                k = k.view(batch_size, seq_len, num_heads, head_dim).permute(0, 2, 1, 3)
+                v = v.view(batch_size, seq_len, num_heads, head_dim).permute(0, 2, 1, 3)
+
+                attn_output = torch.nn.functional.scaled_dot_product_attention(
+                    q, k, v, scale=1.0
+                )
+                attn_output = attn_output.permute(0, 2, 1, 3).reshape(
+                    batch_size, seq_len, num_heads * head_dim
+                )
+            else:
+                attn_output = self.attn(q, k, v, forward_batch, save_kv_cache=True)
+
+        attn_output, _ = self.out_proj(attn_output)
+
+        return attn_output
+
+
+class WhisperEncoderLayer(torch.nn.Module):
+    def __init__(
+        self,
+        config: WhisperConfig,
+        layer_id: Optional[int] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = WhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            is_encoder=True,
+        )
+        self.self_attn_layer_norm = torch.nn.LayerNorm(self.embed_dim)
+
+        self.activation_fn = get_act_fn(
+            config.activation_function, quant_config=quant_config
+        )
+
+        self.fc1 = ColumnParallelLinear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = RowParallelLinear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = torch.nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states = self.self_attn(hidden_states, forward_batch)
+
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+
+        hidden_states, _ = self.fc2(hidden_states)
+
+        hidden_states = residual + hidden_states
+
+        if hidden_states.dtype == torch.float16:
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(
+                hidden_states, min=-clamp_value, max=clamp_value
+            )
+        return hidden_states
+
+
+class WhisperDecoderLayer(torch.nn.Module):
+    def __init__(
+        self,
+        config: WhisperConfig,
+        layer_id: Optional[int] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        # Offset decoder layer IDs to avoid overlap with encoder layers
+        decoder_self_attn_layer_id = config.encoder_layers + layer_id
+        decoder_cross_attn_layer_id = (
+            config.encoder_layers + config.decoder_layers + layer_id
+        )
+
+        self.self_attn = WhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            layer_id=decoder_self_attn_layer_id,
+            quant_config=quant_config,
+        )
+
+        self.activation_fn = get_act_fn(
+            config.activation_function, quant_config=quant_config
+        )
+
+        self.self_attn_layer_norm = torch.nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = WhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            layer_id=decoder_cross_attn_layer_id,
+            quant_config=quant_config,
+            is_cross_attention=True,
+        )
+        self.encoder_attn_layer_norm = torch.nn.LayerNorm(self.embed_dim)
+        self.fc1 = ColumnParallelLinear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = RowParallelLinear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = torch.nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        decoder_hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor],
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+
+        residual = decoder_hidden_states
+        decoder_hidden_states = self.self_attn_layer_norm(decoder_hidden_states)
+        decoder_hidden_states = self.self_attn(decoder_hidden_states, forward_batch)
+        decoder_hidden_states = residual + decoder_hidden_states
+
+        residual = decoder_hidden_states
+        decoder_hidden_states = self.encoder_attn_layer_norm(decoder_hidden_states)
+        decoder_hidden_states = self.encoder_attn(
+            decoder_hidden_states, forward_batch, encoder_hidden_states
+        )
+        decoder_hidden_states = residual + decoder_hidden_states
+
+        residual = decoder_hidden_states
+        decoder_hidden_states = self.final_layer_norm(decoder_hidden_states)
+        decoder_hidden_states, _ = self.fc1(decoder_hidden_states)
+        decoder_hidden_states = self.activation_fn(decoder_hidden_states)
+        decoder_hidden_states, _ = self.fc2(decoder_hidden_states)
+
+        decoder_hidden_states = residual + decoder_hidden_states
+
+        return decoder_hidden_states
+
+
+class WhisperEncoder(torch.nn.Module):
+
+    def __init__(
+        self, config: WhisperConfig, quant_config: Optional[QuantizationConfig] = None
+    ):
+        super().__init__()
+
+        embed_dim = config.d_model
+        self.embed_scale = embed_dim**-0.5 if config.scale_embedding else 1.0
+
+        self.conv1 = torch.nn.Conv1d(
+            config.num_mel_bins, embed_dim, kernel_size=3, padding=1
+        )
+        self.conv2 = torch.nn.Conv1d(
+            embed_dim, embed_dim, kernel_size=3, stride=2, padding=1
+        )
+        self.embed_positions = torch.nn.Embedding(
+            config.max_source_positions, embed_dim
+        )
+
+        self.layers = torch.nn.ModuleList(
+            [
+                WhisperEncoderLayer(config, id, quant_config)
+                for id in range(config.encoder_layers)
+            ]
+        )
+        self.layer_norm = torch.nn.LayerNorm(config.d_model)
+
+    def forward(
+        self,
+        input_features: torch.Tensor,
+        position_ids: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ):
+        inputs_embeds = torch.nn.functional.gelu(self.conv1(input_features))
+        inputs_embeds = torch.nn.functional.gelu(self.conv2(inputs_embeds))
+
+        inputs_embeds = inputs_embeds.mT
+
+        hidden_states = inputs_embeds + self.embed_positions(position_ids)
+
+        for encoder_layer in self.layers:
+            hidden_states = encoder_layer(hidden_states, forward_batch)
+
+        hidden_states = self.layer_norm(hidden_states)
+        return hidden_states
+
+
+class WhisperDecoder(torch.nn.Module):
+
+    def __init__(
+        self, config: WhisperConfig, quant_config: Optional[QuantizationConfig] = None
+    ):
+        super().__init__()
+        self.max_target_positions = config.max_target_positions
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = config.d_model**-0.5 if config.scale_embedding else 1.0
+
+        self.embed_tokens = torch.nn.Embedding(
+            config.vocab_size, config.d_model, padding_idx=config.pad_token_id
+        )
+        self.embed_positions = torch.nn.Embedding(
+            self.max_target_positions, config.d_model
+        )
+
+        self.layers = torch.nn.ModuleList(
+            [
+                WhisperDecoderLayer(config, layer_idx, quant_config)
+                for layer_idx in range(config.decoder_layers)
+            ]
+        )
+
+        self.layer_norm = torch.nn.LayerNorm(config.d_model)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor],
+        forward_batch: ForwardBatch,
+        position_ids=None,
+    ):
+        inputs_embeds = self.embed_tokens(input_ids)
+        positions = self.embed_positions(position_ids)
+        hidden_states = inputs_embeds + positions.to(inputs_embeds.device)
+
+        for decoder_layer in self.layers:
+            hidden_states = decoder_layer(
+                hidden_states, encoder_hidden_states, forward_batch
+            )
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        return hidden_states
+
+
+class WhisperForConditionalGeneration(torch.nn.Module):
+
+    def __init__(
+        self, config: WhisperConfig, quant_config: Optional[QuantizationConfig] = None
+    ):
+        super().__init__()
+        self.encoder = WhisperEncoder(config, quant_config)
+        self.decoder = WhisperDecoder(config, quant_config)
+        self.proj_out = ParallelLMHead(
+            config.vocab_size, config.d_model, quant_config=quant_config
+        )
+        self.logits_processor = LogitsProcessor(config)
+        self.config = config
+        self._encoder_cache = {}
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            (".self_attn.qkv_proj", ".self_attn.q_proj", "q"),
+            (".self_attn.qkv_proj", ".self_attn.k_proj", "k"),
+            (".self_attn.qkv_proj", ".self_attn.v_proj", "v"),
+            (".encoder_attn.kv_proj", ".encoder_attn.k_proj", "k"),
+            (".encoder_attn.kv_proj", ".encoder_attn.v_proj", "v"),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        weights_dict = dict(weights)
+
+        # Whisper has no k_proj bias, create zeros
+        for layer_idx in range(self.config.decoder_layers):
+            layer_prefix = f"model.decoder.layers.{layer_idx}.encoder_attn."
+            k_proj_key = layer_prefix + "k_proj.weight"
+            if k_proj_key in weights_dict:
+                k_proj_weight = weights_dict[k_proj_key]
+                bias_key = layer_prefix + "k_proj.bias"
+                if bias_key not in weights_dict:
+                    weights_dict[bias_key] = torch.zeros(k_proj_weight.size(0))
+
+        weights_dict["proj_out.weight"] = weights_dict[
+            "model.decoder.embed_tokens.weight"
+        ]
+
+        for name, loaded_weight in weights_dict.items():
+            name = name.replace("model.", "")
+
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                if name not in params_dict:
+                    break
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                if name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+    def pad_input_ids(self, input_ids: List[int], _mm_inputs: MultimodalInputs):
+        return input_ids
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        **kwargs: Any,
+    ) -> LogitsProcessorOutput:
+        dtype = self.encoder.conv1.weight.dtype
+        is_decode = forward_batch.forward_mode.is_decode()
+
+        if is_decode:
+            encoder_outputs = None
+            if forward_batch.req_pool_indices is not None:
+                req_indices = forward_batch.req_pool_indices.tolist()
+                encoder_list = []
+                for req_idx in req_indices:
+                    if req_idx in self._encoder_cache:
+                        encoder_list.append(self._encoder_cache[req_idx])
+                if encoder_list:
+                    encoder_outputs = torch.cat(encoder_list, dim=0)
+        else:
+            encoder_list = []
+            mm_inputs_list = forward_batch.mm_inputs if forward_batch.mm_inputs else []
+            req_indices = (
+                forward_batch.req_pool_indices.tolist()
+                if forward_batch.req_pool_indices is not None
+                else []
+            )
+
+            for req_idx, mm_input in zip(req_indices, mm_inputs_list):
+                if mm_input is None or not mm_input.mm_items:
+                    continue
+
+                features = mm_input.mm_items[0].feature
+                if features.ndim == 2:
+                    features = features.unsqueeze(0)
+
+                encoder_len = features.shape[-1] // 2
+                encoder_position_ids = torch.arange(encoder_len).to(
+                    features.device, non_blocking=True
+                )
+
+                req_encoder_outputs = self.encoder(
+                    features.to(dtype), encoder_position_ids, forward_batch
+                )
+                req_encoder_outputs = req_encoder_outputs.squeeze(0)
+
+                self._encoder_cache[req_idx] = req_encoder_outputs
+                encoder_list.append(req_encoder_outputs)
+
+            if encoder_list:
+                encoder_outputs = torch.cat(encoder_list, dim=0)
+            else:
+                encoder_outputs = None
+
+        decoder_outputs = self.decoder(
+            input_ids, encoder_outputs, forward_batch, positions
+        )
+
+        logits = self.logits_processor(
+            input_ids=input_ids,
+            lm_head=self.proj_out,
+            hidden_states=decoder_outputs,
+            logits_metadata=forward_batch,
+        )
+
+        return logits
+
+
+EntryClass = [WhisperForConditionalGeneration]

python/sglang/srt/multimodal/processors/whisper.py

@@ -0,0 +1,191 @@
+import logging
+from typing import Any, Dict, Optional
+
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
+from sglang.srt.models.whisper import WhisperForConditionalGeneration
+from sglang.srt.multimodal.processors.base_processor import BaseMultimodalProcessor
+from sglang.srt.utils import load_audio
+
+logger = logging.getLogger(__name__)
+
+# ISO 639-1 supported languages for Whisper
+# From https://platform.openai.com/docs/guides/speech-to-text/supported-languages
+# Maps ISO 639-1 code -> Full language name
+ISO639_1_SUPPORTED_LANGS = {
+    "af": "Afrikaans",
+    "ar": "Arabic",
+    "hy": "Armenian",
+    "az": "Azerbaijani",
+    "be": "Belarusian",
+    "bs": "Bosnian",
+    "bg": "Bulgarian",
+    "ca": "Catalan",
+    "zh": "Chinese",
+    "hr": "Croatian",
+    "cs": "Czech",
+    "da": "Danish",
+    "nl": "Dutch",
+    "en": "English",
+    "et": "Estonian",
+    "fi": "Finnish",
+    "fr": "French",
+    "gl": "Galician",
+    "de": "German",
+    "el": "Greek",
+    "he": "Hebrew",
+    "hi": "Hindi",
+    "hu": "Hungarian",
+    "is": "Icelandic",
+    "id": "Indonesian",
+    "it": "Italian",
+    "ja": "Japanese",
+    "kn": "Kannada",
+    "kk": "Kazakh",
+    "ko": "Korean",
+    "lv": "Latvian",
+    "lt": "Lithuanian",
+    "mk": "Macedonian",
+    "ms": "Malay",
+    "mr": "Marathi",
+    "mi": "Maori",
+    "ne": "Nepali",
+    "no": "Norwegian",
+    "fa": "Persian",
+    "pl": "Polish",
+    "pt": "Portuguese",
+    "ro": "Romanian",
+    "ru": "Russian",
+    "sr": "Serbian",
+    "sk": "Slovak",
+    "sl": "Slovenian",
+    "es": "Spanish",
+    "sw": "Swahili",
+    "sv": "Swedish",
+    "tl": "Tagalog",
+    "ta": "Tamil",
+    "th": "Thai",
+    "tr": "Turkish",
+    "uk": "Ukrainian",
+    "ur": "Urdu",
+    "vi": "Vietnamese",
+    "cy": "Welsh",
+}
+
+# Reverse mapping: Full language name (lowercase) -> ISO 639-1 code
+LANG_NAME_TO_CODE = {
+    name.lower(): code for code, name in ISO639_1_SUPPORTED_LANGS.items()
+}
+
+
+def normalize_language_to_code(language: Optional[str]) -> Optional[str]:
+    """Convert a language input (full name or code) to ISO 639-1 code.
+
+    Args:
+        language: Language as full name (e.g., 'English', 'Spanish') or
+                  ISO 639-1 code (e.g., 'en', 'es')
+
+    Returns:
+        ISO 639-1 code or None if input is None
+    """
+    if language is None:
+        return None
+
+    language_lower = language.lower().strip()
+
+    # Check if it's already a valid ISO code
+    if language_lower in ISO639_1_SUPPORTED_LANGS:
+        return language_lower
+
+    # Check if it's a full language name
+    if language_lower in LANG_NAME_TO_CODE:
+        return LANG_NAME_TO_CODE[language_lower]
+
+    # Not recognized
+    raise ValueError(
+        f"Language '{language}' not recognized. "
+        f"Use full name (e.g., 'English') or ISO 639-1 code (e.g., 'en')."
+    )
+
+
+class WhisperProcessor(BaseMultimodalProcessor):
+    models = [WhisperForConditionalGeneration]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+        # Cache tokenizer for language token lookup
+        self._tokenizer = getattr(self._processor, "tokenizer", None)
+
+    def _extract_language_from_request(self, request_obj) -> Optional[str]:
+        sampling_params = getattr(request_obj, "sampling_params", None) or {}
+        language = sampling_params.pop("language", None)
+        return normalize_language_to_code(language)
+
+    def _get_language_token_id(self, language: Optional[str]) -> int:
+        # Default to English if not specified
+        if language is None:
+            language = "en"  # Default to English
+        language_token = f"<|{language}|>"
+        return self._tokenizer.convert_tokens_to_ids(language_token)
+
+    async def process_mm_data_async(
+        self,
+        image_data,
+        audio_data,
+        input_text,
+        request_obj,
+        **kwargs,
+    ) -> Optional[Dict[str, Any]]:
+        if not audio_data:
+            return None
+
+        if len(audio_data) != 1:
+            raise ValueError(
+                f"Whisper expects exactly 1 audio input, got {len(audio_data)}"
+            )
+
+        audios = [load_audio(audio) for audio in audio_data]
+
+        # For Whisper, ALWAYS use the proper transcription token sequence
+        # and IGNORE any text prompt - Whisper is a pure speech-to-text model
+        # The decoder_start_token_id and forced_decoder_ids from generation config
+        # set up: <|startoftranscript|> <|lang|> <|task|> [<|notimestamps|>]
+
+        # Extract language from request and get token ID
+        language = self._extract_language_from_request(request_obj)
+        language_token_id = self._get_language_token_id(language)
+
+        # Build decoder input tokens
+        # <|startoftranscript|> + <|lang|> + <|transcribe|> + <|notimestamps|>
+        decoder_start_token_id = getattr(
+            self.hf_config, "decoder_start_token_id", 50258
+        )
+        transcribe_token_id = self._tokenizer.convert_tokens_to_ids("<|transcribe|>")
+        notimestamps_token_id = self._tokenizer.convert_tokens_to_ids(
+            "<|notimestamps|>"
+        )
+
+        input_ids = [
+            decoder_start_token_id,
+            language_token_id,
+            transcribe_token_id,
+            notimestamps_token_id,
+        ]
+
+        # Whisper expects input features padded to max_length (3000 frames = 30 seconds)
+        # This is the standard context length for Whisper
+        input_features = self._processor.feature_extractor(
+            audios[0],
+            sampling_rate=16000,
+            padding="max_length",  # Pad to 3000 frames
+            return_tensors="pt",
+        )["input_features"][0]
+
+        return {
+            "input_ids": input_ids,
+            "mm_items": [
+                MultimodalDataItem(
+                    feature=input_features,
+                    modality=Modality.AUDIO,
+                )
+            ],
+        }

python/sglang/srt/parser/conversation.py

@@ -1027,6 +1027,23 @@ register_conv_template(
     )
 )
 
+# Whisper speech-to-text template
+# Whisper uses special tokens: <|startoftranscript|>, <|en|>, <|transcribe|>, etc.
+# Audio features are processed by encoder separately, not inserted into text
+# The decoder start tokens (task, language) should be set via generation config
+register_conv_template(
+    Conversation(
+        name="whisper",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        sep_style=SeparatorStyle.NO_COLON_SINGLE,
+        sep="",
+        stop_str=["<|endoftext|>"],
+        audio_token="",  # Empty - audio is handled by encoder, not as text token
+    )
+)
+
 MODEL_TYPE_TO_TEMPLATE = {
     "internvl_chat": "internvl-2-5",
     "deepseek_vl_v2": "deepseek-vl2",
@@ -1036,6 +1053,7 @@ MODEL_TYPE_TO_TEMPLATE = {
     "minicpmo": "minicpmo",
     "deepseek-ocr": "deepseek-ocr",
     "paddleocr_vl": "paddle-ocr",
+    "whisper": "whisper",
 }
 
 
@@ -1129,3 +1147,11 @@ def match_paddle_ocr(model_path: str):
         return "paddle-ocr"
     model_type = get_model_type(model_path)
     return MODEL_TYPE_TO_TEMPLATE.get(model_type)
+
+
+@register_conv_template_matching_function
+def match_whisper(model_path: str):
+    if "whisper" in model_path.lower():
+        return "whisper"
+    model_type = get_model_type(model_path)
+    return MODEL_TYPE_TO_TEMPLATE.get(model_type)

python/sglang/srt/server_args.py

@@ -1840,6 +1840,13 @@ class ServerArgs:
                 self.speculative_algorithm is None
             ), "Speculative decoding is currently not supported with Flex Attention backend"
 
+        # Encoder-decoder models (e.g., Whisper)
+        if model_config.is_encoder_decoder:
+            logger.warning(
+                "Cuda graph is disabled for encoder-decoder models (e.g., Whisper)"
+            )
+            self.disable_cuda_graph = True
+
         # Major NVIDIA platforms backends
         if (
             self.attention_backend == "flashmla"