v1: Add Whisper encoder-decoder model support

Implements Whisper mdoel support in the V1 engine. Key changes include:

- Add encoder-decoder architecture support with cross-attention KV cache management
- Add CrossAttentionManager and CrossAttentionSpec for encoder-decoder KV cache
- Update scheduler to handle cross-attention block allocation and disable prefix caching
- Modify GPU model runner for encoder input processing and attention metadata
- Disable BART / other enc-dec tests/examples (Whisper-only support for now)
- Optimize test performance and fix various integration issues

This closes a major feature gap between V0 and V1, enabling Whisper transcription
in the new engine architecture while maintaining backward compatibility.

Related to V0 deprecation (#18571) and 2025 Q3 roadmap (#20336).

Signed-off-by: Russell Bryant <[email protected]>
Co-authored-by: NickLucche <[email protected]>
Signed-off-by: Russell Bryant <[email protected]>

Author: russellb

.buildkite/test-pipeline.yaml

@@ -295,7 +295,6 @@ steps:
     - python3 offline_inference/vision_language_pooling.py --seed 0
     - python3 offline_inference/vision_language_multi_image.py --seed 0
     - VLLM_USE_V1=0 python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference/encoder_decoder.py
     - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
     - python3 offline_inference/basic/classify.py
     - python3 offline_inference/basic/embed.py
@@ -580,7 +579,7 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
-    - cd .. && pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
+    - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
 - label: Multi-Modal Models Test (Extended) 1
   mirror_hardwares: [amdexperimental]

examples/offline_inference/encoder_decoder.py

@@ -5,6 +5,8 @@
 encoder/decoder models, specifically BART and mBART.
 
 This script is refactored to allow model selection via command-line arguments.
+
+NOTE: This example is not yet supported in V1.
 """
 
 import argparse

tests/encoder_decoder/test_e2e_correctness.py

@@ -63,6 +63,7 @@ def clear_cache():
     current_platform.is_cpu(),
     reason="CPU backend is not currently supported with encoder/decoder models"
 )
+@pytest.mark.skip(reason="bart not supported in V1")
 def test_encoder_decoder_e2e(
     hf_runner,
     vllm_runner,

tests/entrypoints/openai/correctness/test_transcription_api_correctness.py

@@ -49,8 +49,7 @@ async def transcribe_audio(client, tokenizer, y, sr):
     return latency, num_output_tokens, transcription.text
 
 
-async def bound_transcribe(model_name, sem, client, audio, reference):
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
+async def bound_transcribe(sem, client, tokenizer, audio, reference):
     # Use semaphore to limit concurrent requests.
     async with sem:
         result = await transcribe_audio(client, tokenizer, *audio)
@@ -63,15 +62,19 @@ async def bound_transcribe(model_name, sem, client, audio, reference):
 async def process_dataset(model, client, data, concurrent_request):
     sem = asyncio.Semaphore(concurrent_request)
 
+    # Load tokenizer once outside the loop
+    tokenizer = AutoTokenizer.from_pretrained(model)
+
     # Warmup call as the first `librosa.load` server-side is quite slow.
     audio, sr = data[0]["audio"]["array"], data[0]["audio"]["sampling_rate"]
-    _ = await bound_transcribe(model, sem, client, (audio, sr), "")
+    _ = await bound_transcribe(sem, client, tokenizer, (audio, sr), "")
 
     tasks: list[asyncio.Task] = []
     for sample in data:
         audio, sr = sample["audio"]["array"], sample["audio"]["sampling_rate"]
         task = asyncio.create_task(
-            bound_transcribe(model, sem, client, (audio, sr), sample["text"]))
+            bound_transcribe(sem, client, tokenizer, (audio, sr),
+                             sample["text"]))
         tasks.append(task)
     return await asyncio.gather(*tasks)
 

tests/entrypoints/openai/test_encoder_decoder.py

@@ -30,6 +30,7 @@ async def client(server):
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.skip(reason="bart is not yet supported in V1")
 async def test_single_completion(client: openai.AsyncOpenAI, model_name: str):
     completion = await client.completions.create(model=model_name,
                                                  prompt="Hello, my name is",

tests/models/language/generation/test_bart.py

@@ -178,6 +178,7 @@ def run_test(
 @pytest.mark.parametrize("max_tokens", [64])
 @pytest.mark.parametrize("num_logprobs", [5])
 @pytest.mark.parametrize("decoder_prompt_type", list(DecoderPromptType))
+@pytest.mark.skip(reason="bart not supported in V1")
 def test_models(hf_runner, vllm_runner, example_encoder_decoder_prompts, model,
                 dtype, max_tokens, num_logprobs, decoder_prompt_type) -> None:
 
@@ -201,6 +202,7 @@ def test_models(hf_runner, vllm_runner, example_encoder_decoder_prompts, model,
 @pytest.mark.parametrize("max_tokens", [64])
 @pytest.mark.parametrize("num_logprobs", [5])
 @pytest.mark.parametrize("decoder_prompt_type", [DecoderPromptType.CUSTOM])
+@pytest.mark.skip(reason="bart not supported in V1")
 def test_models_distributed(hf_runner, vllm_runner,
                             example_encoder_decoder_prompts,
                             distributed_executor_backend, model, dtype,

tests/models/multimodal/processing/test_tensor_schema.py

@@ -30,6 +30,7 @@
 
 ARCH_TO_SKIP = {
     "MolmoForCausalLM": "incompatible requirements",
+    "Florence2ForConditionalGeneration": "not supported in V1",
 }
 ARCH_NEEDS_EXTRAS = [
     "InternVLChatModel",

tests/models/test_initialization.py

@@ -68,6 +68,12 @@ def _initialize_kv_caches_v1(self, vllm_config):
             # has cc==8.9 which hasn't supported FA3 yet. Remove this hack when
             # L4 supports FA3.
             m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN_VLLM_V1")
+        if model_arch == "Florence2ForConditionalGeneration":
+            # An encoder-decoder model that's V0-only. Just skip it
+            # since V0 is about to be removed.
+            pytest.skip("Skipping Florence2ForConditionalGeneration")
+        if model_arch == "WhisperForConditionalGeneration":
+            m.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
         LLM(
             model_info.default,
             tokenizer=model_info.tokenizer,

tests/v1/test_oracle.py

@@ -10,7 +10,6 @@
 from vllm.engine.async_llm_engine import AsyncLLMEngine
 
 UNSUPPORTED_MODELS_V1 = [
-    "openai/whisper-large-v3",  # transcription
     "facebook/bart-large-cnn",  # encoder decoder
 ]
 

vllm/attention/layers/cross_attention.py

@@ -0,0 +1,89 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import functools
+from copy import copy
+from typing import Optional
+
+import torch
+from transformers import CacheConfig
+
+from vllm import envs
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.layer import Attention
+from vllm.attention.selector import get_attn_backend
+from vllm.v1.attention.backends.utils import (CommonAttentionMetadata,
+                                              subclass_attention_backend)
+
+
+@functools.lru_cache
+def create_cross_attention_backend(
+    underlying_attn_backend: AttentionBackend, ) -> type[AttentionBackend]:
+    prefix = "CrossAttention_"
+    underlying_builder = underlying_attn_backend.get_builder_cls()
+
+    class CrossAttentionBuilder(underlying_builder):  # type: ignore
+
+        def build(self,
+                  common_prefix_len: int,
+                  common_attn_metadata: CommonAttentionMetadata,
+                  fast_build: bool = False) -> AttentionMetadata:
+            # Cross-attention metadata is built in GPU model runner
+            # We just ensure it's non-causal and pass through
+            new_common_attn_metadata = copy(common_attn_metadata)
+            new_common_attn_metadata.causal = False
+            return super().build(common_prefix_len, new_common_attn_metadata,
+                                 fast_build)
+
+    attn_backend = subclass_attention_backend(
+        name_prefix=prefix,
+        attention_backend_cls=underlying_attn_backend,
+        builder_cls=CrossAttentionBuilder)
+
+    return attn_backend
+
+
+class CrossAttention(Attention):
+    """
+    Cross-attention for encoder-decoder models.
+    Handles attention between decoder queries and encoder keys/values.
+    """
+
+    def __init__(self,
+                 num_heads: int,
+                 head_size: int,
+                 scale: float,
+                 cache_config: Optional[CacheConfig] = None,
+                 attn_type: Optional[str] = None,
+                 **kwargs):
+        dtype = torch.get_default_dtype()
+
+        if cache_config is not None:
+            kv_cache_dtype = cache_config.cache_dtype
+            block_size = cache_config.block_size
+        else:
+            kv_cache_dtype = "auto"
+            block_size = 16
+
+        if envs.VLLM_USE_V1:
+            underlying_attn_backend = get_attn_backend(head_size, dtype,
+                                                       kv_cache_dtype,
+                                                       block_size)
+
+            attn_backend = create_cross_attention_backend(
+                underlying_attn_backend)
+        else:
+            # in v0 cross attention is handled inside the backends
+            attn_backend = None
+
+        if attn_type is not None:
+            assert attn_type == AttentionType.ENCODER_DECODER, (
+                "CrossAttention only supports AttentionType.ENCODER_DECODER")
+
+        super().__init__(num_heads=num_heads,
+                         head_size=head_size,
+                         scale=scale,
+                         cache_config=cache_config,
+                         attn_backend=attn_backend,
+                         attn_type=AttentionType.ENCODER_DECODER,
+                         **kwargs)