[v1] Add Whisper model support (encoder-decoder) (#21088)

Signed-off-by: Russell Bryant <rbryant@redhat.com>
Co-authored-by: NickLucche <nlucches@redhat.com>
This commit is contained in:
Russell Bryant
2025-09-10 13:53:35 -07:00
committed by GitHub
co-authored by NickLucche
parent 4db4426404
commit 37e8182bfe
31 changed files with 429 additions and 92 deletions
@@ -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,
@@ -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",
@@ -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,
@@ -122,8 +122,7 @@ def run_test(
@pytest.mark.core_model
@pytest.mark.parametrize(
"model", ["openai/whisper-small", "openai/whisper-large-v3-turbo"])
@pytest.mark.parametrize("model", ["openai/whisper-large-v3-turbo"])
@create_new_process_for_each_test()
def test_models(vllm_runner, model) -> None:
run_test(
@@ -31,6 +31,7 @@ from ...utils import dummy_hf_overrides
ARCH_TO_SKIP = {
"MolmoForCausalLM": "incompatible requirements",
"Florence2ForConditionalGeneration": "not supported in V1",
}
ARCH_NEEDS_EXTRAS = [
"InternVLChatModel",
+6
View File
@@ -68,6 +68,12 @@ def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
# 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,
-1
View File
@@ -10,7 +10,6 @@ from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.async_llm_engine import AsyncLLMEngine
UNSUPPORTED_MODELS_V1 = [
"openai/whisper-large-v3", # transcription
"facebook/bart-large-cnn", # encoder decoder
]