diff --git a/tests/conftest.py b/tests/conftest.py index b21cfd5ba85c4..a03f40a9a72ac 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -702,10 +702,16 @@ class HfRunner: **kwargs, ) + # Encoder-decoder models return decoder_hidden_states instead of + # hidden_states + hidden_states = ( + getattr(output, "hidden_states", None) or output.decoder_hidden_states + ) + ( seq_logprobs_lst, output_len, - ) = self._hidden_states_to_logprobs(output.hidden_states, num_logprobs) + ) = self._hidden_states_to_logprobs(hidden_states, num_logprobs) all_logprobs.append(seq_logprobs_lst) seq_ids = output.sequences[0] diff --git a/tests/models/multimodal/generation/test_whisper.py b/tests/models/multimodal/generation/test_whisper.py index 592862c2a0bb0..b206995a9cecc 100644 --- a/tests/models/multimodal/generation/test_whisper.py +++ b/tests/models/multimodal/generation/test_whisper.py @@ -1,150 +1,146 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project +from collections.abc import Sequence +from typing import Any + +import librosa import pytest +from transformers import AutoModelForSpeechSeq2Seq -from vllm import SamplingParams from vllm.assets.audio import AudioAsset +from vllm.platforms import current_platform -from ....conftest import VllmRunner +from ....conftest import HfRunner, PromptAudioInput, VllmRunner from ....utils import create_new_process_for_each_test, multi_gpu_test +from ...registry import HF_EXAMPLE_MODELS +from ...utils import check_logprobs_close -PROMPTS = [ - { - "prompt": "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>", - "multi_modal_data": { - "audio": AudioAsset("mary_had_lamb").audio_and_sample_rate, - }, - }, - { # Test explicit encoder/decoder prompt - "encoder_prompt": { - "prompt": "", - "multi_modal_data": { - "audio": AudioAsset("winning_call").audio_and_sample_rate, - }, - }, - "decoder_prompt": "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>", - }, -] +VLLM_PROMPT = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>" +HF_PROMPT = "" +# Whisper expects 16kHz audio +WHISPER_SAMPLE_RATE = 16000 -EXPECTED = { - "openai/whisper-tiny": [ - " He has birth words I spoke in the original corner of that. And a" - " little piece of black coat poetry. Mary had a little sandwich," - " sweet, with white and snow. And everyone had it very went the last" - " would sure to go.", - " >> And the old one, fit John the way to Edgar Martinez. >> One more" - " to line down the field line for our base camp. Here comes joy. Here" - " is June and the third base. They're going to wave him in. The throw" - " to the plate will be late. The Mariners are going to play for the" - " American League Championship. I don't believe it. It just continues" - " by all five.", - ], - "openai/whisper-small": [ - " The first words I spoke in the original pornograph. A little piece" - " of practical poetry. Mary had a little lamb, its fleece was quite a" - " slow, and everywhere that Mary went the lamb was sure to go.", - " And the old one pitch on the way to Edgar Martinez one month. Here" - " comes joy. Here is Junior to third base. They're gonna wave him" - " in. The throw to the plate will be late. The Mariners are going to" - " play for the American League Championship. I don't believe it. It" - " just continues. My, oh my.", - ], - "openai/whisper-medium": [ - " The first words I spoke in the original phonograph, a little piece" - " of practical poetry. Mary had a little lamb, its fleece was quite as" - " slow, and everywhere that Mary went the lamb was sure to go.", - " And the 0-1 pitch on the way to Edgar Martinez swung on the line" - " down the left field line for Obeyshev. Here comes Joy. Here is" - " Jorgen at third base. They're going to wave him in. The throw to the" - " plate will be late. The Mariners are going to play for the American" - " League Championship. I don't believe it. It just continues. My, oh" - " my.", - ], - "openai/whisper-large-v3": [ - " The first words I spoke in the original phonograph, a little piece" - " of practical poetry. Mary had a little lamb, its feet were quite as" - " slow, and everywhere that Mary went, the lamb was sure to go.", - " And the 0-1 pitch on the way to Edgar Martinez. Swung on the line." - " Now the left field line for a base hit. Here comes Joy. Here is" - " Junior to third base. They're going to wave him in. The throw to the" - " plate will be late. The Mariners are going to play for the American" - " League Championship. I don't believe it. It just continues. My, oh," - " my.", - ], - "openai/whisper-large-v3-turbo": [ - " The first words I spoke in the original phonograph, a little piece" - " of practical poetry. Mary had a little lamb, its streets were quite" - " as slow, and everywhere that Mary went the lamb was sure to go.", - " And the 0-1 pitch on the way to Edgar Martinez. Swung on the line" - " down the left field line for a base hit. Here comes Joy. Here is" - " Junior to third base. They're going to wave him in. The throw to the" - " plate will be late. The Mariners are going to play for the American" - " League Championship. I don't believe it. It just continues. My, oh," - " my.", - ], -} + +@pytest.fixture(autouse=True) +def use_spawn_for_whisper(monkeypatch): + """Whisper has issues with forked workers, use spawn instead.""" + monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn") def run_test( + hf_runner: type[HfRunner], vllm_runner: type[VllmRunner], + inputs: Sequence[tuple[list[str], list[str], PromptAudioInput]], model: str, *, + max_model_len: int, + dtype: str, + max_tokens: int, + num_logprobs: int, tensor_parallel_size: int, distributed_executor_backend: str | None = None, - dtype: str = "half", + enforce_eager: bool = True, ) -> None: - prompt_list = PROMPTS * 10 - expected_list = EXPECTED[model] * 10 + """Inference result should be the same between hf and vllm. + All the audio fixtures for the test are from AudioAsset. + For huggingface runner, we provide the audio as input. + For vllm runner, we provide MultiModalDataDict objects + and corresponding MultiModalConfig as input. + """ with vllm_runner( model, dtype=dtype, - max_model_len=448, + max_model_len=max_model_len, tensor_parallel_size=tensor_parallel_size, distributed_executor_backend=distributed_executor_backend, - # TODO (NickLucche) figure out output differences with non-eager and re-enable - enforce_eager=True, + limit_mm_per_prompt={"audio": 2}, + enforce_eager=enforce_eager, + disable_custom_all_reduce=True, ) as vllm_model: - llm = vllm_model.llm + vllm_outputs_per_case = [ + vllm_model.generate_greedy_logprobs( + vllm_prompts, + max_tokens, + num_logprobs=num_logprobs, + audios=audios, + ) + for vllm_prompts, _, audios in inputs + ] - sampling_params = SamplingParams( - temperature=0, - top_p=1.0, - max_tokens=200, + with hf_runner(model, dtype=dtype, auto_cls=AutoModelForSpeechSeq2Seq) as hf_model: + hf_outputs_per_case = [ + hf_model.generate_greedy_logprobs_limit( + hf_prompts, + max_tokens, + num_logprobs=num_logprobs, + audios=audios, + ) + for _, hf_prompts, audios in inputs + ] + + for hf_outputs, vllm_outputs in zip(hf_outputs_per_case, vllm_outputs_per_case): + check_logprobs_close( + outputs_0_lst=hf_outputs, + outputs_1_lst=vllm_outputs, + name_0="hf", + name_1="vllm", ) - outputs = llm.generate(prompt_list, sampling_params) - for output, expected in zip(outputs, expected_list): - print(output.outputs[0].text) - assert output.outputs[0].text == expected +@pytest.fixture +def input_audios() -> list[tuple[list[str], list[str], list[tuple[Any, int]]]]: + audio_assets = [AudioAsset("mary_had_lamb"), AudioAsset("winning_call")] + inputs = [] + for asset in audio_assets: + audio, orig_sr = asset.audio_and_sample_rate + # Resample to Whisper's expected sample rate (16kHz) + if orig_sr != WHISPER_SAMPLE_RATE: + audio = librosa.resample( + audio, orig_sr=orig_sr, target_sr=WHISPER_SAMPLE_RATE + ) + # vLLM prompts, HF prompts, audio inputs + inputs.append(([VLLM_PROMPT], [HF_PROMPT], [(audio, WHISPER_SAMPLE_RATE)])) + return inputs + + +def check_model_available(model: str) -> None: + model_info = HF_EXAMPLE_MODELS.find_hf_info(model) + model_info.check_available_online(on_fail="skip") + model_info.check_transformers_version(on_fail="skip") @pytest.mark.core_model -@pytest.mark.parametrize("model", ["openai/whisper-large-v3-turbo"]) -@pytest.mark.parametrize("dtype", ["half"]) -@create_new_process_for_each_test() -def test_models(vllm_runner, model, dtype) -> None: - run_test( - vllm_runner, - model, - tensor_parallel_size=1, - dtype=dtype, - ) - - @pytest.mark.cpu_model @pytest.mark.parametrize("model", ["openai/whisper-large-v3-turbo"]) @pytest.mark.parametrize("dtype", ["half"]) -def test_models_cpu(vllm_runner, model, dtype) -> None: - # @create_new_process_for_each_test() does not work for some runners - # TODO: to fix cpu privilege issues in run-cpu-test-arm.sh +@pytest.mark.parametrize("num_logprobs", [5]) +@pytest.mark.parametrize("enforce_eager", [True, False]) +@create_new_process_for_each_test("spawn") +def test_models( + hf_runner, + vllm_runner, + model: str, + dtype: str, + num_logprobs: int, + input_audios, + enforce_eager: bool, +) -> None: + check_model_available(model) + if current_platform.is_cpu() and not enforce_eager: + pytest.skip("Skipping test for CPU with non-eager mode") run_test( + hf_runner, vllm_runner, + input_audios, model, - tensor_parallel_size=1, dtype=dtype, + max_model_len=448, + max_tokens=200, + num_logprobs=num_logprobs, + tensor_parallel_size=1, + enforce_eager=enforce_eager, ) @@ -152,15 +148,31 @@ def test_models_cpu(vllm_runner, model, dtype) -> None: @pytest.mark.core_model @pytest.mark.parametrize("model", ["openai/whisper-large-v3-turbo"]) @pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"]) -@create_new_process_for_each_test() +@pytest.mark.parametrize("dtype", ["half"]) +@pytest.mark.parametrize("max_tokens", [200]) +@pytest.mark.parametrize("num_logprobs", [5]) +@create_new_process_for_each_test("spawn") def test_models_distributed( + hf_runner, vllm_runner, - model, - distributed_executor_backend, + model: str, + distributed_executor_backend: str, + dtype: str, + max_tokens: int, + num_logprobs: int, + input_audios, ) -> None: + check_model_available(model) run_test( + hf_runner, vllm_runner, + input_audios, model, + dtype=dtype, + max_model_len=448, + max_tokens=max_tokens, + num_logprobs=num_logprobs, tensor_parallel_size=2, distributed_executor_backend=distributed_executor_backend, + enforce_eager=False, ) diff --git a/tests/models/registry.py b/tests/models/registry.py index 769b33d877983..ca50785b46a1a 100644 --- a/tests/models/registry.py +++ b/tests/models/registry.py @@ -840,7 +840,10 @@ _MULTIMODAL_EXAMPLE_MODELS = { is_available_online=False, ), # [Encoder-decoder] - "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"), + "WhisperForConditionalGeneration": _HfExamplesInfo( + "openai/whisper-large-v3-turbo", + extras={"v3": "openai/whisper-large-v3"}, + ), # [Cross-encoder] "JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"), }