[MM][V2] Enable encoder-only ViT CUDA graph capture (#56922)

Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
jiangkuaixue123
2026-09-17 16:12:17 +08:00
committed by GitHub
co-authored by Codex
parent 4e5ffda19d
commit de24e51908
7 changed files with 362 additions and 5 deletions
+54 -1
View File
@@ -9,12 +9,15 @@ and tolerated (token-embedding fallback) when it is not, while a miss within
the processed range still fails loudly.
"""
from unittest.mock import MagicMock
from types import SimpleNamespace
from unittest.mock import MagicMock, patch
import numpy as np
import pytest
import torch
from vllm.config.compilation import CUDAGraphMode
from vllm.model_executor.models.interfaces import SupportsEncoderCudaGraph
from vllm.multimodal.inputs import (
MultiModalFeatureSpec,
MultiModalFieldElem,
@@ -31,6 +34,56 @@ pytestmark = pytest.mark.cpu_test
HIDDEN = 4
@pytest.mark.parametrize(
"enabled,enforce_eager,supported,expected",
[
(True, False, True, True),
(False, False, True, False),
(True, True, True, False),
(True, False, False, False),
],
)
def test_encoder_graph_configuration(enabled, enforce_eager, supported, expected):
"""Encoder graph eligibility is independent of decoder graph mode."""
config = MagicMock()
config.model_config.enforce_eager = enforce_eager
config.model_config.dtype = torch.float32
config.model_config.get_inputs_embeds_size.return_value = HIDDEN
config.scheduler_config.max_num_batched_tokens = 8
config.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
config.compilation_config.cudagraph_mm_encoder = enabled
model = MagicMock(spec=SupportsEncoderCudaGraph) if supported else torch.nn.Module()
state = _model_state(EncoderCache())
with patch("vllm.v1.worker.gpu.model_states.interface.EncoderCudaGraphManager"):
ModelState.__init__(state, config, model, state.encoder_cache, state.device)
assert state.encoder_runner.has_cudagraph() is expected
def test_encoder_graph_unsupported_modality_uses_eager_output():
"""An image graph must not intercept another modality's encoder output."""
from vllm.v1.worker.encoder_cudagraph import EncoderCudaGraphManager
runner = _make_runner([], [])
output = torch.ones(2, HIDDEN)
runner.model = MagicMock()
runner.model.embed_multimodal.return_value = [output]
manager = MagicMock(spec=EncoderCudaGraphManager)
manager.config = SimpleNamespace(modalities=["image"])
manager.is_captured.return_value = True
manager.supports_modality.side_effect = lambda modality: (
EncoderCudaGraphManager.supports_modality(manager, modality)
)
runner.cudagraph_manager = manager
with patch(
"vllm.v1.worker.gpu.mm.encoder_runner.group_and_batch_mm_kwargs",
return_value=[("audio", 1, {"audio_values": torch.zeros(2)})],
):
result = runner.execute_mm_encoder([])
assert len(result) == 1
assert result[0] is output
manager.execute.assert_not_called()
def _model_state(cache: EncoderCache) -> MagicMock:
"""A mock ModelState backed by a real EncoderCache."""
state = MagicMock()