[Core] Check for GPU<->CPU syncs during CI (#43107)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Nick Hill
2026-08-14 17:42:05 -07:00
committed by GitHub
co-authored by Claude Opus 4.8
parent 7b544ecb52
commit 615d4cfade
44 changed files with 600 additions and 357 deletions
@@ -54,6 +54,7 @@ from vllm.multimodal.processing import (
TimingContext,
)
from vllm.sequence import IntermediateTensors
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
if TYPE_CHECKING:
from transformers import BatchFeature, PreTrainedModel
@@ -726,12 +727,17 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
kwargs.pop("token_type_ids", None)
kwargs.pop("mm_token_type_ids", None)
audio_output = self.model.get_audio_features(
input_features, return_dict=True, **kwargs
)
# HuggingFace's `get_audio_features` implementations branch on
# per-sample feature lengths internally.
with gpu_sync_allowed():
audio_output = self.model.get_audio_features(
input_features, return_dict=True, **kwargs
)
audio_embeddings = audio_output.pooler_output
split_sizes = num_audio_tokens.flatten().tolist()
# Per-audio token counts are needed as Python ints to split.
with gpu_sync_allowed():
split_sizes = num_audio_tokens.flatten().tolist()
return self._split_embeddings(audio_embeddings, split_sizes)
def _process_image_input(self, **kwargs) -> list[torch.Tensor] | None:
@@ -749,7 +755,12 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
num_image_patches = kwargs.pop("num_image_patches")
vision_embeddings = self.model.get_image_features(pixel_values, **kwargs)
# The underlying HuggingFace `get_image_features` implementations
# contain model-internal syncs (e.g. Idefics3 filters all-zero
# padding images via boolean-mask indexing, LlavaOnevision
# branches on per-sample batch counts).
with gpu_sync_allowed():
vision_embeddings = self.model.get_image_features(pixel_values, **kwargs)
# Transformers `v5`, `self.get_image_features` returns a tuple
# containing the features and optionally attentions/hidden_states