[Perf] Avoid more h2d copies from non-pinned tensors (#54660)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Nick Hill
2026-09-01 20:37:09 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 3ba9907a1d
commit a566ea7e8f
5 changed files with 72 additions and 64 deletions
@@ -60,6 +60,7 @@ from vllm.multimodal.processing import (
)
from vllm.sequence import IntermediateTensors
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
from vllm.utils.torch_utils import async_tensor_h2d
if TYPE_CHECKING:
from transformers import BatchFeature, PreTrainedModel
@@ -1180,8 +1181,8 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
# grid_thw fields are registered keep_on_cpu; restore the on-device
# placement that HF get_image_features implementations expect.
for key, value in kwargs.items():
if isinstance(value, torch.Tensor):
kwargs[key] = value.to(pixel_values.device, non_blocking=True)
if isinstance(value, torch.Tensor) and value.is_cpu:
kwargs[key] = async_tensor_h2d(value, pixel_values.device)
# The underlying HuggingFace `get_image_features` implementations
# contain model-internal syncs (e.g. Idefics3 filters all-zero