[Core] Ensure memory is pinned prior to async h2d copy (#45424)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-06-20 20:02:24 -07:00
committed by GitHub
parent 8dd1b702f2
commit 7df3d7dada
49 changed files with 254 additions and 264 deletions
+14 -2
View File
@@ -488,7 +488,13 @@ class MultiModalBatchedField(BaseMultiModalField):
# An optimization when `batch` contains only one tensor:
# - produce exactly same result as `torch.stack(batch)`
# - will achieve zero-copy if the tensor is contiguous
return batch[0].unsqueeze(0).contiguous()
out = batch[0].unsqueeze(0)
if not pin_memory:
return out.contiguous()
# Avoid extra copy - pinning unpinned memory will make it contiguous
if not out.is_contiguous() and out.is_pinned():
out = out.contiguous()
return out.pin_memory()
first_shape = batch[0].shape
if all(elem.shape == first_shape for elem in batch):
out = torch.empty(
@@ -538,7 +544,13 @@ class MultiModalFlatField(BaseMultiModalField):
# An optimization when `batch` contains only one tensor:
# - produce exactly same result as `torch.concat(batch)`
# - will achieve zero-copy if the tensor is contiguous
return batch[0].contiguous()
out = batch[0]
if not pin_memory:
return out.contiguous()
# Avoid extra copy - pinning unpinned memory will make it contiguous
if not out.is_contiguous() and out.is_pinned():
out = out.contiguous()
return out.pin_memory()
dim = self.dim + (self.dim < 0) * len(batch[0].shape)