[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
+5 -6
View File
@@ -36,10 +36,9 @@ def async_copy_to_gpu(
assert device is not None
out = torch.empty_like(x, device=device)
# Copy directly to GPU — explicit pin_memory() causes sporadic stalls
# under high concurrency due to CUDA driver contention. The driver
# handles the transfer efficiently without manual pinning.
return out.copy_(x, non_blocking=True)
# pin_memory() is no-op if the memory is already pinned.
pinned = x.pin_memory()
return out.copy_(pinned, non_blocking=True)
class UvaBuffer:
@@ -183,7 +182,7 @@ class StagedWriteTensor:
# Special handling for write_contents
write_contents = async_tensor_h2d(
self._staged_write_contents, self.dtype, self.device
self._staged_write_contents, device=self.device, dtype=self.dtype
)
# Write diffs to the GPU buffer
@@ -255,7 +254,7 @@ class FusedStagedWriter:
indices_uva = self.indices.copy_to_uva(indices)
starts_uva = self.starts.copy_to_uva(starts)
cu_lens_uva = self.cu_lens.copy_to_uva(cu_lens)
contents_gpu = async_tensor_h2d(contents, torch.int32, self.device)
contents_gpu = async_tensor_h2d(contents, device=self.device, dtype=torch.int32)
_apply_write_kernel[(len(group_ids),)](
output_ptrs,