[V1] Remove pin_memory() in async_copy_to_gpu to fix sporadic stalls (#37006)

Signed-off-by: Sebastien Beurnier <sbeurnier@together.ai>
This commit is contained in:
sbeurnier
2026-03-14 01:37:32 +00:00
committed by GitHub
parent 092ace9e3a
commit a116f96930
+4 -6
View File
@@ -27,12 +27,10 @@ def async_copy_to_gpu(
assert device is not None assert device is not None
out = torch.empty_like(x, device=device) out = torch.empty_like(x, device=device)
# CPU-to-CPU copy # Copy directly to GPU — explicit pin_memory() causes sporadic stalls
tmp = x.pin_memory() # under high concurrency due to CUDA driver contention. The driver
assert tmp is not x # handles the transfer efficiently without manual pinning.
return out.copy_(x, non_blocking=True)
# CPU-to-GPU copy
return out.copy_(tmp, non_blocking=True)
class UvaBuffer: class UvaBuffer: