[Perf][MRV2] Avoid stream syncs in NonUvaBuffer host-to-device copies (#60215)

Signed-off-by: cheesecake <farzanaman99@gmail.com>
Co-authored-by: William Joyner <133391209+joynerwk03@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Cheese Cake
2026-10-07 01:51:12 +00:00
committed by GitHub
co-authored by William Joyner Claude Opus 5.5
parent 283d76d56a
commit 956960f85f
2 changed files with 21 additions and 2 deletions
+19
View File
@@ -234,6 +234,25 @@ def test_non_pinned_cpu_tensor(device):
assert gpu_view[9, 9] == 198
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA-only sync check.")
def test_non_uva_buffer_copies_without_stream_sync():
buffer = buffer_utils.NonUvaBuffer(8, torch.int32)
buffer.np[:] = np.arange(8)
torch.cuda.set_sync_debug_mode("error")
try:
prefix = buffer.uva(4)
full = buffer.uva()
finally:
torch.cuda.set_sync_debug_mode(0)
# The next step overwrites the host buffer right after the copy is issued.
buffer.np[:] = -1
torch.accelerator.synchronize()
assert prefix.tolist() == list(range(4))
assert full.tolist() == list(range(8))
@pytest.mark.skipif(not is_uva_available(), reason="UVA is not available.")
@pytest.mark.skipif(
not current_platform.is_xpu(), reason="XPU non-contiguous UVA test."
+2 -2
View File
@@ -54,8 +54,8 @@ class NonUvaBuffer:
def uva(self, n: int | None = None) -> torch.Tensor:
if n is None:
return self._uva.copy_(self.cpu)
return self._uva[:n].copy_(self.cpu[:n])
return self._uva.copy_(self.cpu, non_blocking=True)
return self._uva[:n].copy_(self.cpu[:n], non_blocking=True)
class UvaBufferPool: