[ROCm][Bugfix] Reduce CUDA graph divergences (#57229)

Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
This commit is contained in:
Matt
2026-09-17 05:12:13 +00:00
committed by GitHub
parent 0bfc7a15d0
commit 2d8a5c5741
5 changed files with 8 additions and 90 deletions
+2 -10
View File
@@ -1632,10 +1632,7 @@ def get_pcp_group() -> GroupCoordinator:
@contextmanager
def graph_capture(
device: torch.device,
graph_capture_context: GraphCaptureContext | None = None,
):
def graph_capture(device: torch.device):
"""`graph_capture` is a context manager which should surround the code that
is capturing the CUDA graph. Its main purpose is to ensure that some
operations will be run after the graph is captured, before the graph
@@ -1647,13 +1644,8 @@ def graph_capture(
the graph capture is running on a separate stream from the default stream,
in order to explicitly distinguish the kernels to capture
from other kernels possibly launched on background in the default stream.
A caller may pass an explicit ``graph_capture_context`` to control the
stream used (e.g. to capture on the default stream).
"""
context = graph_capture_context or GraphCaptureContext(
torch.cuda.Stream(device=device)
)
context = GraphCaptureContext(torch.cuda.Stream(device=device))
with (
get_tp_group().graph_capture(context),
get_pp_group().graph_capture(context),