[ROCm] Re-enable cudagraph memory profiling, captured on the current stream (#48526)

Signed-off-by: pei.zhang <pei.zhang@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
peizhang56
2026-07-15 10:03:48 -05:00
committed by GitHub
co-authored by Claude Andreas Karatzas
parent 5810e884f1
commit 05eed72aec
3 changed files with 37 additions and 8 deletions
+7 -5
View File
@@ -492,11 +492,14 @@ class Worker(WorkerBase):
)
# Profile CUDA graph memory if graphs will be captured.
# Skip on ROCm/HIP/XPU as graph pool handles and get_memory_info
# behave differently and can produce incorrect/negative estimates.
# ROCm is included: #44825 moved the profiler to
# torch.accelerator.get_memory_info (reliable on ROCm, as used by
# the AMD-CI mem tests), and graph_pool_handle resolves to the same
# torch.cuda handle the live capture path already uses on ROCm.
# XPU stays excluded (see #39977).
cudagraph_memory_estimate = 0
if (
current_platform.is_cuda()
current_platform.is_cuda_alike()
and self.vllm_config.compilation_config.cudagraph_mode
!= CUDAGraphMode.NONE
):
@@ -512,8 +515,7 @@ class Worker(WorkerBase):
+ profile_result.weights_memory
)
# On ROCm, cudagraph_memory_estimate is always 0 so this is a no-op.
# On CUDA, respect the opt-in flag as originally designed.
# Respect the opt-in flag as originally designed.
cudagraph_memory_estimate_applied = (
cudagraph_memory_estimate
if envs.VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS