Re-disable CUDA graph memory profiling on ROCm (#48440)
Signed-off-by: Rohan Potdar <rohan.potdar@amd.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rohan Potdar committed
Jul 12, 2026 at 23:59 UTC
d973cce3ca809b44e0794ee044519e8ba91502e1
1 file changed
+5
-7
vllm/v1/worker/gpu_worker.py
+5
-7
@@ -492,14 +492,11 @@ class Worker(WorkerBase):
492
)
493
494
# Profile CUDA graph memory if graphs will be captured.
495
- # ROCm is included: #44825 moved the profiler to
496
- # torch.accelerator.get_memory_info (reliable on ROCm, as used by
497
- # the AMD-CI mem tests), and graph_pool_handle resolves to the same
498
- # torch.cuda handle the live capture path already uses on ROCm.
499
- # XPU stays excluded (see #39977).
495
+ # Skip on ROCm/HIP/XPU as graph pool handles and get_memory_info
496
+ # behave differently and can produce incorrect/negative estimates.
497
cudagraph_memory_estimate = 0
498
if (
502
- current_platform.is_cuda_alike()
499
+ current_platform.is_cuda()
500
and self.vllm_config.compilation_config.cudagraph_mode
501
!= CUDAGraphMode.NONE
502
):
@@ -515,7 +512,8 @@ class Worker(WorkerBase):
512
+ profile_result.weights_memory
513
)
514
518
- # Respect the opt-in flag as originally designed.
515
+ # On ROCm, cudagraph_memory_estimate is always 0 so this is a no-op.
516
+ # On CUDA, respect the opt-in flag as originally designed.
517
cudagraph_memory_estimate_applied = (
518
cudagraph_memory_estimate
519
if envs.VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS