A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 23:45 UTC
[XPU] [torch.compile] Skipping CUDA graph memory estimation to avoid startup errors. (#39977)
Signed-off-by: chaojun-zhang <chaojun.zhang@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Chaojun Zhang committed
Apr 20, 2026 at 13:04 UTC
4f4713f96e01ebc799125a670c0bca9a2ae7d59d
1 file changed
+6
-2
vllm/v1/worker/gpu_worker.py
+6
-2
@@ -374,10 +374,14 @@ class Worker(WorkerBase):
374
)
375
376
# Profile CUDA graph memory if graphs will be captured.
377
- # Skip on ROCm/HIP as graph pool handles and mem_get_info behave
377
+ # Skip on ROCm/HIP/XPU as graph pool handles and mem_get_info behave
378
# differently and can produce incorrect/negative estimates.
379
cudagraph_memory_estimate = 0
380
- if not self.model_config.enforce_eager and not current_platform.is_rocm():
380
+ if (
381
+ not current_platform.is_rocm()
382
+ and self.vllm_config.compilation_config.cudagraph_mode
383
+ != CUDAGraphMode.NONE
384
+ ):
385
cudagraph_memory_estimate = self.model_runner.profile_cudagraph_memory()
386
387
# Use the pre-cudagraph torch peak to avoid double-counting.