Unable to capture GPU stream via start_profile on P node in separated PD deployment (1P2D), while D node works fine

[ysay-d]

I am encountering an issue with the profiling functionality (start_profile) in a separated Prefill-Decode (PD) deployment architecture. Specifically, I am unable to capture GPU stream traces on the Prefill (P) node, whereas the Decode (D) nodes work as expected.

• SGLang Version 0.5.9
• Launch Command (P Node):
  CUDA_VISIBLE_DEVICES=4 sglang serve --model-path /mnt/Qwen3-32B/ --port 36666 --disaggregation-mode prefill --disaggregation-ib-device mlx5_bond_4
• Launch Command (D Node):
  CUDA_VISIBLE_DEVICES=7 sglang serve --model-path /mnt/Qwen3-32B/ --port 36667 --disaggregation-mode decode --disaggregation-ib-device mlx5_bond_4

[ysay-d]

I resolved this issue by extending the profiling duration, which suggests the root cause was that the CUPTI buffer did not fill up within the shorter time window, preventing the collected GPU stream data from being flushed and returned; increasing the capture time allows the buffer to reach its threshold and successfully commit the trace on the P-node.