EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Remove Unintended Capture Batch Sizes in AMD HIP Graph Runner (#4638)

Browse Source
This commit is contained in:
AinL
2025-03-28 00:41:33 +09:00
committed by GitHub
parent 668ecc6c5b
commit 17000d2b3a
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/model_executor/cuda_graph_runner.py
View File

@@ -124,8 +124,8 @@ def get_batch_sizes_to_capture(model_runner: ModelRunner):
# capture less.
capture_bs = list(range(1, 9)) + list(range(9, 33, 2)) + [64, 96, 128, 160]
if _is_hip:
capture_bs += [i * 8 for i in range(21, 33)]
if _is_hip:
capture_bs += [i * 8 for i in range(21, 33)]
if max(capture_bs) > model_runner.req_to_token_pool.size:
# In some case (e.g., with a small GPU or --max-running-requests), the #max-running-requests