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

[fix] reduce dp capture bs (#5634)

Browse Source 
Co-authored-by: alcanerian <alcanerian@gmail.com>
This commit is contained in:
JieXin Liang
2025-04-23 02:08:45 +08:00
committed by GitHub
parent 2ed96c7a8a
commit 917324862e
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
python/sglang/srt/model_executor/cuda_graph_runner.py
View File

@@ -134,7 +134,8 @@ def get_batch_sizes_to_capture(model_runner: ModelRunner):
)
gpu_mem = get_device_memory_capacity()
if gpu_mem is not None and gpu_mem > 81920:
# Batch size of each rank will not become so large when DP is on
if gpu_mem is not None and gpu_mem > 81920 and server_args.dp_size == 1:
capture_bs += list(range(160, 257, 8))
if max(capture_bs) > model_runner.req_to_token_pool.size: