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

Extend cuda graph capture bs for B200 (#6937)

Browse Source
This commit is contained in:
Baizhou Zhang
2025-06-08 05:13:22 -07:00
committed by GitHub
parent 2fc1299562
commit 971a0dfa32
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/model_executor/cuda_graph_runner.py
View File

@@ -139,6 +139,8 @@ def get_batch_sizes_to_capture(model_runner: ModelRunner):
gpu_mem = get_device_memory_capacity() gpu_mem = get_device_memory_capacity()
if gpu_mem is not None and gpu_mem > 96 * 1024: if gpu_mem is not None and gpu_mem > 96 * 1024:
capture_bs += list(range(160, 257, 8)) capture_bs += list(range(160, 257, 8))
if gpu_mem is not None and gpu_mem > 180 * 1000:
capture_bs += list(range(256, 513, 16))
if max(capture_bs) > model_runner.req_to_token_pool.size: if max(capture_bs) > model_runner.req_to_token_pool.size:
# In some cases (e.g., with a small GPU or --max-running-requests), the #max-running-requests # In some cases (e.g., with a small GPU or --max-running-requests), the #max-running-requests