chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Bugfix] Fix Qwen3/DSV3/DSV3.2 model support (#11510)

Browse Source
This commit is contained in:
Even Zhou
2025-10-16 15:14:09 +08:00
committed by GitHub
parent b0d20cdec7
commit 3cceaa381a
12 changed files with 102 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+20
View File
@@ -628,6 +628,16 @@ class ServerArgs:
self.chunked_prefill_size = 2048
if self.cuda_graph_max_bs is None:
self.cuda_graph_max_bs = 8
elif is_npu() and gpu_mem < 32 * 1024:
# Atlas A2B4
# (chunked_prefill_size 32k, cuda_graph_max_bs 16 if tp < 4 else 64)
if self.chunked_prefill_size is None:
self.chunked_prefill_size = 32768
if self.cuda_graph_max_bs is None:
if self.tp_size < 4:
self.cuda_graph_max_bs = 16
else:
self.cuda_graph_max_bs = 64
elif gpu_mem < 35 * 1024:
# A10, 4090, 5090
# (chunked_prefill_size 2k, cuda_graph_max_bs 16 if tp < 4 else 80)
@@ -651,6 +661,16 @@ class ServerArgs:
self.cuda_graph_max_bs = 32
else:
self.cuda_graph_max_bs = 160
elif is_npu() and gpu_mem < 64 * 1024:
# Atlas A2 and Atlas A3
# (chunked_prefill_size 32k, cuda_graph_max_bs 64 if tp < 4 else 128)
if self.chunked_prefill_size is None:
self.chunked_prefill_size = 32768
if self.cuda_graph_max_bs is None:
if self.tp_size < 4:
self.cuda_graph_max_bs = 64
else:
self.cuda_graph_max_bs = 128
elif gpu_mem < 90 * 1024:
# H100, A100
# (chunked_prefill_size 8k, cuda_graph_max_bs 256 if tp < 4 else 512)