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

remove redundant max_running_reqs calculation in r3 (#16629)

Browse Source
This commit is contained in:
Junrong Lin
2026-01-08 09:45:21 +08:00
committed by GitHub
parent 109fe03ad1
commit eec7dbd31e
1 changed files with 1 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/model_runner.py
+1 -16
View File
@@ -585,21 +585,6 @@ class ModelRunner(ModelRunnerKVCacheMixin):
self.init_piecewise_cuda_graphs() self.init_piecewise_cuda_graphs()
def init_routed_experts_capturer(self): def init_routed_experts_capturer(self):
# TODO: the redundant logic with TpModelWorker
max_running_requests = min(
(
self.max_total_num_tokens // 2
if self.server_args.max_running_requests is None
else self.server_args.max_running_requests
// (
self.server_args.dp_size
if self.server_args.enable_dp_attention
else 1
)
),
self.req_to_token_pool.size,
)
if not self.server_args.disable_shared_experts_fusion and hasattr( if not self.server_args.disable_shared_experts_fusion and hasattr(
self.model, "num_fused_shared_experts" self.model, "num_fused_shared_experts"
): ):
@@ -613,7 +598,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
model_config=self.model_config, model_config=self.model_config,
num_fused_shared_experts=num_fused_shared_experts, num_fused_shared_experts=num_fused_shared_experts,
num_tokens=self.max_total_num_tokens + self.page_size, num_tokens=self.max_total_num_tokens + self.page_size,
max_running_requests=max_running_requests, max_running_requests=self.max_running_requests,
device=self.device, device=self.device,
) )
) )