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

Revert "[server_args] compute mem_fraction_static after dp chunked-prefill division" (#28991)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-06-22 20:48:16 -07:00
committed by GitHub
parent fc0d8fa262
commit 4740f23e1f
1 changed files with 1 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+1 -4
View File
@@ -3266,12 +3266,9 @@ class ServerArgs:
if self.mem_fraction_static is None: if self.mem_fraction_static is None:
# Constant meta data (e.g., from attention backend) # Constant meta data (e.g., from attention backend)
reserved_mem = 512 reserved_mem = 512
effective_chunked_prefill_size = self.chunked_prefill_size // (
self.dp_size if self.enable_dp_attention and self.dp_size > 1 else 1
)
# For activation during large prefill # For activation during large prefill
if self.chunked_prefill_size > 0: if self.chunked_prefill_size > 0:
reserved_mem += max(effective_chunked_prefill_size, 2048) * 1.5 reserved_mem += max(self.chunked_prefill_size, 2048) * 1.5
else: else:
reserved_mem += max(self.max_prefill_tokens, 2048) * 1.5 reserved_mem += max(self.max_prefill_tokens, 2048) * 1.5
# For cuda graphs # For cuda graphs