chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

enable piecewise cuda graph for prefill server (#13377)

Browse Source 
Co-authored-by: serverance.fu <serverance.fu@temu.com>
This commit is contained in:
fjybiocs
2025-11-29 12:09:26 +08:00
committed by GitHub
parent f446b51c41
commit 143b57b805
2 changed files with 92 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/server_args.py
View File

@@ -1825,8 +1825,12 @@ class ServerArgs:
self.disaggregation_prefill_pp = self.pp_size
self.validate_disagg_tp_size(self.tp_size, self.disaggregation_decode_tp)
self.disable_cuda_graph = True
logger.warning("Cuda graph is disabled for prefill server")
if not self.enable_piecewise_cuda_graph:
self.disable_cuda_graph = True
logger.warning(
"Cuda graph is disabled for prefill server when piecewise cuda graph is not enabled."
)
def _handle_tokenizer_batching(self):
if self.enable_tokenizer_batch_encode and self.enable_dynamic_batch_tokenizer: