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

[Bug Fix] Reject incompatible combination of --disable-cuda-graph-padding and --enable-torch-compile (#23903)

Browse Source
This commit is contained in:
Praneth Paruchuri
2026-05-12 16:18:36 +08:00
committed by GitHub
parent 8d27ce7371
commit 1efe9e2164
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+8
View File
@@ -7067,6 +7067,14 @@ class ServerArgs:
f"Got: {self.pp_max_micro_batch_size}" f"Got: {self.pp_max_micro_batch_size}"
) )
assert not (self.disable_cuda_graph_padding and self.enable_torch_compile), (
"--disable-cuda-graph-padding is incompatible with --enable-torch-compile. "
"With padding disabled, every distinct batch size gets its own torch.compile + "
"Triton autotune cycle (O(max_batch_size) compilations) instead of the small fixed "
"set of padded bucket sizes, causing engine initialisation to stall for many minutes. "
"Remove --disable-cuda-graph-padding or --enable-torch-compile."
)
if self.pp_size > 1: if self.pp_size > 1:
assert ( assert (
self.disable_overlap_schedule and self.speculative_algorithm is None self.disable_overlap_schedule and self.speculative_algorithm is None