A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI] Only capture a single CUDA graph size in CI by default (#25951)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Oct 1, 2025 at 10:03 UTC
a332b84578cdc0706e040f6a765954c8a289904f
1 file changed
+3
tests/conftest.py
+3
@@ -731,6 +731,9 @@ class VllmRunner:
731
init_ctx = (nullcontext() if default_torch_num_threads is None else
732
set_default_torch_num_threads(default_torch_num_threads))
733
734
+ if not kwargs.get("compilation_config", None):
735
+ kwargs["compilation_config"] = {"cudagraph_capture_sizes": [8]}
736
+
737
with init_ctx:
738
self.llm = LLM(
739
model=model_name,