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] Add comment about the single cudagraph capture size that is used (#26252)
Thomas Parnell committed
Oct 6, 2025 at 04:35 UTC
d3c84297c38c79e2da6cd0a4ba8be58d95d13d97
1 file changed
+4
tests/conftest.py
+4
@@ -740,6 +740,10 @@ class VllmRunner:
740
)
741
742
if not kwargs.get("compilation_config", None):
743
+ # Note(@tdoublep): This is set to 4 because some tests (e.g., hybrid
744
+ # model tests) may set max_num_seqs=4. If min cudagraph_capture_size is
745
+ # set to larger than max_num_seqs, then it will lead to *no* graphs
746
+ # being captured which can trigger edge cases that we don't handle yet.
747
kwargs["compilation_config"] = {"cudagraph_capture_sizes": [4]}
748
749
with init_ctx: