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 13, 14:30 UTC
[Performance][gpt-oss] Revert gpt-oss max cudagraph size to 1024 (#28345)
Signed-off-by: Mohammad Miadh Angkad <MAngkad.BSDSBA2027@aim.edu>
Mohammad Miadh Angkad committed
Nov 9, 2025 at 06:50 UTC
404d7a9d1452d571c52fdcc6795ad18add57483c
1 file changed
+4
-6
vllm/model_executor/models/config.py
+4
-6
@@ -258,9 +258,9 @@ class GptOssForCausalLMConfig(VerifyAndUpdateConfig):
258
if structured_outputs_config.reasoning_parser == "":
259
structured_outputs_config.reasoning_parser = "openai_gptoss"
260
261
- # Increase the max capture size from 512 to 992 for performance.
261
+ # Increase the max capture size from 512 to 1024 for performance.
262
# NOTE(woosuk): This will increase the number of CUDA graphs
263
- # from 67 to 81.
263
+ # from 67 to 83.
264
compilation_config = vllm_config.compilation_config
265
# Only override when the user has not set either of
266
# cudagraph_capture_sizes or max_cudagraph_capture_size.
@@ -268,11 +268,9 @@ class GptOssForCausalLMConfig(VerifyAndUpdateConfig):
268
compilation_config.cudagraph_capture_sizes is None
269
and compilation_config.max_cudagraph_capture_size is None
270
):
271
- # FIXME(woosuk): When using full cuda graph with FA3, the max
272
- # supported size is 992.
273
- compilation_config.max_cudagraph_capture_size = 992
271
+ compilation_config.max_cudagraph_capture_size = 1024
272
logger.info(
275
- "Overriding max cuda graph capture size to %d for performance.", 992
273
+ "Overriding max cuda graph capture size to %d for performance.", 1024
274
)
275
276