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
[Bugfix] Lower gpt-oss max cudagraph size to 992 to be compatible with FA3 (#25508)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Sep 23, 2025 at 15:49 UTC
a8ffc4f0f2d02aa4e505dcc3c974d5ec6e00737c
1 file changed
+5
-5
vllm/model_executor/models/config.py
+5
-5
@@ -266,24 +266,24 @@ class GptOssForCausalLMConfig(VerifyAndUpdateConfig):
266
if structured_outputs_config.reasoning_parser == "":
267
structured_outputs_config.reasoning_parser = "openai_gptoss"
268
269
- # Increase the max capture size from 512 to 1024 for performance.
269
+ # Increase the max capture size from 512 to 992 for performance.
270
# NOTE(woosuk): This will increase the number of CUDA graphs
271
- # from 67 to 83.
271
+ # from 67 to 81.
272
scheduler_config = vllm_config.scheduler_config
273
if len(scheduler_config.cuda_graph_sizes) == 1:
274
max_capture_size = scheduler_config.cuda_graph_sizes[0]
275
# FIXME(woosuk): When using full cuda graph with FA3, the max
276
# supported size is 992.
277
- if max_capture_size < 1024:
277
+ if max_capture_size < 992:
278
cuda_graph_sizes = [1, 2, 4]
279
# Step size 8 for small batch sizes
280
cuda_graph_sizes += [i for i in range(8, 256, 8)]
281
# Step size 16 for larger batch sizes
282
- cuda_graph_sizes += [i for i in range(256, 1025, 16)]
282
+ cuda_graph_sizes += [i for i in range(256, 993, 16)]
283
scheduler_config.cuda_graph_sizes = cuda_graph_sizes
284
logger.info(
285
"Overriding max cuda graph capture size to "
286
- "%d for performance.", 1024)
286
+ "%d for performance.", 992)
287
288
289
class MambaModelConfig(VerifyAndUpdateConfig):