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][DCP] Set default CUDAGraphMode to PIECEWISE for DCP (#26574)
Signed-off-by: FENP <32334296+FENP@users.noreply.github.com>
Jaya Yuan committed
Oct 12, 2025 at 17:58 UTC
b91d8db873a5f4d639a5cb57288cd94ed1614bb0
1 file changed
+9
vllm/config/vllm.py
+9
@@ -350,6 +350,15 @@ class VllmConfig:
350
or self.model_config.is_encoder_decoder
351
):
352
self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
353
+
354
+ # decode context parallel do not support full cudagraphs now.
355
+ if self.parallel_config.decode_context_parallel_size > 1:
356
+ logger.warning(
357
+ "Decode context parallel (DCP) is enabled, which is "
358
+ "incompatible with full CUDA graphs. Set "
359
+ "cudagraph_mode to PIECEWISE."
360
+ )
361
+ self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
362
else:
363
self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
364