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
[Logging] Improve log for when DeepEP HT disables CUDA Graphs (#25531)
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Tyler Michael Smith committed
Sep 24, 2025 at 18:43 UTC
fea8006062866900c1a2bee9dd4f18381777e332
1 file changed
+6
-5
vllm/platforms/cuda.py
+6
-5
@@ -186,11 +186,12 @@ class CudaPlatformBase(Platform):
186
# if torch compile cache key issue fixed
187
# See https://github.com/vllm-project/vllm/pull/25093
188
logger.info(
189
- "Data Parallel: disabling cudagraphs since DP "
190
- "with DeepEP high-throughput kernels are not CUDA Graph "
191
- "compatible. The DeepEP low-latency kernels are CUDA Graph "
192
- "compatible. Set the all_to_all backend to deepep_low_latency "
193
- "to use those kernels instead.")
189
+ "WideEP: Disabling CUDA Graphs since DeepEP high-throughput "
190
+ "kernels are optimized for prefill and are incompatible with "
191
+ "CUDA Graphs. "
192
+ "In order to use CUDA Graphs for decode-optimized workloads, "
193
+ "set VLLM_ALL2ALL_BACKEND to another option, such as "
194
+ "deepep_low_latency, pplx, or allgather_reducescatter.")
195
compilation_config.cudagraph_mode = CUDAGraphMode.NONE
196
197
@classmethod