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 7, 02:45 UTC
[Bugfix] Fix `cudagraph_mode:FULL` dispatch (This does not impact `FULL_AND_PIECEWISE` (default)) (#36165)
Raphaël Rialland committed
Mar 6, 2026 at 15:15 UTC
39f9ea0da4a45e9638937b062f86f03db313a0d8
1 file changed
+2
-4
vllm/v1/cudagraph_dispatcher.py
+2
-4
@@ -293,16 +293,14 @@ class CudagraphDispatcher:
293
)
294
effective_num_active_loras = self.vllm_config.lora_config.max_loras + 1
295
296
+ normalized_uniform = uniform_decode and self.cudagraph_mode.separate_routine()
297
batch_desc = self._create_padded_batch_descriptor(
297
- num_tokens, uniform_decode, has_lora, effective_num_active_loras
298
+ num_tokens, normalized_uniform, has_lora, effective_num_active_loras
299
)
300
301
if CUDAGraphMode.FULL in allowed_modes:
302
# check if key exists for full cudagraph
302
- # For pure FULL mode, keys are registered with uniform=False.
303
batch_desc_to_check = batch_desc
304
- if self.cudagraph_mode == CUDAGraphMode.FULL:
305
- batch_desc_to_check = replace(batch_desc, uniform=False)
304
if batch_desc_to_check in self.cudagraph_keys[CUDAGraphMode.FULL]:
305
return CUDAGraphMode.FULL, batch_desc_to_check
306