From 39f9ea0da4a45e9638937b062f86f03db313a0d8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Rapha=C3=ABl=20Rialland?= <36076211+TQCB@users.noreply.github.com> Date: Fri, 6 Mar 2026 15:15:31 +0100 Subject: [PATCH] [Bugfix] Fix `cudagraph_mode:FULL` dispatch (This does not impact `FULL_AND_PIECEWISE` (default)) (#36165) --- vllm/v1/cudagraph_dispatcher.py | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/vllm/v1/cudagraph_dispatcher.py b/vllm/v1/cudagraph_dispatcher.py index be459cd29..b852808ec 100644 --- a/vllm/v1/cudagraph_dispatcher.py +++ b/vllm/v1/cudagraph_dispatcher.py @@ -293,16 +293,14 @@ class CudagraphDispatcher: ) effective_num_active_loras = self.vllm_config.lora_config.max_loras + 1 + normalized_uniform = uniform_decode and self.cudagraph_mode.separate_routine() batch_desc = self._create_padded_batch_descriptor( - num_tokens, uniform_decode, has_lora, effective_num_active_loras + num_tokens, normalized_uniform, has_lora, effective_num_active_loras ) if CUDAGraphMode.FULL in allowed_modes: # check if key exists for full cudagraph - # For pure FULL mode, keys are registered with uniform=False. batch_desc_to_check = batch_desc - if self.cudagraph_mode == CUDAGraphMode.FULL: - batch_desc_to_check = replace(batch_desc, uniform=False) if batch_desc_to_check in self.cudagraph_keys[CUDAGraphMode.FULL]: return CUDAGraphMode.FULL, batch_desc_to_check