From b7950e798f2094b1163c22787c0ba2e3231bf01b Mon Sep 17 00:00:00 2001 From: avalliappan-nvidia Date: Wed, 15 Jul 2026 12:09:00 -0700 Subject: [PATCH] [Bugfix] Initialize draft CUDA-graph keys for the native draft_model proposer (#47460) Signed-off-by: Alagappan Valliappan Co-authored-by: Claude --- vllm/v1/worker/gpu_model_runner.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index ffc596417f..b104e4c5ca 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -7061,12 +7061,14 @@ class GPUModelRunner( # Initialize drafter's cudagraph dispatcher if using spec decode. if self.speculative_config and ( self.speculative_config.use_eagle() + or self.speculative_config.uses_draft_model() or self.speculative_config.uses_extract_hidden_states() ): assert isinstance( self.drafter, EagleProposer | DFlashProposer + | DraftModelProposer | ExtractHiddenStatesProposer | Gemma4Proposer, )