[Bugfix] Initialize draft CUDA-graph keys for the native draft_model proposer (#47460)
Signed-off-by: Alagappan Valliappan <avalliappan@nvidia.com> Co-authored-by: Claude <noreply@anthropic.com>
avalliappan-nvidia committed
Jul 15, 2026 at 12:09 UTC
b7950e798f2094b1163c22787c0ba2e3231bf01b
1 file changed
+2
vllm/v1/worker/gpu_model_runner.py
+2
@@ -7061,12 +7061,14 @@ class GPUModelRunner(
7061
# Initialize drafter's cudagraph dispatcher if using spec decode.
7062
if self.speculative_config and (
7063
self.speculative_config.use_eagle()
7064
+ or self.speculative_config.uses_draft_model()
7065
or self.speculative_config.uses_extract_hidden_states()
7066
):
7067
assert isinstance(
7068
self.drafter,
7069
EagleProposer
7070
| DFlashProposer
7071
+ | DraftModelProposer
7072
| ExtractHiddenStatesProposer
7073
| Gemma4Proposer,
7074
)