[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>
This commit is contained in:
avalliappan-nvidia
2026-07-15 15:09:00 -04:00
committed by GitHub
co-authored by Claude
parent de100ffb62
commit b7950e798f
+2
View File
@@ -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,
)