[nvbug/5354946][fix] Fix mtp vanilla draft inputs (NVIDIA#5568)

lfr-0531 · dominicshanshan · commit 6265a5e87184 · 2025-07-09T01:40:23.000-07:00
Signed-off-by: Fanrong Li &lt;23290157+lfr-0531@users.noreply.github.com&gt;
diff --git a/tensorrt_llm/_torch/speculative/mtp.py b/tensorrt_llm/_torch/speculative/mtp.py
@@ -522,7 +522,6 @@ def forward(
                 "position_ids": draft_inputs["position_ids"],
                 "hidden_states": draft_hidden_states,
                 "attn_metadata": draft_inputs["attn_metadata"],
-                "spec_metadata": draft_inputs["spec_metadata"],
             }
         next_draft_tokens = torch.stack(next_draft_tokens, dim=1)
 

Original file line number	Diff line number	Diff line change
`@@ -522,7 +522,6 @@ def forward(`
`522`	`522`	`"position_ids": draft_inputs["position_ids"],`
`523`	`523`	`"hidden_states": draft_hidden_states,`
`524`	`524`	`"attn_metadata": draft_inputs["attn_metadata"],`
`525`		`- "spec_metadata": draft_inputs["spec_metadata"],`
`526`	`525`	`}`
`527`	`526`	`next_draft_tokens = torch.stack(next_draft_tokens, dim=1)`
`528`	`527`