A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 00:15 UTC
[MTP][Misc] Clean up dead code (#36507)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Mar 9, 2026 at 14:43 UTC
4e571ce6433b6768950becda40d55cb4f24741ce
1 file changed
+1
-4
vllm/v1/spec_decode/eagle.py
+1
-4
@@ -481,10 +481,7 @@ class SpecDecodeBaseProposer:
481
positions = self.mrope_positions[:, token_indices_to_sample]
482
else:
483
positions = self.positions[token_indices_to_sample]
484
- if self.method == "mtp":
485
- hidden_states = self.hidden_states[token_indices_to_sample]
486
- else:
487
- hidden_states = hidden_states[token_indices_to_sample]
484
+ hidden_states = hidden_states[token_indices_to_sample]
485
486
if isinstance(attn_metadata, TreeAttentionMetadata):
487
# Draft using tree attention - requires full logits for top-k