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 14, 05:30 UTC
[bugfix]fix MTP hidden states (#24056)
Signed-off-by: Lu Fang <fanglu@fb.com>
Lucia Fang committed
Sep 1, 2025 at 14:09 UTC
a344a5aa0a58cc1758d9721e848ce1f5ca4b6c7f
1 file changed
+1
vllm/v1/spec_decode/eagle.py
+1
@@ -220,6 +220,7 @@ class EagleProposer:
220
)
221
if self.method in ("deepseek_mtp", "ernie_mtp"):
222
last_hidden_states = ret_hidden_states
223
+ hidden_states = last_hidden_states
224
else:
225
last_hidden_states, hidden_states = ret_hidden_states
226
sample_hidden_states = last_hidden_states[last_token_indices]