tiny fix chain-style multi layer eagle comments (#22206)
This commit is contained in:
@@ -429,7 +429,7 @@ class MultiLayerEagleDraftExtendCudaGraphRunner:
|
||||
forward_batch,
|
||||
)
|
||||
|
||||
# Chain-style MTP: overwrite self.hidden_states with the draft model's
|
||||
# Chain-style MTP: overwrite buffers.hidden_states with the draft model's
|
||||
# output (hidden_states_before_norm) so that assign_new_state_triton
|
||||
# propagates each MTP layer's own output to the next MTP layer,
|
||||
# rather than always feeding the target model's hidden states.
|
||||
|
||||
Reference in New Issue
Block a user