minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

tiny fix chain-style multi layer eagle comments (#22206)

Browse Source
This commit is contained in:
Qiaolin Yu
2026-04-06 13:49:03 -07:00
committed by GitHub
parent 8f337682bd
commit 93f38fe410
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py
+1 -1
View File
@@ -429,7 +429,7 @@ class MultiLayerEagleDraftExtendCudaGraphRunner:
forward_batch, 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 # output (hidden_states_before_norm) so that assign_new_state_triton
# propagates each MTP layer's own output to the next MTP layer, # propagates each MTP layer's own output to the next MTP layer,
# rather than always feeding the target model's hidden states. # rather than always feeding the target model's hidden states.