From 93f38fe410e4d18b550a0f07c09a136e98e8f160 Mon Sep 17 00:00:00 2001 From: Qiaolin Yu Date: Mon, 6 Apr 2026 13:49:03 -0700 Subject: [PATCH] tiny fix chain-style multi layer eagle comments (#22206) --- .../multi_layer_eagle_draft_extend_cuda_graph_runner.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py b/python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py index 4eb3d55c4..f036ed371 100644 --- a/python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py +++ b/python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py @@ -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.