From 99aefa037edf44b855a22bf350adac7c559cded8 Mon Sep 17 00:00:00 2001 From: Jay Zhou <50169346+Ja1Zhou@users.noreply.github.com> Date: Sun, 20 Jul 2025 00:28:06 -0700 Subject: [PATCH] Fix eagle3 cuda graph (#8163) --- .../eagle_draft_extend_cuda_graph_runner.py | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py b/python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py index b4ffde60e..7057c502d 100644 --- a/python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +++ b/python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py @@ -84,7 +84,15 @@ class EAGLEDraftExtendCudaGraphRunner: self.hidden_states = torch.zeros( ( self.max_num_token, - self.model_runner.model_config.hidden_size * 3, + ( + self.model_runner.model_config.hf_config.target_hidden_size + * 3 + if hasattr( + self.model_runner.model_config.hf_config, + "target_hidden_size", + ) + else self.model_runner.model_config.hidden_size * 3 + ), ), dtype=self.model_runner.dtype, )