diff --git a/python/sglang/srt/debug_utils/pr_fix_toggle.py b/python/sglang/srt/debug_utils/pr_fix_toggle.py index 3eee2837d..2f602fdc9 100644 --- a/python/sglang/srt/debug_utils/pr_fix_toggle.py +++ b/python/sglang/srt/debug_utils/pr_fix_toggle.py @@ -66,6 +66,17 @@ patches: """ +_PR_REVERT_YAML_27338 = """ +patches: + - target: sglang.srt.layers.attention.flashinfer_backend.FlashInferMultiStepDraftBackend.init_cuda_graph_state + edits: + - match: | + (self.speculative_num_steps, max_bs * self.topk * self.max_context_len), + replacement: | + (self.speculative_num_steps, max_bs * self.max_context_len), +""" + + _PR_REVERT_YAML_27360 = """ patches: - target: sglang.srt.layers.attention.flashattention_backend.FlashAttentionBackend._apply_cuda_graph_metadata @@ -79,6 +90,7 @@ patches: _PR_FIX_REVERT_YAML: Dict[int, str] = { 25015: _PR_REVERT_YAML_25015, 26329: _PR_REVERT_YAML_26329, + 27338: _PR_REVERT_YAML_27338, 27360: _PR_REVERT_YAML_27360, }