From c394a90c2b4d82483ed908d2afbca1901d9d59d2 Mon Sep 17 00:00:00 2001 From: Alex Nails Date: Thu, 22 Oct 2026 15:37:19 -0700 Subject: [PATCH] [Perf] gemma4: Capture single-step frozen-KV drafts (#45771) --- python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py b/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py index 3527bbc3ed..a8cdd5ac67 100644 --- a/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py +++ b/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py @@ -373,7 +373,8 @@ class FrozenKVMTPDraftWorker(EagleDraftWorkerBase, TpModelWorker): self.draft_attn_backend.init_forward_metadata_out_graph(fb_view) def _capture_cuda_graphs(self) -> None: - if cuda_graph_fully_disabled() or self.speculative_num_steps <= 1: + # Even a one-step frozen draft runs the assistant seed forward. + if cuda_graph_fully_disabled() or self.speculative_num_steps <= 0: return if self.target_worker.device != "cuda": logger.info(