From 28f3a2d8ed6daf3ea4f42eabe5dcfa61950422e2 Mon Sep 17 00:00:00 2001 From: Tarushii Goel Date: Mon, 20 Apr 2026 16:22:16 -0700 Subject: [PATCH] [sgl] multilayereagleworkerv2 fix (#22954) --- .../sglang/srt/speculative/multi_layer_eagle_worker_v2.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py b/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py index 34ce26ee4..3c96a5315 100644 --- a/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py +++ b/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py @@ -394,6 +394,9 @@ class MultiLayerEagleDraftWorker(BaseDraftWorker): topk_p_list = [] topk_index_list = [] for step in range(self.speculative_num_steps): + forward_batch.req_to_token_pool = self.draft_runner_list[ + step + ].req_to_token_pool output: ModelRunnerOutput = self.draft_runner_list[step].forward( forward_batch ) @@ -498,6 +501,9 @@ class MultiLayerEagleDraftWorker(BaseDraftWorker): draft_logits_output.topk_index, ) else: + forward_batch.req_to_token_pool = self.draft_runner_list[ + step + ].req_to_token_pool draft_logits_output = self.draft_runner_list[step].forward( forward_batch, skip_attn_backend_init=True )