From 429ac2d82ccc931712a866e6e0991cf0413e82d4 Mon Sep 17 00:00:00 2001 From: Duyi-Wang Date: Thu, 3 Sep 2026 17:46:49 +0800 Subject: [PATCH] [AMD] Fix DSv4 draft extend taking the target compression path during prefill (#37713) --- .../srt/layers/attention/deepseek_v4_backend_hip_radix.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend_hip_radix.py b/python/sglang/srt/layers/attention/deepseek_v4_backend_hip_radix.py index d1caca5ca..3f3fde172 100644 --- a/python/sglang/srt/layers/attention/deepseek_v4_backend_hip_radix.py +++ b/python/sglang/srt/layers/attention/deepseek_v4_backend_hip_radix.py @@ -481,9 +481,9 @@ class DeepseekV4HipRadixBackend( self.mtp_enabled = self.topk > 0 self.speculative_num_steps = speculative_num_steps self.speculative_num_draft_tokens: int = get_spec().speculative_num_draft_tokens + self.is_draft_worker = getattr(model_runner, "is_draft_worker", False) self.is_dspark_draft = ( - getattr(model_runner, "is_draft_worker", False) - and model_runner.spec_algorithm.is_dspark() + self.is_draft_worker and model_runner.spec_algorithm.is_dspark() ) self.target_verify_num_draft_tokens = self.speculative_num_draft_tokens if self.is_dspark_draft: @@ -1149,7 +1149,9 @@ class DeepseekV4HipRadixBackend( and extend_seq_lens is not None and extend_seq_lens_cpu is not None ) - is_draft = forward_batch.forward_mode.is_draft_extend_v2() + is_draft = ( + forward_batch.forward_mode.is_draft_extend_v2() or self.is_draft_worker + ) metadata = self.init_forward_metadata_prefill( max_seq_len=max_seq_len, req_pool_indices=req_pool_indices,