From 58f0d250ecf462dc80eb495a2a813497cbe315d1 Mon Sep 17 00:00:00 2001 From: jiaryang <96409713+jiaryang@users.noreply.github.com> Date: Fri, 2 Oct 2026 09:45:29 +0800 Subject: [PATCH] [ROCm] Use the fused MLA absorb + RoPE + KV-write kernel for decode-sized forward modes only (#41533) Co-authored-by: Thomas Wang --- .../attention_forward_methods/forward_mla_rocm.py | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_rocm.py b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_rocm.py index 0ee2a3cb9b..21c12c8288 100644 --- a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_rocm.py +++ b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_rocm.py @@ -618,8 +618,12 @@ class DeepseekMLARocmForwardMixin: q_nope, q_pe, k_pe = self._split_q_nope_pe(q, latent_cache) - fuse_bmm_rope_cache = not q_replicate_active and ( - _can_fuse_bmm_rope_cat_and_cache(self) + # The fused kernel wins at decode-sized batches (decode, target verify, + # draft extend); prefill shapes run faster as the two separate launches. + fuse_bmm_rope_cache = ( + not q_replicate_active + and not forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed() + and _can_fuse_bmm_rope_cat_and_cache(self) ) if q_replicate_active: