diff --git a/python/sglang/multimodal_gen/runtime/layers/attention/backends/xpu_backend.py b/python/sglang/multimodal_gen/runtime/layers/attention/backends/xpu_backend.py index 229a7b17d5..f4095cfeb0 100644 --- a/python/sglang/multimodal_gen/runtime/layers/attention/backends/xpu_backend.py +++ b/python/sglang/multimodal_gen/runtime/layers/attention/backends/xpu_backend.py @@ -26,6 +26,10 @@ from sglang.multimodal_gen.runtime.layers.attention.backends.flash_attn import ( class XPUAttentionBackend(AttentionBackend): + @classmethod + def supports_ring_rotation(cls) -> bool: + return True + accept_output_buffer: bool = True @staticmethod @@ -115,6 +119,8 @@ class XPUAttentionImpl(AttentionImpl): if return_softmax_lse: out_tensor, softmax_lse = out[:2] result = out_tensor.reshape(bsz, seqlen_q, nheads_q, d) + # The kernel returns lse as (H, B * S); ring attention needs [B, H, S]. + softmax_lse = softmax_lse.view(nheads_q, bsz, seqlen_q).permute(1, 0, 2) return result, softmax_lse result = out.reshape(bsz, seqlen_q, nheads_q, d)