[XPU] Enable ring attention on XPUAttentionBackend (#41406)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
df1e4f3f2b
commit
fedab5537b
@@ -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)
|
||||
|
||||
Reference in New Issue
Block a user