[Diffusion] Fix Ring Parallel bug with FA4 (sgl-project#18062)

BBuf · Johnsonms · commit 8ac57b5b11f8 · 2026-02-14T01:41:21.000Z
diff --git a/python/sglang/jit_kernel/flash_attention/cute/interface.py b/python/sglang/jit_kernel/flash_attention/cute/interface.py
@@ -1364,6 +1364,7 @@ def forward(
             pack_gqa=pack_gqa,
             score_mod=score_mod,
             aux_tensors=aux_tensors,
+            return_lse=True,
         )
         ctx.save_for_backward(q, k, v, out, lse, cu_seqlens_q, cu_seqlens_k, seqused_q, seqused_k)
         ctx.softmax_scale = softmax_scale

Original file line number	Diff line number	Diff line change
`@@ -1364,6 +1364,7 @@ def forward(`
`1364`	`1364`	`pack_gqa=pack_gqa,`
`1365`	`1365`	`score_mod=score_mod,`
`1366`	`1366`	`aux_tensors=aux_tensors,`
	`1367`	`+ return_lse=True,`
`1367`	`1368`	`)`
`1368`	`1369`	`ctx.save_for_backward(q, k, v, out, lse, cu_seqlens_q, cu_seqlens_k, seqused_q, seqused_k)`
`1369`	`1370`	`ctx.softmax_scale = softmax_scale`