From 590969ee9c6af96ea484623241ca8530577986f7 Mon Sep 17 00:00:00 2001
From: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>
Date: Thu, 22 Jan 2026 08:23:53 +0800
Subject: [PATCH] [Diffusion] Support select fa2 backend in hopper (#17514)

---
 python/sglang/multimodal_gen/runtime/platforms/cuda.py | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/python/sglang/multimodal_gen/runtime/platforms/cuda.py b/python/sglang/multimodal_gen/runtime/platforms/cuda.py
index cebd7dd53..d59326de1 100644
--- a/python/sglang/multimodal_gen/runtime/platforms/cuda.py
+++ b/python/sglang/multimodal_gen/runtime/platforms/cuda.py
@@ -246,6 +246,13 @@ class CudaPlatformBase(Platform):
         elif selected_backend == AttentionBackendEnum.SAGE_SLA_ATTN:
             logger.info("Using Sage Sparse Linear Attention backend")
             return "sglang.multimodal_gen.runtime.layers.attention.backends.sparse_linear_attn.SageSparseLinearAttentionBackend"
+        elif selected_backend == AttentionBackendEnum.FA2:
+            from sglang.multimodal_gen.runtime.layers.attention.backends.flash_attn_2 import (  # noqa: F401
+                FlashAttention2Backend,
+            )
+
+            logger.info("Using FlashAttention2 backend")
+            return "sglang.multimodal_gen.runtime.layers.attention.backends.flash_attn_2.FlashAttention2Backend"
         elif selected_backend in [
             AttentionBackendEnum.FA,
         ]: