[diffusion] refactor: move SLA to attention_backend folder (#17020)

2026-01-15 21:36:48 +08:00
parent e997995037
commit e7df8bdc5c
6 changed files with 422 additions and 296 deletions
@@ -245,6 +245,9 @@ class CudaPlatformBase(Platform):
        elif selected_backend == AttentionBackendEnum.TORCH_SDPA:
            logger.info("Using Torch SDPA backend")
            return "sglang.multimodal_gen.runtime.layers.attention.backends.sdpa.SDPABackend"
+        elif selected_backend == AttentionBackendEnum.SLA_ATTN:
+            logger.info("Using Sparse Linear Attention backend")
+            return "sglang.multimodal_gen.runtime.layers.attention.backends.sparse_linear_attn.SparseLinearAttentionBackend"
        elif selected_backend in [
            AttentionBackendEnum.FA,
        ]:
@@ -33,6 +33,7 @@ class AttentionBackendEnum(enum.Enum):
    VIDEO_SPARSE_ATTN = enum.auto()
    VMOBA_ATTN = enum.auto()
    AITER = enum.auto()
+    SLA_ATTN = enum.auto()
    NO_ATTENTION = enum.auto()

    def __str__(self):