Add SDPA fallback in flash_attention for GPUs without flash-attn

wan/modules/attention.py

@@ -192,6 +192,23 @@ def flash_attention(
     assert dtype in half_dtypes
     assert q.device.type == 'cuda' and q.size(-1) <= 256
 
+    # Fallback: on hardware without FlashAttention (e.g. ZeroGPU / Blackwell
+    # without a prebuilt flash-attn wheel), use exact SDPA. Inputs here are
+    # [B, L, N, D]; SDPA expects [B, N, L, D]. This matches the math of the
+    # flash path (full, non-varlen attention over the given sequences).
+    if not FLASH_ATTN_3_AVAILABLE and not FLASH_ATTN_2_AVAILABLE:
+        out_dtype = q.dtype
+        qs = q.transpose(1, 2).to(dtype)
+        ks = k.transpose(1, 2).to(dtype)
+        vs = v.transpose(1, 2).to(dtype)
+        if q_scale is not None:
+            qs = qs * q_scale
+        x = torch.nn.functional.scaled_dot_product_attention(
+            qs, ks, vs, attn_mask=None, is_causal=causal,
+            dropout_p=dropout_p, scale=softmax_scale,
+        )
+        return x.transpose(1, 2).contiguous().type(out_dtype)
+
     # params
     b, lq, lk, out_dtype = q.size(0), q.size(1), k.size(1), q.dtype