fix: tiny fix cli (#12744)

python/sglang/multimodal_gen/docs/cli.md

@@ -143,7 +143,7 @@ SERVER_ARGS=(
   --ring-degree=2
 )
 
-sglang serve $SERVER_ARGS
+sglang serve"${SERVER_ARGS[@]}"
 ```
 
 - **--model-path**: Which model to load. The example uses `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`.
@@ -265,7 +265,7 @@ SAMPLING_ARGS=(
   --output-file-name "A curious raccoon.mp4"
 )
 
-sglang generate $SERVER_ARGS $SAMPLING_ARGS
+sglang generate "${SERVER_ARGS[@]}" "${SAMPLING_ARGS[@]}"
 ```
 
 Once the generation task has finished, the server will shut down automatically.

python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py

@@ -258,10 +258,10 @@ class DiffGenerator:
         data_type = (
             DataType.IMAGE
             if self.server_args.pipeline_config.is_image_gen
-            or sampling_params.num_frames == 1
+            or pretrained_sampling_params.num_frames == 1
             else DataType.VIDEO
         )
-        sampling_params.data_type = data_type
+        pretrained_sampling_params.data_type = data_type
         pretrained_sampling_params.set_output_file_name()
 
         requests: list[Req] = []

python/sglang/multimodal_gen/runtime/platforms/cuda.py

@@ -217,6 +217,7 @@ class CudaPlatformBase(Platform):
         elif selected_backend == AttentionBackendEnum.FA3:
             if is_blackwell():
                 raise ValueError("The 'fa3' backend is not supported on Blackwell GPUs")
+            target_backend = AttentionBackendEnum.FA3
         elif selected_backend:
             raise ValueError(f"Invalid attention backend for {cls.device_name}")
         else:

python/sglang/multimodal_gen/runtime/server_args.py

@@ -777,6 +777,13 @@ class ServerArgs:
                 )
             self.sp_degree = self.ulysses_degree = self.ring_degree = 1
 
+        if (
+            self.ring_degree is not None
+            and self.ring_degree > 1
+            and self.attention_backend != "fa3"
+        ):
+            raise ValueError("Ring Attention is only supported for fa3 backend for now")
+
         if self.sp_degree == -1:
             # assume we leave all remaining gpus to sp
             num_gpus_per_group = self.dp_size * self.tp_size