diff --git a/Dockerfile.bi100 b/Dockerfile.bi100
new file mode 100644
index 0000000..be3de26
--- /dev/null
+++ b/Dockerfile.bi100
@@ -0,0 +1,7 @@
+FROM git.modelhub.org.cn:980/enginex-iluvatar/bi100-3.2.1-x86-ubuntu20.04-py3.10-poc-llm-infer:v1.2.2
+
+WORKDIR /workspace
+
+RUN pip install diffusers==0.34.0
+RUN pip install imageio[ffmpeg] einops datasets==3.2.0 simplejson addict open_clip_torch==2.24.0 sortedcontainers modelscope==1.28.2 av==11.0.0 pytorch-lightning
+COPY . /workspace/
diff --git a/diffusers_video.py b/diffusers_video.py
new file mode 100644
index 0000000..babf2b0
--- /dev/null
+++ b/diffusers_video.py
@@ -0,0 +1,13 @@
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.utils import export_to_video
+
+model_path = "/mnt/models/AI-ModelScope/text-to-video-ms-1.7b"
+pipe = DiffusionPipeline.from_pretrained(
+            model_path, torch_dtype=torch.float16, variant="fp16"
+            )
+pipe.enable_model_cpu_offload()       # 省显存
+pipe.enable_vae_slicing()
+
+frames = pipe("Spiderman is surfing", num_frames=16).frames[0]
+export_to_video(frames, "output.mp4")    # 默认约2秒/8fps
diff --git a/main.py b/main.py
index ad053b0..72a8ef8 100644
--- a/main.py
+++ b/main.py
@@ -13,6 +13,8 @@ import patch
 from modelscope.pipelines import pipeline
 from modelscope.outputs import OutputKeys
 import torch
+from diffusers import DiffusionPipeline
+from diffusers.utils import export_to_video
 
 
 def safe_stem(text: str, maxlen: int = 60) -> str:
@@ -47,12 +49,19 @@ def load_prompts(json_path: Path):
     return prompts
 
 
-def build_pipeline(model_path: str, device: str = "cuda", dtype=torch.float16):
-    pipe = pipeline('text-to-video-synthesis', model_path, device=device)
+def build_pipeline(model_path: str, device: str = "cuda", dtype=torch.float16, model_type: str = "text-to-video-synthesis"):
+    if model_type == "text-to-video-synthesis":
+        pipe = pipeline('text-to-video-synthesis', model_path, device=device)
+    elif model_type == "text-to-video-ms":
+        pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=dtype)
+        pipe.enable_model_cpu_offload()       # 省显存
+        pipe.enable_vae_slicing()
+    else:
+        raise ValueError(f"不支持的模型类型: {model_type}")
     return pipe
 
 
-def generate_one(pipe, cfg: dict, out_dir: Path, index: int):
+def generate_one(pipe, cfg: dict, out_dir: Path, index: int, model_type: str = "text-to-video-synthesis"):
     """
     依据 cfg 生成一张图并返回 (保存路径, 耗时秒, 详细参数)
     支持字段：
@@ -65,7 +74,13 @@ def generate_one(pipe, cfg: dict, out_dir: Path, index: int):
     out_path = out_dir / filename
 
     start = time.time()
-    output_video_path = pipe({"text": prompt}, output_video=str(out_path))[OutputKeys.OUTPUT_VIDEO]
+    if model_type == "text-to-video-synthesis":
+        output_video_path = pipe({"text": prompt}, output_video=str(out_path))[OutputKeys.OUTPUT_VIDEO]
+    elif model_type == "text-to-video-ms":
+        frames = pipe(prompt, num_frames=16).frames[0]
+        export_to_video(frames, str(out_path))
+    else:
+        raise ValueError(f"不支持的模型类型: {model_type}")
     elapsed = time.time() - start
 
     detail = {
@@ -87,6 +102,7 @@ def main():
     parser.add_argument("--outdir", required=True, help="图片输出目录")
     parser.add_argument("--device", default="cuda", help="推理设备")
     parser.add_argument("--dtype", default="fp16", choices=["fp16", "fp32"], help="推理精度")
+    parser.add_argument("--model_type", default="text-to-video-synthesis", choices=["text-to-video-synthesis", "text-to-video-ms"], help="模型类型")
     args, _ = parser.parse_known_args()
 
     model_path = args.model
@@ -103,12 +119,12 @@ def main():
     if not prompts:
         raise ValueError("测试列表为空。")
 
-    pipe = build_pipeline(model_path=model_path, device=args.device, dtype=dtype)
+    pipe = build_pipeline(model_path=model_path, device=args.device, dtype=dtype, model_type=args.model_type)
 
     records = []
     total_start = time.time()
     for i, cfg in enumerate(prompts, 1):
-        out_path, elapsed, detail = generate_one(pipe, cfg, out_dir, i)
+        out_path, elapsed, detail = generate_one(pipe, cfg, out_dir, i, model_type=args.model_type)
         print(f"[{i}/{len(prompts)}] saved: {out_path.name}  elapsed: {elapsed:.3f}s")
         records.append(detail)
     total_elapsed = round(time.time() - total_start, 6)
diff --git a/patch.py b/patch.py
index 04786e2..2de0d3f 100644
--- a/patch.py
+++ b/patch.py
@@ -22,7 +22,7 @@ try:
   if torch.npu.is_available() and not torch.cuda.is_available():
       enable_cuda_to_npu_shim()
 except: 
-  print("exception")
+  print("no npu. use native cuda")
 
 # 1) 可选：如果你的权重来自 lightning 的 ckpt，放行其类（仅在可信来源时）
 try:
diff --git a/run_in_docker_bi100.sh b/run_in_docker_bi100.sh
new file mode 100755
index 0000000..c0ee0eb
--- /dev/null
+++ b/run_in_docker_bi100.sh
@@ -0,0 +1,3 @@
+#! /usr/bin/env bash
+image=harbor-contest.4pd.io/zhanghao/t2v:bi100-0.1
+docker run -it -v /data2:/data2 -v /home/zhanghao/workspace:/host -v /mnt:/mnt --device=dev/iluvatar1:/dev/iluvatar0  --device=/dev/iluvatar2:/dev/iluvatar1 $image bash
diff --git a/test_ms.sh b/test_ms.sh
new file mode 100755
index 0000000..34a63a4
--- /dev/null
+++ b/test_ms.sh
@@ -0,0 +1 @@
+python3 main.py --model "/mnt/models/AI-ModelScope/text-to-video-ms-1.7b" --json  "dataset.json" --results "results.json" --outdir "output" --device cuda --dtype fp16 --model_type "text-to-video-ms"
\ No newline at end of file