mlu

Dockerfile

@@ -9,4 +9,4 @@ RUN pip install open_clip_torch==2.24.0 sortedcontainers modelscope av==11.0.0
 RUN sed -i 's|source /torch/venv3/pytorch/bin/activate|source /torch/venv3/pytorch_infer/bin/activate|' /root/.bashrc
 
 COPY . /workspace/
-RUN pip install whls/cambricon_pytorch_lightning-2.5.0+mlu0.7.0-py3-none-any.whl 
+RUN pip install whls.mlu/cambricon_pytorch_lightning-2.5.0+mlu0.7.0-py3-none-any.whl 

Dockerfile.a100

@@ -1,7 +0,0 @@
-FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel
-
-WORKDIR /workspace
-
-RUN pip install opencv-python-headless imageio[ffmpeg] einops datasets==3.2.0 simplejson diffusers==0.34.0 open_clip_torch==2.24.0 sortedcontainers modelscope av==11.0.0 addict -i https://nexus.4pd.io/repository/pypi-all/simple
-
-COPY . /workspace/

Dockerfile.ascend

@@ -1,9 +0,0 @@
-FROM git.modelhub.org.cn:9443/enginex-ascend/vllm-ascend:v0.10.0rc1
-
-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/
-RUN patch /usr/local/python3.11.13/lib/python3.11/site-packages/modelscope/models/multi_modal/video_synthesis/text_to_video_synthesis_model.py patch.ascend/text_to_video_synthesis_model.py.patch
-RUN patch /usr/local/python3.11.13/lib/python3.11/site-packages/modelscope/utils/device.py patch.ascend/device.py.patch

Dockerfile.r200

@@ -1,8 +0,0 @@
-FROM git.modelhub.org.cn:9443/enginex-kunlunxin/text2image/r200_8f-diffuser:v0.21.4
-
-WORKDIR /workspace
-
-RUN source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda && pip install imageio[ffmpeg] einops datasets==3.1.0 simplejson addict sortedcontainers modelscope==1.28.2 av==11.0.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
-RUN source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda && pip install megatron megatron-core urllib3==1.26.20 -i https://pypi.tuna.tsinghua.edu.cn/simple
-RUN source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda && pip install open_clip_torch==2.24.0 pytorch-lightning==2.0.1 -i https://pypi.tuna.tsinghua.edu.cn/simple
-COPY . /workspace/

README.md

@@ -0,0 +1,23 @@
+## Quickstart
+
+### 构建镜像
+```bash
+docker build -t text2video:v0.1 .
+```
+
+### 模型下载
+模型地址：https://modelscope.cn/models/iic/text-to-video-synthesis
+并放到目录：`/mnt/contest_ceph/zhanghao/models/iic/text-to-video-synthesis`（如更改目录，请修改后面的执行脚本中的模型路径）
+
+### 测试程序
+1. 准备输入数据集，可以参考示例`dataset.json`
+2. 在docker镜像里运行测试程序，会根据`dataset.json`内容，在`output`目录下生成视频文件。
+```bash
+./run_in_docker.sh
+```
+
+## 测试结果
+|      | A100 平均生成时间（秒） | MLU-x4 平均生成时间（秒） |   MLU-x8 平均生成时间（秒）|
+|------|-------------------------|----------------------------|---------------------------|
+| 时间 | 12                  | 37                     |   45
+

dataset.json

@@ -1,4 +1,4 @@
 [
-    "An image of a squirrel in Picasso style",
-    "A cozy cabin in the woods, watercolor painting"
-]
+    "A panda is eating burgers and french fries",
+    "A sheep is walking and eating in the grass"
+]

iic.py

@@ -1,16 +0,0 @@
-import os
-import torch
-device = "cuda" if torch.cuda.is_available() else "npu"
-import patch
-
-from modelscope.pipelines import pipeline
-from modelscope.outputs import OutputKeys
-
-model_path = "/mnt/contest_ceph/zhanghao/models/iic/text-to-video-synthesis"
-p = pipeline('text-to-video-synthesis', model_path, device=device)
-
-test_text = {
-        'text': 'A panda eating a burger and french fries on a rock.',
-    }
-output_video_path = p(test_text, device=device, output_video='./output.mp4')[OutputKeys.OUTPUT_VIDEO]
-print('output_video_path:', output_video_path)

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)

patch.ascend/device.py.patch

@@ -1,2 +0,0 @@
-27d26
-<     assert eles[0] in ['cpu', 'cuda', 'gpu'], err_msg

patch.ascend/text_to_video_synthesis_model.py.patch

@@ -1,10 +0,0 @@
-60a61
->         print(f"kwargs: {kwargs}")
-62a64
->         print(f"device: {self.device}")
-129c131
-<             layer='penultimate')
----
->             layer='penultimate', device=self.device)
-224a227
->         print(f"self.device: {self.device}")

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:

run_in_docker.sh

@@ -0,0 +1,5 @@
+#! /usr/bin/env bash
+# cnmon
+image=text2video:v0.1
+device_id=0
+docker run -v `pwd`:/workspace -v /mnt:/mnt --device=/dev/cambricon_dev$device_id:/dev/cambricon_dev0 --device=/dev/cambricon_ctl:/dev/cambricon_ctl $image ./test.sh

run_in_docker_a100.sh

@@ -1,3 +0,0 @@
-#! /usr/bin/env bash
-image=harbor-contest.4pd.io/zhanghao/t2v:a100-0.1
-docker run -it -v /home/zhanghao/workspace:/workspace -v /mnt:/mnt $image bash

run_in_docker_ascend.sh

@@ -1,4 +0,0 @@
-#! /usr/bin/env bash
-image=harbor-contest.4pd.io/zhanghao/t2v:ascend-0.1
-device=0
-docker run -it -v `pwd`:/host -e ASCEND_VISIBLE_DEVICES=$device --device /dev/davinci$device:/dev/davinci0 --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc -v /mnt:/mnt -v /usr/local/dcmi:/usr/local/dcmi -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info -v /etc/ascend_install.info:/etc/ascend_install.info --privileged --entrypoint bash  $image

run_in_docker_mlu370.sh

@@ -1,6 +0,0 @@
-#! /usr/bin/env bash
-# cnmon
-image=harbor-contest.4pd.io/zhanghao/iic:mlu370
-image=harbor-contest.4pd.io/zhanghao/t2v:mlu370-0.1
-device_id=2
-docker run -it -v /root/zhanghao:/workspace -v /mnt:/mnt --device=/dev/cambricon_dev$device_id:/dev/cambricon_dev0 --device=/dev/cambricon_ctl:/dev/cambricon_ctl $image bash

run_in_docker_r200.sh

@@ -1,5 +0,0 @@
-#! /usr/bin/env bash
-# cnmon
-image=harbor-contest.4pd.io/zhanghao/t2v:r200-0.1
-device_id=2
-docker run -it -v /root/zhanghao:/workspace -v /mnt:/mnt --security-opt=seccomp=unconfined --cap-add=SYS_PTRACE --cap-add=SYS_ADMIN --device /dev/fuse --shm-size=32g --ulimit=memlock=-1 --ulimit=nofile=120000 --ulimit=stack=67108864 --device=/dev/xpu$device_id:/dev/xpu0 --device=/dev/xpuctrl:/dev/xpuctrl $image bash