JLNTV_IPD/CogVideo
4
0
Fork 0
mirror of https://github.com/THUDM/CogVideo.git synced 2025-04-06 03:57:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #507 from THUDM/CogVideoX_dev

Browse Source 
diffusers version
This commit is contained in:
Yuxuan.Zhang 2024-11-17 21:54:47 +08:00 committed by GitHub
commit 2fdc59c3ce
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 174 additions and 117 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
.gitignore vendored
View File

@ -7,4 +7,5 @@ logs/
*.DS_Store *.DS_Store
.idea .idea
output* output*
test* test*
venv

29
README.md
View File

@ -22,7 +22,8 @@ Experience the CogVideoX-5B model online at <a href="https://huggingface.co/spac
## Project Updates ## Project Updates
- 🔥🔥 News: ```2024/11/08```: We have released the CogVideoX1.5 model. CogVideoX1.5 is an upgraded version of the open-source model CogVideoX. - 🔥🔥 **News**: ```2024/11/15```: We released the `CogVideoX1.5` model in the diffusers version. Only minor parameter adjustments are needed to continue using previous code.
- 🔥 News: ```2024/11/08```: We have released the CogVideoX1.5 model. CogVideoX1.5 is an upgraded version of the open-source model CogVideoX.
The CogVideoX1.5-5B series supports 10-second videos with higher resolution, and CogVideoX1.5-5B-I2V supports video generation at any resolution. The CogVideoX1.5-5B series supports 10-second videos with higher resolution, and CogVideoX1.5-5B-I2V supports video generation at any resolution.
The SAT code has already been updated, while the diffusers version is still under adaptation. Download the SAT version code [here](https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT). The SAT code has already been updated, while the diffusers version is still under adaptation. Download the SAT version code [here](https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT).
- 🔥 **News**: ```2024/10/13```: A more cost-effective fine-tuning framework for `CogVideoX-5B` that works with a single - 🔥 **News**: ```2024/10/13```: A more cost-effective fine-tuning framework for `CogVideoX-5B` that works with a single
@ -43,11 +44,11 @@ The SAT code has already been updated, while the diffusers version is still unde
model [CogVLM2-Caption](https://huggingface.co/THUDM/cogvlm2-llama3-caption), used in the training process of model [CogVLM2-Caption](https://huggingface.co/THUDM/cogvlm2-llama3-caption), used in the training process of
CogVideoX to convert video data into text descriptions, has been open-sourced. Welcome to download and use it. CogVideoX to convert video data into text descriptions, has been open-sourced. Welcome to download and use it.
- 🔥 ```2024/8/27```: We have open-sourced a larger model in the CogVideoX series, **CogVideoX-5B**. We have - 🔥 ```2024/8/27```: We have open-sourced a larger model in the CogVideoX series, **CogVideoX-5B**. We have
significantly optimized the model's inference performance, greatly lowering the inference threshold. You can run * significantly optimized the model's inference performance, greatly lowering the inference threshold.
*CogVideoX-2B** on older GPUs like `GTX 1080TI`, and **CogVideoX-5B** on desktop GPUs like `RTX 3060`. Please strictly You can run **CogVideoX-2B** on older GPUs like `GTX 1080TI`, and **CogVideoX-5B** on desktop GPUs like `RTX 3060`. Please strictly
follow the [requirements](requirements.txt) to update and install dependencies, and refer follow the [requirements](requirements.txt) to update and install dependencies, and refer
to [cli_demo](inference/cli_demo.py) for inference code. Additionally, the open-source license for the **CogVideoX-2B to [cli_demo](inference/cli_demo.py) for inference code. Additionally, the open-source license for
** model has been changed to the **Apache 2.0 License**. the **CogVideoX-2B** model has been changed to the **Apache 2.0 License**.
- 🔥 ```2024/8/6```: We have open-sourced **3D Causal VAE**, used for **CogVideoX-2B**, which can reconstruct videos with - 🔥 ```2024/8/6```: We have open-sourced **3D Causal VAE**, used for **CogVideoX-2B**, which can reconstruct videos with
almost no loss. almost no loss.
- 🔥 ```2024/8/6```: We have open-sourced the first model of the CogVideoX series video generation models, **CogVideoX-2B - 🔥 ```2024/8/6```: We have open-sourced the first model of the CogVideoX series video generation models, **CogVideoX-2B
@ -193,19 +194,19 @@ models we currently offer, along with their foundational information.
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">Inference Precision</td> <td style="text-align: center;">Inference Precision</td>
<td colspan="2" style="text-align: center;"><b>BF16</b></td> <td colspan="2" style="text-align: center;"><b>BF16 (Recommended)</b>, FP16, FP32, FP8*, INT8, Not supported: INT4</td>
<td style="text-align: center;"><b>FP16*(Recommended)</b>, BF16, FP32, FP8*, INT8, Not supported: INT4</td> <td style="text-align: center;"><b>FP16*(Recommended)</b>, BF16, FP32, FP8*, INT8, Not supported: INT4</td>
<td colspan="2" style="text-align: center;"><b>BF16 (Recommended)</b>, FP16, FP32, FP8*, INT8, Not supported: INT4</td> <td colspan="2" style="text-align: center;"><b>BF16 (Recommended)</b>, FP16, FP32, FP8*, INT8, Not supported: INT4</td>
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">Single GPU Memory Usage<br></td> <td style="text-align: center;">Single GPU Memory Usage<br></td>
<td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 66GB <br></td> <td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 76GB <br><b>diffusers BF16: from 10GB*</b><br><b>diffusers INT8(torchao): from 7GB*</b></td>
<td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> FP16: 18GB <br><b>diffusers FP16: 4GB minimum* </b><br><b>diffusers INT8 (torchao): 3.6GB minimum*</b></td> <td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> FP16: 18GB <br><b>diffusers FP16: 4GB minimum* </b><br><b>diffusers INT8 (torchao): 3.6GB minimum*</b></td>
<td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 26GB <br><b>diffusers BF16 : 5GB minimum* </b><br><b>diffusers INT8 (torchao): 4.4GB minimum* </b></td> <td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 26GB <br><b>diffusers BF16 : 5GB minimum* </b><br><b>diffusers INT8 (torchao): 4.4GB minimum* </b></td>
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">Multi-GPU Memory Usage</td> <td style="text-align: center;">Multi-GPU Memory Usage</td>
<td colspan="2" style="text-align: center;"><b>Not Supported</b><br></td> <td colspan="2" style="text-align: center;"><b>BF16: 24GB* using diffusers</b><br></td>
<td style="text-align: center;"><b>FP16: 10GB* using diffusers</b><br></td> <td style="text-align: center;"><b>FP16: 10GB* using diffusers</b><br></td>
<td colspan="2" style="text-align: center;"><b>BF16: 15GB* using diffusers</b><br></td> <td colspan="2" style="text-align: center;"><b>BF16: 15GB* using diffusers</b><br></td>
</tr> </tr>
@ -243,7 +244,8 @@ models we currently offer, along with their foundational information.
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">Download Link (Diffusers)</td> <td style="text-align: center;">Download Link (Diffusers)</td>
<td colspan="2" style="text-align: center;"> Coming Soon </td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX1.5-5B">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX1.5-5B">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX1.5-5B">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX1.5-5B-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX1.5-5B-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX1.5-5B-I2V">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-2b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-2b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-2b">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-2b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-2b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-2b">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b-I2V">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b-I2V">🟣 WiseModel</a></td>
@ -278,15 +280,12 @@ pipe.vae.enable_tiling()
used to quantize the text encoder, transformer, and VAE modules to reduce the memory requirements of CogVideoX. This used to quantize the text encoder, transformer, and VAE modules to reduce the memory requirements of CogVideoX. This
allows the model to run on free T4 Colabs or GPUs with smaller memory! Also, note that TorchAO quantization is fully allows the model to run on free T4 Colabs or GPUs with smaller memory! Also, note that TorchAO quantization is fully
compatible with `torch.compile`, which can significantly improve inference speed. FP8 precision must be used on compatible with `torch.compile`, which can significantly improve inference speed. FP8 precision must be used on
devices with NVIDIA H100 and above, requiring source installation of `torch`, `torchao`, `diffusers`, and `accelerate` devices with NVIDIA H100 and above, requiring source installation of `torch`, `torchao` Python packages. CUDA 12.4 is recommended.
Python packages. CUDA 12.4 is recommended.
+ The inference speed tests also used the above memory optimization scheme. Without memory optimization, inference speed + The inference speed tests also used the above memory optimization scheme. Without memory optimization, inference speed
increases by about 10%. Only the `diffusers` version of the model supports quantization. increases by about 10%. Only the `diffusers` version of the model supports quantization.
+ The model only supports English input; other languages can be translated into English for use via large model + The model only supports English input; other languages can be translated into English for use via large model
refinement. refinement.
+ The memory usage of model fine-tuning is tested in an `8 * H100` environment, and the program automatically
uses `Zero 2` optimization. If a specific number of GPUs is marked in the table, that number or more GPUs must be used
for fine-tuning.
## Friendly Links ## Friendly Links
@ -319,6 +318,8 @@ works have already been adapted for CogVideoX, and we invite everyone to use the
+ [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio): DiffSynth Studio is a diffusion engine. It has + [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio): DiffSynth Studio is a diffusion engine. It has
restructured the architecture, including text encoders, UNet, VAE, etc., enhancing computational performance while restructured the architecture, including text encoders, UNet, VAE, etc., enhancing computational performance while
maintaining compatibility with open-source community models. The framework has been adapted for CogVideoX. maintaining compatibility with open-source community models. The framework has been adapted for CogVideoX.
+ [CogVideoX-Controlnet](https://github.com/TheDenk/cogvideox-controlnet): A simple ControlNet module code that includes the CogVideoX model.
+ [VideoTuna](https://github.com/VideoVerses/VideoTuna): VideoTuna is the first repo that integrates multiple AI video generation models for text-to-video, image-to-video, text-to-image generation.
## Project Structure ## Project Structure

20
README_ja.md
View File

@ -22,7 +22,8 @@
## 更新とニュース ## 更新とニュース
- 🔥🔥 ニュース: ```2024/11/08```: `CogVideoX1.5` モデルをリリースしました。CogVideoX1.5 は CogVideoX オープンソースモデルのアップグレードバージョンです。 - 🔥🔥 **ニュース**: ```2024/11/15```: `CogVideoX1.5`モデルのdiffusersバージョンをリリースしました。わずかなパラメータ調整で以前のコードをそのまま利用可能です。
- 🔥 **ニュース**: ```2024/11/08```: `CogVideoX1.5` モデルをリリースしました。CogVideoX1.5 は CogVideoX オープンソースモデルのアップグレードバージョンです。
CogVideoX1.5-5B シリーズモデルは、10秒 長の動画とより高い解像度をサポートしており、`CogVideoX1.5-5B-I2V` は任意の解像度での動画生成に対応しています。 CogVideoX1.5-5B シリーズモデルは、10秒 長の動画とより高い解像度をサポートしており、`CogVideoX1.5-5B-I2V` は任意の解像度での動画生成に対応しています。
SAT コードはすでに更新されており、`diffusers` バージョンは現在適応中です。 SAT コードはすでに更新されており、`diffusers` バージョンは現在適応中です。
SAT バージョンのコードは [こちら](https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT) からダウンロードできます。 SAT バージョンのコードは [こちら](https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT) からダウンロードできます。
@ -185,19 +186,19 @@ CogVideoXは、[清影](https://chatglm.cn/video?fr=osm_cogvideox) と同源の
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">推論精度</td> <td style="text-align: center;">推論精度</td>
<td colspan="2" style="text-align: center;"><b>BF16</b></td> <td colspan="2" style="text-align: center;"><b>BF16(推奨)</b>, FP16, FP32，FP8*，INT8，INT4非対応</td>
<td style="text-align: center;"><b>FP16*(推奨)</b>, BF16, FP32，FP8*，INT8，INT4非対応</td> <td style="text-align: center;"><b>FP16*(推奨)</b>, BF16, FP32，FP8*，INT8，INT4非対応</td>
<td colspan="2" style="text-align: center;"><b>BF16(推奨)</b>, FP16, FP32，FP8*，INT8，INT4非対応</td> <td colspan="2" style="text-align: center;"><b>BF16(推奨)</b>, FP16, FP32，FP8*，INT8，INT4非対応</td>
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">単一GPUメモリ消費量<br></td> <td style="text-align: center;">単一GPUメモリ消費量<br></td>
<td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 66GB <br></td> <td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 76GB <br><b>diffusers BF16：10GBから*</b><br><b>diffusers INT8(torchao)：7GBから*</b></td>
<td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> FP16: 18GB <br><b>diffusers FP16: 4GB以上* </b><br><b>diffusers INT8(torchao): 3.6GB以上*</b></td> <td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> FP16: 18GB <br><b>diffusers FP16: 4GB以上* </b><br><b>diffusers INT8(torchao): 3.6GB以上*</b></td>
<td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 26GB <br><b>diffusers BF16 : 5GB以上* </b><br><b>diffusers INT8(torchao): 4.4GB以上* </b></td> <td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 26GB <br><b>diffusers BF16 : 5GB以上* </b><br><b>diffusers INT8(torchao): 4.4GB以上* </b></td>
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">複数GPU推論メモリ消費量</td> <td style="text-align: center;">複数GPU推論メモリ消費量</td>
<td colspan="2" style="text-align: center;"><b>非対応</b><br></td> <td colspan="2" style="text-align: center;"><b>BF16: 24GB* using diffusers</b><br></td>
<td style="text-align: center;"><b>FP16: 10GB* diffusers使用</b><br></td> <td style="text-align: center;"><b>FP16: 10GB* diffusers使用</b><br></td>
<td colspan="2" style="text-align: center;"><b>BF16: 15GB* diffusers使用</b><br></td> <td colspan="2" style="text-align: center;"><b>BF16: 15GB* diffusers使用</b><br></td>
</tr> </tr>
@ -235,7 +236,8 @@ CogVideoXは、[清影](https://chatglm.cn/video?fr=osm_cogvideox) と同源の
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">ダウンロードリンク (Diffusers)</td> <td style="text-align: center;">ダウンロードリンク (Diffusers)</td>
<td colspan="2" style="text-align: center;"> 近日公開 </td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX1.5-5B">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX1.5-5B">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX1.5-5B">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX1.5-5B-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX1.5-5B-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX1.5-5B-I2V">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-2b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-2b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-2b">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-2b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-2b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-2b">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b-I2V">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b-I2V">🟣 WiseModel</a></td>
@ -268,13 +270,11 @@ pipe.vae.enable_tiling()
は、CogVideoXのメモリ要件を削減するためにテキストエンコーダ、トランスフォーマ、およびVAEモジュールを量子化するために使用できます。これにより、無料のT4 は、CogVideoXのメモリ要件を削減するためにテキストエンコーダ、トランスフォーマ、およびVAEモジュールを量子化するために使用できます。これにより、無料のT4
Colabやより少ないメモリのGPUでモデルを実行することが可能になります。同様に重要なのは、TorchAOの量子化は`torch.compile` Colabやより少ないメモリのGPUでモデルを実行することが可能になります。同様に重要なのは、TorchAOの量子化は`torch.compile`
と完全に互換性があり、推論速度を大幅に向上させることができる点です。`NVIDIA H100`およびそれ以上のデバイスでは`FP8` と完全に互換性があり、推論速度を大幅に向上させることができる点です。`NVIDIA H100`およびそれ以上のデバイスでは`FP8`
精度を使用する必要があります。これには、`torch`、`torchao`、`diffusers`、`accelerate` 精度を使用する必要があります。これには、`torch`、`torchao` Pythonパッケージのソースコードからのインストールが必要です。`CUDA 12.4`の使用をお勧めします。
Pythonパッケージのソースコードからのインストールが必要です。`CUDA 12.4`の使用をお勧めします。
+ 推論速度テストも同様に、上記のメモリ最適化方法を使用しています。メモリ最適化を使用しない場合、推論速度は約10％向上します。 + 推論速度テストも同様に、上記のメモリ最適化方法を使用しています。メモリ最適化を使用しない場合、推論速度は約10％向上します。
`diffusers`バージョンのモデルのみが量子化をサポートしています。 `diffusers`バージョンのモデルのみが量子化をサポートしています。
+ モデルは英語入力のみをサポートしており、他の言語は大規模モデルの改善を通じて英語に翻訳できます。 + モデルは英語入力のみをサポートしており、他の言語は大規模モデルの改善を通じて英語に翻訳できます。
+ モデルのファインチューニングに使用されるメモリは`8 * H100`環境でテストされています。プログラムは自動的に`Zero 2`
最適化を使用しています。表に具体的なGPU数が記載されている場合、ファインチューニングにはその数以上のGPUが必要です。
## 友好的リンク ## 友好的リンク
@ -299,6 +299,8 @@ pipe.vae.enable_tiling()
キーフレーム補間生成において、より大きな柔軟性を提供することを目的とした、CogVideoX構造を基にした修正版のパイプライン。 キーフレーム補間生成において、より大きな柔軟性を提供することを目的とした、CogVideoX構造を基にした修正版のパイプライン。
+ [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio): DiffSynth + [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio): DiffSynth
Studioは、拡散エンジンです。テキストエンコーダー、UNet、VAEなどを含むアーキテクチャを再構築し、オープンソースコミュニティモデルとの互換性を維持しつつ、計算性能を向上させました。このフレームワークはCogVideoXに適応しています。 Studioは、拡散エンジンです。テキストエンコーダー、UNet、VAEなどを含むアーキテクチャを再構築し、オープンソースコミュニティモデルとの互換性を維持しつつ、計算性能を向上させました。このフレームワークはCogVideoXに適応しています。
+ [CogVideoX-Controlnet](https://github.com/TheDenk/cogvideox-controlnet): CogVideoXモデルを含むシンプルなControlNetモジュールのコード。
+ [VideoTuna](https://github.com/VideoVerses/VideoTuna): VideoTuna は、テキストからビデオ、画像からビデオ、テキストから画像生成のための複数のAIビデオ生成モデルを統合した最初のリポジトリです。
## プロジェクト構造 ## プロジェクト構造

22
README_zh.md
View File

@ -22,7 +22,8 @@
## 项目更新 ## 项目更新
- 🔥🔥 **News**: ```2024/11/08```: 我们发布 `CogVideoX1.5` 模型。CogVideoX1.5 是 CogVideoX 开源模型的升级版本。 - 🔥🔥 **News**: ```2024/11/15```: 我们发布 `CogVideoX1.5` 模型的diffusers版本，仅需调整部分参数仅可沿用之前的代码。
- 🔥 **News**: ```2024/11/08```: 我们发布 `CogVideoX1.5` 模型。CogVideoX1.5 是 CogVideoX 开源模型的升级版本。
CogVideoX1.5-5B 系列模型支持 **10秒** 长度的视频和更高的分辨率，其中 `CogVideoX1.5-5B-I2V` 支持 **任意分辨率** 的视频生成，SAT代码已经更新。`diffusers`版本还在适配中。SAT版本代码前往 [这里](https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT) 下载。 CogVideoX1.5-5B 系列模型支持 **10秒** 长度的视频和更高的分辨率，其中 `CogVideoX1.5-5B-I2V` 支持 **任意分辨率** 的视频生成，SAT代码已经更新。`diffusers`版本还在适配中。SAT版本代码前往 [这里](https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT) 下载。
- 🔥**News**: ```2024/10/13```: 成本更低，单卡4090可微调 `CogVideoX-5B` - 🔥**News**: ```2024/10/13```: 成本更低，单卡4090可微调 `CogVideoX-5B`
的微调框架[cogvideox-factory](https://github.com/a-r-r-o-w/cogvideox-factory)已经推出，多种分辨率微调，欢迎使用。 的微调框架[cogvideox-factory](https://github.com/a-r-r-o-w/cogvideox-factory)已经推出，多种分辨率微调，欢迎使用。
@ -39,8 +40,7 @@ CogVideoX1.5-5B 系列模型支持 **10秒** 长度的视频和更高的分辨
- 🔥 ```2024/8/27```: 我们开源 CogVideoX 系列更大的模型 **CogVideoX-5B** - 🔥 ```2024/8/27```: 我们开源 CogVideoX 系列更大的模型 **CogVideoX-5B**
。我们大幅度优化了模型的推理性能，推理门槛大幅降低，您可以在 `GTX 1080TI` 等早期显卡运行 **CogVideoX-2B**，在 `RTX 3060` 。我们大幅度优化了模型的推理性能，推理门槛大幅降低，您可以在 `GTX 1080TI` 等早期显卡运行 **CogVideoX-2B**，在 `RTX 3060`
等桌面端甜品卡运行 **CogVideoX-5B** 模型。 请严格按照[要求](requirements.txt) 等桌面端甜品卡运行 **CogVideoX-5B** 模型。 请严格按照[要求](requirements.txt)
更新安装依赖，推理代码请查看 [cli_demo](inference/cli_demo.py)。同时，**CogVideoX-2B** 模型开源协议已经修改为**Apache 2.0 更新安装依赖，推理代码请查看 [cli_demo](inference/cli_demo.py)。同时，**CogVideoX-2B** 模型开源协议已经修改为**Apache 2.0 协议**。
协议**。
- 🔥 ```2024/8/6```: 我们开源 **3D Causal VAE**，用于 **CogVideoX-2B**，可以几乎无损地重构视频。 - 🔥 ```2024/8/6```: 我们开源 **3D Causal VAE**，用于 **CogVideoX-2B**，可以几乎无损地重构视频。
- 🔥 ```2024/8/6```: 我们开源 CogVideoX 系列视频生成模型的第一个模型, **CogVideoX-2B**。 - 🔥 ```2024/8/6```: 我们开源 CogVideoX 系列视频生成模型的第一个模型, **CogVideoX-2B**。
- 🌱 **Source**: ```2022/5/19```: 我们开源了 CogVideo 视频生成模型（现在你可以在 `CogVideo` 分支中看到），这是首个开源的基于 - 🌱 **Source**: ```2022/5/19```: 我们开源了 CogVideo 视频生成模型（现在你可以在 `CogVideo` 分支中看到），这是首个开源的基于
@ -176,19 +176,19 @@ CogVideoX是 [清影](https://chatglm.cn/video?fr=osm_cogvideox) 同源的开源
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">推理精度</td> <td style="text-align: center;">推理精度</td>
<td colspan="2" style="text-align: center;"><b>BF16</b></td> <td colspan="2" style="text-align: center;"><b>BF16(推荐)</b>, FP16, FP32，FP8*，INT8，不支持INT4</td>
<td style="text-align: center;"><b>FP16*(推荐)</b>, BF16, FP32，FP8*，INT8，不支持INT4</td> <td style="text-align: center;"><b>FP16*(推荐)</b>, BF16, FP32，FP8*，INT8，不支持INT4</td>
<td colspan="2" style="text-align: center;"><b>BF16(推荐)</b>, FP16, FP32，FP8*，INT8，不支持INT4</td> <td colspan="2" style="text-align: center;"><b>BF16(推荐)</b>, FP16, FP32，FP8*，INT8，不支持INT4</td>
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">单GPU显存消耗<br></td> <td style="text-align: center;">单GPU显存消耗<br></td>
<td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 66GB <br></td> <td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 76GB <br><b>diffusers BF16 : 10GB起* </b><br><b>diffusers INT8(torchao): 7G起* </b></td>
<td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> FP16: 18GB <br><b>diffusers FP16: 4GB起* </b><br><b>diffusers INT8(torchao): 3.6G起*</b></td> <td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> FP16: 18GB <br><b>diffusers FP16: 4GB起* </b><br><b>diffusers INT8(torchao): 3.6G起*</b></td>
<td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 26GB <br><b>diffusers BF16 : 5GB起* </b><br><b>diffusers INT8(torchao): 4.4G起* </b></td> <td colspan="2" style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 26GB <br><b>diffusers BF16 : 5GB起* </b><br><b>diffusers INT8(torchao): 4.4G起* </b></td>
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">多GPU推理显存消耗</td> <td style="text-align: center;">多GPU推理显存消耗</td>
<td colspan="2" style="text-align: center;"><b>不支持</b><br></td> <td colspan="2" style="text-align: center;"><b>BF16: 24GB* using diffusers</b><br></td>
<td style="text-align: center;"><b>FP16: 10GB* using diffusers</b><br></td> <td style="text-align: center;"><b>FP16: 10GB* using diffusers</b><br></td>
<td colspan="2" style="text-align: center;"><b>BF16: 15GB* using diffusers</b><br></td> <td colspan="2" style="text-align: center;"><b>BF16: 15GB* using diffusers</b><br></td>
</tr> </tr>
@ -226,7 +226,8 @@ CogVideoX是 [清影](https://chatglm.cn/video?fr=osm_cogvideox) 同源的开源
</tr> </tr>
<tr> <tr>
<td style="text-align: center;">下载链接 (Diffusers)</td> <td style="text-align: center;">下载链接 (Diffusers)</td>
<td colspan="2" style="text-align: center;"> 即将推出 </td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX1.5-5B">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX1.5-5B">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX1.5-5B">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX1.5-5B-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX1.5-5B-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX1.5-5B-I2V">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-2b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-2b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-2b">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-2b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-2b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-2b">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b">🟣 WiseModel</a></td>
<td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b-I2V">🟣 WiseModel</a></td> <td style="text-align: center;"><a href="https://huggingface.co/THUDM/CogVideoX-5b-I2V">🤗 HuggingFace</a><br><a href="https://modelscope.cn/models/ZhipuAI/CogVideoX-5b-I2V">🤖 ModelScope</a><br><a href="https://wisemodel.cn/models/ZhipuAI/CogVideoX-5b-I2V">🟣 WiseModel</a></td>
@ -256,11 +257,9 @@ pipe.vae.enable_tiling()
+ [PytorchAO](https://github.com/pytorch/ao) 和 [Optimum-quanto](https://github.com/huggingface/optimum-quanto/) + [PytorchAO](https://github.com/pytorch/ao) 和 [Optimum-quanto](https://github.com/huggingface/optimum-quanto/)
可以用于量化文本编码器、Transformer 和 VAE 模块，以降低 CogVideoX 的内存需求。这使得在免费的 T4 Colab 或更小显存的 GPU 可以用于量化文本编码器、Transformer 和 VAE 模块，以降低 CogVideoX 的内存需求。这使得在免费的 T4 Colab 或更小显存的 GPU
上运行模型成为可能！同样值得注意的是，TorchAO 量化完全兼容 `torch.compile`，这可以显著提高推理速度。在 `NVIDIA H100` 上运行模型成为可能！同样值得注意的是，TorchAO 量化完全兼容 `torch.compile`，这可以显著提高推理速度。在 `NVIDIA H100`
及以上设备上必须使用 `FP8` 精度，这需要源码安装 `torch`、`torchao`、`diffusers` 和 `accelerate` Python 及以上设备上必须使用 `FP8` 精度，这需要源码安装 `torch`、`torchao` Python 包。建议使用 `CUDA 12.4`。
包。建议使用 `CUDA 12.4`。
+ 推理速度测试同样采用了上述显存优化方案，不采用显存优化的情况下，推理速度提升约10%。 只有`diffusers`版本模型支持量化。 + 推理速度测试同样采用了上述显存优化方案，不采用显存优化的情况下，推理速度提升约10%。 只有`diffusers`版本模型支持量化。
+ 模型仅支持英语输入，其他语言可以通过大模型润色时翻译为英语。 + 模型仅支持英语输入，其他语言可以通过大模型润色时翻译为英语。
+ 模型微调所占用的显存是在 `8 * H100` 环境下进行测试，程序已经自动使用`Zero 2` 优化。表格中若有标注具体GPU数量则必须使用大于等于该数量的GPU进行微调。
## 友情链接 ## 友情链接
@ -280,6 +279,9 @@ pipe.vae.enable_tiling()
+ [CogVideoX-Interpolation](https://github.com/feizc/CogvideX-Interpolation): 基于 CogVideoX 结构修改的管道，旨在为关键帧插值生成提供更大的灵活性。 + [CogVideoX-Interpolation](https://github.com/feizc/CogvideX-Interpolation): 基于 CogVideoX 结构修改的管道，旨在为关键帧插值生成提供更大的灵活性。
+ [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio): DiffSynth 工作室是一款扩散引擎。重构了架构，包括文本编码器、UNet、VAE + [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio): DiffSynth 工作室是一款扩散引擎。重构了架构，包括文本编码器、UNet、VAE
等，在保持与开源社区模型兼容性的同时，提升了计算性能。该框架已经适配 CogVideoX。 等，在保持与开源社区模型兼容性的同时，提升了计算性能。该框架已经适配 CogVideoX。
+ [CogVideoX-Controlnet](https://github.com/TheDenk/cogvideox-controlnet): 一个包含 CogvideoX 模型的简单 Controlnet 模块的代码。
+ [VideoTuna](https://github.com/VideoVerses/VideoTuna)：VideoTuna 是首个集成多种 AI 视频生成模型的仓库，支持文本转视频、图像转视频、文本转图像生成。
## 完整项目代码结构 ## 完整项目代码结构

62
inference/cli_demo.py
View File

@ -3,15 +3,15 @@ This script demonstrates how to generate a video using the CogVideoX model with
The script supports different types of video generation, including text-to-video (t2v), image-to-video (i2v), The script supports different types of video generation, including text-to-video (t2v), image-to-video (i2v),
and video-to-video (v2v), depending on the input data and different weight. and video-to-video (v2v), depending on the input data and different weight.
- text-to-video: THUDM/CogVideoX-5b or THUDM/CogVideoX-2b - text-to-video: THUDM/CogVideoX-5b, THUDM/CogVideoX-2b or THUDM/CogVideoX1.5-5b
- video-to-video: THUDM/CogVideoX-5b or THUDM/CogVideoX-2b - video-to-video: THUDM/CogVideoX-5b, THUDM/CogVideoX-2b or THUDM/CogVideoX1.5-5b
- image-to-video: THUDM/CogVideoX-5b-I2V - image-to-video: THUDM/CogVideoX-5b-I2V or THUDM/CogVideoX1.5-5b-I2V
Running the Script: Running the Script:
To run the script, use the following command with appropriate arguments: To run the script, use the following command with appropriate arguments:
```bash ```bash
$ python cli_demo.py --prompt "A girl riding a bike." --model_path THUDM/CogVideoX-5b --generate_type "t2v" $ python cli_demo.py --prompt "A girl riding a bike." --model_path THUDM/CogVideoX1.5-5b --generate_type "t2v"
``` ```
Additional options are available to specify the model path, guidance scale, number of inference steps, video generation type, and output paths. Additional options are available to specify the model path, guidance scale, number of inference steps, video generation type, and output paths.
@ -23,7 +23,6 @@ from typing import Literal
import torch import torch
from diffusers import ( from diffusers import (
CogVideoXPipeline, CogVideoXPipeline,
CogVideoXDDIMScheduler,
CogVideoXDPMScheduler, CogVideoXDPMScheduler,
CogVideoXImageToVideoPipeline, CogVideoXImageToVideoPipeline,
CogVideoXVideoToVideoPipeline, CogVideoXVideoToVideoPipeline,
@ -37,6 +36,9 @@ def generate_video(
model_path: str, model_path: str,
lora_path: str = None, lora_path: str = None,
lora_rank: int = 128, lora_rank: int = 128,
num_frames: int = 81,
width: int = 1360,
height: int = 768,
output_path: str = "./output.mp4", output_path: str = "./output.mp4",
image_or_video_path: str = "", image_or_video_path: str = "",
num_inference_steps: int = 50, num_inference_steps: int = 50,
@ -45,6 +47,7 @@ def generate_video(
dtype: torch.dtype = torch.bfloat16, dtype: torch.dtype = torch.bfloat16,
generate_type: str = Literal["t2v", "i2v", "v2v"], # i2v: image to video, v2v: video to video generate_type: str = Literal["t2v", "i2v", "v2v"], # i2v: image to video, v2v: video to video
seed: int = 42, seed: int = 42,
fps: int = 8,
): ):
""" """
Generates a video based on the given prompt and saves it to the specified path. Generates a video based on the given prompt and saves it to the specified path.
@ -56,11 +59,15 @@ def generate_video(
- lora_rank (int): The rank of the LoRA weights. - lora_rank (int): The rank of the LoRA weights.
- output_path (str): The path where the generated video will be saved. - output_path (str): The path where the generated video will be saved.
- num_inference_steps (int): Number of steps for the inference process. More steps can result in better quality. - num_inference_steps (int): Number of steps for the inference process. More steps can result in better quality.
- num_frames (int): Number of frames to generate. CogVideoX1.0 generates 49 frames for 6 seconds at 8 fps, while CogVideoX1.5 produces either 81 or 161 frames, corresponding to 5 seconds or 10 seconds at 16 fps.
- width (int): The width of the generated video, applicable only for CogVideoX1.5-5B-I2V
- height (int): The height of the generated video, applicable only for CogVideoX1.5-5B-I2V
- guidance_scale (float): The scale for classifier-free guidance. Higher values can lead to better alignment with the prompt. - guidance_scale (float): The scale for classifier-free guidance. Higher values can lead to better alignment with the prompt.
- num_videos_per_prompt (int): Number of videos to generate per prompt. - num_videos_per_prompt (int): Number of videos to generate per prompt.
- dtype (torch.dtype): The data type for computation (default is torch.bfloat16). - dtype (torch.dtype): The data type for computation (default is torch.bfloat16).
- generate_type (str): The type of video generation (e.g., 't2v', 'i2v', 'v2v').· - generate_type (str): The type of video generation (e.g., 't2v', 'i2v', 'v2v').·
- seed (int): The seed for reproducibility. - seed (int): The seed for reproducibility.
- fps (int): The frames per second for the generated video.
""" """
# 1. Load the pre-trained CogVideoX pipeline with the specified precision (bfloat16). # 1. Load the pre-trained CogVideoX pipeline with the specified precision (bfloat16).
@ -97,23 +104,23 @@ def generate_video(
# and enable to("cuda") # and enable to("cuda")
# pipe.to("cuda") # pipe.to("cuda")
pipe.enable_sequential_cpu_offload() pipe.enable_sequential_cpu_offload()
pipe.vae.enable_slicing() pipe.vae.enable_slicing()
pipe.vae.enable_tiling() pipe.vae.enable_tiling()
# 4. Generate the video frames based on the prompt. # 4. Generate the video frames based on the prompt.
# `num_frames` is the Number of frames to generate. # `num_frames` is the Number of frames to generate.
# This is the default value for 6 seconds video and 8 fps and will plus 1 frame for the first frame and 49 frames.
if generate_type == "i2v": if generate_type == "i2v":
video_generate = pipe( video_generate = pipe(
height=height,
width=width,
prompt=prompt, prompt=prompt,
image=image, # The path of the image to be used as the background of the video image=image,
# The path of the image, the resolution of video will be the same as the image for CogVideoX1.5-5B-I2V, otherwise it will be 720 * 480
num_videos_per_prompt=num_videos_per_prompt, # Number of videos to generate per prompt num_videos_per_prompt=num_videos_per_prompt, # Number of videos to generate per prompt
num_inference_steps=num_inference_steps, # Number of inference steps num_inference_steps=num_inference_steps, # Number of inference steps
num_frames=49, # Number of frames to generate，changed to 49 for diffusers version `0.30.3` and after. num_frames=num_frames, # Number of frames to generate
use_dynamic_cfg=True, # This id used for DPM Sechduler, for DDIM scheduler, it should be False use_dynamic_cfg=True, # This id used for DPM scheduler, for DDIM scheduler, it should be False
guidance_scale=guidance_scale, guidance_scale=guidance_scale,
generator=torch.Generator().manual_seed(seed), # Set the seed for reproducibility generator=torch.Generator().manual_seed(seed), # Set the seed for reproducibility
).frames[0] ).frames[0]
@ -122,7 +129,7 @@ def generate_video(
prompt=prompt, prompt=prompt,
num_videos_per_prompt=num_videos_per_prompt, num_videos_per_prompt=num_videos_per_prompt,
num_inference_steps=num_inference_steps, num_inference_steps=num_inference_steps,
num_frames=49, num_frames=num_frames,
use_dynamic_cfg=True, use_dynamic_cfg=True,
guidance_scale=guidance_scale, guidance_scale=guidance_scale,
generator=torch.Generator().manual_seed(seed), generator=torch.Generator().manual_seed(seed),
@ -133,13 +140,12 @@ def generate_video(
video=video, # The path of the video to be used as the background of the video video=video, # The path of the video to be used as the background of the video
num_videos_per_prompt=num_videos_per_prompt, num_videos_per_prompt=num_videos_per_prompt,
num_inference_steps=num_inference_steps, num_inference_steps=num_inference_steps,
# num_frames=49, num_frames=num_frames,
use_dynamic_cfg=True, use_dynamic_cfg=True,
guidance_scale=guidance_scale, guidance_scale=guidance_scale,
generator=torch.Generator().manual_seed(seed), # Set the seed for reproducibility generator=torch.Generator().manual_seed(seed), # Set the seed for reproducibility
).frames[0] ).frames[0]
# 5. Export the generated frames to a video file. fps must be 8 for original video. export_to_video(video_generate, output_path, fps=fps)
export_to_video(video_generate, output_path, fps=8)
if __name__ == "__main__": if __name__ == "__main__":
@ -152,24 +158,20 @@ if __name__ == "__main__":
help="The path of the image to be used as the background of the video", help="The path of the image to be used as the background of the video",
) )
parser.add_argument( parser.add_argument(
"--model_path", type=str, default="THUDM/CogVideoX-5b", help="The path of the pre-trained model to be used" "--model_path", type=str, default="THUDM/CogVideoX1.5-5B", help="Path of the pre-trained model use"
) )
parser.add_argument("--lora_path", type=str, default=None, help="The path of the LoRA weights to be used") parser.add_argument("--lora_path", type=str, default=None, help="The path of the LoRA weights to be used")
parser.add_argument("--lora_rank", type=int, default=128, help="The rank of the LoRA weights") parser.add_argument("--lora_rank", type=int, default=128, help="The rank of the LoRA weights")
parser.add_argument( parser.add_argument("--output_path", type=str, default="./output.mp4", help="The path save generated video")
"--output_path", type=str, default="./output.mp4", help="The path where the generated video will be saved"
)
parser.add_argument("--guidance_scale", type=float, default=6.0, help="The scale for classifier-free guidance") parser.add_argument("--guidance_scale", type=float, default=6.0, help="The scale for classifier-free guidance")
parser.add_argument( parser.add_argument("--num_inference_steps", type=int, default=50, help="Inference steps")
"--num_inference_steps", type=int, default=50, help="Number of steps for the inference process" parser.add_argument("--num_frames", type=int, default=81, help="Number of steps for the inference process")
) parser.add_argument("--width", type=int, default=1360, help="Number of steps for the inference process")
parser.add_argument("--height", type=int, default=768, help="Number of steps for the inference process")
parser.add_argument("--fps", type=int, default=16, help="Number of steps for the inference process")
parser.add_argument("--num_videos_per_prompt", type=int, default=1, help="Number of videos to generate per prompt") parser.add_argument("--num_videos_per_prompt", type=int, default=1, help="Number of videos to generate per prompt")
parser.add_argument( parser.add_argument("--generate_type", type=str, default="t2v", help="The type of video generation")
"--generate_type", type=str, default="t2v", help="The type of video generation (e.g., 't2v', 'i2v', 'v2v')" parser.add_argument("--dtype", type=str, default="bfloat16", help="The data type for computation")
)
parser.add_argument(
"--dtype", type=str, default="bfloat16", help="The data type for computation (e.g., 'float16' or 'bfloat16')"
)
parser.add_argument("--seed", type=int, default=42, help="The seed for reproducibility") parser.add_argument("--seed", type=int, default=42, help="The seed for reproducibility")
args = parser.parse_args() args = parser.parse_args()
@ -180,6 +182,9 @@ if __name__ == "__main__":
lora_path=args.lora_path, lora_path=args.lora_path,
lora_rank=args.lora_rank, lora_rank=args.lora_rank,
output_path=args.output_path, output_path=args.output_path,
num_frames=args.num_frames,
width=args.width,
height=args.height,
image_or_video_path=args.image_or_video_path, image_or_video_path=args.image_or_video_path,
num_inference_steps=args.num_inference_steps, num_inference_steps=args.num_inference_steps,
guidance_scale=args.guidance_scale, guidance_scale=args.guidance_scale,
@ -187,4 +192,5 @@ if __name__ == "__main__":
dtype=dtype, dtype=dtype,
generate_type=args.generate_type, generate_type=args.generate_type,
seed=args.seed, seed=args.seed,
fps=args.fps,
) )

53
inference/cli_demo_quantization.py
View File

@ -3,7 +3,7 @@ This script demonstrates how to generate a video from a text prompt using CogVid
Note: Note:
Must install the `torchao`，`torch`,`diffusers`,`accelerate` library FROM SOURCE to use the quantization feature. Must install the `torchao`，`torch` library FROM SOURCE to use the quantization feature.
Only NVIDIA GPUs like H100 or higher are supported om FP-8 quantization. Only NVIDIA GPUs like H100 or higher are supported om FP-8 quantization.
ALL quantization schemes must use with NVIDIA GPUs. ALL quantization schemes must use with NVIDIA GPUs.
@ -51,6 +51,9 @@ def generate_video(
num_videos_per_prompt: int = 1, num_videos_per_prompt: int = 1,
quantization_scheme: str = "fp8", quantization_scheme: str = "fp8",
dtype: torch.dtype = torch.bfloat16, dtype: torch.dtype = torch.bfloat16,
num_frames: int = 81,
fps: int = 8,
seed: int = 42,
): ):
""" """
Generates a video based on the given prompt and saves it to the specified path. Generates a video based on the given prompt and saves it to the specified path.
@ -65,7 +68,6 @@ def generate_video(
- quantization_scheme (str): The quantization scheme to use ('int8', 'fp8'). - quantization_scheme (str): The quantization scheme to use ('int8', 'fp8').
- dtype (torch.dtype): The data type for computation (default is torch.bfloat16). - dtype (torch.dtype): The data type for computation (default is torch.bfloat16).
""" """
text_encoder = T5EncoderModel.from_pretrained(model_path, subfolder="text_encoder", torch_dtype=dtype) text_encoder = T5EncoderModel.from_pretrained(model_path, subfolder="text_encoder", torch_dtype=dtype)
text_encoder = quantize_model(part=text_encoder, quantization_scheme=quantization_scheme) text_encoder = quantize_model(part=text_encoder, quantization_scheme=quantization_scheme)
transformer = CogVideoXTransformer3DModel.from_pretrained(model_path, subfolder="transformer", torch_dtype=dtype) transformer = CogVideoXTransformer3DModel.from_pretrained(model_path, subfolder="transformer", torch_dtype=dtype)
@ -80,54 +82,38 @@ def generate_video(
torch_dtype=dtype, torch_dtype=dtype,
) )
pipe.scheduler = CogVideoXDPMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing") pipe.scheduler = CogVideoXDPMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
# Using with compile will run faster. First time infer will cost ~30min to compile.
# pipe.transformer.to(memory_format=torch.channels_last)
# for FP8 should remove pipe.enable_model_cpu_offload()
pipe.enable_model_cpu_offload() pipe.enable_model_cpu_offload()
# This is not for FP8 and INT8 and should remove this line
# pipe.enable_sequential_cpu_offload()
pipe.vae.enable_slicing() pipe.vae.enable_slicing()
pipe.vae.enable_tiling() pipe.vae.enable_tiling()
video = pipe( video = pipe(
prompt=prompt, prompt=prompt,
num_videos_per_prompt=num_videos_per_prompt, num_videos_per_prompt=num_videos_per_prompt,
num_inference_steps=num_inference_steps, num_inference_steps=num_inference_steps,
num_frames=49, num_frames=num_frames,
use_dynamic_cfg=True, use_dynamic_cfg=True,
guidance_scale=guidance_scale, guidance_scale=guidance_scale,
generator=torch.Generator(device="cuda").manual_seed(42), generator=torch.Generator(device="cuda").manual_seed(seed),
).frames[0] ).frames[0]
export_to_video(video, output_path, fps=8) export_to_video(video, output_path, fps=fps)
if __name__ == "__main__": if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Generate a video from a text prompt using CogVideoX") parser = argparse.ArgumentParser(description="Generate a video from a text prompt using CogVideoX")
parser.add_argument("--prompt", type=str, required=True, help="The description of the video to be generated") parser.add_argument("--prompt", type=str, required=True, help="The description of the video to be generated")
parser.add_argument("--model_path", type=str, default="THUDM/CogVideoX-5b", help="Path of the pre-trained model")
parser.add_argument("--output_path", type=str, default="./output.mp4", help="Path to save generated video")
parser.add_argument("--num_inference_steps", type=int, default=50, help="Inference steps")
parser.add_argument("--guidance_scale", type=float, default=6.0, help="Classifier-free guidance scale")
parser.add_argument("--num_videos_per_prompt", type=int, default=1, help="Videos to generate per prompt")
parser.add_argument("--dtype", type=str, default="bfloat16", help="Data type (e.g., 'float16', 'bfloat16')")
parser.add_argument( parser.add_argument(
"--model_path", type=str, default="THUDM/CogVideoX-5b", help="The path of the pre-trained model to be used" "--quantization_scheme", type=str, default="fp8", choices=["int8", "fp8"], help="Quantization scheme"
)
parser.add_argument(
"--output_path", type=str, default="./output.mp4", help="The path where the generated video will be saved"
)
parser.add_argument(
"--num_inference_steps", type=int, default=50, help="Number of steps for the inference process"
)
parser.add_argument("--guidance_scale", type=float, default=6.0, help="The scale for classifier-free guidance")
parser.add_argument("--num_videos_per_prompt", type=int, default=1, help="Number of videos to generate per prompt")
parser.add_argument(
"--dtype", type=str, default="bfloat16", help="The data type for computation (e.g., 'float16', 'bfloat16')"
)
parser.add_argument(
"--quantization_scheme",
type=str,
default="bf16",
choices=["int8", "fp8"],
help="The quantization scheme to use (int8, fp8)",
) )
parser.add_argument("--num_frames", type=int, default=81, help="Number of frames in the video")
parser.add_argument("--fps", type=int, default=16, help="Frames per second for output video")
parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
args = parser.parse_args() args = parser.parse_args()
dtype = torch.float16 if args.dtype == "float16" else torch.bfloat16 dtype = torch.float16 if args.dtype == "float16" else torch.bfloat16
@ -140,4 +126,7 @@ if __name__ == "__main__":
num_videos_per_prompt=args.num_videos_per_prompt, num_videos_per_prompt=args.num_videos_per_prompt,
quantization_scheme=args.quantization_scheme, quantization_scheme=args.quantization_scheme,
dtype=dtype, dtype=dtype,
num_frames=args.num_frames,
fps=args.fps,
seed=args.seed,
) )

10
requirements.txt
View File

@ -1,14 +1,14 @@
diffusers>=0.31.0 diffusers>=0.31.0
accelerate>=1.0.1 accelerate>=1.1.1
transformers>=4.46.1 transformers>=4.46.2
numpy==1.26.0 numpy==1.26.0
torch>=2.5.0 torch>=2.5.0
torchvision>=0.20.0 torchvision>=0.20.0
sentencepiece>=0.2.0 sentencepiece>=0.2.0
SwissArmyTransformer>=0.4.12 SwissArmyTransformer>=0.4.12
gradio>=5.4.0 gradio>=5.5.0
imageio>=2.35.1 imageio>=2.35.1
imageio-ffmpeg>=0.5.1 imageio-ffmpeg>=0.5.1
openai>=1.53.0 openai>=1.54.0
moviepy>=1.0.3 moviepy>=1.0.3
scikit-video>=1.1.11 scikit-video>=1.1.11

92
tools/convert_weight_sat2hf.py
View File

@ -92,6 +92,8 @@ TRANSFORMER_KEYS_RENAME_DICT = {
"post_attn1_layernorm": "norm2.norm", "post_attn1_layernorm": "norm2.norm",
"time_embed.0": "time_embedding.linear_1", "time_embed.0": "time_embedding.linear_1",
"time_embed.2": "time_embedding.linear_2", "time_embed.2": "time_embedding.linear_2",
"ofs_embed.0": "ofs_embedding.linear_1",
"ofs_embed.2": "ofs_embedding.linear_2",
"mixins.patch_embed": "patch_embed", "mixins.patch_embed": "patch_embed",
"mixins.final_layer.norm_final": "norm_out.norm", "mixins.final_layer.norm_final": "norm_out.norm",
"mixins.final_layer.linear": "proj_out", "mixins.final_layer.linear": "proj_out",
@ -146,12 +148,13 @@ def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key:
def convert_transformer( def convert_transformer(
ckpt_path: str, ckpt_path: str,
num_layers: int, num_layers: int,
num_attention_heads: int, num_attention_heads: int,
use_rotary_positional_embeddings: bool, use_rotary_positional_embeddings: bool,
i2v: bool, i2v: bool,
dtype: torch.dtype, dtype: torch.dtype,
init_kwargs: Dict[str, Any],
): ):
PREFIX_KEY = "model.diffusion_model." PREFIX_KEY = "model.diffusion_model."
@ -161,11 +164,13 @@ def convert_transformer(
num_layers=num_layers, num_layers=num_layers,
num_attention_heads=num_attention_heads, num_attention_heads=num_attention_heads,
use_rotary_positional_embeddings=use_rotary_positional_embeddings, use_rotary_positional_embeddings=use_rotary_positional_embeddings,
use_learned_positional_embeddings=i2v, ofs_embed_dim=512 if (i2v and init_kwargs["patch_size_t"] is not None) else None, # CogVideoX1.5-5B-I2V
use_learned_positional_embeddings=i2v and init_kwargs["patch_size_t"] is None, # CogVideoX-5B-I2V
**init_kwargs,
).to(dtype=dtype) ).to(dtype=dtype)
for key in list(original_state_dict.keys()): for key in list(original_state_dict.keys()):
new_key = key[len(PREFIX_KEY):] new_key = key[len(PREFIX_KEY) :]
for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items(): for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
new_key = new_key.replace(replace_key, rename_key) new_key = new_key.replace(replace_key, rename_key)
update_state_dict_inplace(original_state_dict, key, new_key) update_state_dict_inplace(original_state_dict, key, new_key)
@ -175,13 +180,18 @@ def convert_transformer(
if special_key not in key: if special_key not in key:
continue continue
handler_fn_inplace(key, original_state_dict) handler_fn_inplace(key, original_state_dict)
transformer.load_state_dict(original_state_dict, strict=True) transformer.load_state_dict(original_state_dict, strict=True)
return transformer return transformer
def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype): def convert_vae(ckpt_path: str, scaling_factor: float, version: str, dtype: torch.dtype):
init_kwargs = {"scaling_factor": scaling_factor}
if version == "1.5":
init_kwargs.update({"invert_scale_latents": True})
original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True)) original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
vae = AutoencoderKLCogVideoX(scaling_factor=scaling_factor).to(dtype=dtype) vae = AutoencoderKLCogVideoX(**init_kwargs).to(dtype=dtype)
for key in list(original_state_dict.keys()): for key in list(original_state_dict.keys()):
new_key = key[:] new_key = key[:]
@ -199,6 +209,34 @@ def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
return vae return vae
def get_transformer_init_kwargs(version: str):
if version == "1.0":
vae_scale_factor_spatial = 8
init_kwargs = {
"patch_size": 2,
"patch_size_t": None,
"patch_bias": True,
"sample_height": 480 // vae_scale_factor_spatial,
"sample_width": 720 // vae_scale_factor_spatial,
"sample_frames": 49,
}
elif version == "1.5":
vae_scale_factor_spatial = 8
init_kwargs = {
"patch_size": 2,
"patch_size_t": 2,
"patch_bias": False,
"sample_height": 768 // vae_scale_factor_spatial,
"sample_width": 1360 // vae_scale_factor_spatial,
"sample_frames": 81,
}
else:
raise ValueError("Unsupported version of CogVideoX.")
return init_kwargs
def get_args(): def get_args():
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument( parser.add_argument(
@ -214,6 +252,12 @@ def get_args():
parser.add_argument( parser.add_argument(
"--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory" "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
) )
parser.add_argument(
"--typecast_text_encoder",
action="store_true",
default=False,
help="Whether or not to apply fp16/bf16 precision to text_encoder",
)
# For CogVideoX-2B, num_layers is 30. For 5B, it is 42 # For CogVideoX-2B, num_layers is 30. For 5B, it is 42
parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks") parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks")
# For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48 # For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48
@ -226,7 +270,18 @@ def get_args():
parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE") parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
# For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0 # For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE") parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
parser.add_argument("--i2v", action="store_true", default=False, help="Whether to save the model weights in fp16") parser.add_argument(
"--i2v",
action="store_true",
default=False,
help="Whether the model to be converted is the Image-to-Video version of CogVideoX.",
)
parser.add_argument(
"--version",
choices=["1.0", "1.5"],
default="1.0",
help="Which version of CogVideoX to use for initializing default modeling parameters.",
)
return parser.parse_args() return parser.parse_args()
@ -242,6 +297,7 @@ if __name__ == "__main__":
dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32 dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32
if args.transformer_ckpt_path is not None: if args.transformer_ckpt_path is not None:
init_kwargs = get_transformer_init_kwargs(args.version)
transformer = convert_transformer( transformer = convert_transformer(
args.transformer_ckpt_path, args.transformer_ckpt_path,
args.num_layers, args.num_layers,
@ -249,14 +305,19 @@ if __name__ == "__main__":
args.use_rotary_positional_embeddings, args.use_rotary_positional_embeddings,
args.i2v, args.i2v,
dtype, dtype,
init_kwargs,
) )
if args.vae_ckpt_path is not None: if args.vae_ckpt_path is not None:
vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, dtype) # Keep VAE in float32 for better quality
vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, args.version, torch.float32)
text_encoder_id = "/share/official_pretrains/hf_home/t5-v1_1-xxl" text_encoder_id = "google/t5-v1_1-xxl"
tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH) tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir) text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
if args.typecast_text_encoder:
text_encoder = text_encoder.to(dtype=dtype)
# Apparently, the conversion does not work anymore without this :shrug: # Apparently, the conversion does not work anymore without this :shrug:
for param in text_encoder.parameters(): for param in text_encoder.parameters():
param.data = param.data.contiguous() param.data = param.data.contiguous()
@ -288,11 +349,6 @@ if __name__ == "__main__":
scheduler=scheduler, scheduler=scheduler,
) )
if args.fp16:
pipe = pipe.to(dtype=torch.float16)
if args.bf16:
pipe = pipe.to(dtype=torch.bfloat16)
# We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird # We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
# for users to specify variant when the default is not fp32 and they want to run with the correct default (which # for users to specify variant when the default is not fp32 and they want to run with the correct default (which
# is either fp16/bf16 here). # is either fp16/bf16 here).