From 8c2a2fc955cc44aa62ccf3f255c9c67be4484b10 Mon Sep 17 00:00:00 2001 From: Cherrytest Date: Tue, 2 Dec 2025 16:18:39 +0000 Subject: [PATCH] update README (batch 1/1) --- README.md | 53 ++++++++++++++++++++++++++++++++++++++++++---------- README_CN.md | 52 ++++++++++++++++++++++++++++++++++++++++++--------- 2 files changed, 86 insertions(+), 19 deletions(-) diff --git a/README.md b/README.md index cb2cab3..a415fd8 100644 --- a/README.md +++ b/README.md @@ -57,7 +57,9 @@ HunyuanVideo-1.5 is a video generation model that delivers top-tier quality with

## 🔥🔥🔥 News -* 🚀 Nov 24, 2025: We now support cache inference, achieving approximately 2x speedup! Pull the latest code to try it. 🔥🔥🔥🆕 +* 📚 Training code is coming soon. HunyuanVideo-1.5 is trained using the Muon optimizer, which we have open-sourced in the in [Training](#-training) section. **If you would like to continue training our model or fine-tune it with LoRA, please use the Muon optimizer.** +* 🚀 Nov 27, 2025: We now support cache inference (deepcache, teacache, taylorcache), achieving significant speedup! Pull the latest code to try it. 🔥🔥🔥🆕 +* 🚀 Nov 24, 2025: We now support deepcache inference. * 👋 Nov 20, 2025: We release the inference code and model weights of HunyuanVideo-1.5. @@ -78,6 +80,8 @@ If you develop/use HunyuanVideo-1.5 in your projects, welcome to let us know. - **Wan2GP v9.62** - [Wan2GP](https://github.com/deepbeepmeep/Wan2GP): WanGP is a very low VRAM app (as low 6 GB of VRAM for Hunyuan Video 1.5) supports Lora Accelerator for a 8 steps generation and offers tools to facilitate Video Generation. +- **ComfyUI-MagCache** - [ComfyUI-MagCache](https://github.com/Zehong-Ma/ComfyUI-MagCache): MagCache is a training-free caching approach that accelerates video generation by estimating fluctuating differences among model outputs across timesteps. It achieves 1.7x speedup for HunyuanVideo-1.5 with 20 inference steps. + ## 📑 Open-source Plan - HunyuanVideo-1.5 (T2V/I2V) @@ -105,6 +109,7 @@ If you develop/use HunyuanVideo-1.5 in your projects, welcome to let us know. - [Command Line Arguments](#command-line-arguments) - [Optimal Inference Configurations](#optimal-inference-configurations) - [🧱 Models Cards](#-models-cards) +- [🎓 Training](#-training) - [🎬 More Examples](#-more-examples) - [📊 Evaluation](#-evaluation) - [📚 Citation](#-citation) @@ -226,20 +231,22 @@ export I2V_REWRITE_MODEL_NAME="" PROMPT='A girl holding a paper with words "Hello, world!"' -IMAGE_PATH=./data/reference_image.png # Optional, 'none' or +IMAGE_PATH=none # Optional, none or to enable i2v mode SEED=1 ASPECT_RATIO=16:9 RESOLUTION=480p OUTPUT_PATH=./outputs/output.mp4 # Configuration +REWRITE=true # Enable prompt rewriting. Please ensure rewrite vLLM server is deployed and configured. N_INFERENCE_GPU=8 # Parallel inference GPU count CFG_DISTILLED=true # Inference with CFG distilled model, 2x speedup SPARSE_ATTN=false # Inference with sparse attention (only 720p models are equipped with sparse attention). Please ensure flex-block-attn is installed SAGE_ATTN=true # Inference with SageAttention -REWRITE=true # Enable prompt rewriting. Please ensure rewrite vLLM server is deployed and configured. OVERLAP_GROUP_OFFLOADING=true # Only valid when group offloading is enabled, significantly increases CPU memory usage but speeds up inference ENABLE_CACHE=true # Enable feature cache during inference. Significantly speeds up inference. +CACHE_TYPE=deepcache # Support: deepcache, teacache, taylorcache +ENABLE_SR=true # Enable super resolution MODEL_PATH=ckpts # Path to pretrained model torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \ @@ -248,14 +255,13 @@ torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \ --resolution $RESOLUTION \ --aspect_ratio $ASPECT_RATIO \ --seed $SEED \ - --cfg_distilled $CFG_DISTILLED \ - --sparse_attn $SPARSE_ATTN \ - --use_sageattn $SAGE_ATTN \ - --enable_cache $ENABLE_CACHE \ --rewrite $REWRITE \ - --output_path $OUTPUT_PATH \ + --cfg_distilled $CFG_DISTILLED \ + --sparse_attn $SPARSE_ATTN --use_sageattn $SAGE_ATTN \ + --enable_cache $ENABLE_CACHE --cache_type $CACHE_TYPE \ --overlap_group_offloading $OVERLAP_GROUP_OFFLOADING \ - --save_pre_sr_video \ + --sr $ENABLE_SR --save_pre_sr_video \ + --output_path $OUTPUT_PATH \ --model_path $MODEL_PATH ``` @@ -295,8 +301,9 @@ torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \ | `--dtype` | str | No | `bf16` | Data type for transformer: `bf16` (faster, lower memory) or `fp32` (better quality, slower, higher memory) | | `--use_sageattn` | bool | No | `false` | Enable SageAttention (use `--use_sageattn` or `--use_sageattn true/1` to enable, `--use_sageattn false/0` to disable) | | `--sage_blocks_range` | str | No | `0-53` | SageAttention blocks range (e.g., `0-5` or `0,1,2,3,4,5`) | -| `--enable_torch_compile` | bool | No | `false` | Enable torch compile for transformer (use `--enable_torch_compile` or `--enable_torch_compile true/1` to enable, `--enable_torch_compile false/0` to disable) | | `--enable_cache` | bool | No | `false` | Enable cache for transformer (use `--enable_cache` or `--enable_cache true/1` to enable, `--enable_cache false/0` to disable) | +| `--cache_type` | str | No | `deepcache` | Cache type for transformer (e.g., `deepcache, teacache, taylorcache`) | +| `--no_cache_block_id` | str | No | `53` | Blocks to exclude from deepcache (e.g., `0-5` or `0,1,2,3,4,5`) | | `--cache_start_step` | int | No | `11` | Start step to skip when using cache | | `--cache_end_step` | int | No | `45` | End step to skip when using cache | | `--total_steps` | int | No | `50` | Total inference steps | @@ -344,6 +351,32 @@ The following table provides the optimal inference configurations (CFG scale, em +## 🎓 Training + +> 💡 Training code is coming soon. We will release the complete training pipeline in the future. + +HunyuanVideo-1.5 is trained using the **Muon optimizer**, which accelerates convergence and improves training stability. The Muon optimizer combines momentum-based updates with Newton-Schulz orthogonalization for efficient optimization of large-scale video generation models. + +### Creating a Muon Optimizer + +Here's how to create a Muon optimizer for your model: + +```python +from hyvideo.optim.muon import get_muon_optimizer + +# Create Muon optimizer for your model +optimizer = get_muon_optimizer( + model=your_model, + lr=lr, # Learning rate + weight_decay=weight_decay, # Weight decay + momentum=momentum, # Momentum coefficient + adamw_betas=adamw_betas, # AdamW betas for 1D parameters + adamw_eps=adamw_eps # AdamW epsilon +) +``` + +> 📝 **To be continued**: More training details and the complete training pipeline will be released soon. Stay tuned! + ## 🎬 More Examples |Features|Demo1|Demo2| |------|------|------| diff --git a/README_CN.md b/README_CN.md index 88132c6..594184e 100644 --- a/README_CN.md +++ b/README_CN.md @@ -40,7 +40,9 @@ HunyuanVideo-1.5作为一款轻量级视频生成模型，仅需83亿参数即

## 🔥🔥🔥 最新动态 -* 🚀 Nov 24, 2025: 我们现已支持 cache 推理，可实现约两倍加速！请 pull 最新代码体验。 🔥🔥🔥🆕 +* 📚 训练代码即将发布。HunyuanVideo-1.5 使用 Muon 优化器进行训练，我们在[Training](#-training) 部分开源。**如果您希望继续训练我们的模型，或使用 LoRA 进行微调，请使用 Muon 优化器。** +* 🚀 Nov 27, 2025: 我们现已支持 cache 推理（deepcache, teacache, taylorcache），可极大加速推理！请 pull 最新代码体验。 🔥🔥🔥🆕 +* 🚀 Nov 24, 2025: 我们现已支持 deepcache 推理。 * 👋 Nov 20, 2025: 我们开源了 HunyuanVideo-1.5的代码和推理权重 ## 🎥 演示视频 @@ -60,6 +62,8 @@ HunyuanVideo-1.5作为一款轻量级视频生成模型，仅需83亿参数即 - **Wan2GP v9.62** - [Wan2GP](https://github.com/deepbeepmeep/Wan2GP): Wan2GP 是一款对显存要求非常低的应用（在 Hunyuan Video 1.5 下最低仅需 6GB 显存），支持 Lora 加速器实现 8 步生成，并且提供多种视频生成辅助工具。 +- **ComfyUI-MagCache** - [ComfyUI-MagCache](https://github.com/Zehong-Ma/ComfyUI-MagCache): MagCache 是一种无需训练的缓存方法，通过估计模型输出在不同时间步之间的波动差异来加速视频生成。在 20 步推理下，可为 HunyuanVideo-1.5 实现 1.7 倍加速。 + ## 📑 开源计划 - HunyuanVideo-1.5 (文生视频/图生视频) @@ -86,6 +90,7 @@ HunyuanVideo-1.5作为一款轻量级视频生成模型，仅需83亿参数即 - [命令行参数](#命令行参数) - [最优推理配置](#最优推理配置) - [🧱 模型卡片](#-模型卡片) +- [🎓 训练](#-训练) - [🎬 更多示例](#-更多示例) - [📊 性能评估](#-性能评估) - [📚 引用](#-引用) @@ -212,20 +217,22 @@ export I2V_REWRITE_MODEL_NAME="" PROMPT='A girl holding a paper with words "Hello, world!"' -IMAGE_PATH=./data/reference_image.png # 可选，'none' 或 <图像路径> +IMAGE_PATH=none # 可选，none 或 <图像路径> 以启用 i2v 模式 SEED=1 ASPECT_RATIO=16:9 RESOLUTION=480p OUTPUT_PATH=./outputs/output.mp4 # 配置 +REWRITE=true # 启用提示词重写。请确保 rewrite vLLM server 已部署和配置。 N_INFERENCE_GPU=8 # 并行推理 GPU 数量 CFG_DISTILLED=true # 使用 CFG 蒸馏模型进行推理，2倍加速 SPARSE_ATTN=false # 使用稀疏注意力进行推理（仅 720p 模型配备了稀疏注意力）。请确保 flex-block-attn 已安装 SAGE_ATTN=true # 使用 SageAttention 进行推理 -REWRITE=true # 启用提示词重写。请确保 rewrite vLLM server 已部署和配置。 OVERLAP_GROUP_OFFLOADING=true # 仅在组卸载启用时有效，会显著增加 CPU 内存占用，但能够提速 ENABLE_CACHE=true # 启用特征缓存进行推理。显著提升推理速度 +CACHE_TYPE=deepcache # 支持：deepcache, teacache, taylorcache +ENABLE_SR=true # 启用超分辨率 MODEL_PATH=ckpts # 预训练模型路径 torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \ @@ -234,14 +241,13 @@ torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \ --resolution $RESOLUTION \ --aspect_ratio $ASPECT_RATIO \ --seed $SEED \ - --cfg_distilled $CFG_DISTILLED \ - --sparse_attn $SPARSE_ATTN \ - --use_sageattn $SAGE_ATTN \ - --enable_cache $ENABLE_CACHE \ --rewrite $REWRITE \ - --output_path $OUTPUT_PATH \ + --cfg_distilled $CFG_DISTILLED \ + --sparse_attn $SPARSE_ATTN --use_sageattn $SAGE_ATTN \ + --enable_cache $ENABLE_CACHE --cache_type $CACHE_TYPE \ --overlap_group_offloading $OVERLAP_GROUP_OFFLOADING \ - --save_pre_sr_video \ + --sr $ENABLE_SR --save_pre_sr_video \ + --output_path $OUTPUT_PATH \ --model_path $MODEL_PATH ``` @@ -282,6 +288,8 @@ torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \ | `--sage_blocks_range` | str | 否 | `0-53` | SageAttention 块范围（例如：`0-5` 或 `0,1,2,3,4,5`） | | `--enable_torch_compile` | bool | 否 | `false` | 启用 torch compile 以优化 transformer（使用 `--enable_torch_compile` 或 `--enable_torch_compile true/1` 来启用，`--enable_torch_compile false/0` 来禁用） | | `--enable_cache` | bool | 否 | `false` | 启用 transformer 缓存（使用 `--enable_cache` 或 `--enable_cache true/1` 来启用，`--enable_cache false/0` 来禁用） | +| `--cache_type` | str | 否 | `deepcache` | Transformer 的缓存类型（例如：`deepcache, teacache, taylorcache`） | +| `--no_cache_block_id` | str | 否 | `53` | 从 deepcache 中排除的块（例如：`0-5` 或 `0,1,2,3,4,5`） | | `--cache_start_step` | int | 否 | `11` | 使用缓存时跳过的起始步数 | | `--cache_end_step` | int | 否 | `45` | 使用缓存时跳过的结束步数 | | `--total_steps` | int | 否 | `50` | 总推理步数 | @@ -329,6 +337,32 @@ torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \ +## 🎓 训练 + +> 💡 训练代码即将发布。我们将在未来发布完整的训练流程。 + +HunyuanVideo-1.5 使用 **Muon 优化器**进行训练，该优化器能够加速收敛并提高训练稳定性。Muon 优化器结合了基于动量的更新和 Newton-Schulz 正交化方法，可高效优化大规模视频生成模型。 + +### 创建 Muon 优化器 + +以下是如何为您的模型创建 Muon 优化器： + +```python +from hyvideo.optim.muon import get_muon_optimizer + +# 为您的模型创建 Muon 优化器 +optimizer = get_muon_optimizer( + model=your_model, + lr=lr, # 学习率 + weight_decay=weight_decay, # 权重衰减 + momentum=momentum, # 动量系数 + adamw_betas=adamw_betas, # 1D 参数的 AdamW betas + adamw_eps=adamw_eps # AdamW epsilon +) +``` + +> 📝 **未完待续**：更多训练细节和完整的训练流程即将发布，敬请期待！ + ## 🎬 更多示例 |特性|示例1|示例2| |------|------|------|