中文 · English
ChituDiffusion 基于 Diffusers 生命周期提供模型生成、上下文并行、缓存加速和常驻服务。
项目处于开发者预览阶段。服务不保证任意 rank 故障后的恢复。
|
提供 Fast AGKV 与 Fast Ulysses，并保留 NCCL fallback。面向具备 GPU P2P、 NVSHMEM 和目标架构扩展的兼容单机环境优化通信路径。 |
根据实测代价和端到端 SLO，在 pulse 边界调整 CP lane，并通过统一 executor 处理 队列、状态迁移和 worker 生命周期。 |
|
提供 MagCache、MeanCache、TeaCache、TaylorSeer 与 PAB 的统一 API，可叠加单卡或 静态 CP。查看 速度与质量评测。 |
保留 tokenizer、文本编码器、scheduler、VAE 和输出处理流程，只适配模型相关的 DiT 执行与 tensor 布局。 |
支持模型
| 模型 | 生成能力 | 并行与服务能力 |
|---|---|---|
| FLUX.1 | 文生图 | NCCL/Fast CP、EPE executor、FlexCache model spec |
| FLUX.2-klein | 文生图 | 固定 full-world 静态 CP |
| Hunyuan Image 3 | 固定尺寸文生图与图生图基础服务链路 | 可配置 TP×CFG×CP×EP decoder（72 GiB 参考配置 TP2×CFG2×CP2×EP2）、NCCL/Fast AGKV/Fast Ulysses、EPE static CP、变长 expert dispatch、独立 VAEP |
| MiniMax-H3 | T2VA、first/last-frame FL2VA 基础服务链路 | TP×CP DiT、EPE、Fast Ulysses/NCCL CP、独立 VAEP、音视频解码与 MP4 输出 |
| Qwen-Image | 文生图 | NCCL/Fast CP、EPE executor |
| Wan 2.1 T2V | 文生视频 | NCCL/Fast CP、EPE executor、FlexCache profiles |
| Z-Image | 文生图 | NCCL/Fast CP、统一 EPE 服务入口、FlexCache integration |
要求 Linux、Python 3.12 或 3.13、NVIDIA GPU 和可用的 CUDA 环境。按机器 CUDA
版本选择 pyproject.toml 中的 PyTorch index，然后运行：
uv sync --group dev
source .venv/bin/activate
chitu --helppyproject.toml 是唯一构建入口。
chitu generate \
--model zimage \
--model-path /path/to/Z-Image \
--prompt "a red cube on a white table" \
--output outputs/zimage.png静态多卡 NCCL CP：
torchrun --standalone --nproc-per-node=4 -m chitu_diffusion.cli \
generate \
--model flux1 \
--model-path /path/to/FLUX.1-dev \
--output outputs/flux1.png单机 TP×CFG×CP×EP MoE，例如 Hunyuan Image 3 的默认 8 卡拓扑：
torchrun --standalone --nproc-per-node=8 -m chitu_diffusion.cli \
generate \
--model hunyuan-image3 \
--model-path /path/to/HunyuanImage-3 \
--steps 50 \
--output outputs/hunyuan_image3.pngFast CP 需要兼容的单机 GPU、GPU P2P、NVSHMEM 和针对目标环境编译的扩展。参数和安装 步骤见 Fast CP README。
chitu generate \
--model wan \
--model-path /path/to/Wan2.1-T2V-1.3B \
--steps 50 \
--cache-strategy magcache \
--output outputs/wan-magcache.mp4缓存支持与模型、步数和 profile 绑定。不支持的组合会在运行前报错。原理、支持范围和 测试结果见 FlexCache 文档。
chitu serve --stage-config examples/stage-zimage.yaml所有 rank 进入服务生命周期，leader 提供 HTTP，其他 rank 执行 worker loop。配置和 限制见运行说明与 EPE 文档。
- 文档首页：
docs/index.md - 安装：
docs/usage/installation.md - 运行：
docs/usage/running.md - 模型 API：
docs/features/diffusers-api.md - 用户示例：
examples/README.md
python -m pytest -q
python -m ruff check chitu_diffusion tests examples
python -m build
mkdocs build --strict
中文 · English
ChituDiffusion provides model generation, context parallelism, cache acceleration, and persistent serving while preserving the Diffusers pipeline lifecycle.
The project is a developer preview. The service does not recover from an arbitrary rank failure.
|
Fast AGKV and Fast Ulysses with an NCCL fallback. The fast paths target compatible single-node systems with GPU P2P, NVSHMEM, and extensions built for the target architecture. |
Adjusts CP lanes at pulse boundaries using measured costs and end-to-end SLOs, with one executor lifecycle for queues, state migration, and workers. |
|
One API for MagCache, MeanCache, TeaCache, TaylorSeer, and PAB, composable with single-GPU or static CP generation. See the speed and quality evaluations. |
Preserves tokenizers, text encoders, schedulers, VAEs, and output processing. Adapters only supply model-specific DiT execution and tensor layouts. |
Supported models
| Model | Generation | Parallelism and serving |
|---|---|---|
| FLUX.1 | Text to image | NCCL/Fast CP, EPE executor, and FlexCache model spec |
| FLUX.2-klein | Text to image | Fixed full-world static CP |
| Hunyuan Image 3 | Foundational fixed-size text-to-image and image-to-image service path | Configurable TP×CFG×CP×EP decoder (72 GiB reference: TP2×CFG2×CP2×EP2), NCCL/Fast AGKV/Fast Ulysses, EPE static CP, variable-length expert dispatch, and independent VAEP |
| MiniMax-H3 | Foundational T2VA and first/last-frame FL2VA service path | TP×CP DiT, EPE, Fast Ulysses/NCCL CP, independent VAEP, audio/video decode, and MP4 output |
| Qwen-Image | Text to image | NCCL/Fast CP and EPE executor |
| Wan 2.1 T2V | Text to video | NCCL/Fast CP, EPE executor, and FlexCache profiles |
| Z-Image | Text to image | NCCL/Fast CP, unified EPE serving, and FlexCache integration |
ChituDiffusion requires Linux, Python 3.12 or 3.13, an NVIDIA GPU, and a working
CUDA environment. Select the matching PyTorch index in pyproject.toml, then run:
uv sync --group dev
source .venv/bin/activate
chitu --helppyproject.toml is the only build entry point.
chitu generate \
--model zimage \
--model-path /path/to/Z-Image \
--prompt "a red cube on a white table" \
--output outputs/zimage.pngStatic multi-GPU NCCL CP:
torchrun --standalone --nproc-per-node=4 -m chitu_diffusion.cli \
generate \
--model flux1 \
--model-path /path/to/FLUX.1-dev \
--output outputs/flux1.pngA single-node TP×CFG×CP×EP MoE stage, for example Hunyuan Image 3 on its default eight-GPU topology:
torchrun --standalone --nproc-per-node=8 -m chitu_diffusion.cli \
generate \
--model hunyuan-image3 \
--model-path /path/to/HunyuanImage-3 \
--steps 50 \
--output outputs/hunyuan_image3.pngFast CP requires compatible single-node GPUs, GPU P2P, NVSHMEM, and extensions built for the target environment. See the Fast CP README.
chitu generate \
--model wan \
--model-path /path/to/Wan2.1-T2V-1.3B \
--steps 50 \
--cache-strategy magcache \
--output outputs/wan-magcache.mp4Cache support is tied to the model, step count, and profile. Unsupported combinations fail before execution. See the FlexCache documentation for principles, support, and measured results.
chitu serve --stage-config examples/stage-zimage.yamlAll ranks enter the service lifecycle. The leader hosts HTTP while the remaining ranks execute worker loops. See the running guide and EPE documentation.
- Documentation:
docs/index.md - Installation:
docs/usage/installation.md - Running:
docs/usage/running.md - Model API:
docs/features/diffusers-api.md - Examples:
examples/README.md
python -m pytest -q
python -m ruff check chitu_diffusion tests examples
python -m build
mkdocs build --strict