[diffusion] doc: rewrite stale diffusion compatibility matrix (#36987)
This commit is contained in:
Mick
@@ -40,7 +40,7 @@ FLUX models are optimized for high-quality image generation. The recommended lau
|
||||
|
||||
### 3.2 Configuration Tips
|
||||
|
||||
Currently supported optimizations are listed [here](/docs/sglang-diffusion/compatibility_matrix).
|
||||
See [Performance Optimization](/docs/sglang-diffusion/performance-optimization) for acceleration features and their runtime requirements.
|
||||
|
||||
- `--vae-path`: Path to a custom VAE model or HuggingFace model ID (e.g., fal/FLUX.2-Tiny-AutoEncoder). If not specified, the VAE will be loaded from the main model path.
|
||||
- `--num-gpus`: Number of GPUs to use
|
||||
@@ -193,6 +193,14 @@ sglang serve --model-path black-forest-labs/FLUX.1-dev
|
||||
- `--vae-cpu-offload`: Use CPU offload for VAE.
|
||||
- `--pin-cpu-memory`: Pin memory for CPU offload. Only added as a temp workaround if it throws "CUDA error: invalid argument".
|
||||
|
||||
#### 4.2.3 Known LoRA examples
|
||||
|
||||
Use `--lora-path` at startup or the [LoRA management API](/docs/sglang-diffusion/api/openai_api#lora-management) to load an adapter. Known FLUX examples include:
|
||||
|
||||
- [`dvyio/flux-lora-simple-illustration`](https://huggingface.co/dvyio/flux-lora-simple-illustration)
|
||||
- [`XLabs-AI/flux-furry-lora`](https://huggingface.co/XLabs-AI/flux-furry-lora)
|
||||
- [`XLabs-AI/flux-RealismLora`](https://huggingface.co/XLabs-AI/flux-RealismLora)
|
||||
|
||||
## 5. Benchmark
|
||||
|
||||
### 5.1 Speedup Benchmark
|
||||
|
||||
@@ -44,7 +44,7 @@ The step count and guidance scale are **request-time** settings (see [API Usage]
|
||||
|
||||
### 3.2 Configuration Tips
|
||||
|
||||
Currently supported optimizations are listed [here](/docs/sglang-diffusion/compatibility_matrix).
|
||||
See [Performance Optimization](/docs/sglang-diffusion/performance-optimization) for acceleration features and their runtime requirements.
|
||||
|
||||
- `--num-gpus`: Number of GPUs to use.
|
||||
- Multi-GPU (tensor and/or sequence parallelism): see [Section 3.3](#3-3-multi-gpu-tensor-and-sequence-parallelism).
|
||||
|
||||
@@ -41,7 +41,7 @@ import { MOVADeployment } from '/src/snippets/diffusion/mova-deployment.jsx'
|
||||
|
||||
### 3.2 Configuration Tips
|
||||
|
||||
Currently supported optimizations are listed [here](/docs/sglang-diffusion/compatibility_matrix).
|
||||
See [Performance Optimization](/docs/sglang-diffusion/performance-optimization) for acceleration features and their runtime requirements.
|
||||
|
||||
- `--num-gpus`: Number of GPUs to use
|
||||
- `--tp`: Tensor parallelism size (should not be larger than 1 if text encoder offload is enabled, as layer-wise offload plus prefetch is faster)
|
||||
|
||||
@@ -35,7 +35,7 @@ Qwen-Image-Edit-2511 is a 20B parameter model optimized for image editing tasks.
|
||||
|
||||
### 3.2 Configuration Tips
|
||||
|
||||
Currently supported optimizations are listed [here](/docs/sglang-diffusion/compatibility_matrix).
|
||||
See [Performance Optimization](/docs/sglang-diffusion/performance-optimization) for acceleration features and their runtime requirements.
|
||||
|
||||
- `--vae-path`: Path to a custom VAE model or HuggingFace model ID (e.g., fal/FLUX.2-Tiny-AutoEncoder). If not specified, the VAE will be loaded from the main model path.
|
||||
- `--num-gpus`: Number of GPUs to use
|
||||
@@ -189,6 +189,13 @@ sglang serve --model-path Qwen/Qwen-Image-Edit-2511
|
||||
- `--vae-cpu-offload`: Use CPU offload for VAE.
|
||||
- `--pin-cpu-memory`: Pin memory for CPU offload. Only added as a temp workaround if it throws "CUDA error: invalid argument".
|
||||
|
||||
#### 4.2.3 Known LoRA examples
|
||||
|
||||
Use `--lora-path` at startup or the [LoRA management API](/docs/sglang-diffusion/api/openai_api#lora-management) to load an adapter. Known Qwen-Image-Edit examples include:
|
||||
|
||||
- [`ostris/qwen_image_edit_inpainting`](https://huggingface.co/ostris/qwen_image_edit_inpainting)
|
||||
- [`lightx2v/Qwen-Image-Edit-2511-Lightning`](https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning)
|
||||
|
||||
## 5. Benchmark
|
||||
|
||||
Test Environment:
|
||||
|
||||
@@ -95,7 +95,7 @@ also tested on this profile and did not improve steady-state latency.
|
||||
|
||||
### 3.3 Configuration Tips
|
||||
|
||||
Currently supported optimizations are listed [here](/docs/sglang-diffusion/compatibility_matrix).
|
||||
See [Performance Optimization](/docs/sglang-diffusion/performance-optimization) for acceleration features and their runtime requirements.
|
||||
|
||||
- `--vae-path`: Path to a custom VAE model or HuggingFace model ID (e.g., fal/FLUX.2-Tiny-AutoEncoder). If not specified, the VAE will be loaded from the main model path.
|
||||
- `--num-gpus`: Number of GPUs to use
|
||||
@@ -249,6 +249,15 @@ sglang serve --model-path Qwen/Qwen-Image
|
||||
- `--vae-cpu-offload`: Use CPU offload for VAE.
|
||||
- `--pin-cpu-memory`: Pin memory for CPU offload. Only added as a temp workaround if it throws "CUDA error: invalid argument".
|
||||
|
||||
#### 4.2.3 Known LoRA examples
|
||||
|
||||
Use `--lora-path` at startup or the [LoRA management API](/docs/sglang-diffusion/api/openai_api#lora-management) to load an adapter. Known Qwen-Image examples include:
|
||||
|
||||
- [`lightx2v/Qwen-Image-Lightning`](https://huggingface.co/lightx2v/Qwen-Image-Lightning)
|
||||
- [`flymy-ai/qwen-image-realism-lora`](https://huggingface.co/flymy-ai/qwen-image-realism-lora)
|
||||
- [`prithivMLmods/Qwen-Image-HeadshotX`](https://huggingface.co/prithivMLmods/Qwen-Image-HeadshotX)
|
||||
- [`starsfriday/Qwen-Image-EVA-LoRA`](https://huggingface.co/starsfriday/Qwen-Image-EVA-LoRA)
|
||||
|
||||
## 5. Benchmark
|
||||
|
||||
Test Environment:
|
||||
|
||||
@@ -69,4 +69,4 @@ validation then catches MDX, navigation, and internal-link problems.
|
||||
|
||||
For runtime concepts and platform support, see the
|
||||
[SGLang Diffusion documentation](/docs/sglang-diffusion/index) and
|
||||
[compatibility matrix](/docs/sglang-diffusion/compatibility_matrix).
|
||||
[supported model catalog](/docs/sglang-diffusion/compatibility_matrix).
|
||||
|
||||
@@ -15,6 +15,12 @@ import { Wan22Deployment } from '/src/snippets/diffusion/wan22-deployment.jsx';
|
||||
|
||||
Choose the A14B MoE checkpoints for maximum T2V or I2V capacity and the 5B TI2V model for a smaller unified 720p-at-24-fps path. MoE reduces active compute relative to total capacity but does not remove the memory cost of loading expert weights, so hardware selection still matters.
|
||||
|
||||
<Warning>
|
||||
The Wan2.2 TI2V 5B checkpoint currently has known quality issues when it is used
|
||||
for image-to-video generation. Use `Wan-AI/Wan2.2-I2V-A14B-Diffusers` when I2V
|
||||
quality is the priority.
|
||||
</Warning>
|
||||
|
||||
## 2. SGLang-diffusion Installation
|
||||
|
||||
SGLang-diffusion offers multiple installation methods. You can choose the most suitable installation method based on your hardware platform and requirements.
|
||||
@@ -35,7 +41,7 @@ The Wan2.2 series offers models in various sizes, architectures and input types,
|
||||
|
||||
### 3.2 Configuration Tips
|
||||
|
||||
Currently supported optimizations are listed [here](/docs/sglang-diffusion/compatibility_matrix).
|
||||
See [Performance Optimization](/docs/sglang-diffusion/performance-optimization) for acceleration features and their runtime requirements.
|
||||
|
||||
- `--vae-path`: Path to a custom VAE model or HuggingFace model ID (e.g., fal/FLUX.2-Tiny-AutoEncoder). If not specified, the VAE will be loaded from the main model path.
|
||||
- `--num-gpus {NUM_GPUS}`: Number of GPUs to use
|
||||
|
||||
@@ -35,7 +35,7 @@ Z-Image-Turbo is optimized for high-quality image generation with only 8 inferen
|
||||
|
||||
### 3.2 Configuration Tips
|
||||
|
||||
Currently supported optimizations are listed [here](/docs/sglang-diffusion/compatibility_matrix).
|
||||
See [Performance Optimization](/docs/sglang-diffusion/performance-optimization) for acceleration features and their runtime requirements.
|
||||
|
||||
- `--vae-path`: Path to a custom VAE model or HuggingFace model ID (e.g., fal/FLUX.2-Tiny-AutoEncoder). If not specified, the VAE will be loaded from the main model path.
|
||||
- `--num-gpus`: Number of GPUs to use
|
||||
@@ -189,6 +189,13 @@ sglang serve --model-path Tongyi-MAI/Z-Image-Turbo
|
||||
- `--vae-cpu-offload`: Use CPU offload for VAE.
|
||||
- `--pin-cpu-memory`: Pin memory for CPU offload. Only added as a temp workaround if it throws "CUDA error: invalid argument".
|
||||
|
||||
#### 4.2.3 Known LoRA examples
|
||||
|
||||
Use `--lora-path` at startup or the [LoRA management API](/docs/sglang-diffusion/api/openai_api#lora-management) to load an adapter. Known Z-Image-Turbo examples include:
|
||||
|
||||
- [`tarn59/pixel_art_style_lora_z_image_turbo`](https://huggingface.co/tarn59/pixel_art_style_lora_z_image_turbo)
|
||||
- [`wcde/Z-Image-Turbo-DeJPEG-Lora`](https://huggingface.co/wcde/Z-Image-Turbo-DeJPEG-Lora)
|
||||
|
||||
## 5. Benchmark
|
||||
|
||||
Test Environment:
|
||||
|
||||
Reference in New Issue
Block a user