Update README.md of compile api (#1070)

src/onediff/infer_compiler/README.md

@@ -1,9 +1,22 @@
 # OneDiff compiler for inference
 
-## With nexfort compiler backend
+## With nexfort compiler backend(Beta release)
+OneDiff Nexfort is a lightweight [torch 2.0 compiler backend](https://pytorch.org/docs/stable/torch.compiler.html) specially optimized for Diffusion Models.
+
+Currently, it is especially for DiT(Diffusion Transformer) models which is the backbone of [SD3](https://stability.ai/news/stable-diffusion-3) and [Sora](https://openai.com/sora/).
+
 ### Installation
-1. Install nexfort: https://github.com/siliconflow/onediff?tab=readme-ov-file#nexfort
-2. Install onediff: https://github.com/siliconflow/onediff?tab=readme-ov-file#3-install-onediff
+####  Dependency
+```
+pip3 install -U torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 torchao==0.1
+```
+The current version of nexfort is compatible with `torch 2.3.0` and `torch 2.4.0`.
+
+#### Install nexfort
+Reference: Install nexfort: https://github.com/siliconflow/onediff?tab=readme-ov-file#nexfort
+
+#### Install onediff
+Reference: https://github.com/siliconflow/onediff?tab=readme-ov-file#3-install-onediff
 
 ### Usage
 ```python
@@ -14,9 +27,89 @@ options = {"mode": "O3"}  # mode can be O2 or O3
 compiled = compile(module, backend="nexfort", options=options)
 ```
 
-### Suggested Modes
+If you are using [onediffx for HF diffusers, you can use compile_pipe](https://github.com/siliconflow/onediff/tree/main/onediff_diffusers_extensions#compile-save-and-load-pipeline) like this:
+```
+from onediffx import compile_pipe
+options = '{"mode": "O3", "memory_format": "channels_last"}'
+pipe = compile_pipe(pipe, backend="nexfort", options=options, fuse_qkv_projections=True)
+```
+
+
+### Available Compiler Modes
+Config Example:
+```
+options = {"mode": "max-optimize:max-autotune:low-precision:cache-all"}
+```
+
+| Mode | Description |
+| - | - |
+| `cache-all` | Cache all the compiled stuff to speed up reloading and recompiling. |
+| `max-autotune` | Enable all the kernel autotuning options to find out the best kernels, this might slow down the compilation. |
+| `max-optimize` | Enable the ***most*** extreme optimization strategies like the most aggressive fusion kernels to maximize the performance, this might slow down the compilation and require long autotuning. |
+| `cudagraphs` | Enable CUDA Graphs to reduce CPU overhead. |
+| `freezing` | Freezing will attempt to inline weights as constants in optimization and run constant folding and other optimizations on them. After freezing, weights can no longer be updated. |
+| `low-precision` | Enable low precision mode. This will allow some math computations happen in low precision to speed up the overall performance. |
+
+### Suggested Combination Modes
+Config Example:
+```
+options = {"mode": "O3"}
+```
 
 | Combination | Description |
 | - | - |
 | `O2` | This is the most suggested combination of compiler modes. This mode requires support for most models, ensuring model accuracy, and supporting dynamic resolution. |
 | `O3` | This aims for efficiency. |
+
+`O2` and `O3` are approximately equal to `options = {"mode": "max-optimize:max-autotune:low-precision:cache-all"}`, but `O2` has higher precision.
+
+### Run pixart alpha (with nexfort backend)
+
+Details at: https://github.com/siliconflow/onediff/tree/main/onediff_diffusers_extensions/examples/pixart_alpha
+
+### Compilation cache to speed up recompilation
+
+Setting cache:
+```
+# Enabled Inductor - FX Graph Cache. Default Off
+export TORCHINDUCTOR_FX_GRAPH_CACHE=1
+
+# Setting Inductor - Autotuning Cache DIR. This cache is enabled by default.
+export TORCHINDUCTOR_CACHE_DIR=~/torchinductor
+```
+
+Clear Cache:
+```
+python3 -m nexfort.utils.clear_inductor_cache
+```
+
+Advanced cache functionality is currently in development.
+
+### Dynamic shape
+Onediff's nexfort backend also supports out-of-the-box dynamic shape inference. You just need to enable `dynamic` during compilation, as in `'{"mode": "max-autotune", "dynamic": true}'`. To understand how dynamic shape support works, please refer to the <https://pytorch.org/docs/stable/generated/torch.compile.html> and <https://github.com/pytorch/pytorch/blob/main/docs/source/torch.compiler_dynamic_shapes.rst> page. To avoid over-specialization and re-compilation, you need to initially call your model with a non-typical shape. For example: you can first call your Stable Diffusion model with a shape of 512x768 (height != width).
+
+Test SDXL:
+```
+# The best practice mode configuration for dynamic shape is `max-optimize:max-autotune:low-precision`.
+python3 ./onediff_diffusers_extensions/examples/text_to_image_sdxl.py \
+--height 512 \
+--width 768 \
+--compiler nexfort \
+--compiler-config '{"mode": "max-optimize:max-autotune:low-precision", "memory_format": "channels_last", "dynamic": true}' \
+--run_multiple_resolutions 1 \
+--run_rare_resolutions 1
+```
+
+Test PixArt alpha:
+```
+python3 ./benchmarks/text_to_image.py \
+--model PixArt-alpha/PixArt-XL-2-1024-MS \
+--scheduler none \
+--steps 20 \
+--height 512 \
+--width 768 \
+--compiler nexfort \
+--compiler-config '{"mode": "max-optimize:max-autotune:low-precision", "memory_format": "channels_last", "dynamic": true}' \
+--run_multiple_resolutions 1
+```
+