Refine

benchmarks/text_to_image.py

@@ -36,6 +36,7 @@
 from diffusers.utils import load_image
 
 from onediffx import compile_pipe, quantize_pipe # quantize_pipe currently only supports the nexfort backend.
+from onediff.infer_compiler import oneflow_compile
 
 
 def parse_args():
@@ -244,7 +245,13 @@ def main():
         pass
     elif args.compiler == "oneflow":
         print("Oneflow backend is now active...")
-        pipe = compile_pipe(pipe)
+        # Note: The compile_pipe() based on the oneflow backend is incompatible with T5EncoderModel.
+        # pipe = compile_pipe(pipe)
+        if hasattr(pipe, "unet"):
+            pipe.unet = oneflow_compile(pipe.unet)
+        if hasattr(pipe, "transformer"):
+            pipe.transformer = oneflow_compile(pipe.transformer)
+        pipe.vae.decoder = oneflow_compile(pipe.vae.decoder)
     elif args.compiler == "nexfort":
         print("Nexfort backend is now active...")
         if args.quantize:

onediff_diffusers_extensions/examples/pixart/README.md

@@ -1,4 +1,4 @@
-# Run PixArt with nexfort backend(Beta Release)
+# Run PixArt with nexfort backend (Beta Release)
 
 
 1. [Environment Setup](#environment-setup)
@@ -27,8 +27,8 @@ https://github.com/siliconflow/onediff/tree/main/src/onediff/infer_compiler/back
 
 HF model: 
 
- - PixArt-sigma: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS
  - PixArt-alpha: https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS
+ - PixArt-sigma: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS
 
 HF pipeline: https://huggingface.co/docs/diffusers/main/en/api/pipelines/pixart
 
@@ -44,7 +44,7 @@ Compared to PixArt-alpha, PixArt-sigma extends the token length of the text enco
 cd onediff
 ```
 
-### Run 1024*1024 without compile(the original pytorch HF diffusers pipeline)
+### Run 1024*1024 without compile (the original pytorch HF diffusers pipeline)
 ```
 # To test sigma, specify the --model parameter as `PixArt-alpha/PixArt-Sigma-XL-2-1024-MS`.
 python3 ./benchmarks/text_to_image.py \
@@ -56,7 +56,19 @@ python3 ./benchmarks/text_to_image.py \
 --prompt "product photography, world of warcraft orc warrior, white background"
 ```
 
-### Run 1024*1024 with compile
+### Run 1024*1024 with oneflow backend compile
+
+```
+python3 ./benchmarks/text_to_image.py \
+--model PixArt-alpha/PixArt-XL-2-1024-MS \
+--scheduler none \
+--steps 20 \
+--compiler oneflow \
+--output-image ./pixart_alpha_compile.png \
+--prompt "product photography, world of warcraft orc warrior, white background"
+```
+
+### Run 1024*1024 with nexfort backend compile
 ```
 python3 ./benchmarks/text_to_image.py \
 --model PixArt-alpha/PixArt-XL-2-1024-MS \

src/onediff/infer_compiler/backends/nexfort/README.md

@@ -42,8 +42,7 @@ 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).
+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:
 ```