PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Update diffusion README (#1439)
Wang, Mengni6e786c90357b
diff --git a/auto_round/compressors/diffusion/README.md b/auto_round/compressors/diffusion/README.mdindex 907bbaacb..16dc87d59 100644--- a/auto_round/compressors/diffusion/README.md+++ b/auto_round/compressors/diffusion/README.md@@ -17,15 +17,10 @@ Quantization for diffusion models is limited:```pythonimport torchfrom auto_round import AutoRound-from diffusers import AutoPipelineForText2Image--# Load the model-model_name = "black-forest-labs/FLUX.1-dev"-pipe = AutoPipelineForText2Image.from_pretrained(model_name, torch_dtype=torch.bfloat16)# Quantize the modelautoround = AutoRound(- pipe,+ "black-forest-labs/FLUX.1-dev",scheme="MXFP8",dataset="coco2014",num_inference_steps=10,@@ -33,12 +28,11 @@ autoround = AutoRound(generator_seed=None,batch_size=1,)-autoround.quantize()# Save the quantized modeloutput_dir = "./tmp_autoround"# Currently loading the quantized diffusion model is not supported, so use fake format-autoround.save_quantized(output_dir, format="fake", inplace=True)+autoround.quantize_and_save(output_dir, format="fake", inplace=True)```- `dataset`: the dataset for quantization training. Currently only support coco2014 and user customized .tsv file.@@ -62,6 +56,7 @@ auto-round \--scheme MXFP8 \--format fake \--batch_size 1 \+ --dataset coco2014 \--output_dir ./tmp_autoround```