Update mini.py

mini.py

@@ -3,12 +3,10 @@ import torch
 import spaces
 from PIL import Image
 import os
-from transformers import CLIPTokenizer, CLIPTextModel, AutoProcessor, T5EncoderModel, T5TokenizerFast, BitsAndBytesConfig
+from transformers import CLIPTokenizer, CLIPTextModel, AutoProcessor, T5EncoderModel, T5TokenizerFast
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler
 from flux.transformer_flux_simple import FluxTransformer2DModel
 from flux.pipeline_flux_chameleon_og import FluxPipeline
-from flux.pipeline_flux_img2img import FluxImg2ImgPipeline
-
 import torch.nn as nn
 import math
 import logging
@@ -31,9 +29,6 @@ MODEL_CACHE_DIR = "model_cache"
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 DTYPE = torch.bfloat16
 
-quant_config = BitsAndBytesConfig(load_in_8bit=True,)
-
-
 # Aspect ratio options
 ASPECT_RATIOS = {
     "1:1": (1024, 1024),
@@ -86,12 +81,11 @@ tokenizer_two = T5TokenizerFast.from_pretrained(
 
 # Load larger models to CPU
 vae = AutoencoderKL.from_pretrained(
-    os.path.join(MODEL_CACHE_DIR, "flux/vae"), 
+    os.path.join(MODEL_CACHE_DIR, "flux/vae")
 ).to(DTYPE).cpu()
 
 transformer = FluxTransformer2DModel.from_pretrained(
-    os.path.join(MODEL_CACHE_DIR, "flux/transformer"), 
-    quantization_config=quant_config,
+    os.path.join(MODEL_CACHE_DIR, "flux/transformer")
 ).to(DTYPE).cpu()
 
 scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
@@ -101,8 +95,7 @@ scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
 
 # Load Qwen2VL to CPU
 qwen2vl = Qwen2VLSimplifiedModel.from_pretrained(
-    os.path.join(MODEL_CACHE_DIR, "qwen2-vl"), 
-    quantization_config=quant_config,
+    os.path.join(MODEL_CACHE_DIR, "qwen2-vl")
 ).to(DTYPE).cpu()
 
 # Load connector and embedder