Update app.py

app.py

@@ -13,7 +13,8 @@ from diffusers import DiffusionPipeline
 import bitsandbytes
 from diffusers.quantizers import PipelineQuantizationConfig
 from PIL import Image
-from diffusers import FluxKontextPipeline
+from transformers import CLIPModel, CLIPProcessor, CLIPTextModel, CLIPTokenizer, CLIPConfig, T5EncoderModel, T5Tokenizer
+from diffusers import FluxKontextPipeline, DiffusionPipeline
 from diffusers.utils import load_image
 from huggingface_hub import login, hf_hub_download, HfFileSystem, ModelCard
 from huggingface_hub.utils._runtime import dump_environment_info
@@ -41,7 +42,7 @@ quant_config = PipelineQuantizationConfig(
 try:
     # Set max memory usage for ZeroGPU
     torch.cuda.set_per_process_memory_fraction(1.0)
-    torch.set_float32_matmul_precision("high")
+    torch.set_float32_matmul_precision("medium")
 except Exception as e:
     print(f"Error setting memory usage: {e}")
 
@@ -52,6 +53,15 @@ pipe = FluxKontextPipeline.from_pretrained(
     torch_dtype=torch.bfloat16
     ).to(DEVICE)
 
+model_id = ("zer0int/LongCLIP-GmP-ViT-L-14")
+config = CLIPConfig.from_pretrained(model_id)
+config.text_config.max_position_embeddings = 248
+clip_model = CLIPModel.from_pretrained(model_id, torch_dtype=torch.bfloat16, config=config, ignore_mismatched_sizes=True)
+clip_processor = CLIPProcessor.from_pretrained(model_id, padding="max_length", max_length=248)
+pipe.tokenizer = clip_processor.tokenizer
+pipe.text_encoder = clip_model.text_model
+pipe.tokenizer_max_length = 248
+pipe.text_encoder.dtype = torch.bfloat16
 # Load LoRA data (you'll need to create this JSON file or modify to load your LoRAs)
 
 with open("flux_loras.json", "r") as file: