Update app.py

app.py

@@ -403,6 +403,8 @@ def http_gen_edit_bot(state, temperature, top_k, top_p, image_gen_temperature,
     generated_image = None
     generated_text = ""
     try:
+        from transformers import set_seed
+        set_seed(42)
         with torch.inference_mode():  # Ensure no gradients are calculated
             output_ids = model.generate(
                 **inputs,
@@ -456,7 +458,7 @@ def http_gen_edit_bot(state, temperature, top_k, top_p, image_gen_temperature,
                         original_image_sizes) == 1:  # editing task, unpad and resize image to original size
                     original_size = original_image_sizes[0]
                     logging.info(f"original size: {original_size}. Output Image size: {generated_image.size}")
-                    # generated_image = processor.unpad_and_resize_back(generated_image, original_size[0], original_size[1])
+                    generated_image = processor.unpad_and_resize_back(generated_image, original_size[0], original_size[1])
                     logging.info(f"final image size: {generated_image.size}")
                 logging.info("Image successfully generated.")
                 # <image> is placeholder.
@@ -830,9 +832,9 @@ if __name__ == "__main__":
     model = AutoModel.from_pretrained(
         args.model_name,
         torch_dtype=torch.bfloat16,
-        attn_implementation='sdpa',  # OR 'sdpa' for Ascend NPUs
+        # attn_implementation='flash_attention_2',  # OR 'sdpa' for Ascend NPUs
         # torch_dtype=args.torch_dtype,
-        # attn_implementation='sdpa',  # OR 'sdpa' for Ascend NPUs
+        attn_implementation='sdpa',  # OR 'sdpa' for Ascend NPUs
         low_cpu_mem_usage=True,
         trust_remote_code=True).eval().cuda()
     processor = AutoProcessor.from_pretrained(args.model_name, trust_remote_code=True)