Update app.py

app.py

@@ -1,30 +1,30 @@
 import gradio as gr
-import spaces
 from transformers import AutoModelForCausalLM, AutoProcessor
 import torch
 from PIL import Image
-import subprocess
-subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
+# Model ve işlemci yükleme
 models = {
-    "microsoft/Phi-3.5-vision-instruct": AutoModelForCausalLM.from_pretrained("microsoft/Phi-3.5-vision-instruct", trust_remote_code=True, torch_dtype="auto", _attn_implementation="flash_attention_2").cuda().eval()
-
+    "microsoft/Phi-3.5-vision-instruct": AutoModelForCausalLM.from_pretrained(
+        "microsoft/Phi-3.5-vision-instruct", trust_remote_code=True, torch_dtype="auto"
+    ).eval()
 }
 
 processors = {
-    "microsoft/Phi-3.5-vision-instruct": AutoProcessor.from_pretrained("microsoft/Phi-3.5-vision-instruct", trust_remote_code=True)
+    "microsoft/Phi-3.5-vision-instruct": AutoProcessor.from_pretrained(
+        "microsoft/Phi-3.5-vision-instruct", trust_remote_code=True
+    )
 }
 
 DESCRIPTION = "[Phi-3.5-vision Demo](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)"
 
 kwargs = {}
-kwargs['torch_dtype'] = torch.bfloat16
+kwargs['torch_dtype'] = torch.float32  # CPU üzerinde çalıştığı için bfloat16 yerine float32 kullanılıyor
 
 user_prompt = '<|user|>\n'
 assistant_prompt = '<|assistant|>\n'
 prompt_suffix = "<|end|>\n"
 
-@spaces.GPU
 def run_example(image, text_input=None, model_id="microsoft/Phi-3.5-vision-instruct"):
     model = models[model_id]
     processor = processors[model_id]
@@ -32,15 +32,18 @@ def run_example(image, text_input=None, model_id="microsoft/Phi-3.5-vision-instr
     prompt = f"{user_prompt}<|image_1|>\n{text_input}{prompt_suffix}{assistant_prompt}"
     image = Image.fromarray(image).convert("RGB")
 
-    inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")
-    generate_ids = model.generate(**inputs, 
-                                max_new_tokens=1000,
-                                eos_token_id=processor.tokenizer.eos_token_id,
-                                )
+    inputs = processor(prompt, image, return_tensors="pt")  # Cihaz belirtilmedi, varsayılan olarak CPU kullanılır
+    generate_ids = model.generate(
+        **inputs,
+        max_new_tokens=1000,
+        eos_token_id=processor.tokenizer.eos_token_id,
+    )
     generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
-    response = processor.batch_decode(generate_ids, 
-                                    skip_special_tokens=True, 
-                                    clean_up_tokenization_spaces=False)[0]
+    response = processor.batch_decode(
+        generate_ids,
+        skip_special_tokens=True,
+        clean_up_tokenization_spaces=False
+    )[0]
     return response
 
 css = """
@@ -66,4 +69,4 @@ with gr.Blocks(css=css) as demo:
         submit_btn.click(run_example, [input_img, text_input, model_selector], [output_text])
 
 demo.queue(api_open=False)
-demo.launch(debug=True, show_api=False)
+demo.launch(debug=True, show_api=False)