byte-vortex commited on
Commit
5e02539
·
verified ·
1 Parent(s): 32fc3ce

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +19 -17
app.py CHANGED
@@ -49,41 +49,43 @@ def generate_response(image, prompt):
49
 
50
  # Handle optional image input
51
  pil_img = None
52
- image_token_str = ""
53
  if image is not None:
54
  if isinstance(image, str):
55
  pil_img = Image.open(image).convert("RGB")
56
  else:
57
  pil_img = Image.fromarray(image).convert("RGB")
58
-
59
- image_token = getattr(proc, "image_token", "<|image|>")
60
- image_token_str = f"{image_token}\n"
61
 
62
- # Construct Gemma 4 native turn formatting tags directly
63
- formatted_prompt = f"<|turn>user\n{image_token_str}{prompt}<|turn>\n<|turn>model\n"
 
 
 
 
 
64
 
65
- inputs = proc(
66
- text=[formatted_prompt],
67
- images=[pil_img] if pil_img is not None else None,
 
 
68
  return_tensors="pt",
69
- padding=True
70
  ).to(mod.device)
71
 
72
- input_len = inputs["input_ids"].shape[1]
73
 
74
  with torch.no_grad():
75
  outputs = mod.generate(
76
  **inputs,
77
- max_new_tokens=256,
78
- do_sample=False, # Greedy decoding ensures deterministic, clean responses
79
  eos_token_id=proc.tokenizer.eos_token_id,
80
  pad_token_id=proc.tokenizer.pad_token_id
81
  )
82
 
83
- # Slice out prompt tokens to show only the model's reply
84
- generated_tokens = outputs[0][input_len:]
85
- response = proc.decode(generated_tokens, skip_special_tokens=True)
86
- return response.strip()
87
 
88
  except Exception as e:
89
  error_msg = str(e).lower()
 
49
 
50
  # Handle optional image input
51
  pil_img = None
 
52
  if image is not None:
53
  if isinstance(image, str):
54
  pil_img = Image.open(image).convert("RGB")
55
  else:
56
  pil_img = Image.fromarray(image).convert("RGB")
 
 
 
57
 
58
+ # Build official chat message structure
59
+ content = []
60
+ if pil_img is not None:
61
+ content.append({"type": "image", "image": pil_img})
62
+ content.append({"type": "text", "text": prompt})
63
+
64
+ messages = [{"role": "user", "content": content}]
65
 
66
+ # Use official Gemma 4 processor chat template pipeline
67
+ inputs = proc.apply_chat_template(
68
+ messages,
69
+ tokenize=True,
70
+ return_dict=True,
71
  return_tensors="pt",
72
+ add_generation_prompt=True
73
  ).to(mod.device)
74
 
75
+ input_len = inputs["input_ids"].shape[-1]
76
 
77
  with torch.no_grad():
78
  outputs = mod.generate(
79
  **inputs,
80
+ max_new_tokens=512,
81
+ do_sample=False, # Greedy decoding for clean, deterministic responses
82
  eos_token_id=proc.tokenizer.eos_token_id,
83
  pad_token_id=proc.tokenizer.pad_token_id
84
  )
85
 
86
+ # Slice precisely using the official input length
87
+ response = proc.decode(outputs[0][input_len:], skip_special_tokens=True)
88
+ return response.strip() if response.strip() else "⚠️ Model generated an empty response. Please try another prompt."
 
89
 
90
  except Exception as e:
91
  error_msg = str(e).lower()