multi line support

app.py

@@ -5,7 +5,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 # --- Configuration ---
 MODEL_NAME = "TechnoByte/Qwen2.5-7B-VNTL-JP-EN"
-MAX_NEW_TOKENS = 512 # Max length of the generated translation
+MAX_NEW_TOKENS = 512 # Max length of the generated translation per line
 
 # --- Load Model and Tokenizer ---
 # Load the model and tokenizer only once when the app starts
@@ -26,47 +26,57 @@ except Exception as e:
 @spaces.GPU(duration=20)
 def translate_japanese_to_english(input_text):
     """
-    Translates Japanese text to English using the loaded model.
+    Translates Japanese text to English using the loaded model, processing line by line.
     """
     if not input_text:
         return "Please enter some Japanese text to translate."
 
-    print(f"Received input: {input_text}")
-
-    # Prepare the input using the chat template
-    messages = [
-        {"role": "user", "content": input_text}
-    ]
-    try:
-        # Apply chat template
-        prompt_text = tokenizer.apply_chat_template(
-            messages,
-            tokenize=False,
-            add_generation_prompt=True
-        )
-
-        # Tokenize the input
-        model_inputs = tokenizer([prompt_text], return_tensors="pt").to(model.device)
-
-        print("Generating translation...")
-        # Generate the translation
-        generated_ids = model.generate(
-            **model_inputs,
-            max_new_tokens=MAX_NEW_TOKENS,
-            do_sample=False
-        )
-
-        # Decode the generated text, skipping the prompt part
-        # Find the length of the input prompt tokens
-        input_ids_len = model_inputs.input_ids.shape[1]
-        # Slice the generated_ids to get only the new tokens
-        output_ids = generated_ids[0][input_ids_len:]
-
-        # Decode the output tokens
-        response = tokenizer.decode(output_ids, skip_special_tokens=True)
-
-        print(f"Generated response: {response}")
-        return response
+    print(f"Received input:\n{input_text}")
+
+    lines = input_text.splitlines() # Split input into lines
+    translated_lines = []
+
+    try: # Wrap the entire multi-line processing
+        for line in lines:
+            if not line.strip(): # If the line is empty or just whitespace
+                translated_lines.append("") # Keep the empty line structure
+                continue # Skip processing for this empty line
+
+            print(f"Translating line: {line}")
+
+            # Prepare the input for the current line using the chat template
+            messages = [
+                {"role": "user", "content": line}
+            ]
+            # Apply chat template
+            prompt_text = tokenizer.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+
+            # Tokenize the input for the current line
+            model_inputs = tokenizer([prompt_text], return_tensors="pt").to(model.device)
+
+            # Generate the translation for the current line
+            generated_ids = model.generate(
+                **model_inputs,
+                max_new_tokens=MAX_NEW_TOKENS,
+                do_sample=False # Use greedy decoding for consistency
+            )
+
+            # Decode the generated text, skipping the prompt part
+            input_ids_len = model_inputs.input_ids.shape[1]
+            output_ids = generated_ids[0][input_ids_len:]
+            response = tokenizer.decode(output_ids, skip_special_tokens=True).strip() # Strip leading/trailing whitespace from the translation
+
+            print(f"Generated response for line: {response}")
+            translated_lines.append(response)
+
+        # Join the translated lines back together with newline characters
+        final_translation = "\n".join(translated_lines)
+        print(f"Final combined translation:\n{final_translation}")
+        return final_translation
 
     except Exception as e:
         print(f"Error during translation: {e}")
@@ -83,7 +93,7 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
     with gr.Row():
         with gr.Column(scale=1):
             input_textbox = gr.Textbox(
-                lines=5,
+                lines=5, # Keep initial size, but it can grow
                 label="Japanese Input Text"
             )
             translate_button = gr.Button("Translate", variant="primary")
@@ -108,6 +118,8 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
             ["このソフトウェアの使い方がよくわかりません。"],
             ["明日の天気はどうなりますか？"],
             ["日本の文化についてもっと知りたいです。"],
+            ["こんにちは。\n元気ですか？\n私は元気です。"], # Multi-line example
+            ["これは最初の行です。\n\nこれは３行目です。空行を挟みます。"] # Example with empty line
         ],
         inputs=input_textbox,
         outputs=output_textbox,