might be done now?

app.py

@@ -8,79 +8,75 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 model_id = "thrishala/mental_health_chatbot"
 
 try:
-    # Load model with appropriate device_map and settings
+    # Load model with appropriate settings
     model = AutoModelForCausalLM.from_pretrained(
         model_id,
-        device_map="auto",  # Use "auto" for device_map instead of device name
+        device_map="auto",
         torch_dtype=torch.float16,
         low_cpu_mem_usage=True,
         max_memory={0: "15GiB"} if torch.cuda.is_available() else None,
         offload_folder="offload",
-    ).eval()  # Set model to evaluation mode
+    ).eval()
 
     tokenizer = AutoTokenizer.from_pretrained(model_id)
-    tokenizer.pad_token = tokenizer.eos_token  # Set padding token if missing
-
-    # Perform a dummy generation to initialize model (if needed)
-    dummy_input = tokenizer("This is a test.", return_tensors="pt").to(model.device)
-    model.generate(
-        input_ids=dummy_input.input_ids,
-        max_new_tokens=1,
-        return_dict_in_generate=True  # Correct parameter name
-    )
+    tokenizer.pad_token = tokenizer.eos_token
+    tokenizer.model_max_length = 4096  # Set to model's actual context length
 
 except Exception as e:
     print(f"Error loading model: {e}")
     exit()
 
-def generate_text(prompt, max_new_tokens=128):
-    input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.device)
-
-    with torch.no_grad():
-        output = model.generate(
-            input_ids=input_ids,
-            max_new_tokens=max_new_tokens,
-            do_sample=False,
-            eos_token_id=tokenizer.eos_token_id,
-            return_dict_in_generate=True  # Correct parameter name
-        )
-
-    generated_text = tokenizer.decode(output.sequences[0], skip_special_tokens=True)
-    return generated_text
-
 def generate_text_streaming(prompt, max_new_tokens=128):
-    input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.device)
-
+    inputs = tokenizer(
+        prompt, 
+        return_tensors="pt", 
+        truncation=True, 
+        max_length=4096  # Match model's context length
+    ).to(model.device)
+    
+    generated_tokens = []
     with torch.no_grad():
         for _ in range(max_new_tokens):
-            output = model.generate(
-                input_ids=input_ids,
+            outputs = model.generate(
+                **inputs,
                 max_new_tokens=1,
                 do_sample=False,
                 eos_token_id=tokenizer.eos_token_id,
-                return_dict_in_generate=True  # Correct parameter name
+                return_dict_in_generate=True
             )
-
-            # Get the last generated token
-            generated_token_id = output.sequences[0, -1]
-            generated_token = tokenizer.decode([generated_token_id], skip_special_tokens=True)
-            yield generated_token
-
-            # Append new token to input_ids
-            input_ids = torch.cat([input_ids, output.sequences[:, -1:]], dim=-1)
-
-            if generated_token_id == tokenizer.eos_token_id:
+            
+            new_token = outputs.sequences[0, -1]
+            generated_tokens.append(new_token)
+            
+            # Update inputs for next iteration
+            inputs = {
+                "input_ids": torch.cat([inputs["input_ids"], new_token.unsqueeze(0).unsqueeze(0)], dim=-1),
+                "attention_mask": torch.cat([inputs["attention_mask"], torch.ones(1, 1, device=model.device)], dim=-1)
+            }
+            
+            # Decode the accumulated tokens
+            current_text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
+            yield current_text  # Yield the full text so far
+            
+            if new_token == tokenizer.eos_token_id:
                 break
 
 def respond(message, history, system_message, max_tokens):
+    # Build prompt with full history
     prompt = f"{system_message}\n"
     for user_msg, bot_msg in history:
         prompt += f"User: {user_msg}\nAssistant: {bot_msg}\n"
     prompt += f"User: {message}\nAssistant:"
-
+    
+    # Keep track of the full response
+    full_response = ""
+    
     try:
-        for token in generate_text_streaming(prompt, max_tokens):
-            yield token
+        for token_chunk in generate_text_streaming(prompt, max_tokens):
+            # Update the full response and yield incremental changes
+            full_response = token_chunk
+            yield full_response
+            
     except Exception as e:
         print(f"Error during generation: {e}")
         yield "An error occurred."
@@ -92,7 +88,7 @@ demo = gr.ChatInterface(
             value="You are a friendly and helpful mental health chatbot.",
             label="System message",
         ),
-        gr.Slider(minimum=1, maximum=128, value=32, step=1, label="Max new tokens"),
+        gr.Slider(minimum=1, maximum=512, value=128, step=1, label="Max new tokens"),
     ],
 )