idk anymore is more a vibe

app.py

@@ -12,14 +12,14 @@ model_id = "thrishala/mental_health_chatbot"
 try:
     model = AutoModelForCausalLM.from_pretrained(
         model_id,
-        device_map=device,  # Use the determined device
+        device_map=device,
         torch_dtype=torch.float16,
         low_cpu_mem_usage=True,
-        max_memory={device: "15GB"},  # Use device-specific memory management
+        max_memory={device: "15GB"},
         offload_folder="offload",
     )
     tokenizer = AutoTokenizer.from_pretrained(model_id)
-    tokenizer.model_max_length = 512  # Set maximum length
+    tokenizer.model_max_length = 512
 
 except Exception as e:
     print(f"Error loading model: {e}")
@@ -32,11 +32,12 @@ def generate_text(prompt, max_new_tokens=128):
         output = model.generate(
             input_ids=input_ids,
             max_new_tokens=max_new_tokens,
-            do_sample=False,  # Or True for sampling
+            do_sample=False,
             eos_token_id=tokenizer.eos_token_id,
+            return_dict=True,  # Explicitly set return_dict=True
         )
 
-    generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
+    generated_text = tokenizer.decode(output.sequences[0], skip_special_tokens=True)  # Decode from sequences
     return generated_text
 
 def generate_text_streaming(prompt, max_new_tokens=128):
@@ -46,20 +47,20 @@ def generate_text_streaming(prompt, max_new_tokens=128):
         for i in range(max_new_tokens):
             output = model.generate(
                 input_ids=input_ids,
-                max_new_tokens=1,  # Generate only 1 new token at a time
-                do_sample=False,  # Or True for sampling
+                max_new_tokens=1,
+                do_sample=False,
                 eos_token_id=tokenizer.eos_token_id,
-                return_dict=True, #Return a dictionary
-                output_scores=True #Return the scores
+                return_dict=True,
+                output_scores=True,
             )
 
-            generated_token = tokenizer.decode(output.logits[0][-1].argmax(), skip_special_tokens=True) #Decode the last token only
-            yield generated_token #Yield the last token
+            generated_token = tokenizer.decode(output.logits[0][-1].argmax(), skip_special_tokens=True)
+            yield generated_token
 
-            input_ids = torch.cat([input_ids, output.sequences[:, -1:]], dim=-1) #Append the new token to the input
+            input_ids = torch.cat([input_ids, output.sequences[:, -1:]], dim=-1)
 
-            if output.sequences[0][-1] == tokenizer.eos_token_id: #Check if the end of sequence token was generated
-                break #Break the loop
+            if output.sequences[0][-1] == tokenizer.eos_token_id:
+                break
 
 def respond(message, history, system_message, max_tokens):
     prompt = f"{system_message}\n"
@@ -68,9 +69,8 @@ def respond(message, history, system_message, max_tokens):
     prompt += f"User: {message}\nAssistant:"
 
     try:
-        for token in generate_text_streaming(prompt, max_tokens): #Iterate over the generator
-            yield token #Yield each token individually
-
+        for token in generate_text_streaming(prompt, max_tokens):
+            yield token  # Yield each token individually
     except Exception as e:
         print(f"Error during generation: {e}")
         yield "An error occurred."
@@ -82,7 +82,7 @@ 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=10, label="Max new tokens"),
+        gr.Slider(minimum=1, maximum=128, value=32, step=1, label="Max new tokens"),
     ],
 )