Update app.py

app.py

@@ -1,74 +1,78 @@
 import gradio as gr
 from huggingface_hub import InferenceClient
+import tensorflow as tf
+from huggingface_hub import login, create_repo, upload_file
+from transformers import AutoTokenizer, TFAutoModelForCausalLM
+policy = tf.keras.mixed_precision.Policy('mixed_bfloat16')
+tf.keras.mixed_precision.set_global_policy(policy)
+strategy = tf.distribute.MultiWorkerMirroredStrategy()
 
-"""
-For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-"""
-client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
+login(os.environ.get("hf_token"))
 
-# Simple rule-based response function
-def rule_based_response(history, message):
-    if "hello" in message.lower():
-        return "Hello! How can I help you today?"
-    elif "how are you" in message.lower():
-        return "I'm doing great, thanks for asking! How about you?"
-    elif "bye" in message.lower():
-        return "Goodbye! Have a nice day!"
-    else:
-        return "Sorry, I don't understand. Can you ask something else?"
+name = "WICKED4950/GPT2-InstEsther0.21eV3.1"
+tokenizer = AutoTokenizer.from_pretrained(name)
+tokenizer.pad_token = tokenizer.eos_token
+with strategy.scope():
+    model = TFAutoModelForCausalLM.from_pretrained(name)
 
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    # Rule-based response logic (for now)
-    response = rule_based_response(history, message)
-    
-    # If the rule-based model cannot respond, fall back to the HuggingFace model
-    if response:
-        return response, history
+def raw_pred(input, model, tokenizer, max_length=50, temperature=0.2):
+    input_ids = tokenizer.encode(input, return_tensors='tf')
 
-    # Otherwise, use the HuggingFace model
-    messages = [{"role": "system", "content": system_message}]
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
+    # Initialize variables
+    generated_ids = input_ids
+    stop_token_id = tokenizer.encode("<|SOH|>", add_special_tokens=False)[0]  # ID for <|SOH|>
+    all_generated_tokens = []  # To store generated token IDs
+    tokens_yielded = []  # To store tokens as they are yielded
 
-    messages.append({"role": "user", "content": message})
+    with strategy.scope():
+        for _ in range(max_length // 1):  # Generate in chunks of 3 tokens
+            # Generate three tokens at a time
+            outputs = model.generate(
+                generated_ids,
+                max_length=generated_ids.shape[1] + 1,  # Increment max length by 3
+                temperature=temperature,
+                pad_token_id=tokenizer.eos_token_id,
+                eos_token_id=stop_token_id,  # Stop generation at <|SOH|>
+                do_sample=True,
+                num_return_sequences=1
+            )
 
-    response = ""
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
-        response += token
-        yield response
+            # Get the newly generated tokens (last 3 tokens)
+            new_tokens = outputs[0, -1:]
+            generated_ids = outputs  # Update the generated_ids with the new tokens
+
+            # Store the generated tokens as numbers (IDs)
+            all_generated_tokens.extend(new_tokens.numpy().tolist())
+
+            # Decode and yield the tokens as they are generated (as numbers)
+            tokens_text = tokenizer.decode(new_tokens, skip_special_tokens=False)
+            tokens_yielded.append(tokens_text)
+            yield tokens_text
 
+            # Stop if the generated tokens include <|SOH|>
+            if stop_token_id in new_tokens.numpy():
+                final_text = tokenizer.decode(all_generated_tokens, skip_special_tokens=False)
+                yield ("<|Clean|>" + final_text)
+                break
+
+def respond(message, history):
+    # Prepare input for the model
+    give_mod = ""
+    for chunk in history:
+        give_mod = give_mod + "<|SOH|>" + chunk[0] + "<|SOB|>" + chunk[1]
+    give_mod = give_mod + "<|SOH|>" + message + "<|SOB|>"
+    print(give_mod)
+    response = ""
+    for token in raw_pred(give_mod, model, tokenizer):
+        if "<|Clean|>" in token:
+            response = token
+        else:
+            response += token
+        yield response.replace("<|SOH|>","").replace("<|Clean|>","")
+    print(response)
 # Gradio Chat Interface Setup
 demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
+    respond
 )
 
 if __name__ == "__main__":