Update space

app.py

@@ -1,64 +1,110 @@
+import torch
 import gradio as gr
-from huggingface_hub import InferenceClient
+from tokenizers import Tokenizer
+from transformers import PreTrainedTokenizerFast
 
-"""
-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")
+from transformer_chat import TransformerChatbot
 
+# Load tokenizer & wrap for HF API
+tokenizer_obj = Tokenizer.from_file("tokenizer.json")
+hf_tok = PreTrainedTokenizerFast(
+    tokenizer_object=tokenizer_obj,
+    unk_token="[UNK]",
+    pad_token="[PAD]",
+    cls_token="[CLS]",
+    sep_token="[SEP]",
+    mask_token="[MASK]"
+)
 
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    messages = [{"role": "system", "content": system_message}]
+# Load model
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model = TransformerChatbot(
+    vocab_size=hf_tok.vocab_size,
+    d_model=512, num_heads=8, d_ff=2048,
+    num_encoder_layers=6, num_decoder_layers=6,
+    num_roles=2, max_turns=16, num_slots=22,
+    dropout=0.1
+).to(device)
+model.load_state_dict(torch.load("atis_transformer.pt", map_location=device))
+model.eval()
 
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
+# Generation function
+def chat_fn(prompt):
+    # Encode user input
+    enc = hf_tok(prompt, return_tensors="pt", padding=True, truncation=True, max_length=128)
+    src_ids  = enc.input_ids.to(device)
+    # For cross-attention, we don't need to mask the encoder output
+    src_mask = None
 
-    messages.append({"role": "user", "content": message})
+    # Roles & turns (user=0)
+    roles = torch.zeros_like(src_ids)
+    turns = torch.zeros_like(src_ids)
 
-    response = ""
+    # Encode
+    with torch.no_grad():
+        enc_out = model.encode(src_ids, roles, turns, src_mask)
 
-    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
+        # Generate reply token-by-token
+        cls_id = hf_tok.cls_token_id
+        sep_id = hf_tok.sep_token_id
+        dec_input = torch.tensor([[cls_id]], device=device)
+        dec_roles = torch.zeros_like(dec_input)
+        dec_turns = torch.zeros_like(dec_input)
 
-        response += token
-        yield response
+        generated = []
+        for step in range(50):
+            T = dec_input.size(1)
+            # Create causal mask for decoder (upper triangular = masked)
+            # PyTorch's MultiheadAttention expects a 2D mask where True = masked
+            causal_mask = torch.triu(torch.ones((T, T), device=device), diagonal=1).bool()
+            tgt_mask = causal_mask
 
+            logits = model.decode(dec_input, enc_out, dec_roles, dec_turns, src_mask, tgt_mask)
+            
+            # Get the last token's logits
+            last_logits = logits[0, -1, :]
+            
+            # Apply repetition penalty
+            if generated:
+                for token_id in set(generated):
+                    last_logits[token_id] *= 0.7  # Penalize repeated tokens
+            
+            # Sample with temperature instead of greedy decoding
+            temperature = 0.8
+            probs = torch.softmax(last_logits / temperature, dim=-1)
+            next_id = torch.multinomial(probs, 1)
+            
+            # Debug: print the token being generated
+            token_text = hf_tok.decode([next_id.item()])
+            print(f"Step {step}: Generated token ID {next_id.item()} -> '{token_text}'")
+            
+            if next_id.item() == sep_id:
+                print("Found SEP token, stopping generation")
+                break
+                
+            generated.append(next_id.item())
+            dec_input = torch.cat([dec_input, next_id.unsqueeze(0)], dim=1)
+            dec_roles = torch.cat([dec_roles, torch.zeros_like(next_id).unsqueeze(0)], dim=1)
+            dec_turns = torch.cat([dec_turns, torch.zeros_like(next_id).unsqueeze(0)], dim=1)
+            
+            # Early stopping if we're stuck in a loop
+            if len(generated) >= 3 and len(set(generated[-3:])) == 1:
+                print("Detected repetition loop, stopping generation")
+                break
 
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
-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)",
-        ),
-    ],
-)
+        output_ids = [cls_id] + generated + [sep_id]
+        reply = hf_tok.decode(output_ids, skip_special_tokens=True)
 
+    return reply
+
+# Build Gradio interface
+interface = gr.Interface(
+    fn=chat_fn,
+    inputs=gr.Textbox(lines=2, placeholder="Enter your question here..."),
+    outputs="text",
+    title="Transformer Chatbot Demo (currently trained with ATIS dataset)",
+    description="Ask flight-related questions and get an answer."
+)
 
 if __name__ == "__main__":
-    demo.launch()
+    interface.launch(share=True)

requirements.txt

@@ -1 +1,5 @@
-huggingface_hub==0.25.2
+huggingface_hub==0.25.2
+torch
+transformers
+tokenizers
+datasets