tt

app.py

@@ -1,64 +1,65 @@
 import gradio as gr
-from huggingface_hub import InferenceClient
+import torch
+import torch.nn.functional as F
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-"""
-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("arnir0/Tiny-LLM")
+# Model selection
+MODEL_ID = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
 
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.float16, device_map="auto")
 
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    messages = [{"role": "system", "content": system_message}]
+def chat_blend(systemA, systemB, wA, wB, user_message, max_new_tokens=50, temperature=1.0, top_p=0.95):
+    # Build messages
+    promptA = systemA + "\nUser: " + user_message + "\nAssistant:"
+    promptB = systemB + "\nUser: " + user_message + "\nAssistant:"
+    input_ids_A = tokenizer(promptA, return_tensors="pt").input_ids.to(model.device)
+    input_ids_B = tokenizer(promptB, return_tensors="pt").input_ids.to(model.device)
+    output_ids = input_ids_A  # share history; or keep separate
 
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
+    for _ in range(max_new_tokens):
+        # forward last token
+        logitsA = model(input_ids=input_ids_A).logits[:, -1, :]
+        logitsB = model(input_ids=input_ids_B).logits[:, -1, :]
+        probsA = F.softmax(logitsA / temperature, dim=-1)
+        probsB = F.softmax(logitsB / temperature, dim=-1)
+        blended = wA * probsA + wB * probsB
+        # apply top_p
+        sorted_probs, sorted_idx = torch.sort(blended, descending=True)
+        cum = torch.cumsum(sorted_probs, dim=-1)
+        mask = cum > top_p
+        sorted_probs[mask] = 0
+        sorted_probs /= sorted_probs.sum(dim=-1, keepdim=True)
+        # repeat mapping back
+        new_id = sorted_idx.gather(-1, torch.multinomial(sorted_probs, num_samples=1))
+        output_ids = torch.cat([output_ids, new_id], dim=-1)
+        # append to each history
+        input_ids_A = torch.cat([input_ids_A, new_id], dim=-1)
+        input_ids_B = torch.cat([input_ids_B, new_id], dim=-1)
+        # stop on EOS
+        if new_id.item() == tokenizer.eos_token_id:
+            break
 
-    messages.append({"role": "user", "content": message})
+    decoded = tokenizer.decode(output_ids[0], skip_special_tokens=True)
+    # strip system + user prompts
+    return decoded.split("Assistant:")[-1].strip()
 
-    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
-
-
-"""
-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)",
-        ),
+iface = gr.Interface(
+    fn=chat_blend,
+    inputs=[
+        gr.Textbox(label="System Prompt A", value="You are assistant A."),
+        gr.Textbox(label="System Prompt B", value="You are assistant B."),
+        gr.Slider(label="wA", minimum=-2.0, maximum=2.0, step=0.1, value=1.0),
+        gr.Slider(label="wB", minimum=-2.0, maximum=2.0, step=0.1, value=1.0),
+        gr.Textbox(label="User message"),
+        gr.Slider(label="Max new tokens", minimum=1, maximum=200, step=1, value=50),
+        gr.Slider(label="Temperature", minimum=0.1, maximum=2.0, step=0.1, value=1.0),
+        gr.Slider(label="Top‑p", minimum=0.1, maximum=1.0, step=0.05, value=0.95),
     ],
+    outputs="text",
+    title="Blended‑LLM Chat (TinyLlama)",
+    description="Uses two system prompts and blends their token distributions using wA*p1 + wB*p2."
 )
 
-
 if __name__ == "__main__":
-    demo.launch()
+    iface.launch()

requirements.txt

@@ -1 +1,3 @@
-huggingface_hub==0.25.2
+transformers>=4.31
+torch
+gradio