app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import re, json
from html import escape

# ─── Configuration ─────────────────────────────────────────────────────────
PRIMARY_MODEL = "Smilyai-labs/Sam-reason-A3"
FALLBACK_MODEL = "Smilyai-labs/Sam-reason-A1"
USAGE_LIMIT   = 5
RESET_MS      = 20 * 60 * 1000   # 20 minutes in milliseconds
device        = "cuda" if torch.cuda.is_available() else "cpu"

primary_model = primary_tokenizer = None
fallback_model = fallback_tokenizer = None

# ─── Model Loading ─────────────────────────────────────────────────────────
def load_models():
    global primary_model, primary_tokenizer, fallback_model, fallback_tokenizer
    primary_tokenizer = AutoTokenizer.from_pretrained(PRIMARY_MODEL, trust_remote_code=True)
    primary_model     = AutoModelForCausalLM.from_pretrained(
        PRIMARY_MODEL, torch_dtype=torch.float16
    ).to(device).eval()
    fallback_tokenizer= AutoTokenizer.from_pretrained(FALLBACK_MODEL, trust_remote_code=True)
    fallback_model    = AutoModelForCausalLM.from_pretrained(
        FALLBACK_MODEL, torch_dtype=torch.float16
    ).to(device).eval()
    return f"✅ Loaded {PRIMARY_MODEL} (fallback: {FALLBACK_MODEL})"

# ─── Build Chat Prompt ──────────────────────────────────────────────────────
def build_chat_prompt(history, user_input, reasoning_enabled):
    system_flag = "/think" if reasoning_enabled else "/no_think"
    prompt = f"<|system|>\n{system_flag}\n"
    for u, a in history:
        prompt += f"<|user|>\n{u}\n<|assistant|>\n{a}\n"
    prompt += f"<|user|>\n{user_input}\n<|assistant|>\n"
    return prompt

# ─── Collapse <think> Blocks ────────────────────────────────────────────────
def format_thinking(text):
    match = re.search(r"<think>(.*?)</think>", text, re.DOTALL)
    if not match:
        return escape(text)
    reasoning = escape(match.group(1).strip())
    visible   = re.sub(r"<think>.*?</think>", "[thinking...]", text, flags=re.DOTALL).strip()
    return (
        escape(visible)
        + "<br><details><summary>🧠 Show reasoning</summary>"
        + f"<pre>{reasoning}</pre></details>"
    )

# ─── Token-by-Token Streaming (Stops on <|user|>) ─────────────────────────
def generate_stream(prompt, use_fallback=False,
                    max_length=100, temperature=0.2, top_p=0.9):
    model     = fallback_model if use_fallback else primary_model
    tokenizer = fallback_tokenizer if use_fallback else primary_tokenizer
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
    generated = input_ids
    assistant_text = ""

    for _ in range(max_length):
        # 1) Get next-token logits and apply top-p
        logits = model(generated).logits[:, -1, :] / temperature
        sorted_logits, idxs = torch.sort(logits, descending=True)
        probs = torch.softmax(sorted_logits, dim=-1).cumsum(dim=-1)
        mask = probs > top_p
        mask[..., 1:] = mask[..., :-1].clone()
        mask[..., 0]  = 0
        filtered = logits.clone()
        filtered[:, idxs[mask]] = -float("Inf")

        # 2) Sample and append
        next_token = torch.multinomial(torch.softmax(filtered, dim=-1), 1)
        generated = torch.cat([generated, next_token], dim=-1)
        new_text = tokenizer.decode(next_token[0], skip_special_tokens=False)
        assistant_text += new_text

        # 3) Remove any leading assistant tag
        if assistant_text.startswith("<|assistant|>"):
            assistant_text = assistant_text[len("<|assistant|>"):]

        # 4) If we see a user‐turn tag, truncate and bail
        if "<|user|>" in assistant_text:
            assistant_text = assistant_text.split("<|user|>")[0]
            yield assistant_text
            break

        # 5) Otherwise stream clean assistant text
        yield assistant_text

        # 6) End if EOS
        if next_token.item() == tokenizer.eos_token_id:
            break

# ─── Main Chat Handler ──────────────────────────────────────────────────────
def respond(message, history, reasoning_enabled, limit_json):
    # parse client-side usage info
    info  = json.loads(limit_json) if limit_json else {"count": 0}
    count = info.get("count", 0)
    use_fallback = count > USAGE_LIMIT
    remaining    = max(0, USAGE_LIMIT - count)
    model_label  = "A3" if not use_fallback else "Fallback A1"

    # initial yield to set "Generating…"
    prompt = build_chat_prompt(history, message.strip(), reasoning_enabled)
    history = history + [[message, ""]]
    yield history, history, f"🧠 A3 left: {remaining}", "Generating…"

    # stream assistant reply
    for chunk in generate_stream(prompt, use_fallback):
        formatted = format_thinking(chunk)
        history[-1][1] = (
            f"{formatted}<br><sub style='color:gray'>({model_label})</sub>"
        )
        yield history, history, f"🧠 A3 left: {remaining}", "Generating…"

    # final yield resets button text
    yield history, history, f"🧠 A3 left: {remaining}", "Send"

# ─── Clear Chat ─────────────────────────────────────────────────────────────
def clear_chat():
    return [], [], "🧠 A3 left: 5", "Send"

# ─── Gradio UI ──────────────────────────────────────────────────────────────
with gr.Blocks() as demo:
    # Inject client-side JS + CSS
    gr.HTML(f"""
<script>
  function updateUsageLimit() {{
    const key = "samai_limit";
    const now = Date.now();
    let rec = JSON.parse(localStorage.getItem(key) || "null");
    if (!rec || now - rec.lastSeen > {RESET_MS}) {{
      rec = {{count:0, lastSeen: now}};
    }}
    rec.count += 1;
    rec.lastSeen = now;
    localStorage.setItem(key, JSON.stringify(rec));
    document.getElementById("limit_json").value = JSON.stringify(rec);
  }}
  document.addEventListener("DOMContentLoaded", () => {{
    const btn = document.getElementById("send_btn");
    btn.addEventListener("click", () => {{
      updateUsageLimit();
      btn.innerText = "Generating…";
    }});
  }});
</script>
<style>
  .send-circle {{
    border-radius: 50%;
    height: 40px;
    width: 40px;
    padding: 0;
    font-size: 12px;
    text-align: center;
  }}
</style>
    """)

    gr.Markdown("# 🤖 SamAI – Chat Reasoning (Final)")

    # Hidden textbox ferrying usage JSON from JS → Python
    limit_json    = gr.Textbox(visible=False, elem_id="limit_json")
    model_status  = gr.Textbox(interactive=False, label="Model Status")
    usage_counter = gr.Textbox("🧠 A3 left: 5", interactive=False, show_label=False)

    chat_box   = gr.Chatbot(type="tuples")
    chat_state = gr.State([])

    with gr.Row():
        user_input    = gr.Textbox(placeholder="Ask anything...", show_label=False, scale=6)
        reason_toggle = gr.Checkbox(label="Reason", value=True, scale=1)
        send_btn      = gr.Button("Send", elem_id="send_btn", elem_classes=["send-circle"], scale=1)

    clear_btn = gr.Button("Clear")

    model_status.value = load_models()

    # Bind Send button -> respond()
    send_btn.click(
        fn=respond,
        inputs=[user_input, chat_state, reason_toggle, limit_json],
        outputs=[chat_box, chat_state, usage_counter, send_btn]
    )

    clear_btn.click(
        fn=clear_chat,
        inputs=[], 
        outputs=[chat_box, chat_state, usage_counter, send_btn]
    )

demo.queue()
demo.launch()