Delete app.py

app.py

@@ -1,96 +0,0 @@
-import time, threading
-import gradio as gr
-import torch, spaces
-from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
-
-# ---------- Config ----------
-MODEL_ID = "WeiboAI/VibeThinker-1.5B"
-SYSTEM_PROMPT = "You are a concise solver. Give one clear final answer."
-
-MAX_INPUT_TOKENS = 384        # clip context so first token comes fast
-MAX_NEW_TOKENS   = 96         # keep short to finish inside the GPU slice
-DO_SAMPLE        = False      # deterministic decode = smoother & faster on ZeroGPU
-TEMPERATURE      = 0.4        # only used if DO_SAMPLE=True
-TOP_P            = 0.9
-FIRST_TOKEN_TIMEOUT = 3       # if no token arrives, likely no GPU slice yet
-
-# ---------- Load model once ----------
-print(f"⏳ Loading {MODEL_ID} …", flush=True)
-tok = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(
-    MODEL_ID,
-    trust_remote_code=True,
-    low_cpu_mem_usage=True,
-    dtype=torch.bfloat16,      # (use dtype, not torch_dtype)
-    device_map="auto",
-).eval()
-print("✅ Model ready.", flush=True)
-
-
-def _to_inputs(messages_text: str):
-    ids = tok([messages_text], return_tensors="pt")
-    if ids["input_ids"].shape[-1] > MAX_INPUT_TOKENS:
-        ids = {k: v[:, -MAX_INPUT_TOKENS:] for k, v in ids.items()}
-    return {k: v.to(model.device) for k, v in ids.items()}
-
-
-@spaces.GPU(duration=60)  # shorter slice => scheduled more reliably
-def respond(message, history):
-    history = history or []
-    msgs = [{"role": "system", "content": SYSTEM_PROMPT}, *history,
-            {"role": "user", "content": str(message)}]
-
-    prompt_text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
-    inputs = _to_inputs(prompt_text)
-
-    # fast, fine-grained streaming
-    streamer = TextIteratorStreamer(tok, skip_prompt=True, skip_special_tokens=True, timeout=0.05)
-
-    gen_kwargs = dict(
-        **inputs,
-        streamer=streamer,
-        do_sample=DO_SAMPLE,
-        temperature=TEMPERATURE,
-        top_p=TOP_P,
-        repetition_penalty=1.15,          # tame short loops
-        max_new_tokens=MAX_NEW_TOKENS,
-        pad_token_id=tok.eos_token_id,
-        use_cache=True,
-        eos_token_id=tok.eos_token_id,
-    )
-
-    th = threading.Thread(target=model.generate, kwargs=gen_kwargs, daemon=True)
-    th.start()
-
-    out = list(history) + [{"role": "assistant", "content": ""}]
-    got_first = False
-    start = time.time()
-
-    # yield every token (no bursting)
-    for chunk in streamer:
-        got_first = True
-        out[-1]["content"] += chunk
-        yield out
-
-    # If no token ever arrived, likely no GPU window was granted
-    if not got_first and (time.time() - start) >= FIRST_TOKEN_TIMEOUT:
-        out[-1]["content"] = "(No GPU slot yet — press Send again.)"
-        yield out
-
-
-# ---------- UI ----------
-with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("## 💡 VibeThinker-1.5B — ZeroGPU slice (smooth streaming)")
-    chat = gr.Chatbot(type="messages", height=520)
-    box  = gr.Textbox(placeholder="Ask a question…")
-    send = gr.Button("Send", variant="primary")
-
-    def pipeline(msg, hist):
-        for hist in respond(msg, hist):
-            yield "", hist
-
-    box.submit(pipeline, [box, chat], [box, chat])
-    send.click(pipeline, [box, chat], [box, chat])
-
-if __name__ == "__main__":
-    demo.queue(max_size=16).launch()