Add tinyLLM demo

README.md

@@ -1,13 +1,25 @@
 ---
-title: TinyLLM Demo
-emoji: ⚡
-colorFrom: yellow
-colorTo: blue
+title: tinyLLM 0.51B Demo
+emoji: 🧩
+colorFrom: blue
+colorTo: indigo
 sdk: gradio
-sdk_version: 6.26.0
-python_version: '3.12'
+sdk_version: 5.49.1
 app_file: app.py
 pinned: false
+license: other
+models:
+- chris0809/tinyLLM-0.51B-SFT
+- chris0809/tinyLLM-0.51B-ARC-GRPO
+- chris0809/tinyLLM-0.51B-IFEval-OPD
+- chris0809/tinyLLM-0.51B-VLM
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# tinyLLM 0.51B Demo
+
+This CPU Space loads the public tinyLLM SFT model and switches the ARC GRPO or
+IFEval OPD LoRA without keeping multiple base models in memory.
+
+The first response is slower because the 0.51B checkpoint is downloaded and
+loaded on demand. The full multimodal demo remains available in the GitHub
+repository; the free CPU Space focuses on text and adapter switching.

app.py

@@ -0,0 +1,120 @@
+"""Small CPU-friendly Hugging Face Space for the public tinyLLM checkpoints."""
+from __future__ import annotations
+
+import threading
+
+import gradio as gr
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+
+
+BASE_REPO = "chris0809/tinyLLM-0.51B-SFT"
+ADAPTERS = {
+    "SFT base": None,
+    "ARC-Easy GRPO": "chris0809/tinyLLM-0.51B-ARC-GRPO",
+    "IFEval OPD": "chris0809/tinyLLM-0.51B-IFEval-OPD",
+}
+
+_model = None
+_tokenizer = None
+_loaded_adapters: set[str] = set()
+_lock = threading.RLock()
+
+
+def load_runtime():
+    global _model, _tokenizer
+    with _lock:
+        if _model is None:
+            _tokenizer = AutoTokenizer.from_pretrained(BASE_REPO)
+            _model = AutoModelForCausalLM.from_pretrained(
+                BASE_REPO,
+                trust_remote_code=True,
+                dtype=torch.float32,
+            ).eval()
+        return _model, _tokenizer
+
+
+def select_adapter(model, mode: str) -> None:
+    adapter_repo = ADAPTERS[mode]
+    if adapter_repo is None:
+        model.activate_single_lora(None)
+        return
+    if adapter_repo not in _loaded_adapters:
+        model.load_lora_pretrained(adapter_repo)
+        _loaded_adapters.add(adapter_repo)
+    else:
+        name = "arc" if mode == "ARC-Easy GRPO" else "opd_ifeval_rank64"
+        model.activate_single_lora(name)
+
+
+def respond(message: str, history: list[dict], mode: str, max_new_tokens: int):
+    if not message.strip():
+        yield "请输入一个问题。"
+        return
+    with _lock:
+        model, tokenizer = load_runtime()
+        select_adapter(model, mode)
+        messages = [
+            {"role": row["role"], "content": row["content"]}
+            for row in history
+            if row.get("role") in {"user", "assistant"} and isinstance(row.get("content"), str)
+        ]
+        messages.append({"role": "user", "content": message.strip()})
+        prompt = tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True,
+        )
+        inputs = tokenizer(prompt, return_tensors="pt")
+        streamer = TextIteratorStreamer(
+            tokenizer,
+            skip_prompt=True,
+            skip_special_tokens=True,
+        )
+        kwargs = {
+            **inputs,
+            "streamer": streamer,
+            "max_new_tokens": int(max_new_tokens),
+            "do_sample": False,
+            "repetition_penalty": 1.10,
+            "no_repeat_ngram_size": 12,
+            "use_cache": True,
+        }
+        worker = threading.Thread(target=model.generate, kwargs=kwargs, daemon=True)
+        worker.start()
+        text = ""
+        for piece in streamer:
+            text += piece
+            yield text
+        worker.join()
+
+
+with gr.Blocks(title="tinyLLM 0.51B") as demo:
+    gr.Markdown(
+        "# tinyLLM 0.51B\n"
+        "从零训练的中英双语小模型。可在同一个基座上切换 SFT、ARC GRPO 和 IFEval OPD LoRA。"
+    )
+    with gr.Row():
+        mode = gr.Dropdown(list(ADAPTERS), value="SFT base", label="Model / LoRA")
+        max_tokens = gr.Slider(32, 512, value=160, step=32, label="Maximum new tokens")
+    gr.ChatInterface(
+        fn=respond,
+        additional_inputs=[mode, max_tokens],
+        type="messages",
+        examples=[
+            ["用两句话介绍一下你自己。", "SFT base", 128],
+            ["Which material is attracted to a magnet? A. Wood B. Iron C. Plastic D. Glass", "ARC-Easy GRPO", 96],
+            ["Write exactly three bullet points about small language models.", "IFEval OPD", 160],
+        ],
+        cache_examples=False,
+        concurrency_limit=1,
+    )
+    gr.Markdown(
+        "多模态权重已发布在 [tinyLLM-0.51B-VLM]"
+        "(https://huggingface.co/chris0809/tinyLLM-0.51B-VLM)。"
+        "完整图片演示建议使用 GitHub 中的本地 Demo；免费 CPU Space 主要用于快速验证文本模型和 LoRA。"
+    )
+
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,5 @@
+transformers==4.57.1
+huggingface-hub>=0.34.4
+safetensors>=0.4.5
+sentencepiece>=0.2.0
+protobuf>=4.25