Keep ZeroGPU demo on stable SFT base

README.md

@@ -17,9 +17,10 @@ models:
 
 # tinyLLM 0.51B Demo
 
-This free ZeroGPU Space loads the public tinyLLM SFT model and switches the ARC
-GRPO or IFEval OPD LoRA without keeping multiple base models in memory.
+This free ZeroGPU Space loads the public tinyLLM 0.51B SFT checkpoint and offers
+a small streaming chat demo.
 
 The first start is slower while the 0.51B checkpoint is downloaded. The full
 multimodal demo remains available in the GitHub repository; this small hosted
-Space focuses on text and adapter switching.
+Space focuses on the stable text base. ARC GRPO and IFEval OPD adapters are
+linked from the app and can be loaded locally with `load_lora_pretrained()`.

app.py

@@ -10,15 +10,9 @@ from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStream
 
 
 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()
 
 
@@ -32,36 +26,16 @@ def load_runtime():
                 trust_remote_code=True,
                 dtype=torch.bfloat16,
             ).to("cuda").eval()
-            # Register both adapters before ZeroGPU packs the model tensors.
-            # Adding CUDA parameters later inside a request is less reliable.
-            for adapter_repo in (repo for repo in ADAPTERS.values() if repo):
-                _model.load_lora_pretrained(adapter_repo, activate=False)
-                _loaded_adapters.add(adapter_repo)
-            _model.activate_single_lora(None)
         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)
-
-
 @spaces.GPU(duration=120)
-def respond(message: str, history: list[dict], mode: str, max_new_tokens: int):
+def respond(message: str, history: list[dict], 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
@@ -108,19 +82,17 @@ load_runtime()
 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。"
+        "从零训练的中英双语小模型。这里在线运行公开的 0.51B SFT checkpoint。"
     )
-    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")
+    max_tokens = gr.Slider(32, 512, value=160, step=32, label="Maximum new tokens")
     gr.ChatInterface(
         fn=respond,
-        additional_inputs=[mode, max_tokens],
+        additional_inputs=[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],
+            ["用两句话介绍一下你自己。", 128],
+            ["What is 12 minus 5? Give a short answer.", 96],
+            ["Write three bullet points about small language models.", 160],
         ],
         cache_examples=False,
         concurrency_limit=1,
@@ -128,7 +100,9 @@ with gr.Blocks(title="tinyLLM 0.51B") as demo:
     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。"
+        "[ARC GRPO](https://huggingface.co/chris0809/tinyLLM-0.51B-ARC-GRPO) 和 "
+        "[IFEval OPD](https://huggingface.co/chris0809/tinyLLM-0.51B-IFEval-OPD) LoRA 也可单独下载。"
+        "完整图片演示建议使用 GitHub 中的本地 Demo；这个免费 Space 主要用于快速验证文本基座。"
     )