Create app.py

app.py

@@ -0,0 +1,81 @@
+import gradio as gr
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+from bitsandbytes.nn import Int8Params, Int8Linear
+from transformers.utils.quantization_config import BitsAndBytesConfig
+
+# モデルとトークナイザの読み込み
+model_name = "Qwen/Qwen2.5-7B-Instruct"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+# 4bit量子化設定
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_quant_type="nf4"
+)
+
+# モデルの読み込みと量子化
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    device_map="auto",
+    torch_dtype=torch.bfloat16,
+    quantization_config=quantization_config
+)
+
+def respond(
+    message,
+    history: list[tuple[str, str]],
+    system_message,
+    max_tokens,
+    temperature,
+    top_p,
+):
+    # メッセージの準備
+    messages = [{"role": "system", "content": system_message}]
+
+    for val in history:
+        if val[0]:
+            messages.append({"role": "user", "content": val[0]})
+        if val[1]:
+            messages.append({"role": "assistant", "content": val[1]})
+
+    messages.append({"role": "user", "content": message})
+
+    # メッセージをトークナイザに通す
+    input_ids = tokenizer([message], return_tensors="pt").input_ids.to(model.device)
+
+    # モデルの推論
+    output_ids = model.generate(
+        input_ids,
+        max_length=max_tokens,
+        temperature=temperature,
+        top_p=top_p,
+        do_sample=True,
+    )
+
+    response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
+    response = response[len(message):]  # 入力メッセージを削除
+    return response
+
+# インターフェース
+demo = gr.ChatInterface(
+    respond,
+    additional_inputs=[
+        gr.Textbox(value="ユーザーの応答と依頼に答えてください。ポジティブに", label="システムメッセージ"),
+        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="新規トークン最大"),
+        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="温度"),
+        gr.Slider(
+            minimum=0.1,
+            maximum=1.0,
+            value=0.95,
+            step=0.05,
+            label="Top-p (核 sampling)",
+        ),
+    ],
+    concurrency_limit=30  # 例: 同時に4つのリクエストを処理
+)
+
+if __name__ == "__main__":
+    demo.launch()