GilbertAkham/deepseek-R1-multitask-lora

1import torch2from transformers import AutoTokenizer, AutoModelForCausalLM3from peft import PeftModel4from huggingface_hub import snapshot_download5 6# === Base & adapter config ===7BASE_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"8ADAPTER_PATH = "GilbertAkham/deepseek-R1-multitask-lora"9 10# === System message ===11SYSTEM_PROMPT = (12    "You are Chat-Bot, a helpful and logical assistant trained for reasoning, "13    "email, chatting, summarization, story continuation, and report writing.\n\n"14)15 16class EndpointHandler:17    def __init__(self, path=""):18        print("🚀 Loading base model...")19        self.tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)20 21        base_model = AutoModelForCausalLM.from_pretrained(22            BASE_MODEL,23            torch_dtype=torch.float16,24            device_map="auto",25            trust_remote_code=True26        )27 28        print(f"🔗 Downloading LoRA adapter from {ADAPTER_PATH}...")29        adapter_local_path = snapshot_download(repo_id=ADAPTER_PATH, allow_patterns=["*adapter*"])30        print(f"📁 Adapter files cached at {adapter_local_path}")31 32        print("🧩 Attaching LoRA adapter...")33        self.model = PeftModel.from_pretrained(base_model, adapter_local_path)34        self.model.eval()35 36        print("✅ Model + LoRA adapter loaded successfully.")37 38    def __call__(self, data):39        # === Combine system + user prompt ===40        user_prompt = data.get("inputs", "")41        full_prompt = SYSTEM_PROMPT + user_prompt42 43        params = data.get("parameters", {})44        max_new_tokens = params.get("max_new_tokens", 512)45        temperature = params.get("temperature", 0.7)46        top_p = params.get("top_p", 0.9)47 48        # === Tokenize and run generation ===49        inputs = self.tokenizer(full_prompt, return_tensors="pt").to(self.model.device)50        with torch.no_grad():51            outputs = self.model.generate(52                **inputs,53                max_new_tokens=max_new_tokens,54                temperature=temperature,55                top_p=top_p,56                do_sample=True,57                pad_token_id=self.tokenizer.eos_token_id,58                eos_token_id=self.tokenizer.eos_token_id,59            )60 61        # === Decode and strip system message ===62        text = self.tokenizer.decode(outputs[0], skip_special_tokens=True)63        if text.startswith(SYSTEM_PROMPT):64            text = text[len(SYSTEM_PROMPT):].strip()65 66        return {"generated_text": text}67