Skywork-Reward-V2-Qwen3-8B 在昇腾 NPU 上的部署

python3 inference.py

import torch
import torch_npu
from transformers import AutoModelForSequenceClassification, AutoTokenizer

device = "npu:0"
model = AutoModelForSequenceClassification.from_pretrained(
    model_path,
    dtype=torch.bfloat16,
    num_labels=1,
).to(device)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(model_path)

# 构建对话并获取奖励分数
conv = [{"role": "user", "content": prompt}, {"role": "assistant", "content": response}]
formatted = tokenizer.apply_chat_template(conv, tokenize=False)
if tokenizer.bos_token is not None and formatted.startswith(tokenizer.bos_token):
    formatted = formatted[len(tokenizer.bos_token):]
inputs = tokenizer(formatted, return_tensors="pt").to(device)
with torch.no_grad():
    score = model(**inputs).logits[0][0].item()

python3 -c "
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch, torch_npu

model_path = 'model/Skywork/Skywork-Reward-V2-Qwen3-8B'
device = 'npu:0'
model = AutoModelForSequenceClassification.from_pretrained(model_path, dtype=torch.bfloat16, num_labels=1).to(device)
tok = AutoTokenizer.from_pretrained(model_path)

prompt = 'What is 2+2?'
resp = '2+2=4'
conv = tok.apply_chat_template([{'role':'user','content':prompt},{'role':'assistant','content':resp}], tokenize=False)
if tok.bos_token and conv.startswith(tok.bos_token):
    conv = conv[len(tok.bos_token):]
inp = tok(conv, return_tensors='pt').to(device)
with torch.no_grad():
    s = model(**inp).logits[0][0].item()
print(f'Reward score: {s:.4f}')
print('Smoke test PASSED')
"