Views
No views yet
value_head.safetensors.1import os
2import torch
3from transformers import AutoModelForCausalLM, AutoTokenizer
4from safetensors.torch import load_file
5
6model_dir = "/home/huimin/New-Proj/value_model-1.5b/hf_converted_model"
7base = AutoModelForCausalLM.from_pretrained(model_dir, torch_dtype=torch.bfloat16)
8value_head = torch.nn.Linear(base.config.hidden_size, 1, bias=False)
9state = load_file(os.path.join(model_dir, "value_head.safetensors"))
10value_head.load_state_dict({"weight": state["value_head.weight"]})
11
12tok = AutoTokenizer.from_pretrained(model_dir)
13inputs = tok("Hello", return_tensors="pt")
14outputs = base(**inputs, output_hidden_states=True)
15last = outputs.hidden_states[-1]
16values = value_head(last).squeeze(-1)
17print(values.shape)