File size: 2,296 Bytes
015300c
 
 
5211e0c
015300c
5211e0c
015300c
 
 
5211e0c
015300c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5211e0c
015300c
5211e0c
015300c
 
 
 
 
5211e0c
015300c
 
 
 
5211e0c
 
 
 
 
 
 
 
 
015300c
5211e0c
 
 
 
015300c
 
 
 
 
 
5211e0c
 
 
 
 
 
 
015300c
 
 
 
 
5211e0c
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
#!/usr/bin/env python3
import os
from transformers import AutoTokenizer, GPT2Tokenizer
#from megatron.initialize import initialize_megatron
from metaseq import checkpoint_utils
from transformers import OPTForCausalLM
import torch

path = "./model"
hf_path = "/home/patrick/facebook/opt-30b"


vocab_file = os.path.join(path, "gpt2-vocab.json")
merges_file = os.path.join(path, "gpt2-merges.txt")

tokenizer = GPT2Tokenizer(vocab_file, merges_file)
tokenizer.save_pretrained(path)

checkpoint = checkpoint_utils.load_model_ensemble_and_task(
    [os.path.join(path, "restored.pt")],
    arg_overrides={
        "vocab_filename": vocab_file,
        "merges_filename": merges_file,
    }
)

model = checkpoint[0][0].eval()
model = model

hf_model = OPTForCausalLM.from_pretrained(hf_path)

# forward passes
def single_batch_forward_logits(prompts):
    input_ids = tokenizer(prompts, return_tensors="pt").input_ids
    input_ids = torch.cat([torch.tensor([[0]]), input_ids], dim=-1)
    input_ids = input_ids
    with torch.no_grad():
        logits = model(input_ids)[0]
    return logits

# forward hf
def forward_hf(prompts):
    input_ids = tokenizer(prompts, return_tensors="pt").input_ids
    input_ids = torch.cat([torch.tensor([[0]]), input_ids], dim=-1)
    input_ids = input_ids
    with torch.no_grad():
        logits = hf_model(input_ids)[0]
    return logits

prompts = [
   "Today is a beautiful day and I want to",
   "In the city of",
   "Paris is the capital of France and",
   "Computers and mobile phones have taken",
]

print("Next word generation")
for prompt in prompts:
    print("-------------")
    print(f"Prompt: {prompt}...\n")
    logits_fsq = single_batch_forward_logits(prompt)
    pred_next_token = torch.argmax(logits_fsq[0, -1], -1)
    next_token = tokenizer.convert_ids_to_tokens([pred_next_token])
    next_token = next_token[0].replace("Ġ", "")
    print(f"Next word: {next_token}")
    print("-------------")
    logits = forward_hf(prompt)
    pred_next_token = torch.argmax(logits[0, -1], -1)
    next_token = tokenizer.convert_ids_to_tokens([pred_next_token])
    next_token = next_token[0].replace("Ġ", "")
    print(f"Next word: {next_token}")
    print("-------------")


print("Is equal:", torch.allclose(logits_fsq.cpu(), logits.cpu(), atol=1e-3))