import time
import torch
import tqdm
import torch.nn as nn
from datasets import load_dataset
from transformers import AutoTokenizer, LlamaForCausalLM, AutoModelForCausalLM
def build_enc(model_path):
enc = AutoTokenizer.from_pretrained(
model_path, use_fast=False, trust_remote_code=True
)
return enc
def get_llama2(model_path, seqlen=2048):
print(f'Getting official pretrained {model_path}')
model = LlamaForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, offload_folder="offload/")
model.seqlen = seqlen
enc = AutoTokenizer.from_pretrained(
model_path, use_fast=False, trust_remote_code=True
)
return model, enc
def get_qwen(model_path, seqlen=2048):
print(f'Getting official pretrained {model_path}')
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", torch_dtype=torch.bfloat16)
model.seqlen = seqlen
enc = AutoTokenizer.from_pretrained(
model_path, use_fast=False, trust_remote_code=True
)
return model, enc
def get_test_dataset(enc, seqlen):
print('Loading dataset: Wikitext2')
testenc = load_dataset('wikitext', 'wikitext-2-raw-v1', split='test')
testenc = enc("\n\n".join(testenc["text"]), return_tensors="pt")
return testenc
def get_calib_dataset(tokenizer=None, n_samples=512, block_size=512):
print('Loading dataset: pileval')
dataset = load_dataset("mit-han-lab/pile-val-backup")
dataset = dataset["validation"]
dataset = dataset.shuffle(seed=42)
samples = []
n_run = 0
for data in dataset:
line = data["text"]
line = line.strip()
line_encoded = tokenizer.encode(line)
if len(line_encoded) > 512:
continue
sample = torch.tensor([line_encoded])
if sample.numel() == 0:
continue
samples.append(sample)
n_run += 1
if n_run == n_samples:
break
cat_samples = torch.cat(samples, dim=1)
n_split = cat_samples.shape[1] // block_size
print(f" * Split into {n_split} blocks")
return [
cat_samples[:, i * block_size : (i + 1) * block_size] for i in range(n_split)
]
def infer_model(model, testenc):
test_start_time = time.time()
with torch.no_grad():
model(testenc[:, :model.seqlen].to(next(model.parameters()).device))
test_end_time = time.time()
total_time = test_end_time - test_start_time
print('Calibration time taken: ', total_time // 60, 'min ', total_time % 60, 's')
def test_ppl(model, testenc):
nlls = []
nsamples = testenc.numel() // model.seqlen
test_start_time = time.time()
for i in tqdm.tqdm(range(nsamples), desc="evaluating..."):
batch = testenc[:, (i * model.seqlen) : ((i + 1) * model.seqlen)].to(
model.device
)
with torch.no_grad():
lm_logits = model(batch).logits
shift_logits = lm_logits[:, :-1, :].contiguous().float()
shift_labels = testenc[:, (i * model.seqlen) : ((i + 1) * model.seqlen)][:, 1:]
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(
shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
)
neg_log_likelihood = loss.float() * model.seqlen
nlls.append(neg_log_likelihood)
test_end_time = time.time()
ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * model.seqlen))
total_time = test_end_time - test_start_time
print('Test time taken: ', total_time // 60, 'min ', total_time % 60, 's')
print('Score: ', ppl.item())