2023-07-10 19:19:16 +00:00
|
|
|
import pytest
|
|
|
|
import torch
|
|
|
|
import math
|
|
|
|
|
2023-07-11 22:40:20 +00:00
|
|
|
from itertools import product
|
|
|
|
|
|
|
|
import transformers
|
2023-07-10 19:19:16 +00:00
|
|
|
from transformers import (
|
|
|
|
AutoConfig,
|
|
|
|
AutoModelForCausalLM,
|
|
|
|
AutoTokenizer,
|
|
|
|
BitsAndBytesConfig,
|
|
|
|
GenerationConfig,
|
|
|
|
set_seed,
|
|
|
|
|
|
|
|
)
|
2023-07-11 22:40:20 +00:00
|
|
|
|
2023-07-10 19:19:16 +00:00
|
|
|
|
|
|
|
|
|
|
|
def get_4bit_config():
|
|
|
|
return BitsAndBytesConfig(
|
|
|
|
load_in_4bit=True,
|
|
|
|
load_in_8bit=False,
|
|
|
|
llm_int8_threshold=6.0,
|
|
|
|
llm_int8_has_fp16_weight=False,
|
|
|
|
bnb_4bit_compute_dtype=torch.float16,
|
|
|
|
bnb_4bit_use_double_quant=True,
|
|
|
|
bnb_4bit_quant_type='nf4',
|
|
|
|
)
|
|
|
|
|
|
|
|
|
2023-07-11 22:40:20 +00:00
|
|
|
def get_model_and_tokenizer(config):
|
|
|
|
model_name_or_path, quant_type = config
|
|
|
|
bnb_config = get_4bit_config()
|
|
|
|
if quant_type == '16bit':
|
|
|
|
bnb_config.load_in_4bit = False
|
|
|
|
else:
|
|
|
|
bnb_config.bnb_4bit_quant_type= quant_type
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
|
|
|
|
quantization_config=bnb_config,
|
|
|
|
max_memory={0:'48GB'},
|
|
|
|
device_map='auto',
|
|
|
|
torch_dtype=torch.bfloat16
|
|
|
|
).eval()
|
2023-07-10 19:19:16 +00:00
|
|
|
|
2023-07-11 22:40:20 +00:00
|
|
|
tokenizer = transformers.AutoTokenizer.from_pretrained(model_name_or_path)
|
|
|
|
|
|
|
|
return model, tokenizer
|
2023-07-10 19:19:16 +00:00
|
|
|
|
|
|
|
def get_prompt_for_generation_eval(text, add_roles=True):
|
|
|
|
description = (
|
|
|
|
"A chat between a curious human and an artificial intelligence assistant. "
|
|
|
|
"The assistant gives helpful, detailed, and polite answers to the user's questions."
|
|
|
|
)
|
|
|
|
if add_roles:
|
|
|
|
prompt = f'{description} ### Human: {text} ### Assistant:'
|
|
|
|
else:
|
|
|
|
prompt = f'{description} {text}'
|
|
|
|
return prompt
|
|
|
|
|
|
|
|
def generate(model, tokenizer, text, generation_config, prompt_func=get_prompt_for_generation_eval):
|
|
|
|
text = prompt_func(text)
|
|
|
|
inputs = tokenizer(text, return_tensors="pt").to('cuda:0')
|
|
|
|
outputs = model.generate(inputs=inputs['input_ids'], generation_config=generation_config)
|
|
|
|
return tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|
|
|
|
2023-07-11 22:40:20 +00:00
|
|
|
models = ['huggyllama/llama-7b', 'bigscience/bloom-1b7']
|
2023-07-12 00:13:33 +00:00
|
|
|
dtypes = ['nf4', 'fp4']
|
2023-07-11 22:40:20 +00:00
|
|
|
load_in_4bit = [True, False]
|
|
|
|
values = list(product(models, dtypes))
|
|
|
|
strfunc = lambda lst: [str(x) for x in lst]
|
|
|
|
ids = ['_'.join(strfunc(x)) for x in values]
|
|
|
|
@pytest.fixture(scope='session', params=values, ids=ids)
|
|
|
|
def model_and_tokenizer(request):
|
|
|
|
model, tokenizer = get_model_and_tokenizer(request.param)
|
2023-07-12 00:13:33 +00:00
|
|
|
yield request.param, model, tokenizer
|
2023-07-11 22:40:20 +00:00
|
|
|
del model
|
|
|
|
|
2023-07-12 00:13:33 +00:00
|
|
|
@pytest.mark.parametrize("DQ", [True, False], ids=['DQ_True', 'DQ_False'])
|
2023-07-11 22:40:20 +00:00
|
|
|
@pytest.mark.parametrize("inference_kernel", [True, False], ids=['inference_kernel_True', 'inference_kernel_False'])
|
2023-07-12 00:13:33 +00:00
|
|
|
#@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32], ids=['fp16', 'bf16', 'fp32'])
|
|
|
|
def test_pi(model_and_tokenizer, inference_kernel, DQ):
|
|
|
|
print('')
|
|
|
|
dtype = torch.float16
|
2023-07-11 22:40:20 +00:00
|
|
|
|
2023-07-12 00:13:33 +00:00
|
|
|
fixture_config, model, tokenizer = model_and_tokenizer
|
2023-07-10 19:19:16 +00:00
|
|
|
|
|
|
|
generation_config = transformers.GenerationConfig(
|
2023-07-11 22:40:20 +00:00
|
|
|
max_new_tokens=20,
|
2023-07-10 19:19:16 +00:00
|
|
|
do_sample=True,
|
|
|
|
top_p=0.9,
|
|
|
|
temperature=0.7,
|
|
|
|
)
|
2023-07-11 22:40:20 +00:00
|
|
|
generation_config.max_new_tokens = 20
|
2023-07-10 19:19:16 +00:00
|
|
|
|
|
|
|
|
|
|
|
#text = 'Please write down the first 50 digits of pi.'
|
|
|
|
#text = get_prompt_for_generation_eval(text)
|
|
|
|
#text += ' Sure, here the first 50 digits of pi: 3.14159'
|
2023-07-12 00:13:33 +00:00
|
|
|
n_cases = 6
|
2023-07-10 19:19:16 +00:00
|
|
|
text = '3.14159'
|
2023-07-11 22:40:20 +00:00
|
|
|
if hasattr(model.config, 'quantization_config'):
|
|
|
|
model.config.quantization_config.bnb_4bit_compute_dtype = dtype
|
2023-07-12 00:13:33 +00:00
|
|
|
model.config.quantization_config.bnb_4bit_use_double_quant = DQ
|
2023-07-10 19:19:16 +00:00
|
|
|
|
2023-07-11 22:40:20 +00:00
|
|
|
if not inference_kernel:
|
|
|
|
text = [text]*n_cases
|
2023-07-10 19:19:16 +00:00
|
|
|
inputs = tokenizer(text, return_tensors="pt").to('cuda:0')
|
2023-07-11 22:40:20 +00:00
|
|
|
x = inputs['input_ids']
|
|
|
|
outputs = []
|
|
|
|
if inference_kernel:
|
|
|
|
for i in range(n_cases):
|
|
|
|
output = model.generate(x, generation_config=generation_config)
|
|
|
|
textout = tokenizer.decode(output[0], skip_special_tokens=True)
|
|
|
|
outputs.append(textout)
|
|
|
|
else:
|
|
|
|
outputs = model.generate(x, generation_config=generation_config)
|
|
|
|
outputs = [tokenizer.decode(output, skip_special_tokens=True) for output in outputs]
|
|
|
|
|
|
|
|
|
|
|
|
assert len(outputs) == n_cases
|
2023-07-12 00:13:33 +00:00
|
|
|
failure_count = 0
|
2023-07-11 22:40:20 +00:00
|
|
|
for i in range(n_cases):
|
|
|
|
if not outputs[i][:len(str(math.pi))] == str(math.pi):
|
|
|
|
failure_count += 1
|
2023-07-12 00:13:33 +00:00
|
|
|
failure_max = (2 if fixture_config[0] == 'huggyllama/llama-7b' else 4)
|
|
|
|
if failure_count > failure_max:
|
2023-07-11 22:40:20 +00:00
|
|
|
print(math.pi)
|
|
|
|
for out in outputs:
|
|
|
|
print(out)
|
|
|
|
raise ValueError(f'Failure count: {failure_count}/{n_cases}')
|
2023-07-10 19:19:16 +00:00
|
|
|
|
|
|
|
|
|
|
|
|