File size: 5,800 Bytes
7d52396 |
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 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 |
# This mimics GPTQ's evaluation metrics: https://github.com/IST-DASLab/gptq/
# Thanks to E. Frantar et al GPTQ: Accurate Post-training Compression for GPT, arXiv:2210.17323
import math
import sys
import time
from pathlib import Path
from typing import Optional
import lightning as L
import torch
import tqdm
# support running without installing as a package
wd = Path(__file__).parent.parent.resolve()
sys.path.append(str(wd))
from lit_llama import Tokenizer
from lit_llama.adapter import LLaMA
from lit_llama.utils import EmptyInitOnDevice, lazy_load, llama_model_lookup
from lit_llama.adapter_v2 import add_adapter_v2_parameters_to_linear_layers
from scripts.prepare_alpaca import generate_prompt
from datasets import load_dataset
def load_eval_data(dataset_name: str) -> str:
# this mimics gptq datautils
if dataset_name == "wikitext":
# traindata = load_dataset('wikitext', 'wikitext-2-raw-v1', split='train')
testdata = load_dataset("wikitext", "wikitext-2-raw-v1", split="test")
testdata = "\n\n".join(testdata["text"])
elif dataset_name == "ptb":
testdata = load_dataset("ptb_text_only", "penn_treebank", split="test")
testdata = "\n\n".join(testdata["sentence"])
elif dataset_name == "c4":
testdata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"validation": "en/c4-validation.00000-of-00008.json.gz"},
split="validation",
)
testdata = " ".join(testdata[:1100]["text"])
else:
raise ValueError("invalid dataset name (wikitext, ptb, c4 are allowed)")
return testdata
@torch.inference_mode()
def main(
datasets: str = "wikitext,ptb,c4",
*,
accelerator: str = "auto",
adapter_path: Path = Path("out/adapter_v2/alpaca/lit-llama-adapter-finetuned.pth"),
checkpoint_path: Path = Path("checkpoints/lit-llama/7B/lit-llama.pth"),
tokenizer_path: Path = Path("checkpoints/lit-llama/tokenizer.model"),
dtype: str = "float32",
quantize: Optional[str] = None,
) -> None:
"""Generates text samples based on a pre-trained LLaMA model and tokenizer.
Args:
datasets: The datasets to use as a comma separated string
accelerator: The hardware to run on. Possible choices are:
``"cpu"``, ``"cuda"``, ``"mps"``, ``"gpu"``, ``"tpu"``, ``"auto"``.
adapter_path: Path to the checkpoint with trained adapter weights, which are the output of
`finetune_adapter_v2.py`.
checkpoint_path: The checkpoint path to load.
tokenizer_path: The tokenizer path to load.
dtype: The tensor dtype for choosing the floating-point precision
quantize: Whether to quantize the model and using which method:
``"llm.int8"``: LLM.int8() mode,
``"gptq.int4"``: GPTQ 4-bit mode.
"""
assert adapter_path.is_file()
assert checkpoint_path.is_file()
assert tokenizer_path.is_file()
fabric = L.Fabric(accelerator=accelerator, devices=1)
dt = getattr(torch, dtype, None)
if not isinstance(dt, torch.dtype):
raise ValueError(f"{dtype} is not a valid dtype.")
dtype = dt
print("Loading model ...", file=sys.stderr)
t0 = time.time()
with lazy_load(checkpoint_path) as pretrained_checkpoint, lazy_load(adapter_path) as adapter_checkpoint:
name = llama_model_lookup(pretrained_checkpoint)
with EmptyInitOnDevice(
device=fabric.device, dtype=dtype, quantization_mode=quantize
):
model = LLaMA.from_name(name)
add_adapter_v2_parameters_to_linear_layers(model)
# 1. Load the pretrained weights
model.load_state_dict(pretrained_checkpoint, strict=False)
# 2. Load the fine-tuned adapter weights
model.load_state_dict(adapter_checkpoint, strict=False)
print(f"Time to load model: {time.time() - t0:.02f} seconds.", file=sys.stderr)
model.eval()
# if compile:
# model = torch.compile(model)
total_toks = 0
model = fabric.setup_module(model)
tokenizer = Tokenizer(tokenizer_path)
for dsname in datasets.split(","):
test_string = load_eval_data(dsname)
sample = {"instruction": test_string, "input": input}
test_string = generate_prompt(sample)
encoded_text = tokenizer.encode(
test_string, bos=True, eos=False, device=fabric.device
)
encoded_text = encoded_text[
None, : 256 * model.config.block_size
] # add batch dimension, trim like gptq implementation
t0 = time.perf_counter()
nlls = 0
toks = 0
block_size = 2048 # this is for compat with gptq, and indeed we get much worse beyond this (https://github.com/facebookresearch/llama/blob/57b0eb62de0636e75af471e49e2f1862d908d9d8/llama/model.py#L30)
for i in tqdm.tqdm(range(0, encoded_text.shape[1], block_size)):
inp = encoded_text[:, i : i + block_size]
logits = model(inp)[0]
nll = torch.nn.functional.cross_entropy(
logits[:-1], inp[0, 1:].to(dtype=torch.long), reduction="sum"
)
toks += inp.size(1) - 1
nlls += nll.item()
print(encoded_text.shape, logits.shape)
ppl = math.exp(nlls / toks)
print(f"Perplexity on {dsname}: {ppl:.2f}")
total_toks += toks
t = time.perf_counter() - t0
print(
f"\n\nTime for inference: {t:.02f} sec total, {total_toks / t:.02f} tokens/sec",
file=sys.stderr,
)
print(
f"Memory used: {torch.cuda.max_memory_reserved() / 1e9:.02f} GB",
file=sys.stderr,
)
if __name__ == "__main__":
from jsonargparse import CLI
torch.set_float32_matmul_precision("high")
CLI(main)
|