Spaces:
Runtime error
Runtime error
File size: 7,655 Bytes
90559ad |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 |
from pathlib import Path
import torch
from llama_cookbook.inference.model_utils import load_model as load_model_llamarecipes
from llama_cookbook.inference.model_utils import load_peft_model
from transformers import AutoTokenizer
from src.utils import RankedLogger
log = RankedLogger(__name__, rank_zero_only=True)
def load_model(
ckpt_path, quantization=None, use_fast_kernels=False, peft_model=False, **kwargs
):
model = load_model_llamarecipes(
model_name=ckpt_path,
quantization=quantization,
use_fast_kernels=use_fast_kernels,
device_map="auto",
**kwargs,
)
if peft_model:
model = load_peft_model(model, peft_model)
tokenizer = AutoTokenizer.from_pretrained(ckpt_path)
tokenizer.pad_token = tokenizer.eos_token
# special_tokens = {"additional_special_tokens": ["<image>"]}
# tokenizer.add_special_tokens(special_tokens)
return model, tokenizer
@torch.no_grad()
def inference(
model,
tokenizer: AutoTokenizer,
prompt: str,
add_special_tokens: bool = True,
temperature: float = 1.0,
max_new_tokens=1024,
top_p: float = 1.0,
top_k: int = 50,
use_cache: bool = True,
max_padding_length: int = None,
do_sample: bool = False,
min_length: int = None,
repetition_penalty: float = 1.0,
length_penalty: int = 1,
max_prompt_tokens: int = 35_000,
**kwargs,
):
"""
temperature: float, optional (default=1.0) The value used to module the next token probabilities.
max_new_tokens: int, optional (default=1024) The maximum number of tokens to generate.
top_p: float, optional (default=1.0) If set to float < 1 only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation.
top_k: int, optional (default=50) The number of highest probability vocabulary tokens to keep for top-k-filtering.
use_cache: bool, optional (default=True) Whether or not the model should use the past last key/values attentions Whether or not the model should use the past last key/values attentions (if applicable to the model) to speed up decoding.
max_padding_length: int, optional (default=None) the max padding length to be used with tokenizer padding the prompts.
do_sample: bool, optional (default=True) Whether or not to use sampling ; use greedy decoding otherwise.
min_length: int, optional (default=None) The minimum length of the sequence to be generated input prompt + min_new_tokens
repetition_penalty: float, optional (default=1.0) The parameter for repetition penalty. 1.0 means no penalty.
length_penalty: int, optional (default=1) Exponential penalty to the length that is used with beam-based generation.
"""
if add_special_tokens:
prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nCutting Knowledge Date: December 2023\nToday Date: 26 Jul 2024\n\n<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>"
# prompt = f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>"
batch = tokenizer(
prompt,
truncation=True,
max_length=max_padding_length,
return_tensors="pt",
)
# if the input is too long, return the length of the input
n_tokens = len(batch["input_ids"][0])
if max_prompt_tokens is not None and n_tokens > max_prompt_tokens:
return n_tokens
batch = {k: v.to("cuda") for k, v in batch.items()}
terminators = [
tokenizer.eos_token_id,
tokenizer.convert_tokens_to_ids("<|eot_id|>"),
]
try:
outputs = model.generate(
**batch,
max_new_tokens=max_new_tokens,
do_sample=do_sample,
top_p=top_p,
temperature=temperature,
min_length=min_length,
use_cache=use_cache,
top_k=top_k,
repetition_penalty=repetition_penalty,
length_penalty=length_penalty,
eos_token_id=terminators,
pad_token_id=tokenizer.eos_token_id,
**kwargs,
)
output_text = tokenizer.decode(outputs[0], skip_special_tokens=False)
output = output_text.split("<|start_header_id|>assistant<|end_header_id|>")[1]
output = output.strip()
output = output.removesuffix("<|eot_id|>")
except torch.cuda.OutOfMemoryError as e:
log.error(f"CUDA out of memory error: {e}")
torch.cuda.empty_cache()
return n_tokens
return output
class LlamaInference:
def __init__(
self,
ckpt_path,
quantization=None,
use_fast_kernels=False,
peft_model=False,
add_special_tokens: bool = True,
temperature: float = 1.0,
max_new_tokens: int = 1024,
top_p: float = 1.0,
top_k: int = 50,
use_cache: bool = True,
max_padding_length: int = None,
do_sample: bool = False,
min_length: int = None,
repetition_penalty: float = 1.0,
length_penalty: int = 1,
max_prompt_tokens: int = 35_000,
**kwargs,
):
# Check if LLaMA model exists
# if not Path(ckpt_path).exists():
# log.warning(f"Model checkpoint does not exist at {ckpt_path}")
# return None
# If PEFT model is specified, check if it exists
if peft_model and not Path(peft_model).exists():
log.warning(f"PEFT model does not exist at {peft_model}")
return None
if peft_model:
log.info(f"PEFT model found at {peft_model}")
model = load_model_llamarecipes(
model_name=ckpt_path,
quantization=quantization,
use_fast_kernels=use_fast_kernels,
device_map="auto",
**kwargs,
)
if peft_model:
model = load_peft_model(model, peft_model)
model.eval()
tokenizer = AutoTokenizer.from_pretrained(ckpt_path)
tokenizer.pad_token = tokenizer.eos_token
self.model = model
self.tokenizer = tokenizer
self.add_special_tokens = add_special_tokens
self.temperature = temperature
self.max_new_tokens = max_new_tokens
self.top_p = top_p
self.top_k = top_k
self.use_cache = use_cache
self.max_padding_length = max_padding_length
self.do_sample = do_sample
self.min_length = min_length
self.repetition_penalty = repetition_penalty
self.length_penalty = length_penalty
self.max_prompt_tokens = max_prompt_tokens
def __call__(self, prompt: str, **kwargs):
# Create a dict of default parameters from instance attributes
params = {
"model": self.model,
"tokenizer": self.tokenizer,
"prompt": prompt,
"add_special_tokens": self.add_special_tokens,
"temperature": self.temperature,
"max_new_tokens": self.max_new_tokens,
"top_p": self.top_p,
"top_k": self.top_k,
"use_cache": self.use_cache,
"max_padding_length": self.max_padding_length,
"do_sample": self.do_sample,
"min_length": self.min_length,
"repetition_penalty": self.repetition_penalty,
"length_penalty": self.length_penalty,
"max_prompt_tokens": self.max_prompt_tokens,
}
# Update with any overrides passed in kwargs
params.update(kwargs)
return inference(**params)
|