Spaces:
Sleeping
Sleeping
File size: 7,055 Bytes
19dc0f3 |
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 |
import os
import traceback
from pathlib import Path
from typing import Any, Dict, Optional, Union
import torch
from exllamav2 import (
ExLlamaV2,
ExLlamaV2Cache,
ExLlamaV2Cache_8bit,
ExLlamaV2Cache_Q4,
ExLlamaV2Config
)
from torch.nn import CrossEntropyLoss
from transformers import GenerationConfig, PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast
from modules import shared
from modules.logging_colors import logger
try:
import flash_attn
except ModuleNotFoundError:
logger.warning(
'You are running ExLlamaV2 without flash-attention. This will cause the VRAM usage '
'to be a lot higher than it could be.\n'
'Try installing flash-attention following the instructions here: '
'https://github.com/Dao-AILab/flash-attention#installation-and-features'
)
pass
except Exception:
logger.warning('Failed to load flash-attention due to the following error:\n')
traceback.print_exc()
class Exllamav2HF(PreTrainedModel):
def __init__(self, config: ExLlamaV2Config):
super().__init__(PretrainedConfig())
self.ex_config = config
self.loras = None
self.generation_config = GenerationConfig()
self.ex_model = ExLlamaV2(config)
if not shared.args.autosplit:
split = None
if shared.args.gpu_split:
split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
self.ex_model.load(split)
if shared.args.cache_8bit:
self.ex_cache = ExLlamaV2Cache_8bit(self.ex_model, lazy=shared.args.autosplit)
elif shared.args.cache_4bit:
self.ex_cache = ExLlamaV2Cache_Q4(self.ex_model, lazy=shared.args.autosplit)
else:
self.ex_cache = ExLlamaV2Cache(self.ex_model, lazy=shared.args.autosplit)
if shared.args.autosplit:
self.ex_model.load_autosplit(self.ex_cache)
self.past_seq = None
if shared.args.cfg_cache:
if shared.args.cache_8bit:
self.ex_cache_negative = ExLlamaV2Cache_8bit(self.ex_model)
elif shared.args.cache_4bit:
self.ex_cache_negative = ExLlamaV2Cache_Q4(self.ex_model)
else:
self.ex_cache_negative = ExLlamaV2Cache(self.ex_model)
self.past_seq_negative = None
def _validate_model_class(self):
pass
def _validate_model_kwargs(self, model_kwargs: Dict[str, Any]):
pass
def prepare_inputs_for_generation(self, input_ids, **kwargs):
return {'input_ids': input_ids, **kwargs}
@property
def device(self) -> torch.device:
return torch.device(0)
def __call__(self, *args, **kwargs):
use_cache = kwargs.get('use_cache', True)
labels = kwargs.get('labels', None)
past_key_values = kwargs.get('past_key_values', None)
if len(args) > 0:
if not shared.args.cfg_cache:
logger.error("Please enable the cfg-cache option to use CFG with ExLlamav2_HF.")
return
input_ids = args[0]
is_negative = True
past_seq = self.past_seq_negative
ex_cache = self.ex_cache_negative
else:
input_ids = kwargs['input_ids']
is_negative = False
past_seq = self.past_seq
ex_cache = self.ex_cache
seq = input_ids[0].tolist()
if is_negative and past_key_values is not None:
seq = past_key_values + seq
seq_tensor = torch.tensor(seq)
reset = True
# Make the forward call
if labels is None:
if past_seq is not None:
min_length = min(past_seq.shape[0], seq_tensor.shape[0])
indices = torch.nonzero(~torch.eq(past_seq[:min_length], seq_tensor[:min_length]))
if len(indices) > 0:
longest_prefix = indices[0].item()
else:
longest_prefix = min_length
if longest_prefix > 0:
reset = False
ex_cache.current_seq_len = longest_prefix
if len(seq_tensor) - longest_prefix > 1:
self.ex_model.forward(seq_tensor[longest_prefix:-1].view(1, -1), ex_cache, preprocess_only=True, loras=self.loras)
elif len(seq_tensor) == longest_prefix:
# Very tricky: if the prefix we are reusing *is* the input_ids, then we have to back up the cache pointer by one,
# because we feed input_ids[-1] to forward() below, but that last token is already in the cache!
ex_cache.current_seq_len -= 1
if reset:
ex_cache.current_seq_len = 0
if len(seq_tensor) > 1:
self.ex_model.forward(seq_tensor[:-1].view(1, -1), ex_cache, preprocess_only=True, loras=self.loras)
logits = self.ex_model.forward(seq_tensor[-1:].view(1, -1), ex_cache, loras=self.loras).to(input_ids.device).float()
else:
ex_cache.current_seq_len = 0
logits = self.ex_model.forward(seq_tensor.view(1, -1), ex_cache, last_id_only=False, loras=self.loras).float()
if is_negative:
self.past_seq_negative = seq_tensor
else:
self.past_seq = seq_tensor
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
shift_logits = shift_logits.view(-1, logits.shape[-1])
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
loss = loss_fct(shift_logits, shift_labels)
return CausalLMOutputWithPast(logits=logits, past_key_values=seq if use_cache else None, loss=loss)
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
assert len(model_args) == 0 and len(kwargs) == 0, "extra args is currently not supported"
if isinstance(pretrained_model_name_or_path, str):
pretrained_model_name_or_path = Path(pretrained_model_name_or_path)
pretrained_model_name_or_path = Path(f'{shared.args.model_dir}') / Path(pretrained_model_name_or_path)
config = ExLlamaV2Config()
config.model_dir = str(pretrained_model_name_or_path)
config.prepare()
config.max_seq_len = shared.args.max_seq_len
config.scale_pos_emb = shared.args.compress_pos_emb
config.scale_alpha_value = shared.args.alpha_value
config.no_flash_attn = shared.args.no_flash_attn
config.num_experts_per_token = int(shared.args.num_experts_per_token)
return Exllamav2HF(config)
|