Spaces:
Sleeping
Sleeping
File size: 7,029 Bytes
dc12c31 |
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 |
import os
from pathlib import Path
from typing import Any, Dict, Optional, Union
import torch
from torch.nn import CrossEntropyLoss
from transformers import GenerationConfig, PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast
from modules import RoPE, shared
from modules.logging_colors import logger
import llama_cpp
if torch.cuda.is_available() and not torch.version.hip:
try:
import llama_cpp_cuda
except:
llama_cpp_cuda = None
else:
llama_cpp_cuda = None
def llama_cpp_lib():
if shared.args.cpu or llama_cpp_cuda is None:
return llama_cpp
else:
return llama_cpp_cuda
class LlamacppHF(PreTrainedModel):
def __init__(self, model, path):
super().__init__(PretrainedConfig())
self.model = model
self.generation_config = GenerationConfig()
self.past_seq = None
self.llamacpp_cache = {
'n_tokens': self.model.n_tokens,
'input_ids': self.model.input_ids,
'scores': self.model.scores,
'ctx': self.model.ctx
}
if shared.args.cfg_cache:
self.past_seq_negative = None
self.llamacpp_cache_negative = {
'n_tokens': self.model.n_tokens,
'input_ids': self.model.input_ids.copy(),
'scores': self.model.scores.copy(),
'ctx': llama_cpp_lib().llama_new_context_with_model(model.model, model.params)
}
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}
def save_cache(self):
self.llamacpp_cache.update({
'n_tokens': self.model.n_tokens,
'input_ids': self.model.input_ids,
'scores': self.model.scores,
'ctx': self.model.ctx
})
def save_negative_cache(self):
self.llamacpp_cache_negative.update({
'n_tokens': self.model.n_tokens,
'input_ids': self.model.input_ids,
'scores': self.model.scores,
'ctx': self.model.ctx
})
def load_cache(self):
self.model.n_tokens = self.llamacpp_cache['n_tokens']
self.model.input_ids = self.llamacpp_cache['input_ids']
self.model.scores = self.llamacpp_cache['scores']
self.model.ctx = self.llamacpp_cache['ctx']
def load_negative_cache(self):
self.model.n_tokens = self.llamacpp_cache_negative['n_tokens']
self.model.input_ids = self.llamacpp_cache_negative['input_ids']
self.model.scores = self.llamacpp_cache_negative['scores']
self.model.ctx = self.llamacpp_cache_negative['ctx']
@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 llamacpp_HF.")
return
input_ids = args[0]
is_negative = True
past_seq = self.past_seq_negative
self.load_negative_cache()
else:
input_ids = kwargs['input_ids']
is_negative = False
past_seq = self.past_seq
self.load_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)
# Make the forward call
if labels is None:
if past_seq is None or not torch.equal(past_seq, seq_tensor[:-1]):
self.model.reset()
self.model.eval(seq)
else:
self.model.eval([seq[-1]])
logits = torch.tensor(self.model.scores[self.model.n_tokens - 1, :]).view(1, 1, -1).to(input_ids.device)
else:
self.model.reset()
self.model.eval(seq)
logits = torch.tensor(self.model.eval_logits)
logits = logits.view(1, logits.shape[0], logits.shape[1]).to(input_ids.device)
if is_negative:
self.save_negative_cache()
self.past_seq_negative = seq_tensor
else:
self.save_cache()
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)
path = Path(f'{shared.args.model_dir}') / Path(pretrained_model_name_or_path)
if path.is_file():
model_file = path
else:
model_file = list(path.glob('*.gguf'))[0]
logger.info(f"llama.cpp weights detected: {model_file}\n")
if shared.args.tensor_split is None or shared.args.tensor_split.strip() == '':
tensor_split_list = None
else:
tensor_split_list = [float(x) for x in shared.args.tensor_split.strip().split(",")]
params = {
'model_path': str(model_file),
'n_ctx': shared.args.n_ctx,
'seed': int(shared.args.llama_cpp_seed),
'n_threads': shared.args.threads or None,
'n_batch': shared.args.n_batch,
'use_mmap': not shared.args.no_mmap,
'use_mlock': shared.args.mlock,
'mul_mat_q': shared.args.mul_mat_q,
'low_vram': shared.args.low_vram,
'n_gpu_layers': shared.args.n_gpu_layers,
'rope_freq_base': RoPE.get_rope_freq_base(shared.args.alpha_value, shared.args.rope_freq_base),
'tensor_split': tensor_split_list,
'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
'logits_all': True,
}
Llama = llama_cpp_lib().Llama
model = Llama(**params)
return LlamacppHF(model, model_file)
|