mirror of
https://github.com/oobabooga/text-generation-webui.git
synced 2024-12-25 05:48:55 +01:00
82 lines
3.1 KiB
Python
82 lines
3.1 KiB
Python
import os
|
|
import sys
|
|
from pathlib import Path
|
|
from typing import *
|
|
|
|
import torch
|
|
from transformers import (
|
|
GenerationConfig,
|
|
LlamaTokenizer,
|
|
PretrainedConfig,
|
|
PreTrainedModel
|
|
)
|
|
from transformers.modeling_outputs import CausalLMOutputWithPast
|
|
|
|
from modules import shared
|
|
from modules.logging_colors import logger
|
|
from modules.relative_imports import RelativeImport
|
|
|
|
with RelativeImport("repositories/exllama"):
|
|
from model import ExLlama, ExLlamaCache, ExLlamaConfig
|
|
|
|
|
|
class ExllamaHF(PreTrainedModel):
|
|
def __init__(self, config: ExLlamaConfig):
|
|
super().__init__(PretrainedConfig())
|
|
self.ex_config = config
|
|
self.ex_model = ExLlama(self.ex_config)
|
|
self.generation_config = GenerationConfig()
|
|
|
|
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:
|
|
# TODO: May cause problem on multi-gpu inference?
|
|
return torch.device(0)
|
|
|
|
def __call__(self, *args, **kwargs):
|
|
# TODO: Some decoding methods (such as Contrastive Search) may not work at this time
|
|
assert len(args) == 0, 'no *args should be passed to forward'
|
|
use_cache = kwargs['use_cache']
|
|
seq = kwargs['input_ids'][0].tolist()
|
|
cache = kwargs['past_key_values'] if 'past_key_values' in kwargs else None
|
|
if cache is None:
|
|
cache = ExLlamaCache(self.ex_model)
|
|
self.ex_model.forward(torch.tensor([seq[:-1]], dtype=torch.long), cache, preprocess_only=True)
|
|
logits = self.ex_model.forward(torch.tensor([seq[-1:]], dtype=torch.long), cache).to(self.device)
|
|
return CausalLMOutputWithPast(logits=logits, past_key_values=cache if use_cache else None)
|
|
|
|
@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 = ExLlamaConfig(pretrained_model_name_or_path / 'config.json')
|
|
|
|
# from 'oobabooga/text-generation-webui/modules/exllama.py'
|
|
weight_path = None
|
|
for ext in ['.safetensors', '.pt', '.bin']:
|
|
found = list(pretrained_model_name_or_path.glob(f"*{ext}"))
|
|
if len(found) > 0:
|
|
weight_path = found[-1]
|
|
break
|
|
assert weight_path is not None, f'could not find weight in "{pretrained_model_name_or_path}"'
|
|
|
|
config.model_path = str(weight_path)
|
|
|
|
# This slowes down a bit but align better with autogptq generation.
|
|
# TODO: Should give user choice to tune the exllama config
|
|
config.act_order = True
|
|
config.fused_attn = False
|
|
config.fused_mlp_thd = 0
|
|
|
|
return ExllamaHF(config) |