text-generation-webui/modules/llamacpp_hf.py

214 lines
7.6 KiB
Python
Raw Normal View History

2023-07-16 07:21:13 +02:00
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
2023-07-16 07:21:13 +02:00
from modules.logging_colors import logger
try:
import llama_cpp
except:
llama_cpp = None
try:
import llama_cpp_cuda
except:
llama_cpp_cuda = None
2023-09-11 16:30:56 +02:00
def llama_cpp_lib():
if (shared.args.cpu and llama_cpp is not None) or llama_cpp_cuda is None:
2023-09-11 16:30:56 +02:00
return llama_cpp
else:
2023-09-11 16:30:56 +02:00
return llama_cpp_cuda
2023-07-16 07:21:13 +02:00
2023-07-16 07:21:13 +02:00
class LlamacppHF(PreTrainedModel):
2023-08-27 07:15:06 +02:00
def __init__(self, model, path):
2023-07-16 07:21:13 +02:00
super().__init__(PretrainedConfig())
self.model = model
self.generation_config = GenerationConfig()
2023-08-24 21:27:36 +02:00
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
2023-08-24 21:27:36 +02:00
}
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.context_params)
2023-08-24 21:27:36 +02:00
}
2023-07-16 07:21:13 +02:00
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}
2023-08-24 21:27:36 +02:00
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
2023-08-24 21:27:36 +02:00
})
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
2023-08-24 21:27:36 +02:00
})
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']
2023-08-24 21:27:36 +02:00
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']
2023-08-24 21:27:36 +02:00
2023-07-16 07:21:13 +02:00
@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)
2023-08-24 21:27:36 +02:00
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()
2023-08-24 21:27:36 +02:00
if is_negative and past_key_values is not None:
seq = past_key_values + seq
2023-07-16 07:21:13 +02:00
seq_tensor = torch.tensor(seq)
2023-09-17 20:50:47 +02:00
reset = True
2023-08-24 21:27:36 +02:00
2023-09-17 20:50:47 +02:00
# Make the forward call. The prefix-match code has been adapted from
# https://github.com/abetlen/llama-cpp-python/commit/f4090a0bb2a2a25acfe28d31c82cc1aa273bedee
2023-07-16 07:21:13 +02:00
if labels is None:
2023-09-17 20:50:47 +02:00
if past_seq is not None:
2023-09-18 20:02:45 +02:00
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
2023-09-17 20:50:47 +02:00
if longest_prefix > 0:
reset = False
self.model.n_tokens = longest_prefix
if len(seq_tensor) - longest_prefix > 0:
self.model.eval(seq[longest_prefix:])
2023-09-17 20:50:47 +02:00
if reset:
2023-07-16 07:21:13 +02:00
self.model.reset()
self.model.eval(seq)
2023-08-24 21:27:36 +02:00
logits = torch.tensor(self.model.scores[self.model.n_tokens - 1, :]).view(1, 1, -1).to(input_ids.device)
2023-07-16 07:21:13 +02:00
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)
2023-07-16 07:21:13 +02:00
2023-08-24 21:27:36 +02:00
if is_negative:
self.save_negative_cache()
self.past_seq_negative = seq_tensor
else:
self.save_cache()
self.past_seq = seq_tensor
2023-07-17 05:49:48 +02:00
2023-07-16 07:21:13 +02:00
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)
2023-08-24 21:27:36 +02:00
return CausalLMOutputWithPast(logits=logits, past_key_values=seq if use_cache else None, loss=loss)
2023-07-16 07:21:13 +02:00
@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"
2023-07-16 07:21:13 +02:00
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:
2023-09-11 16:30:56 +02:00
model_file = list(path.glob('*.gguf'))[0]
2023-07-16 07:21:13 +02:00
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(",")]
2023-07-16 07:21:13 +02:00
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,
2023-10-02 06:27:04 +02:00
'n_threads_batch': shared.args.threads_batch or None,
2023-07-16 07:21:13 +02:00
'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,
'numa': shared.args.numa,
2023-07-16 07:21:13 +02:00
'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,
2023-07-16 07:21:13 +02:00
'logits_all': True,
}
2023-08-27 07:11:07 +02:00
2023-09-11 16:30:56 +02:00
Llama = llama_cpp_lib().Llama
2023-07-16 07:21:13 +02:00
model = Llama(**params)
2023-08-27 07:15:06 +02:00
return LlamacppHF(model, model_file)