text-generation-webui/modules/models.py

303 lines
13 KiB
Python
Raw Normal View History

2023-04-08 00:36:04 +00:00
import gc
2023-02-23 16:28:30 +00:00
import json
2023-05-04 00:43:17 +00:00
import logging
2023-02-23 16:28:30 +00:00
import os
import re
2023-02-23 16:28:30 +00:00
import time
import zipfile
from pathlib import Path
import numpy as np
import torch
import transformers
2023-03-16 16:34:23 +00:00
from accelerate import infer_auto_device_map, init_empty_weights
from transformers import (AutoConfig, AutoModel, AutoModelForCausalLM,
AutoModelForSeq2SeqLM, AutoTokenizer,
BitsAndBytesConfig, LlamaTokenizer)
2023-02-23 17:41:42 +00:00
import modules.shared as shared
2023-04-10 02:08:40 +00:00
from modules import llama_attn_hijack
2023-02-23 16:28:30 +00:00
transformers.logging.set_verbosity_error()
2023-02-23 16:28:30 +00:00
if shared.args.flexgen:
2023-03-16 13:18:34 +00:00
from flexgen.flex_opt import CompressionConfig, ExecutionEnv, OptLM, Policy
2023-02-23 16:28:30 +00:00
2023-04-08 00:36:04 +00:00
local_rank = None
2023-02-23 16:28:30 +00:00
if shared.args.deepspeed:
import deepspeed
2023-02-23 17:41:42 +00:00
from transformers.deepspeed import (HfDeepSpeedConfig,
is_deepspeed_zero3_enabled)
2023-02-23 16:28:30 +00:00
from modules.deepspeed_parameters import generate_ds_config
# Distributed setup
local_rank = shared.args.local_rank if shared.args.local_rank is not None else int(os.getenv("LOCAL_RANK", "0"))
world_size = int(os.getenv("WORLD_SIZE", "1"))
torch.cuda.set_device(local_rank)
deepspeed.init_distributed()
ds_config = generate_ds_config(shared.args.bf16, 1 * world_size, shared.args.nvme_offload_dir)
dschf = HfDeepSpeedConfig(ds_config) # Keep this object alive for the Transformers integration
2023-02-23 16:28:30 +00:00
2023-03-13 17:00:38 +00:00
def find_model_type(model_name):
2023-04-26 04:55:40 +00:00
model_name_lower = model_name.lower()
if 'rwkv-' in model_name_lower:
return 'rwkv'
elif len(list(Path(f'{shared.args.model_dir}/{model_name}').glob('*ggml*.bin'))) > 0:
return 'llamacpp'
2023-04-26 04:55:40 +00:00
elif re.match('.*ggml.*\.bin', model_name_lower):
return 'llamacpp'
2023-04-26 04:55:40 +00:00
elif 'chatglm' in model_name_lower:
return 'chatglm'
2023-04-26 04:55:40 +00:00
elif 'galactica' in model_name_lower:
return 'galactica'
2023-04-26 04:55:40 +00:00
elif 'llava' in model_name_lower:
2023-04-23 23:32:22 +00:00
return 'llava'
elif 'oasst' in model_name_lower:
return 'oasst'
2023-04-26 04:55:40 +00:00
elif any((k in model_name_lower for k in ['gpt4chan', 'gpt-4chan'])):
return 'gpt4chan'
else:
config = AutoConfig.from_pretrained(Path(f'{shared.args.model_dir}/{model_name}'), trust_remote_code=shared.args.trust_remote_code)
# Not a "catch all", but fairly accurate
if config.to_dict().get("is_encoder_decoder", False):
return 'HF_seq2seq'
else:
return 'HF_generic'
2023-02-23 16:28:30 +00:00
def load_model(model_name):
2023-05-04 00:43:17 +00:00
logging.info(f"Loading {model_name}...")
2023-02-23 16:28:30 +00:00
t0 = time.time()
shared.model_type = find_model_type(model_name)
trust_remote_code = shared.args.trust_remote_code
if shared.model_type == 'chatglm':
LoaderClass = AutoModel
elif shared.model_type == 'HF_seq2seq':
LoaderClass = AutoModelForSeq2SeqLM
else:
LoaderClass = AutoModelForCausalLM
2023-02-28 02:03:35 +00:00
# Load the model in simple 16-bit mode by default
if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.wbits, shared.args.auto_devices, shared.args.disk, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None, shared.args.deepspeed, shared.args.flexgen, shared.model_type in ['rwkv', 'llamacpp']]):
model = LoaderClass.from_pretrained(Path(f"{shared.args.model_dir}/{model_name}"), low_cpu_mem_usage=True, torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16, trust_remote_code=trust_remote_code)
if torch.has_mps:
device = torch.device('mps')
model = model.to(device)
2023-03-18 01:27:26 +00:00
else:
model = model.cuda()
2023-03-18 01:27:26 +00:00
2023-02-23 16:28:30 +00:00
# FlexGen
elif shared.args.flexgen:
2023-02-26 19:53:41 +00:00
# Initialize environment
env = ExecutionEnv.create(shared.args.disk_cache_dir)
2023-02-23 16:28:30 +00:00
# Offloading policy
policy = Policy(1, 1,
shared.args.percent[0], shared.args.percent[1],
shared.args.percent[2], shared.args.percent[3],
shared.args.percent[4], shared.args.percent[5],
2023-03-04 04:04:02 +00:00
overlap=True, sep_layer=True, pin_weight=shared.args.pin_weight,
2023-02-23 16:28:30 +00:00
cpu_cache_compute=False, attn_sparsity=1.0,
compress_weight=shared.args.compress_weight,
comp_weight_config=CompressionConfig(
num_bits=4, group_size=64,
group_dim=0, symmetric=False),
compress_cache=False,
comp_cache_config=CompressionConfig(
num_bits=4, group_size=64,
group_dim=2, symmetric=False))
model = OptLM(f"facebook/{model_name}", env, shared.args.model_dir, policy)
2023-02-23 16:28:30 +00:00
# DeepSpeed ZeRO-3
elif shared.args.deepspeed:
model = LoaderClass.from_pretrained(Path(f"{shared.args.model_dir}/{model_name}"), torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16)
2023-02-23 16:28:30 +00:00
model = deepspeed.initialize(model=model, config_params=ds_config, model_parameters=None, optimizer=None, lr_scheduler=None)[0]
model.module.eval() # Inference
2023-05-04 00:43:17 +00:00
logging.info(f"DeepSpeed ZeRO-3 is enabled: {is_deepspeed_zero3_enabled()}")
2023-02-23 16:28:30 +00:00
2023-02-28 02:03:35 +00:00
# RMKV model (not on HuggingFace)
elif shared.model_type == 'rwkv':
2023-03-06 11:45:49 +00:00
from modules.RWKV import RWKVModel, RWKVTokenizer
2023-02-28 02:03:35 +00:00
model = RWKVModel.from_pretrained(Path(f'{shared.args.model_dir}/{model_name}'), dtype="fp32" if shared.args.cpu else "bf16" if shared.args.bf16 else "fp16", device="cpu" if shared.args.cpu else "cuda")
2023-03-28 02:42:29 +00:00
tokenizer = RWKVTokenizer.from_pretrained(Path(shared.args.model_dir))
2023-03-01 15:08:55 +00:00
2023-03-06 11:45:49 +00:00
return model, tokenizer
2023-02-28 02:03:35 +00:00
# llamacpp model
elif shared.model_type == 'llamacpp':
from modules.llamacpp_model import LlamaCppModel
path = Path(f'{shared.args.model_dir}/{model_name}')
if path.is_file():
model_file = path
else:
model_file = list(Path(f'{shared.args.model_dir}/{model_name}').glob('*ggml*.bin'))[0]
2023-05-04 00:43:17 +00:00
logging.info(f"llama.cpp weights detected: {model_file}\n")
model, tokenizer = LlamaCppModel.from_pretrained(model_file)
return model, tokenizer
2023-03-13 17:00:38 +00:00
# Quantized model
elif shared.args.wbits > 0:
2023-04-17 02:26:52 +00:00
# Monkey patch
if shared.args.monkey_patch:
2023-05-04 01:06:46 +00:00
logging.warning("Applying the monkey patch for using LoRAs in 4-bit mode. It may cause undefined behavior outside its intended scope.")
2023-04-17 02:26:52 +00:00
from modules.monkey_patch_gptq_lora import load_model_llama
model, _ = load_model_llama(model_name)
2023-04-17 02:26:52 +00:00
# No monkey patch
else:
from modules.GPTQ_loader import load_quantized
model = load_quantized(model_name)
2023-03-09 18:50:26 +00:00
2023-02-23 16:28:30 +00:00
# Custom
else:
params = {
2023-05-10 01:49:39 +00:00
"low_cpu_mem_usage": True,
"trust_remote_code": trust_remote_code
}
2023-05-10 01:49:39 +00:00
2023-03-18 01:56:46 +00:00
if not any((shared.args.cpu, torch.cuda.is_available(), torch.has_mps)):
2023-05-04 01:06:46 +00:00
logging.warning("torch.cuda.is_available() returned False. This means that no GPU has been detected. Falling back to CPU mode.")
2023-02-23 16:28:30 +00:00
shared.args.cpu = True
if shared.args.cpu:
params["torch_dtype"] = torch.float32
2023-02-23 16:28:30 +00:00
else:
params["device_map"] = 'auto'
if shared.args.load_in_8bit and any((shared.args.auto_devices, shared.args.gpu_memory)):
params['quantization_config'] = BitsAndBytesConfig(load_in_8bit=True, llm_int8_enable_fp32_cpu_offload=True)
elif shared.args.load_in_8bit:
params['quantization_config'] = BitsAndBytesConfig(load_in_8bit=True)
elif shared.args.bf16:
params["torch_dtype"] = torch.bfloat16
else:
params["torch_dtype"] = torch.float16
2023-02-23 16:28:30 +00:00
if shared.args.gpu_memory:
memory_map = list(map(lambda x: x.strip(), shared.args.gpu_memory))
max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
2023-03-16 16:34:23 +00:00
max_memory = {}
for i in range(len(memory_map)):
max_memory[i] = f'{memory_map[i]}GiB' if not re.match('.*ib$', memory_map[i].lower()) else memory_map[i]
2023-05-04 00:43:17 +00:00
2023-05-12 09:17:07 +00:00
max_memory['cpu'] = f'{max_cpu_memory}GiB' if not re.match('.*ib$', max_cpu_memory.lower()) else max_cpu_memory
params['max_memory'] = max_memory
elif shared.args.auto_devices:
total_mem = (torch.cuda.get_device_properties(0).total_memory / (1024 * 1024))
suggestion = round((total_mem - 1000) / 1000) * 1000
if total_mem - suggestion < 800:
2023-02-23 16:28:30 +00:00
suggestion -= 1000
2023-05-04 00:43:17 +00:00
suggestion = int(round(suggestion / 1000))
logging.warning(f"\033[1;32;1mAuto-assiging --gpu-memory {suggestion} for your GPU to try to prevent out-of-memory errors.\nYou can manually set other values.\033[0;37;0m")
2023-03-16 16:34:23 +00:00
max_memory = {0: f'{suggestion}GiB', 'cpu': f'{shared.args.cpu_memory or 99}GiB'}
params['max_memory'] = max_memory
2023-02-23 16:28:30 +00:00
if shared.args.disk:
params["offload_folder"] = shared.args.disk_cache_dir
checkpoint = Path(f'{shared.args.model_dir}/{model_name}')
if shared.args.load_in_8bit and params.get('max_memory', None) is not None and params['device_map'] == 'auto':
config = AutoConfig.from_pretrained(checkpoint, trust_remote_code=trust_remote_code)
with init_empty_weights():
model = LoaderClass.from_config(config, trust_remote_code=trust_remote_code)
2023-05-04 00:43:17 +00:00
model.tie_weights()
params['device_map'] = infer_auto_device_map(
model,
dtype=torch.int8,
max_memory=params['max_memory'],
no_split_module_classes=model._no_split_modules
)
model = LoaderClass.from_pretrained(checkpoint, **params)
2023-02-23 16:28:30 +00:00
2023-04-10 02:08:40 +00:00
# Hijack attention with xformers
if any((shared.args.xformers, shared.args.sdp_attention)):
llama_attn_hijack.hijack_llama_attention()
2023-02-23 16:28:30 +00:00
# Loading the tokenizer
if shared.model_type == 'gpt4chan' and Path(f"{shared.args.model_dir}/gpt-j-6B/").exists():
tokenizer = AutoTokenizer.from_pretrained(Path(f"{shared.args.model_dir}/gpt-j-6B/"))
elif type(model) is transformers.LlamaForCausalLM:
2023-04-20 00:23:51 +00:00
tokenizer = None
# Try to load an universal LLaMA tokenizer
if shared.model_type not in ['llava', 'oasst']:
2023-04-23 23:32:22 +00:00
for p in [Path(f"{shared.args.model_dir}/llama-tokenizer/"), Path(f"{shared.args.model_dir}/oobabooga_llama-tokenizer/")]:
if p.exists():
2023-05-04 00:43:17 +00:00
logging.info(f"Loading the universal LLaMA tokenizer from {p}...")
2023-04-23 23:32:22 +00:00
tokenizer = LlamaTokenizer.from_pretrained(p, clean_up_tokenization_spaces=True)
break
2023-04-20 00:23:51 +00:00
# Otherwise, load it from the model folder and hope that these
# are not outdated tokenizer files.
if tokenizer is None:
tokenizer = LlamaTokenizer.from_pretrained(Path(f"{shared.args.model_dir}/{model_name}/"), clean_up_tokenization_spaces=True)
2023-04-20 00:23:51 +00:00
try:
tokenizer.eos_token_id = 2
tokenizer.bos_token_id = 1
tokenizer.pad_token_id = 0
except:
pass
2023-02-23 16:28:30 +00:00
else:
tokenizer = AutoTokenizer.from_pretrained(Path(f"{shared.args.model_dir}/{model_name}/"), trust_remote_code=trust_remote_code)
2023-02-23 16:28:30 +00:00
2023-05-10 14:59:59 +00:00
logging.info(f"Loaded the model in {(time.time()-t0):.2f} seconds.\n")
2023-02-23 16:28:30 +00:00
return model, tokenizer
2023-04-08 00:36:04 +00:00
def clear_torch_cache():
gc.collect()
if not shared.args.cpu:
torch.cuda.empty_cache()
def unload_model():
shared.model = shared.tokenizer = None
clear_torch_cache()
def reload_model():
2023-04-08 00:37:41 +00:00
unload_model()
2023-04-08 00:36:04 +00:00
shared.model, shared.tokenizer = load_model(shared.model_name)
2023-02-23 16:28:30 +00:00
def load_soft_prompt(name):
if name == 'None':
shared.soft_prompt = False
shared.soft_prompt_tensor = None
else:
with zipfile.ZipFile(Path(f'softprompts/{name}.zip')) as zf:
zf.extract('tensor.npy')
zf.extract('meta.json')
j = json.loads(open('meta.json', 'r').read())
2023-05-04 00:43:17 +00:00
logging.info(f"\nLoading the softprompt \"{name}\".")
2023-02-23 16:28:30 +00:00
for field in j:
if field != 'name':
if type(j[field]) is list:
2023-05-04 00:43:17 +00:00
logging.info(f"{field}: {', '.join(j[field])}")
2023-02-23 16:28:30 +00:00
else:
2023-05-04 00:43:17 +00:00
logging.info(f"{field}: {j[field]}")
2023-05-10 01:49:39 +00:00
2023-05-04 00:43:17 +00:00
logging.info()
2023-02-23 16:28:30 +00:00
tensor = np.load('tensor.npy')
Path('tensor.npy').unlink()
Path('meta.json').unlink()
2023-05-04 00:43:17 +00:00
2023-02-23 16:28:30 +00:00
tensor = torch.Tensor(tensor).to(device=shared.model.device, dtype=shared.model.dtype)
tensor = torch.reshape(tensor, (1, tensor.shape[0], tensor.shape[1]))
shared.soft_prompt = True
shared.soft_prompt_tensor = tensor
return name