modules/models.py

import gc
import os
import re
import time
from pathlib import Path

import torch
import transformers
from accelerate import infer_auto_device_map, init_empty_weights
from transformers import (
    AutoConfig,
    AutoModel,
    AutoModelForCausalLM,
    AutoModelForSeq2SeqLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    LlamaTokenizer
)

import modules.shared as shared
from modules import llama_attn_hijack, sampler_hijack
from modules.logging_colors import logger
from modules.models_settings import infer_loader

transformers.logging.set_verbosity_error()

local_rank = None
if shared.args.deepspeed:
    import deepspeed
    from transformers.deepspeed import (
        HfDeepSpeedConfig,
        is_deepspeed_zero3_enabled
    )

    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

sampler_hijack.hijack_samplers()


def load_model(model_name, loader=None):
    logger.info(f"Loading {model_name}...")
    t0 = time.time()

    shared.is_seq2seq = False
    load_func_map = {
        'Transformers': huggingface_loader,
        'AutoGPTQ': AutoGPTQ_loader,
        'GPTQ-for-LLaMa': GPTQ_loader,
        'llama.cpp': llamacpp_loader,
        'FlexGen': flexgen_loader,
        'RWKV': RWKV_loader,
        'ExLlama': ExLlama_loader,
        'ExLlama_HF': ExLlama_HF_loader
    }

    if loader is None:
        if shared.args.loader is not None:
            loader = shared.args.loader
        else:
            loader = infer_loader(model_name)
            if loader is None:
                logger.error('The path to the model does not exist. Exiting.')
                return None, None

    shared.args.loader = loader
    output = load_func_map[loader](model_name)
    if type(output) is tuple:
        model, tokenizer = output
    else:
        model = output
        if model is None:
            return None, None
        else:
            tokenizer = load_tokenizer(model_name, model)

    # Hijack attention with xformers
    if any((shared.args.xformers, shared.args.sdp_attention)):
        llama_attn_hijack.hijack_llama_attention()

    logger.info(f"Loaded the model in {(time.time()-t0):.2f} seconds.\n")
    return model, tokenizer


def load_tokenizer(model_name, model):
    tokenizer = None
    if any(s in model_name.lower() for s in ['gpt-4chan', '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 model.__class__.__name__ in ['LlamaForCausalLM', 'LlamaGPTQForCausalLM', 'ExllamaHF']:
        # Try to load an universal LLaMA tokenizer
        if not any(s in shared.model_name.lower() for s in ['llava', 'oasst']):
            for p in [Path(f"{shared.args.model_dir}/llama-tokenizer/"), Path(f"{shared.args.model_dir}/oobabooga_llama-tokenizer/")]:
                if p.exists():
                    logger.info(f"Loading the universal LLaMA tokenizer from {p}...")
                    tokenizer = LlamaTokenizer.from_pretrained(p, clean_up_tokenization_spaces=True)
                    return tokenizer

        # Otherwise, load it from the model folder and hope that these
        # are not outdated tokenizer files.
        tokenizer = LlamaTokenizer.from_pretrained(Path(f"{shared.args.model_dir}/{model_name}/"), clean_up_tokenization_spaces=True)
        try:
            tokenizer.eos_token_id = 2
            tokenizer.bos_token_id = 1
            tokenizer.pad_token_id = 0
        except:
            pass
    else:
        path_to_model = Path(f"{shared.args.model_dir}/{model_name}/")
        if path_to_model.exists():
            tokenizer = AutoTokenizer.from_pretrained(path_to_model, trust_remote_code=shared.args.trust_remote_code)

    return tokenizer


def huggingface_loader(model_name):
    path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
    if 'chatglm' in model_name.lower():
        LoaderClass = AutoModel
    else:
        config = AutoConfig.from_pretrained(path_to_model, trust_remote_code=shared.args.trust_remote_code)
        if config.to_dict().get("is_encoder_decoder", False):
            LoaderClass = AutoModelForSeq2SeqLM
            shared.is_seq2seq = True
        else:
            LoaderClass = AutoModelForCausalLM

    # Load the model in simple 16-bit mode by default
    if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.load_in_4bit, shared.args.auto_devices, shared.args.disk, shared.args.deepspeed, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None]):
        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=shared.args.trust_remote_code)
        if torch.has_mps:
            device = torch.device('mps')
            model = model.to(device)
        else:
            model = model.cuda()

    # 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)
        model = deepspeed.initialize(model=model, config_params=ds_config, model_parameters=None, optimizer=None, lr_scheduler=None)[0]
        model.module.eval()  # Inference
        logger.info(f"DeepSpeed ZeRO-3 is enabled: {is_deepspeed_zero3_enabled()}")

    # Custom
    else:
        params = {
            "low_cpu_mem_usage": True,
            "trust_remote_code": shared.args.trust_remote_code
        }

        if not any((shared.args.cpu, torch.cuda.is_available(), torch.has_mps)):
            logger.warning("torch.cuda.is_available() returned False. This means that no GPU has been detected. Falling back to CPU mode.")
            shared.args.cpu = True

        if shared.args.cpu:
            params["torch_dtype"] = torch.float32
        else:
            params["device_map"] = 'auto'
            if shared.args.load_in_4bit:

                # See https://github.com/huggingface/transformers/pull/23479/files
                # and https://huggingface.co/blog/4bit-transformers-bitsandbytes
                quantization_config_params = {
                    'load_in_4bit': True,
                    'bnb_4bit_compute_dtype': eval("torch.{}".format(shared.args.compute_dtype)) if shared.args.compute_dtype in ["bfloat16", "float16", "float32"] else None,
                    'bnb_4bit_quant_type': shared.args.quant_type,
                    'bnb_4bit_use_double_quant': shared.args.use_double_quant,
                }

                logger.warning("Using the following 4-bit params: " + str(quantization_config_params))
                params['quantization_config'] = BitsAndBytesConfig(**quantization_config_params)

            elif 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

            params['max_memory'] = get_max_memory_dict()
            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=shared.args.trust_remote_code)
            with init_empty_weights():
                model = LoaderClass.from_config(config, trust_remote_code=shared.args.trust_remote_code)

            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)

    return model


def flexgen_loader(model_name):
    from flexgen.flex_opt import CompressionConfig, ExecutionEnv, OptLM, Policy

    # Initialize environment
    env = ExecutionEnv.create(shared.args.disk_cache_dir)

    # 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],
                    overlap=True, sep_layer=True, pin_weight=shared.args.pin_weight,
                    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)
    return model


def RWKV_loader(model_name):
    from modules.RWKV import RWKVModel, RWKVTokenizer

    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")
    tokenizer = RWKVTokenizer.from_pretrained(Path(shared.args.model_dir))
    return model, tokenizer


def llamacpp_loader(model_name):
    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]

    logger.info(f"llama.cpp weights detected: {model_file}\n")
    model, tokenizer = LlamaCppModel.from_pretrained(model_file)
    return model, tokenizer


def GPTQ_loader(model_name):

    # Monkey patch
    if shared.args.monkey_patch:
        logger.warning("Applying the monkey patch for using LoRAs with GPTQ models. It may cause undefined behavior outside its intended scope.")
        from modules.monkey_patch_gptq_lora import load_model_llama

        model, _ = load_model_llama(model_name)

    # No monkey patch
    else:
        import modules.GPTQ_loader

        model = modules.GPTQ_loader.load_quantized(model_name)

    return model


def AutoGPTQ_loader(model_name):
    import modules.AutoGPTQ_loader

    return modules.AutoGPTQ_loader.load_quantized(model_name)


def ExLlama_loader(model_name):
    from modules.exllama import ExllamaModel

    model, tokenizer = ExllamaModel.from_pretrained(model_name)
    return model, tokenizer


def ExLlama_HF_loader(model_name):
    from modules.exllama_hf import ExllamaHF

    return ExllamaHF.from_pretrained(model_name)


def get_max_memory_dict():
    max_memory = {}
    if shared.args.gpu_memory:
        memory_map = list(map(lambda x: x.strip(), shared.args.gpu_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]

        max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
        max_memory['cpu'] = f'{max_cpu_memory}GiB' if not re.match('.*ib$', max_cpu_memory.lower()) else max_cpu_memory

    # If --auto-devices is provided standalone, try to get a reasonable value
    # for the maximum memory of device :0
    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:
            suggestion -= 1000

        suggestion = int(round(suggestion / 1000))
        logger.warning(f"Auto-assiging --gpu-memory {suggestion} for your GPU to try to prevent out-of-memory errors. You can manually set other values.")
        max_memory = {0: f'{suggestion}GiB', 'cpu': f'{shared.args.cpu_memory or 99}GiB'}

    return max_memory if len(max_memory) > 0 else None


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():
    unload_model()
    shared.model, shared.tokenizer = load_model(shared.model_name)