modules/GPTQ_loader.py

import inspect
import logging
import re
import sys
from pathlib import Path

import accelerate
import torch
import transformers
from transformers import AutoConfig, AutoModelForCausalLM

import modules.shared as shared

sys.path.insert(0, str(Path("repositories/GPTQ-for-LLaMa")))

try:
    import llama_inference_offload
except ImportError:
    logging.error('Failed to load GPTQ-for-LLaMa')
    logging.error('See https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md')
    sys.exit(-1)

try:
    from modelutils import find_layers
except ImportError:
    from utils import find_layers

try:
    from quant import make_quant
    is_triton = False
except ImportError:
    import quant
    is_triton = True


# This function is a replacement for the load_quant function in the
# GPTQ-for_LLaMa repository. It supports more models and branches.
def _load_quant(model, checkpoint, wbits, groupsize=-1, faster_kernel=False, exclude_layers=None, kernel_switch_threshold=128, eval=True):
    exclude_layers = exclude_layers or ['lm_head']

    def noop(*args, **kwargs):
        pass

    config = AutoConfig.from_pretrained(model, trust_remote_code=shared.args.trust_remote_code)
    torch.nn.init.kaiming_uniform_ = noop
    torch.nn.init.uniform_ = noop
    torch.nn.init.normal_ = noop

    torch.set_default_dtype(torch.half)
    transformers.modeling_utils._init_weights = False
    torch.set_default_dtype(torch.half)
    model = AutoModelForCausalLM.from_config(config, trust_remote_code=shared.args.trust_remote_code)
    torch.set_default_dtype(torch.float)
    if eval:
        model = model.eval()

    layers = find_layers(model)
    for name in exclude_layers:
        if name in layers:
            del layers[name]

    if not is_triton:
        gptq_args = inspect.getfullargspec(make_quant).args

        make_quant_kwargs = {
            'module': model,
            'names': layers,
            'bits': wbits,
        }
        if 'groupsize' in gptq_args:
            make_quant_kwargs['groupsize'] = groupsize
        if 'faster' in gptq_args:
            make_quant_kwargs['faster'] = faster_kernel
        if 'kernel_switch_threshold' in gptq_args:
            make_quant_kwargs['kernel_switch_threshold'] = kernel_switch_threshold

        make_quant(**make_quant_kwargs)
    else:
        quant.make_quant_linear(model, layers, wbits, groupsize)

    del layers
    if checkpoint.endswith('.safetensors'):
        from safetensors.torch import load_file as safe_load
        model.load_state_dict(safe_load(checkpoint), strict=False)
    else:
        model.load_state_dict(torch.load(checkpoint), strict=False)

    if is_triton:
        if shared.args.quant_attn:
            quant.make_quant_attn(model)

        if eval and shared.args.fused_mlp:
            quant.make_fused_mlp(model)

        if shared.args.warmup_autotune:
            quant.autotune_warmup_linear(model, transpose=not eval)
            if eval and shared.args.fused_mlp:
                quant.autotune_warmup_fused(model)

    model.seqlen = 2048
    return model


# Used to locate the .pt/.safetensors quantized file
def find_quantized_model_file(model_name):
    if shared.args.checkpoint:
        return Path(shared.args.checkpoint)

    path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
    pt_path = None
    priority_name_list = [
        Path(f'{shared.args.model_dir}/{model_name}{hyphen}{shared.args.wbits}bit{group}{ext}')
        for group in ([f'-{shared.args.groupsize}g', ''] if shared.args.groupsize > 0 else [''])
        for ext in ['.safetensors', '.pt']
        for hyphen in ['-', f'/{model_name}-', '/']
    ]

    for path in priority_name_list:
        if path.exists():
            pt_path = path
            break

    # If the model hasn't been found with a well-behaved name, pick the last .pt
    # or the last .safetensors found in its folder as a last resort
    if not pt_path:
        for ext in ['.pt', '.safetensors']:
            found = list(path_to_model.glob(f"*{ext}"))
            if len(found) > 0:
                if len(found) > 1:
                    logging.warning(f'More than one {ext} model has been found. The last one will be selected. It could be wrong.')

                pt_path = found[-1]
                break

    return pt_path


# The function that loads the model in modules/models.py
def load_quantized(model_name):
    if shared.args.model_type is None:
        logging.error("The model could not be loaded because its type could not be inferred from its name.")
        logging.error("Please specify the type manually using the --model_type argument.")
        return None

    # Select the appropriate load_quant function
    model_type = shared.args.model_type.lower()
    if shared.args.pre_layer and model_type == 'llama':
        load_quant = llama_inference_offload.load_quant
    elif model_type in ('llama', 'opt', 'gptj'):
        if shared.args.pre_layer:
            logging.warning("Ignoring --pre_layer because it only works for llama model type.")

        load_quant = _load_quant
    else:
        logging.error("Unknown pre-quantized model type specified. Only 'llama', 'opt' and 'gptj' are supported")
        exit()

    # Find the quantized model weights file (.pt/.safetensors)
    path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
    pt_path = find_quantized_model_file(model_name)
    if not pt_path:
        logging.error("Could not find the quantized model in .pt or .safetensors format, exiting...")
        exit()
    else:
        logging.info(f"Found the following quantized model: {pt_path}")

    # qwopqwop200's offload
    if model_type == 'llama' and shared.args.pre_layer:
        if len(shared.args.pre_layer) == 1:
            pre_layer = shared.args.pre_layer[0]
        else:
            pre_layer = shared.args.pre_layer

        model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, pre_layer)
    else:
        threshold = False if model_type == 'gptj' else 128
        model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, kernel_switch_threshold=threshold)

        # accelerate offload (doesn't work properly)
        if shared.args.gpu_memory or torch.cuda.device_count() > 1:
            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'
                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]

                max_memory['cpu'] = f'{max_cpu_memory}GiB' if not re.match('.*ib$', max_cpu_memory.lower()) else max_cpu_memory
            else:
                max_memory = accelerate.utils.get_balanced_memory(model)

            device_map = accelerate.infer_auto_device_map(model, max_memory=max_memory, no_split_module_classes=["LlamaDecoderLayer"])
            logging.info("Using the following device map for the quantized model:", device_map)
            # https://huggingface.co/docs/accelerate/package_reference/big_modeling#accelerate.dispatch_model
            model = accelerate.dispatch_model(model, device_map=device_map, offload_buffers=True)

        # No offload
        elif not shared.args.cpu:
            model = model.to(torch.device('cuda:0'))

    return model