tools/quant_llama.py

import argparse
import time
import numpy as np
import torch
import torch.nn as nn
import quant

from gptq import GPTQ
from datautils import get_loaders

def find_layers(module, layers=[nn.Conv2d, nn.Linear], name=''):
    if type(module) in layers:
        return {name: module}
    res = {}
    for name1, child in module.named_children():
        res.update(find_layers(child, layers=layers, name=name + '.' + name1 if name != '' else name1))
    return res

def get_llama(model):

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

    torch.nn.init.kaiming_uniform_ = skip
    torch.nn.init.uniform_ = skip
    torch.nn.init.normal_ = skip
    from transformers import LlamaForCausalLM
    model = LlamaForCausalLM.from_pretrained(model, torch_dtype=torch.float16)
    model.seqlen = 2048
    return model


@torch.no_grad()
def llama_sequential(model, dataloader, dev):
    print('Starting ...')

    use_cache = model.config.use_cache
    model.config.use_cache = False
    layers = model.model.layers

    model.model.embed_tokens = model.model.embed_tokens.to(dev)
    model.model.norm = model.model.norm.to(dev)
    layers[0] = layers[0].to(dev)

    dtype = next(iter(model.parameters())).dtype
    inps = torch.zeros((args.nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev)
    cache = {'i': 0, 'attention_mask': None}

    class Catcher(nn.Module):

        def __init__(self, module):
            super().__init__()
            self.module = module

        def forward(self, inp, **kwargs):
            inps[cache['i']] = inp
            cache['i'] += 1
            cache['attention_mask'] = kwargs['attention_mask']
            cache['position_ids'] = kwargs['position_ids']
            raise ValueError

    layers[0] = Catcher(layers[0])
    for batch in dataloader:
        try:
            model(batch[0].to(dev))
        except ValueError:
            pass
    layers[0] = layers[0].module

    layers[0] = layers[0].cpu()
    model.model.embed_tokens = model.model.embed_tokens.cpu()
    model.model.norm = model.model.norm.cpu()
    torch.cuda.empty_cache()

    outs = torch.zeros_like(inps)
    attention_mask = cache['attention_mask']
    position_ids = cache['position_ids']

    print('Ready.')

    quantizers = {}
    for i in range(len(layers)):

        print(f'Quantizing layer {i+1}/{len(layers)}..')
        print('+------------------+--------------+------------+-----------+-------+')
        print('|       name       | weight_error | fp_inp_SNR | q_inp_SNR | time  |')
        print('+==================+==============+============+===========+=======+')

        layer = layers[i].to(dev)
        full = find_layers(layer)
        if args.true_sequential:
            sequential = [['self_attn.k_proj', 'self_attn.v_proj', 'self_attn.q_proj'], ['self_attn.o_proj'], ['mlp.up_proj', 'mlp.gate_proj'], ['mlp.down_proj']]
        else:
            sequential = [list(full.keys())]

        for names in sequential:
            subset = {n: full[n] for n in names}
            gptq = {}
            for name in subset:
                gptq[name] = GPTQ(subset[name])
                gptq[name].quantizer.configure(args.wbits, perchannel=True, mse=False)

            def add_batch(name):

                def tmp(_, inp, out):
                    gptq[name].add_batch(inp[0].data, out.data)

                return tmp

            handles = []
            for name in subset:
                handles.append(subset[name].register_forward_hook(add_batch(name)))
            for j in range(args.nsamples):
                outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask, position_ids=position_ids)[0]
            for h in handles:
                h.remove()

            for name in subset:
                scale, zero, g_idx, error = gptq[name].fasterquant(percdamp=args.percdamp, groupsize=args.groupsize, actorder=args.act_order, name=name)
                quantizers['model.layers.%d.%s' % (i, name)] = (gptq[name].quantizer.cpu(), scale.cpu(), zero.cpu(), g_idx.cpu(), args.wbits, args.groupsize)
                gptq[name].free()

        for j in range(args.nsamples):
            outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask, position_ids=position_ids)[0]

        layers[i] = layer.cpu()
        del layer
        del gptq
        torch.cuda.empty_cache()

        inps, outs = outs, inps
        print('+------------------+--------------+------------+-----------+-------+')
        print('\n')

    model.config.use_cache = use_cache

    return quantizers


@torch.no_grad()
def llama_eval(model, testenc, dev):
    print('Evaluating ...')

    testenc = testenc.input_ids
    nsamples = testenc.numel() // model.seqlen

    use_cache = model.config.use_cache
    model.config.use_cache = False
    layers = model.model.layers

    model.model.embed_tokens = model.model.embed_tokens.to(dev)
    layers[0] = layers[0].to(dev)

    dtype = next(iter(model.parameters())).dtype
    inps = torch.zeros((nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev)
    cache = {'i': 0, 'attention_mask': None}

    class Catcher(nn.Module):

        def __init__(self, module):
            super().__init__()
            self.module = module

        def forward(self, inp, **kwargs):
            inps[cache['i']] = inp
            cache['i'] += 1
            cache['attention_mask'] = kwargs['attention_mask']
            cache['position_ids'] = kwargs['position_ids']
            raise ValueError

    layers[0] = Catcher(layers[0])
    for i in range(nsamples):
        batch = testenc[:, (i * model.seqlen):((i + 1) * model.seqlen)].to(dev)
        try:
            model(batch)
        except ValueError:
            pass
    layers[0] = layers[0].module

    layers[0] = layers[0].cpu()
    model.model.embed_tokens = model.model.embed_tokens.cpu()
    torch.cuda.empty_cache()

    outs = torch.zeros_like(inps)
    attention_mask = cache['attention_mask']
    position_ids = cache['position_ids']

    for i in range(len(layers)):
        print(i)
        layer = layers[i].to(dev)

        if args.nearest:
            subset = find_layers(layer)
            for name in subset:
                quantizer = quant.Quantizer()
                quantizer.configure(args.wbits, perchannel=True, sym=args.sym, mse=False)
                W = subset[name].weight.data
                quantizer.find_params(W, weight=True)
                subset[name].weight.data = quantizer.quantize(W).to(next(iter(layer.parameters())).dtype)

        for j in range(nsamples):
            outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask, position_ids=position_ids)[0]
        layers[i] = layer.cpu()
        del layer
        torch.cuda.empty_cache()
        inps, outs = outs, inps

    if model.model.norm is not None:
        model.model.norm = model.model.norm.to(dev)
    model.lm_head = model.lm_head.to(dev)

    testenc = testenc.to(dev)
    nlls = []
    for i in range(nsamples):
        hidden_states = inps[i].unsqueeze(0)
        if model.model.norm is not None:
            hidden_states = model.model.norm(hidden_states)
        lm_logits = model.lm_head(hidden_states)
        shift_logits = lm_logits[:, :-1, :].contiguous()
        shift_labels = testenc[:, (i * model.seqlen):((i + 1) * model.seqlen)][:, 1:]
        loss_fct = nn.CrossEntropyLoss()
        loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
        neg_log_likelihood = loss.float() * model.seqlen
        nlls.append(neg_log_likelihood)
    ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * model.seqlen))
    print(ppl.item())

    model.config.use_cache = use_cache


# TODO: perform packing on GPU
def llama_pack(model, quantizers, wbits, groupsize):
    layers = find_layers(model)
    layers = {n: layers[n] for n in quantizers}
    quant.make_quant_linear(model, quantizers, wbits, groupsize)
    qlayers = find_layers(model, [quant.QuantLinear])
    print('Packing ...')
    for name in qlayers:
        print(name)
        quantizers[name], scale, zero, g_idx, _, _ = quantizers[name]
        qlayers[name].pack(layers[name], scale, zero, g_idx)
    print('Done.')
    return model


def load_quant(model, checkpoint, wbits, groupsize=-1, fused_mlp=True, eval=True, warmup_autotune=True):
    from transformers import LlamaConfig, LlamaForCausalLM, modeling_utils
    config = LlamaConfig.from_pretrained(model)

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

    torch.nn.init.kaiming_uniform_ = noop
    torch.nn.init.uniform_ = noop
    torch.nn.init.normal_ = noop

    torch.set_default_dtype(torch.half)
    modeling_utils._init_weights = False
    torch.set_default_dtype(torch.half)
    model = LlamaForCausalLM(config)
    torch.set_default_dtype(torch.float)
    if eval:
        model = model.eval()
    layers = find_layers(model)
    for name in ['lm_head']:
        if name in layers:
            del layers[name]
    quant.make_quant_linear(model, layers, wbits, groupsize)

    del layers

    print('Loading model ...')
    if checkpoint.endswith('.safetensors'):
        from safetensors.torch import load_file as safe_load
        model.load_state_dict(safe_load(checkpoint))
    else:
        model.load_state_dict(torch.load(checkpoint))

    quant.make_quant_attn(model)
    if eval and fused_mlp:
        quant.make_fused_mlp(model)

    if warmup_autotune:
        quant.autotune_warmup_linear(model, transpose=not (eval))
        if eval and fused_mlp:
            quant.autotune_warmup_fused(model)
    model.seqlen = 2048
    print('Done.')

    return model


if __name__ == '__main__':

    parser = argparse.ArgumentParser()

    parser.add_argument('model', type=str, help='llama model to load')
    parser.add_argument('dataset', type=str, choices=['wikitext2', 'ptb', 'c4'], help='Where to extract calibration data from.')
    parser.add_argument('--seed', type=int, default=0, help='Seed for sampling the calibration data.')
    parser.add_argument('--nsamples', type=int, default=128, help='Number of calibration data samples.')
    parser.add_argument('--percdamp', type=float, default=.01, help='Percent of the average Hessian diagonal to use for dampening.')
    parser.add_argument('--wbits', type=int, default=16, choices=[2, 3, 4, 8, 16], help='#bits to use for quantization; use 16 for evaluating base model.')
    parser.add_argument('--groupsize', type=int, default=-1, help='Groupsize to use for quantization; default uses full row.')
    parser.add_argument('--eval', action='store_true', help='evaluate quantized model.')
    parser.add_argument('--save', type=str, default='', help='Save quantized checkpoint under this name.')
    parser.add_argument('--save_safetensors', type=str, default='', help='Save quantized `.safetensors` checkpoint under this name.')
    parser.add_argument('--quant-directory', type=str, default=None, help='Specify the directory for export quantization parameters to toml format. `None` means no export by default.')
    parser.add_argument('--act-order', action='store_true', help='Whether to apply the activation order GPTQ heuristic')
    parser.add_argument('--true-sequential', action='store_true', help='Whether to run in true sequential model.')
    args = parser.parse_args()

    DEV = torch.device('cuda:0')

    gpu_dist = []
    model = get_llama(args.model)
    model.eval()

    dataloader, testloader = get_loaders(args.dataset, nsamples=args.nsamples, seed=args.seed, model=args.model, seqlen=model.seqlen)

    if args.wbits < 16:
        tick = time.time()
        quantizers = llama_sequential(model, dataloader, DEV)
        print(time.time() - tick)


    if args.eval:
        datasets = ['wikitext2', 'ptb', 'c4']
        if args.new_eval:
            datasets = ['wikitext2', 'ptb-new', 'c4-new']
        for dataset in datasets:
            dataloader, testloader = get_loaders(dataset, seed=args.seed, model=args.model, seqlen=model.seqlen)
            print(dataset)
            llama_eval(model, testloader, DEV)
    
    llama_pack(model, quantizers, args.wbits, args.groupsize)
    torch.save(model.state_dict(), args.save)
# bash : CUDA_VISIBLE_DEVICES=0 proxychains python quant_llama.py ../model/llama7b_hf wikitext2 --wbits 4 --groupsize 128 --save llama7b-4bit-128g.pt