huggingface_opt_convert.py

# Copyright (c) 2021-2022, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

'''
Convert huggingface Meta OPT model. Use https://huggingface.co/facebook/opt-125m as demo.
'''

import argparse
import configparser
import multiprocessing
import numpy as np
import os
import sys
import torch

from datetime import datetime
from pathlib import Path
from tqdm import tqdm
from transformers import OPTForCausalLM, AutoModelForCausalLM, AutoTokenizer # transformers-4.20.0.dev0
from transformers.models.opt.modeling_opt import OPTAttention, OPTDecoderLayer

dir_path = os.path.dirname(os.path.realpath(__file__))
sys.path.append(dir_path + "/../../../..")
sys.path.append(dir_path)

class Evaluator:
    def __init__(self, dataset, tokenizer, device):
        self.dataset = dataset
        self.tokenizer = tokenizer
        self.device = device

        # tokenize the dataset
        def tokenize_function(examples):
            example = self.tokenizer(examples['text'])
            return example

        self.dataset = self.dataset.map(tokenize_function, batched=True)
        self.dataset.set_format(type='torch', columns=['input_ids'])

    @torch.no_grad()
    def evaluate(self, model):
        model.eval()
        # The task is to predict the last word of the input.
        total, hit = 0, 0
        for batch in tqdm(self.dataset):
            input_ids = batch['input_ids'].to(self.device).unsqueeze(0)
            label = input_ids[:, -1]
            outputs = model(input_ids)
            last_token_logits = outputs.logits[:, -2, :]
            pred = last_token_logits.argmax(dim=-1)
            total += label.size(0)
            hit += (pred == label).sum().item()
        acc = hit / total
        return acc

class RangeDetector(torch.nn.Module):
    def __init__(self, layer, name, save_dict):
        super().__init__()
        self.layer = layer
        self.name = f"model.{name}.weight"
        self.save_dict = save_dict

    @torch.no_grad()
    def forward(self, x):
        matmul = x @ self.layer.weight.T

        act_pre_max = x.abs().max()
        act_post_max = matmul.abs().max()

        if self.name in self.save_dict:
            act_pre_max = torch.max(act_pre_max, self.save_dict[self.name][0])
            act_post_max = torch.max(act_post_max, self.save_dict[self.name][1])
        self.save_dict[self.name] = (act_pre_max, act_post_max)

        return matmul + self.layer.bias


def inject_range_detectors(model):
    capture_dict = {}
    for name, m in model.model.named_modules():
        if isinstance(m, OPTDecoderLayer):
            m.fc1 = RangeDetector(m.fc1, f"{name}.fc1", capture_dict)
            m.fc2 = RangeDetector(m.fc2, f"{name}.fc2", capture_dict)
        elif isinstance(m, OPTAttention):
            m.q_proj = RangeDetector(m.q_proj, f"{name}.q_proj", capture_dict)
            m.k_proj = RangeDetector(m.k_proj, f"{name}.k_proj", capture_dict)
            m.v_proj = RangeDetector(m.v_proj, f"{name}.v_proj", capture_dict)
            m.out_proj = RangeDetector(m.out_proj, f"{name}.out_proj", capture_dict)
    return capture_dict


def remove_range_detectors(model):
    for name, m in model.model.named_modules():
        if isinstance(m, OPTDecoderLayer):
            m.fc1 = m.fc1.layer
            m.fc2 = m.fc2.layer
        elif isinstance(m, OPTAttention):
            m.q_proj = m.q_proj.layer
            m.k_proj = m.k_proj.layer
            m.v_proj = m.v_proj.layer
            m.out_proj = m.out_proj.layer


def get_weight_data_type(data_type):
    if data_type == "fp32":
        return np.float32
    elif data_type == "fp16":
        return np.float16
    else:
        assert False, f"Invalid weight data type {data_type}"


def quantize(mat, act_range):
    if mat.ndim == 3 and mat.shape[1] == 3:
        mat_max = np.abs(mat).clip(1e-8, None).max(axis=(0,2))[None, :, None]
    else:
        mat_max = np.abs(mat).clip(1e-8, None).max()

    act_scale_in = 127. / act_range[0].cpu().numpy()
    weight_scales = 127. / mat_max
    act_scale_post = 127. / act_range[1].cpu().numpy()

    mat_quant = (mat * weight_scales).round().astype(np.int8)
    return mat_quant, weight_scales, act_scale_in, act_scale_post


def split_and_convert_process(i, saved_dir, factor, key, args, val, capture_dict, old_name, dtype):

    def save_val(val, key, tp_num=None):
        path = saved_dir + "/model." + key
        if tp_num is not None:
            path += "." + str(tp_num)
        path += ".bin"

        val.tofile(path)

    quantized_out = args.quantize or args.act_scale is not None

    if "input_layernorm.weight" in key or "input_layernorm.bias" in key or \
        "attention.dense.bias" in key or "post_attention_layernorm.weight" in key or \
        "post_attention_layernorm.bias" in key or "mlp.dense_4h_to_h.bias" in key or \
        "final_layernorm.weight" in key or "final_layernorm.bias" in key:

        # shared weights, only need to convert the weights of rank 0
        if i == 0:
            save_val(val, key)

    elif "attention.dense.weight" in key or "mlp.dense_4h_to_h.weight" in key:
        if quantized_out:
            val_q, weight_scales, act_scale_pre, act_scale_post = quantize(val, capture_dict[old_name])
            save_val(act_scale_pre.astype(dtype), key.replace("weight", "scale"))
            scale_inter = (act_scale_post / (act_scale_pre * weight_scales)).astype(dtype)
            save_val(scale_inter, key.replace("weight", "scale_inter"))
            save_val((1. / act_scale_post).astype(dtype), key.replace("weight", "scale_out"))

            split_vals_q = np.split(val_q, factor, axis=0)
            for j in range(factor):
                save_val(split_vals_q[j], key + ".int8", i * factor + j)

        split_vals = np.split(val, factor, axis=0)
        for j in range(factor):
            save_val(split_vals[j], key, i * factor + j)

    elif "mlp.dense_h_to_4h.weight" in key or "mlp.dense_h_to_4h.bias" in key:
        if quantized_out and "weight" in key:
            val_q, weight_scales, act_scale_pre, act_scale_post = quantize(val, capture_dict[old_name])
            save_val(act_scale_pre.astype(dtype), key.replace("weight", "scale"))
            scale_inter = (act_scale_post / (act_scale_pre * weight_scales)).astype(dtype)
            save_val(scale_inter, key.replace("weight", "scale_inter"))
            save_val((1. / act_scale_post).astype(dtype), key.replace("weight", "scale_out"))

            split_vals_q = np.split(val_q, factor, axis=-1)
            for j in range(factor):
                save_val(split_vals_q[j], key + ".int8", i * factor + j)

        split_vals = np.split(val, factor, axis=-1)
        for j in range(factor):
            save_val(split_vals[j], key, i * factor + j)

    elif "attention.query_key_value.bias" in key:
        local_dim = (int)(val.shape[-1] / 3)

        val = val.reshape(3, local_dim)
        split_vals = np.split(val, factor, axis=-1)

        for j in range(factor):
            save_val(split_vals[j], key, i * factor + j)

    elif "attention.query_key_value.weight" in key:
        hidden_dim = val.shape[0]
        local_dim = (int)(val.shape[-1] / 3)

        val = val.reshape(hidden_dim, 3, local_dim)
        if quantized_out:
            val_q, weight_scales, act_scale_pre, act_scale_post = quantize(val, capture_dict[old_name])
            weight_scales = weight_scales[0] * np.ones((3, local_dim // factor))
            save_val(act_scale_pre.astype(dtype), key.replace("weight", "scale"))
            scale_inter = (act_scale_post[:, None] / (act_scale_pre[:, None] * weight_scales)).astype(dtype)
            save_val(scale_inter, key.replace("weight", "scale_inter"))
            save_val((1. / act_scale_post).astype(dtype), key.replace("weight", "scale_out"))

            split_vals_q = np.split(val_q, factor, axis=-1)
            for j in range(factor):
                save_val(split_vals_q[j], key + ".int8", i * factor + j)

        split_vals = np.split(val, factor, axis=-1)
        for j in range(factor):
            save_val(split_vals[j], key, i * factor + j)

    else:
        print("[ERROR] cannot find key '{}'".format(key))

def fuse_qkv_weight(q, k, v):
    if q.dim() == 0:
        qkv = torch.tensor((q.item(), k.item(), v.item()))
    else:
        qkv = torch.cat([q, k, v], dim=-1)
    return qkv

def split_and_convert(args):
    saved_dir = args.saved_dir + "/%d-gpu/" % args.infer_gpu_num

    if(os.path.exists(saved_dir) == False):
        os.makedirs(saved_dir)
    ckpt_name = args.in_file

    t_gpu_num = args.trained_gpu_num
    i_gpu_num = args.infer_gpu_num
    assert(i_gpu_num % t_gpu_num == 0)

    save_int8 = args.quantize or args.act_scale is not None

    factor = (int)(i_gpu_num / t_gpu_num)

    # load position_embedding from rank 0
    model = AutoModelForCausalLM.from_pretrained(args.in_file)

    capture_dict = None
    if args.quantize:
        capture_dict = inject_range_detectors(model)
        from datasets import load_dataset
        Evaluator(load_dataset('lambada', split='validation[:1000]'),
                  AutoTokenizer.from_pretrained("facebook/opt-125m"),
                  "cuda").evaluate(model.to("cuda"))
        remove_range_detectors(model)
        model.to("cpu")
    elif args.act_scale is not None:
        scales = {key: value.max() for (key, value) in torch.load(args.act_scale).items()}
        capture_dict = {}
        for key, value in scales.items():
            if key.endswith(".after"):
                continue
            capture_dict[key + ".weight"] = (value, scales[key + ".after"])

    hf_config = vars(model.config)

    num_layers = hf_config["num_hidden_layers"]

    layer_names = [name for name, param in model.named_parameters()]

    # NOTE: save parameters to config files (loaded by triton backends)
    config = configparser.ConfigParser()
    config["gpt"] = {}
    has_post_decoder_layernorm = "model.decoder.final_layer_norm.bias" in layer_names
    try:
        config["gpt"]["model_name"] = "opt" if hf_config["_name_or_path"] == '' else hf_config["_name_or_path"]
        config["gpt"]["head_num"] = str(hf_config["num_attention_heads"])
        n_embd = hf_config["hidden_size"]
        config["gpt"]["size_per_head"] = str(n_embd // hf_config["num_attention_heads"])
        config["gpt"]["inter_size"] = str(hf_config["ffn_dim"])
        config['gpt']['max_pos_seq_len'] = str(hf_config['max_position_embeddings'])
        config["gpt"]["num_layer"] = str(hf_config["num_hidden_layers"])
        config["gpt"]["layernorm_eps"] = "1e-5";
        config["gpt"]["layernorm_type"] = "pre_layernorm" if hf_config["do_layer_norm_before"] else "post_layernorm"
        config["gpt"]["activation_type"] = "Relu"
        config["gpt"]["has_post_decoder_layernorm"] = "1" if has_post_decoder_layernorm else "0"
        config["gpt"]["vocab_size"] = str(hf_config["vocab_size"])
        config["gpt"]["start_id"] = str(hf_config["bos_token_id"])
        config["gpt"]["end_id"] = str(hf_config["eos_token_id"])
        config['gpt']['weight_data_type'] = args.weight_data_type
        config['gpt']['int8'] = str(save_int8) # really useful?
        with open(saved_dir + "/config.ini", 'w') as configfile:
            config.write(configfile)
    except:
        print(f"Fail to save the config in config.ini.")

    np_weight_data_type = get_weight_data_type(args.weight_data_type)

    huggingface_model_name_pattern = [
        "self_attn_layer_norm.bias",
        "self_attn_layer_norm.weight",
        "self_attn.qkv_proj.bias",
        "self_attn.qkv_proj.weight",
        "self_attn.out_proj.bias",
        "self_attn.out_proj.weight",
        "final_layer_norm.bias",
        "final_layer_norm.weight",
        "fc1.bias",
        "fc1.weight",
        "fc2.bias",
        "fc2.weight",
    ]

    ft_model_name_pattern = [
        "input_layernorm.bias",
        "input_layernorm.weight",
        "attention.query_key_value.bias",
        "attention.query_key_value.weight",
        "attention.dense.bias",
        "attention.dense.weight",
        "post_attention_layernorm.bias",
        "post_attention_layernorm.weight",
        "mlp.dense_h_to_4h.bias",
        "mlp.dense_h_to_4h.weight",
        "mlp.dense_4h_to_h.bias",
        "mlp.dense_4h_to_h.weight",
    ]

    model_named_parameters_iter =  model.named_parameters()
    model_named_parameters = dict()
    for name, param in model_named_parameters_iter:
        if "embed" in name:
            model_named_parameters[name] = param
        elif "project_in" in name:
            model_named_parameters[name] = param.permute(1, 0)
        elif "project_out" in name:
            model_named_parameters[name] = param
        else:
            model_named_parameters[name] = param.permute(1, 0) if len(param.shape) == 2 else param

    for l in range(num_layers):
        prefix = f'model.decoder.layers.{l}.self_attn.'
        q_weight = model_named_parameters[prefix + 'q_proj.weight']
        k_weight = model_named_parameters[prefix + 'k_proj.weight']
        v_weight = model_named_parameters[prefix + 'v_proj.weight']
        q_bias = model_named_parameters[prefix + 'q_proj.bias']
        k_bias = model_named_parameters[prefix + 'k_proj.bias']
        v_bias = model_named_parameters[prefix + 'v_proj.bias']
        qkv_weight = fuse_qkv_weight(q_weight, k_weight, v_weight)
        qkv_bias = fuse_qkv_weight(q_bias, k_bias, v_bias)
        if save_int8:
            qkv_scales = (capture_dict[prefix + 'q_proj.weight'],
                          capture_dict[prefix + 'k_proj.weight'],
                          capture_dict[prefix + 'v_proj.weight'])
            capture_dict[prefix + 'qkv_proj.weight'] = (fuse_qkv_weight(qkv_scales[0][0], qkv_scales[1][0], qkv_scales[2][0]),
                                                        fuse_qkv_weight(qkv_scales[0][1], qkv_scales[1][1], qkv_scales[2][1]))
        model_named_parameters[prefix + 'qkv_proj.weight'] = qkv_weight
        model_named_parameters[prefix + 'qkv_proj.bias'] = qkv_bias

    pool = multiprocessing.Pool(args.processes)
    padding_offset = 2
    for name, param in model_named_parameters.items():
        if name == 'model.decoder.embed_positions.weight':
            param[padding_offset:,...].detach().cpu().numpy().astype(np_weight_data_type).tofile(saved_dir + "model.wpe.bin")
        elif name == 'model.decoder.embed_tokens.weight':
            if 'model.decoder.project_in.weight' in model_named_parameters.keys():
                project_in = model_named_parameters['model.decoder.project_in.weight']
                project_out = model_named_parameters['model.decoder.project_out.weight']
                torch.matmul(param, project_in).detach().cpu().numpy().astype(np_weight_data_type).tofile(saved_dir + "model.wte.bin")
                torch.matmul(param, project_out).detach().cpu().numpy().astype(np_weight_data_type).tofile(saved_dir + "model.lm_head.weight.bin")
            else:
                param.detach().cpu().numpy().astype(np_weight_data_type).tofile(saved_dir + "model.wte.bin")
                param.detach().cpu().numpy().astype(np_weight_data_type).tofile(saved_dir + "model.lm_head.weight.bin")
        elif name == 'model.decoder.final_layer_norm.weight':
            param.detach().cpu().numpy().astype(np_weight_data_type).tofile(saved_dir + "model.final_layernorm.weight.bin")
        elif name == 'model.decoder.final_layer_norm.bias':
            param.detach().cpu().numpy().astype(np_weight_data_type).tofile(saved_dir + "model.final_layernorm.bias.bin")
        elif "project_in" in name or "project_out" in name:
            continue
        else:
            starmap_args = []
            for i in range(len(huggingface_model_name_pattern)):
                if huggingface_model_name_pattern[i] in name:
                    new_name = name.replace("model.decoder.layers.", "layers.").replace(huggingface_model_name_pattern[i], ft_model_name_pattern[i])

                    starmap_args.append((0, saved_dir, factor, new_name, args,
                                        param.detach().cpu().numpy().astype(np_weight_data_type),
                                        capture_dict, name, np_weight_data_type))
            pool.starmap(split_and_convert_process, starmap_args)
    pool.close()
    pool.join()

if __name__ == "__main__":
    torch.multiprocessing.set_start_method("spawn")
    torch.multiprocessing.set_sharing_strategy("file_system")

    parser = argparse.ArgumentParser(formatter_class=argparse.RawTextHelpFormatter)
    parser.add_argument('-saved_dir', '-o', type=str, help='file name of output file', required=True)
    parser.add_argument('-in_file', '-i', type=str, help='file name of input checkpoint file', required=True)
    parser.add_argument('-trained_gpu_num', '-t_g', type=int, help='How many gpus for inference', default=1)
    parser.add_argument('-infer_gpu_num', '-i_g', type=int, help='How many gpus for inference', required=True)
    parser.add_argument("-processes", "-p", type=int, help="How many processes to spawn for conversion (default: 4)", default=4)
    parser.add_argument("-weight_data_type", type=str, default="fp16", choices=["fp32", "fp16"])
    parser.add_argument("-quantize",
                        help="Store selected in int8, run the models to determine scaling factors",
                        action="store_true")
    parser.add_argument("-act_scale", default=None, help="path to activation scalings for int8 conversion")

    args = parser.parse_args()
    print("\n=============== Argument ===============")
    for key in vars(args):
        print(f"{key}: {vars(args)[key]}")
    print("========================================")

    if args.quantize and args.act_scale is not None:
        print("[ERROR] Cannot specify both -quantize and -act_scale")
        sys.exit()

    start_time = datetime.now()
    split_and_convert(args)
    stop_time = datetime.now()
    run_time = (stop_time - start_time)
    print(f"[INFO] Spend {run_time} (h:m:s) to convert the model")