convert_checkpoint.py

import argparse
import copy
import functools
import json
import os
import time
import traceback
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path

import numpy as np
import safetensors
import torch
import torch.nn as nn
from tqdm import tqdm
from transformers import LlamaConfig, LlamaForCausalLM, LlamaTokenizer
from transformers.models.llama.modeling_llama import LlamaDecoderLayer
from transformers.pytorch_utils import Conv1D

import tensorrt_llm
from tensorrt_llm._utils import str_dtype_to_torch
from tensorrt_llm.logger import logger
from tensorrt_llm.mapping import Mapping
from tensorrt_llm.models import PretrainedConfig
from tensorrt_llm.models.convert_utils import load_calib_dataset
from tensorrt_llm.models.llama.convert import load_weights_from_hf_by_shard
from tensorrt_llm.quantization import QuantAlgo

try:
    from transformers import MixtralForCausalLM
except ImportError:
    MixtralForCausalLM = None


def parse_arguments():
    parser = argparse.ArgumentParser()
    parser.add_argument('--model_dir', type=str, default=None)
    parser.add_argument('--meta_ckpt_dir', type=str, default=None)
    parser.add_argument('--tp_size',
                        type=int,
                        default=1,
                        help='N-way tensor parallelism size')
    parser.add_argument('--pp_size',
                        type=int,
                        default=1,
                        help='N-way pipeline parallelism size')
    parser.add_argument('--dtype',
                        type=str,
                        default='float16',
                        choices=['float32', 'bfloat16', 'float16'])
    parser.add_argument('--vocab_size', type=int, default=32000)
    parser.add_argument('--n_positions', type=int, default=2048)
    parser.add_argument('--n_layer', type=int, default=32)

    parser.add_argument(
        '--use_weight_only',
        default=False,
        action="store_true",
        help='Quantize weights for the various GEMMs to INT4/INT8.'
        'See --weight_only_precision to set the precision')
    parser.add_argument(
        '--weight_only_precision',
        const='int8',
        type=str,
        nargs='?',
        default='int8',
        choices=['int8', 'int4', 'int4_gptq'],
        help=
        'Define the precision for the weights when using weight-only quantization.'
        'You must also use --use_weight_only for that argument to have an impact.'
    )
    parser.add_argument(
        '--calib_dataset',
        type=str,
        default='ccdv/cnn_dailymail',
        help=
        "The huggingface dataset name or the local directory of the dataset for calibration."
    )
    parser.add_argument(
        "--smoothquant",
        "-sq",
        type=float,
        default=None,
        help="Set the α parameter (see https://arxiv.org/pdf/2211.10438.pdf)"
        " to Smoothquant the model, and output int8 weights."
        " A good first try is 0.5. Must be in [0, 1]")
    parser.add_argument(
        '--per_channel',
        action="store_true",
        default=False,
        help=
        'By default, we use a single static scaling factor for the GEMM\'s result. '
        'per_channel instead uses a different static scaling factor for each channel. '
        'The latter is usually more accurate, but a little slower.')
    parser.add_argument(
        '--per_token',
        action="store_true",
        default=False,
        help=
        'By default, we use a single static scaling factor to scale activations in the int8 range. '
        'per_token chooses at run time, and for each token, a custom scaling factor. '
        'The latter is usually more accurate, but a little slower.')
    parser.add_argument(
        '--int8_kv_cache',
        default=False,
        action="store_true",
        help=
        'By default, we use dtype for KV cache. int8_kv_cache chooses int8 quantization for KV'
    )
    parser.add_argument(
        '--modelopt_quant_ckpt_path',
        type=str,
        default=None,
        help='Path of a quantized model checkpoint in .npz format')

    parser.add_argument(
        '--per_group',
        default=False,
        action="store_true",
        help=
        'By default, we use a single static scaling factor to scale weights in the int4 range. '
        'per_group chooses at run time, and for each group, a custom scaling factor. '
        'The flag is built for GPTQ/AWQ quantization.')

    parser.add_argument('--load_by_shard',
                        action='store_true',
                        help='Load a pretrained model shard-by-shard.')
    parser.add_argument('--hidden_act', type=str, default='silu')

    parser.add_argument('--rotary_base', type=float, default=10000.0)
    parser.add_argument('--rotary_scaling', nargs=2, type=str, default=None)

    parser.add_argument('--group_size',
                        type=int,
                        default=128,
                        help='Group size used in GPTQ/AWQ quantization.')

    parser.add_argument("--storage-type",
                        "-t",
                        type=str,
                        default="fp32",
                        choices=["fp32", "fp16"])
    parser.add_argument("--dataset-cache-dir",
                        type=str,
                        default=None,
                        help="cache dir to load the hugging face dataset")
    parser.add_argument("--load-model-on-cpu", action="store_true")
    parser.add_argument("--convert-model-on-cpu", action="store_true")
    parser.add_argument(
        '--use_parallel_embedding',
        action="store_true",
        default=False,
        help=
        'By default embedding parallelism is disabled. By setting this flag, embedding parallelism is enabled'
    )
    parser.add_argument(
        '--embedding_sharding_dim',
        type=int,
        default=0,
        choices=[0, 1],
        help=
        'By default the embedding lookup table is sharded along vocab dimension (embedding_sharding_dim=0). '
        'To shard it along hidden dimension, set embedding_sharding_dim=1'
        'Note: embedding sharing is only enabled when embedding_sharding_dim = 0'
    )
    parser.add_argument(
        '--use_embedding_sharing',
        action="store_true",
        default=False,
        help=
        'Try to reduce the engine size by sharing the embedding lookup table between two layers.'
        'Note: the flag might not take effect when the criteria are not met.')
    parser.add_argument('--output_dir',
                        type=str,
                        default='tllm_checkpoint',
                        help='The path to save the TensorRT-LLM checkpoint')
    parser.add_argument(
        '--workers',
        type=int,
        default=1,
        help='The number of workers for converting checkpoint in parallel')

    parser.add_argument('--num_medusa_heads', type=int, default=4)
    parser.add_argument(
        '--fixed_num_medusa_heads',
        type=int,
        default=None,
        help="If exist, fix medusa_num_heads from config.json."
        "num_medusa_heads < medusa_num_heads in config.json < fixed_num_medusa_heads"
    )
    parser.add_argument('--num_medusa_layers', type=int, default=1)
    parser.add_argument('--max_medusa_token_len', type=int, default=63)
    parser.add_argument('--medusa_hidden_act', type=str, default="silu")
    parser.add_argument('--medusa_model_dir', type=str, default=None)
    args = parser.parse_args()
    return args


def generate_int8(weights, act_range, is_qkv=False, multi_query_mode=False):
    """
     This function has two purposes:
      - compute quantized weights, scaled either per-tensor or per-column
      - compute scaling factors

      Depending on the GEMM API (CUTLASS/CUBLAS) the required scaling factors differ.
      CUTLASS uses two sets of scaling factors. One for the activation X, one for the weight W.
      CUBLAS only has one (we can't do per-row scaling). So we must provide pre-multiplied scaling factor.

      Here is the list of what we need (T means per-tensor, C per-column):
        - scale_x_orig_quant puts fp activation into the quantized range (i.e. [-128, 127], for int8). Used before the GEMM. (T)
        - scale_y_quant_orig puts quantized activation into the fp range. Used if the GEMM outputs int8. (T)
        - scale_w_quant_orig puts weights from quant range to fp range (used with CUTLASS) (T, C)
        - scale_y_accum_quant puts the GEMM result (XW) from accumulation range (int32)
          to quant range (int8) (used for CUBLAS) (T, C)

      Note that we don't do anything special about row-parallel GEMM. Theoretically, we could have per-GPU scaling factors too,
      but then the model would change depending on the number of GPUs used.

      For QKV projection, the behavior is special. Even if we have a single matrix to perform QKV projection, we consider it
      as three different matrices: Q, K, and V. So per-tensor actually means one scaling factor for each Q, K and V.
      For our GEMM implementation to respect this behavior, we use per-column mode and replicate values along columns.
    """
    weights = weights.detach().cpu().numpy()

    # compute weight scaling factors for fp->int8 and int8->fp
    if is_qkv and not multi_query_mode:
        scale_w_orig_quant_t = 127. / act_range["w"].reshape(3, -1).max(
            dim=-1, keepdims=True)[0].cpu().numpy()
        scale_w_orig_quant_c = 127. / act_range["w"].reshape(3,
                                                             -1).cpu().numpy()
    elif is_qkv and multi_query_mode:
        hidden_dim = weights.shape[0]
        local_dim = act_range["w"].shape[0]
        kv_dim = (local_dim - hidden_dim) // 2
        scale_w_q = act_range["w"][0:hidden_dim]
        scale_w_k = act_range["w"][hidden_dim:hidden_dim + kv_dim]
        scale_w_v = act_range["w"][-kv_dim:]

        scale_w_qkv_t = torch.concat([
            scale_w_q.max(dim=0, keepdim=True)[0],
            scale_w_k.max(dim=0, keepdim=True)[0],
            scale_w_v.max(dim=0, keepdim=True)[0]
        ])

        scale_w_orig_quant_t = 127. / scale_w_qkv_t.cpu().numpy()
        scale_w_orig_quant_c = 127. / act_range["w"].cpu().numpy()
    else:
        scale_w_orig_quant_t = 127. / act_range["w"].max().cpu().numpy()
        scale_w_orig_quant_c = 127. / act_range["w"].cpu().numpy()
    scale_w_quant_orig_t = 1.0 / scale_w_orig_quant_t
    scale_w_quant_orig_c = 1.0 / scale_w_orig_quant_c

    # compute the rest of needed scaling factors
    scale_x_orig_quant_t = np.array(127. / act_range["x"].max().item())
    scale_y_orig_quant_t = np.array(127. / act_range["y"].max().item())
    scale_y_quant_orig_t = np.array(act_range["y"].max().item() / 127.)
    scale_y_accum_quant_t = scale_y_orig_quant_t / (scale_x_orig_quant_t *
                                                    scale_w_orig_quant_t)
    scale_y_accum_quant_c = scale_y_orig_quant_t / (scale_x_orig_quant_t *
                                                    scale_w_orig_quant_c)
    if is_qkv and not multi_query_mode:
        scale_y_accum_quant_t = np.broadcast_to(scale_y_accum_quant_t,
                                                scale_w_orig_quant_c.shape)
        scale_w_quant_orig_t = np.broadcast_to(scale_w_quant_orig_t,
                                               scale_w_orig_quant_c.shape)
    if is_qkv and multi_query_mode:
        scale_q_y_accum_t = np.broadcast_to(scale_y_accum_quant_t[0],
                                            scale_w_q.shape)
        scale_k_y_accum_t = np.broadcast_to(scale_y_accum_quant_t[1],
                                            scale_w_k.shape)
        scale_v_y_accum_t = np.broadcast_to(scale_y_accum_quant_t[2],
                                            scale_w_v.shape)
        scale_y_accum_quant_t = np.concatenate(
            [scale_q_y_accum_t, scale_k_y_accum_t, scale_v_y_accum_t])
        scale_w_quant_orig_t = np.concatenate([
            np.broadcast_to(scale_w_quant_orig_t[0], scale_w_q.shape),
            np.broadcast_to(scale_w_quant_orig_t[1], scale_w_k.shape),
            np.broadcast_to(scale_w_quant_orig_t[2], scale_w_v.shape)
        ])

    to_i8 = lambda x: x.round().clip(-127, 127).astype(np.int8)

    if is_qkv and multi_query_mode:
        scale_w_quant_orig_t_expand = np.ones([weights.shape[-1]])
        scale_w_quant_orig_t_expand[:hidden_dim] = scale_w_quant_orig_t[0]
        scale_w_quant_orig_t_expand[hidden_dim:hidden_dim +
                                    kv_dim] = scale_w_quant_orig_t[1]
        scale_w_quant_orig_t_expand[-kv_dim:] = scale_w_quant_orig_t[2]
        weight_int8 = to_i8(weights * scale_w_quant_orig_t_expand)
    else:
        weight_int8 = to_i8(weights * scale_w_orig_quant_t)
    return {
        "weight.int8": weight_int8,
        "weight.int8.col": to_i8(weights * scale_w_orig_quant_c),
        "scale_x_orig_quant": scale_x_orig_quant_t.astype(np.float32),
        "scale_w_quant_orig": scale_w_quant_orig_t.astype(np.float32),
        "scale_w_quant_orig.col": scale_w_quant_orig_c.astype(np.float32),
        "scale_y_accum_quant": scale_y_accum_quant_t.astype(np.float32),
        "scale_y_accum_quant.col": scale_y_accum_quant_c.astype(np.float32),
        "scale_y_quant_orig": scale_y_quant_orig_t.astype(np.float32),
    }


@torch.no_grad()
def apply_smoothing(scales,
                    gemm_weights,
                    layernorm_weights=None,
                    layernorm_bias=None,
                    dtype=torch.float32,
                    layernorm_1p=False):
    if not isinstance(gemm_weights, list):
        gemm_weights = [gemm_weights]

    if layernorm_weights is not None:
        assert layernorm_weights.numel() == scales.numel()
        layernorm_weights.div_(scales).to(dtype)
    if layernorm_bias is not None:
        assert layernorm_bias.numel() == scales.numel()
        layernorm_bias.div_(scales).to(dtype)
    if layernorm_1p:
        layernorm_weights += (1 / scales) - 1

    for gemm in gemm_weights:
        gemm.mul_(scales.view(1, -1)).to(dtype)


@torch.no_grad()
def smooth_gemm(gemm_weights,
                act_scales,
                layernorm_weights=None,
                layernorm_bias=None,
                alpha=0.5,
                weight_scales=None):
    if not isinstance(gemm_weights, list):
        gemm_weights = [gemm_weights]
    orig_dtype = gemm_weights[0].dtype

    for gemm in gemm_weights:
        # gemm_weights are expected to be transposed
        assert gemm.shape[1] == act_scales.numel()

    if weight_scales is None:
        weight_scales = torch.cat(
            [gemm.abs().max(dim=0, keepdim=True)[0] for gemm in gemm_weights],
            dim=0)
        weight_scales = weight_scales.max(dim=0)[0]
    weight_scales.to(float).clamp(min=1e-5)
    scales = (act_scales.to(gemm_weights[0].device).to(float).pow(alpha) /
              weight_scales.pow(1 - alpha)).clamp(min=1e-5)

    apply_smoothing(scales, gemm_weights, layernorm_weights, layernorm_bias,
                    orig_dtype)

    return scales


@torch.no_grad()
def smooth_gemm_fc1_gate(fc1_weights,
                         gate_weights,
                         act_scales,
                         layernorm_weights=None,
                         layernorm_bias=None,
                         alpha=0.5,
                         weight_scales=None):
    gemm_weights = []
    if not isinstance(fc1_weights, list):
        fc1_weights = [fc1_weights]
    if not isinstance(gate_weights, list):
        gate_weights = [gate_weights]

    for i in range(len(fc1_weights)):
        gemm_weight = torch.cat([fc1_weights[i], gate_weights[i]], dim=0)
        gemm_weights.append(gemm_weight)

    orig_dtype = gemm_weights[0].dtype

    for gemm in gemm_weights:
        # gemm_weights are expected to be transposed
        assert gemm.shape[1] == act_scales.numel()

    if weight_scales is None:
        weight_scales = torch.cat(
            [gemm.abs().max(dim=0, keepdim=True)[0] for gemm in gemm_weights],
            dim=0)
        weight_scales = weight_scales.max(dim=0)[0]
    weight_scales.to(float).clamp(min=1e-5)
    scales = (act_scales.to(gemm_weights[0].device).to(float).pow(alpha) /
              weight_scales.pow(1 - alpha)).clamp(min=1e-5)

    apply_smoothing(scales, fc1_weights + gate_weights, layernorm_weights,
                    layernorm_bias, orig_dtype)

    return scales


@torch.no_grad()
def smooth_llama_model(model, scales, alpha, llama_qkv_para, llama_smoother):
    # Smooth the activation and weights with smoother = $\diag{s}$
    for name, module in model.named_modules():
        if not isinstance(module, LlamaDecoderLayer):
            continue
        # qkv_proj
        layer_name_q = name + ".self_attn.q_proj"
        layer_name_k = name + ".self_attn.k_proj"
        layer_name_v = name + ".self_attn.v_proj"
        layer_name_qkv = name + ".self_attn.qkv_proj"

        weight = torch.cat([
            module.self_attn.q_proj.weight, module.self_attn.k_proj.weight,
            module.self_attn.v_proj.weight
        ],
                           dim=0)

        smoother = smooth_gemm(weight, scales[layer_name_q]["x"],
                               module.input_layernorm.weight, None, alpha)

        scales[layer_name_qkv]["x"] = scales[layer_name_q]["x"] / smoother
        scales[layer_name_qkv]["w"] = weight.abs().max(dim=1)[0]
        scales[layer_name_qkv]["y"] = torch.cat([
            scales[layer_name_q]["y"], scales[layer_name_k]["y"],
            scales[layer_name_v]["y"]
        ],
                                                dim=0)

        # see transpose_weights function
        llama_qkv_para[layer_name_qkv] = weight.transpose(0, 1)

        # =================================================================
        layer_name = name + ".self_attn.o_proj"
        smoother = smooth_gemm(module.self_attn.o_proj.weight,
                               scales[layer_name]["x"], None, None, alpha)
        llama_smoother[layer_name] = smoother.float()

        scales[layer_name]["x"] = scales[layer_name]["x"] / smoother
        scales[layer_name]["w"] = module.self_attn.o_proj.weight.abs().max(
            dim=1)[0]

        # ==================================================================
        fc1_layer_name = name + ".mlp.gate_proj"
        gate_layer_name = name + ".mlp.up_proj"

        smoother = smooth_gemm_fc1_gate(module.mlp.gate_proj.weight,
                                        module.mlp.up_proj.weight,
                                        scales[fc1_layer_name]["x"],
                                        module.post_attention_layernorm.weight,
                                        None, alpha)

        scales[fc1_layer_name]["x"] = scales[fc1_layer_name]["x"] / smoother
        scales[fc1_layer_name]["w"] = module.mlp.gate_proj.weight.abs().max(
            dim=1)[0]

        scales[gate_layer_name]["x"] = scales[gate_layer_name]["x"] / smoother
        scales[gate_layer_name]["w"] = module.mlp.up_proj.weight.abs().max(
            dim=1)[0]

        # ==================================================================
        layer_name = name + ".mlp.down_proj"
        smoother = smooth_gemm(module.mlp.down_proj.weight,
                               scales[layer_name]["x"], None, None, alpha)
        llama_smoother[layer_name] = smoother.float()
        scales[layer_name]["x"] = scales[layer_name]["x"] / smoother
        scales[layer_name]["w"] = module.mlp.down_proj.weight.abs().max(
            dim=1)[0]


@torch.no_grad()
def capture_activation_range(model,
                             tokenizer,
                             dataset,
                             num_samples=512,
                             seq_len=512):
    model.eval()
    device = next(model.parameters()).device
    act_scales = defaultdict(lambda: {"x": None, "y": None, "w": None})

    tokenizer.pad_token = tokenizer.eos_token

    def stat_tensor(name, tensor, act_scales, key):
        hidden_dim = tensor.shape[-1]
        tensor = tensor.view(-1, hidden_dim).abs().detach()
        comming_max = torch.max(tensor, dim=0)[0].float()

        if act_scales[name][key] is None:
            act_scales[name][key] = comming_max
        else:
            act_scales[name][key] = torch.max(act_scales[name][key],
                                              comming_max)

    def stat_input_hook(m, x, y, name):
        if isinstance(x, tuple):
            x = x[0]
        stat_tensor(name, x, act_scales, "x")
        stat_tensor(name, y, act_scales, "y")

        if act_scales[name]["w"] is None:
            act_scales[name]["w"] = m.weight.abs().clip(1e-8,
                                                        None).max(dim=1)[0]

    hooks = []
    for name, m in model.named_modules():
        if isinstance(m, nn.Linear) or isinstance(m, Conv1D):
            hooks.append(
                m.register_forward_hook(
                    functools.partial(stat_input_hook, name=name)))

    for i in tqdm(range(num_samples), desc="calibrating model"):
        datapoint = dataset[i:i + 1]
        line = copy.copy(datapoint)
        line[0] = line[0] + ' TL;DR: '
        line[0] = line[0].strip()
        line[0] = line[0].replace(" n't", "n't")
        input_ids = tokenizer(line,
                              return_tensors="pt",
                              max_length=seq_len,
                              padding=True,
                              truncation=True).input_ids.to(device)
        model(input_ids)
    for h in hooks:
        h.remove()
    return act_scales


def split(v, tp_size, idx, dim=0):
    if tp_size == 1:
        return v
    if len(v.shape) == 1:
        return torch.chunk(v, tp_size)[idx].contiguous()
    else:
        return torch.chunk(v, tp_size, dim=dim)[idx].contiguous()


def split_qkv_tp(v, n_head, n_hidden, tensor_parallel, rank):
    """
    Splits the QKV matrix according to tensor parallelism
    """
    v = v.reshape(3, n_hidden, n_hidden)
    split_v = split(v, tensor_parallel, rank, dim=1)
    split_v = split_v.reshape(3 * (n_hidden // tensor_parallel), n_hidden)
    return split_v.contiguous()


def split_qkv_bias_tp(v, n_head, n_hidden, tensor_parallel, rank):
    """
    Splits the QKV bias according to tensor parallelism
    """
    v = v.reshape(3, n_hidden)
    split_v = split(v, tensor_parallel, rank, dim=1)
    split_v = split_v.reshape(3 * (n_hidden // tensor_parallel))
    return split_v.contiguous()


def split_matrix_tp(v, tensor_parallel, rank, dim):
    return split(v, tensor_parallel, rank, dim=dim)


def get_weight(config, prefix, dtype):
    if config[prefix + '.weight'].dtype != dtype:
        config[prefix + '.weight'].data = config[prefix + '.weight'].to(dtype)
    return config[prefix + '.weight']


def get_bias(config, prefix, dtype):
    if config[prefix + '.bias'].dtype != dtype:
        config[prefix + '.bias'].data = config[prefix + '.bias'].to(dtype)
    return config[prefix + '.bias']


def get_weight_and_bias(config, prefix, dtype):
    return get_weight(config, prefix, dtype), get_bias(config, prefix, dtype)


def get_tllm_linear_weight(weight,
                           prefix,
                           bias=None,
                           use_weight_only=False,
                           plugin_weight_only_quant_type=torch.int8,
                           postfix='weight'):
    results = {}
    if use_weight_only:
        v = weight.t().contiguous().cpu()
        processed_torch_weights, torch_weight_scales = \
            torch.ops.trtllm.symmetric_quantize_last_axis_of_batched_matrix(
                v, plugin_weight_only_quant_type)
        results[prefix + postfix] = processed_torch_weights
        results[prefix + 'per_channel_scale'] = torch_weight_scales
    else:
        results[prefix + postfix] = weight.contiguous()

    if bias is not None:
        results[prefix + 'bias'] = bias

    return results


def dup_kv_weight(v, num_head, tp_size):
    assert tp_size % num_head == 0
    reps = tp_size // num_head
    head_size = v.shape[0] // num_head
    v = v.reshape(num_head, head_size,
                  -1)[:, None, :, :].expand(num_head, reps, head_size,
                                            v.shape[1])
    return v.reshape(num_head * reps * head_size, -1).clone().detach()


def get_tllm_linear_sq_weight(vals,
                              prefix,
                              shape,
                              tensor_parallel,
                              is_qkv=False,
                              per_token=False,
                              per_channel=False,
                              last_prefix=None,
                              bias=None,
                              smoother_value=None,
                              smoother_shape=None,
                              rank=0,
                              cat_dim=0,
                              multi_query_mode=False):
    results = {}

    def multi_query_split(data, local_dim, head_size, tp_size, cur_rank):
        q, k, v = np.split(data, [local_dim, local_dim + head_size], axis=-1)
        q_split = np.split(q, tp_size, axis=-1)
        k_split = np.split(k, tp_size, axis=-1)
        v_split = np.split(v, tp_size, axis=-1)
        return [
            np.concatenate((q_split[ii], k_split[ii], v_split[ii]), axis=-1)
            for ii in range(tp_size)
        ][cur_rank]

    col_shape = shape if (is_qkv or per_channel) else [1, 1]

    if per_token:
        original_weights = vals["weight.int8.col"]

        local_dim = original_weights.shape[0]
        head_size = (original_weights.shape[1] - local_dim) // 2
        if multi_query_mode:
            cur_weights = multi_query_split(original_weights, local_dim,
                                            head_size, tensor_parallel, rank)
        else:
            cur_weights = np.split(original_weights,
                                   tensor_parallel,
                                   axis=cat_dim)[rank]
        if is_qkv:
            hidden_dim = cur_weights.shape[0]
            cur_weights = cur_weights.reshape(hidden_dim, -1)
        results[prefix +
                'weight'] = torch.from_numpy(cur_weights).t().contiguous()
        if smoother_value is None:
            results[last_prefix] = torch.from_numpy(
                np.array([1.0], dtype=np.float32))

        if smoother_value is None:
            if multi_query_mode:
                cur_per_channel_value = multi_query_split(
                    vals["scale_w_quant_orig.col"], local_dim, head_size,
                    tensor_parallel, rank)
            else:
                cur_per_channel_value = np.split(vals["scale_w_quant_orig.col"],
                                                 tensor_parallel,
                                                 axis=cat_dim)[rank]
        else:
            cur_per_channel_value = vals["scale_w_quant_orig.col"]
        results[prefix + 'per_channel_scale'] = torch.from_numpy(
            np.array(cur_per_channel_value,
                     dtype=np.float32).reshape(col_shape)).contiguous()
    else:
        original_weights = np.array(vals["weight.int8"])
        cur_weights = np.split(original_weights, tensor_parallel,
                               axis=cat_dim)[rank]

        if is_qkv:
            hidden_dim = cur_weights.shape[0]
            cur_weights = cur_weights.reshape(hidden_dim, -1)
        results[prefix +
                'weight'] = torch.from_numpy(cur_weights).t().contiguous()

        cur_per_channel_value = vals["scale_y_accum_quant"]

        results[prefix + 'per_channel_scale'] = torch.from_numpy(
            np.array([cur_per_channel_value],
                     dtype=np.float32).reshape(col_shape)).contiguous()

        results[last_prefix] = torch.from_numpy(
            np.array([vals['scale_x_orig_quant']],
                     dtype=np.float32)).contiguous()

        results[prefix + 'act_scale'] = torch.from_numpy(
            np.array([[vals["scale_y_quant_orig"]]],
                     dtype=np.float32)).contiguous()

    if smoother_value is not None:
        cur_smoother_value = np.split(smoother_value,
                                      tensor_parallel,
                                      axis=cat_dim)[rank]
        results[prefix + 'smoother'] = cur_smoother_value.reshape(
            smoother_shape).contiguous().to(torch.float32)

    if bias is not None:
        results[prefix + 'bias'] = bias

    return results


def convert_hf_llama(hf_model,
                     mapping,
                     rank=0,
                     dtype='float32',
                     use_parallel_embedding=False,
                     sharding_dim=0,
                     use_weight_only=False,
                     share_embedding_table=False,
                     plugin_weight_only_quant_type=torch.int8,
                     use_smooth_quant=False,
                     per_channel=False,
                     per_token=False,
                     int8_kv_cache=False,
                     act_range=[],
                     qkv_para=[],
                     smoother=[],
                     lora_config=None):

    weights = {}
    tik = time.time()
    tensor_parallel = mapping.tp_size
    model_params = dict(hf_model.named_parameters())
    dtype = getattr(torch, dtype)
    num_attention_heads = hf_model.config.num_attention_heads
    hidden_size = hf_model.config.hidden_size
    intermediate_size = hf_model.config.intermediate_size
    num_key_value_heads = hf_model.config.num_key_value_heads
    mha_mode = (num_key_value_heads == num_attention_heads)

    num_hidden_layers = hf_model.config.num_hidden_layers
    layers_range = mapping.pp_layers(num_hidden_layers)
    for l in layers_range:
        layer_idx = l - layers_range[0]
        prefix = f'model.layers.{l}.'
        tllm_prex = f'transformer.layers.{layer_idx}.'

        q_weight = get_weight(model_params, prefix + 'self_attn.q_proj', dtype)
        k_weight = get_weight(model_params, prefix + 'self_attn.k_proj', dtype)
        v_weight = get_weight(model_params, prefix + 'self_attn.v_proj', dtype)

        if not mha_mode:
            head_size = hidden_size // num_attention_heads
            if num_key_value_heads < tensor_parallel:
                # duplicate the KV heads up to tensor_parallel
                k_weight = dup_kv_weight(k_weight, num_key_value_heads,
                                         tensor_parallel)
                v_weight = dup_kv_weight(v_weight, num_key_value_heads,
                                         tensor_parallel)
            assert (k_weight.shape[0] % (mapping.tp_size * head_size)) == 0
            assert (v_weight.shape[0] % (mapping.tp_size * head_size)) == 0

            wq = split(q_weight, mapping.tp_size, mapping.tp_rank)
            wk = split(k_weight, mapping.tp_size, mapping.tp_rank)
            wv = split(v_weight, mapping.tp_size, mapping.tp_rank)

            split_v = torch.concat((wq, wk, wv))

        else:
            qkv_weight = torch.cat([q_weight, k_weight, v_weight], dim=0)

            split_v = split_qkv_tp(qkv_weight, num_attention_heads, hidden_size,
                                   tensor_parallel, mapping.tp_rank)
        if use_smooth_quant:
            qkv_weight = qkv_para[prefix + 'self_attn.qkv_proj']

            if not mha_mode:
                hidden_size = qkv_weight.shape[0]
                local_dim = hidden_size
                head_size = (qkv_weight.shape[-1] - local_dim) // 2
                qkv_weight = qkv_weight.reshape(hidden_size,
                                                local_dim + 2 * head_size)
            else:
                qkv_weight = qkv_weight.reshape(hidden_size, 3, hidden_size)

            int8_weights = generate_int8(qkv_weight,
                                         act_range.get(prefix +
                                                       'self_attn.qkv_proj'),
                                         is_qkv=True,
                                         multi_query_mode=bool(not mha_mode))

            weights.update(
                get_tllm_linear_sq_weight(
                    int8_weights,
                    tllm_prex + 'attention.qkv.', [
                        1, 3 * hidden_size // tensor_parallel
                        if mha_mode else hidden_size // tensor_parallel +
                        (hidden_size // num_key_value_heads) //
                        tensor_parallel * 2
                    ],
                    tensor_parallel,
                    is_qkv=True,
                    per_token=per_token,
                    per_channel=per_channel,
                    last_prefix=tllm_prex + 'input_layernorm.scale_to_int',
                    smoother_value=None,
                    smoother_shape=None,
                    rank=mapping.tp_rank,
                    cat_dim=-1,
                    multi_query_mode=bool(not mha_mode)))
        else:
            weights.update(
                get_tllm_linear_weight(split_v, tllm_prex + 'attention.qkv.',
                                       None, use_weight_only,
                                       plugin_weight_only_quant_type))

        if int8_kv_cache:
            qkv_y = torch.cat([
                act_range.get(prefix + 'self_attn.q_proj')["y"],
                act_range.get(prefix + 'self_attn.k_proj')["y"],
                act_range.get(prefix + 'self_attn.v_proj')["y"]
            ],
                              dim=0)

            int8_kv_scales = qkv_y.max() / 127.

            kv_cache_weights = {}

            kv_cache_weights[
                tllm_prex +
                'attention.kv_cache_scaling_factor'] = int8_kv_scales.reshape(
                    [1])

        attn_dense_weight = get_weight(model_params,
                                       prefix + 'self_attn.o_proj', dtype)
        split_v = split_matrix_tp(attn_dense_weight,
                                  tensor_parallel,
                                  mapping.tp_rank,
                                  dim=1)
        if use_smooth_quant:
            attn_dense_weight = attn_dense_weight.t()
            int8_weights = generate_int8(
                attn_dense_weight, act_range.get(prefix + 'self_attn.o_proj'))
            weights.update(
                get_tllm_linear_sq_weight(
                    int8_weights,
                    tllm_prex + 'attention.dense.', [1, hidden_size],
                    tensor_parallel,
                    is_qkv=False,
                    per_token=per_token,
                    per_channel=per_channel,
                    last_prefix=tllm_prex +
                    'attention.quantization_scaling_factor',
                    smoother_value=smoother[(prefix + 'self_attn.o_proj')],
                    smoother_shape=[1, hidden_size // tensor_parallel],
                    rank=mapping.tp_rank,
                    cat_dim=0))
        else:
            weights.update(
                get_tllm_linear_weight(split_v, tllm_prex + 'attention.dense.',
                                       None, use_weight_only,
                                       plugin_weight_only_quant_type))

        mlp_gate_weight = get_weight(model_params, prefix + 'mlp.up_proj',
                                     dtype)
        split_v = split_matrix_tp(mlp_gate_weight,
                                  tensor_parallel,
                                  mapping.tp_rank,
                                  dim=0)
        if use_smooth_quant:
            mlp_gate_weight = mlp_gate_weight.t()
            int8_weights = generate_int8(mlp_gate_weight,
                                         act_range.get(prefix + 'mlp.up_proj'))

            weights.update(
                get_tllm_linear_sq_weight(
                    int8_weights,
                    tllm_prex + 'mlp.gate.',
                    [1, intermediate_size // tensor_parallel],
                    tensor_parallel,
                    is_qkv=False,
                    per_token=per_token,
                    per_channel=per_channel,
                    last_prefix=tllm_prex + 'post_layernorm.scale_to_int',
                    smoother_value=None,
                    smoother_shape=None,
                    rank=mapping.tp_rank,
                    cat_dim=-1))
        else:
            weights.update(
                get_tllm_linear_weight(split_v, tllm_prex + 'mlp.gate.', None,
                                       use_weight_only,
                                       plugin_weight_only_quant_type))

        mlp_fc_weight = get_weight(model_params, prefix + 'mlp.gate_proj',
                                   dtype)
        split_v = split_matrix_tp(mlp_fc_weight,
                                  tensor_parallel,
                                  mapping.tp_rank,
                                  dim=0)

        if use_smooth_quant:
            mlp_fc_weight = mlp_fc_weight.t()  #verified
            int8_weights = generate_int8(
                mlp_fc_weight, act_range.get(prefix + 'mlp.gate_proj'))
            weights.update(
                get_tllm_linear_sq_weight(
                    int8_weights,
                    tllm_prex + 'mlp.fc.',
                    [1, intermediate_size // tensor_parallel],
                    tensor_parallel,
                    is_qkv=False,
                    per_token=per_token,
                    per_channel=per_channel,
                    last_prefix=tllm_prex + 'post_layernorm.scale_to_int',
                    smoother_value=None,
                    smoother_shape=None,
                    rank=mapping.tp_rank,
                    cat_dim=-1))
        else:
            weights.update(
                get_tllm_linear_weight(split_v, tllm_prex + 'mlp.fc.', None,
                                       use_weight_only,
                                       plugin_weight_only_quant_type))

        mlp_proj_weight = get_weight(model_params, prefix + 'mlp.down_proj',
                                     dtype)
        split_v = split_matrix_tp(mlp_proj_weight,
                                  tensor_parallel,
                                  mapping.tp_rank,
                                  dim=1)

        if use_smooth_quant:
            mlp_proj_weight = mlp_proj_weight.t()
            int8_weights = generate_int8(
                mlp_proj_weight, act_range.get(prefix + 'mlp.down_proj'))
            weights.update(
                get_tllm_linear_sq_weight(
                    int8_weights,
                    tllm_prex + 'mlp.proj.', [1, hidden_size],
                    tensor_parallel,
                    is_qkv=False,
                    per_token=per_token,
                    per_channel=per_channel,
                    last_prefix=tllm_prex + 'mlp.quantization_scaling_factor',
                    smoother_value=smoother[prefix + 'mlp.down_proj'],
                    smoother_shape=[1, intermediate_size // tensor_parallel],
                    rank=mapping.tp_rank,
                    cat_dim=0))
        else:
            weights.update(
                get_tllm_linear_weight(split_v, tllm_prex + 'mlp.proj.', None,
                                       use_weight_only,
                                       plugin_weight_only_quant_type))
        # Layer norms do not use tensor parallelism
        input_ln_weight = get_weight(model_params, prefix + 'input_layernorm',
                                     dtype)
        weights[tllm_prex + 'input_layernorm.weight'] = input_ln_weight

        post_ln_weight = get_weight(model_params,
                                    prefix + 'post_attention_layernorm', dtype)
        weights[tllm_prex + 'post_layernorm.weight'] = post_ln_weight

    v = get_weight(model_params, 'model.embed_tokens', dtype)

    if hf_model.config.tie_word_embeddings:
        # lm_head.weight has the same weights as embedding
        if mapping.is_last_pp_rank():
            weights['lm_head.weight'] = split(v, mapping.tp_size,
                                              mapping.tp_rank)

    if use_parallel_embedding:
        v = split_matrix_tp(v,
                            mapping.tp_size,
                            mapping.tp_rank,
                            dim=sharding_dim)

    if mapping.is_first_pp_rank():
        weights['transformer.vocab_embedding.weight'] = v

    lm_head_weights = get_weight(model_params, 'lm_head', dtype)

    if mapping.is_last_pp_rank():
        weights['lm_head.weight'] = split_matrix_tp(lm_head_weights,
                                                    tensor_parallel,
                                                    mapping.tp_rank,
                                                    dim=0)

        ln_f_w = get_weight(model_params, 'model.norm', dtype)
    weights['transformer.ln_f.weight'] = ln_f_w

    tok = time.time()
    t = time.strftime('%H:%M:%S', time.gmtime(tok - tik))
    print(f'Weights loaded. Total time: {t}')
    return weights


if __name__ == '__main__':
    # TODO(qijun): Currently, the convert script depends on a torch op:
    # torch.ops.fastertransformer.symmetric_quantize_last_axis_of_batched_matrix,
    # which is included in tensorrt_llm Python package. Otherwise, the convert
    # script does not need to import tensorrt_llm. Will remove it after reimplementing
    # the op with PyTorch.
    print(tensorrt_llm.__version__)
    args = parse_arguments()
    world_size = args.tp_size * args.pp_size

    tik = time.time()

    if not os.path.exists(args.output_dir):
        os.makedirs(args.output_dir)
    hf_config = None
    if args.model_dir is not None:
        hf_config = LlamaConfig.from_pretrained(args.model_dir)

        args.model_type = hf_config.model_type
        args.n_head = hf_config.num_attention_heads
        args.inter_size = hf_config.intermediate_size
        args.n_layer = hf_config.num_hidden_layers
        args.n_embd = hf_config.hidden_size
        args.n_kv_head = hf_config.num_key_value_heads
        args.rms_norm_eps = hf_config.rms_norm_eps
        args.vocab_size = hf_config.vocab_size
        args.n_positions = hf_config.max_position_embeddings

    elif args.meta_ckpt_dir is not None:

        with open(Path(args.meta_ckpt_dir, "params.json")) as fp:
            meta_config: dict = json.load(fp)
        args.n_embd = meta_config["dim"]
        args.n_head = meta_config["n_heads"]
        args.n_layer = meta_config["n_layers"]
        args.n_kv_head = meta_config.get("n_kv_heads", args.n_head)

        if "hidden_dim" in meta_config:
            args.inter_size = meta_config["hidden_dim"]
        else:
            args.multiple_of = meta_config.get("multiple_of", 1)
            n_embd = int(4 * args.n_embd * 2 / 3)
            args.ffn_dim_multiplier = meta_config.get("ffn_dim_multiplier", 1)
            args.inter_size = args.multiple_of * (
                (int(n_embd * args.ffn_dim_multiplier) + args.multiple_of - 1)
                // args.multiple_of)
        args.rms_norm_eps = meta_config["norm_eps"]

    if args.rotary_scaling is not None:
        # assert args.use_gpt_attention_plugin, "RoPE scaling is only supported through GPT attention plugin."
        rotary_scaling = {
            "type": args.rotary_scaling[0],
            "factor": float(args.rotary_scaling[1])
        }
        assert rotary_scaling["type"] in ["linear", "dynamic"]
        assert rotary_scaling["factor"] > 1.0
        args.rotary_scaling = rotary_scaling

    config = {
        'architecture': 'MedusaForCausalLM',
        'dtype': args.dtype,
        'logits_dtype': 'float32',
        'num_hidden_layers': args.n_layer,
        'num_attention_heads': args.n_head,
        'hidden_size': args.n_embd,
        'intermediate_size': args.inter_size,
        'num_key_value_heads': args.n_kv_head,
        'vocab_size': args.vocab_size,
        'position_embedding_type': 'rope_gpt_neox',
        'max_position_embeddings': args.n_positions,
        'hidden_act': args.hidden_act,
        'rotary_base': args.rotary_base,
        'rotary_scaling': args.rotary_scaling,
        'norm_epsilon': args.rms_norm_eps,
        'quantization': {
            'quant_algo': None,
            'kv_cache_quant_algo': None,
        },
        'mapping': {
            'world_size': world_size,
            'tp_size': args.tp_size,
            'pp_size': args.pp_size,
        },
        'use_parallel_embedding': args.use_parallel_embedding,
        'embedding_sharding_dim': args.embedding_sharding_dim,
        'share_embedding_table': args.use_embedding_sharing,
        'max_draft_len': args.max_medusa_token_len,
        'num_medusa_heads': args.num_medusa_heads,
        'num_medusa_layers': args.num_medusa_layers
    }

    if args.use_weight_only:
        if args.weight_only_precision == 'int8':
            config['quantization']['quant_algo'] = QuantAlgo.W8A16
        elif args.weight_only_precision == 'int4':
            config['quantization']['quant_algo'] = QuantAlgo.W4A16
    elif args.smoothquant:
        if args.per_channel:
            if args.per_token:
                config['quantization'][
                    'quant_algo'] = QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TOKEN_PLUGIN
            else:
                config['quantization'][
                    'quant_algo'] = QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TENSOR_PLUGIN
        else:
            if args.per_token:
                config['quantization'][
                    'quant_algo'] = QuantAlgo.W8A8_SQ_PER_TENSOR_PER_TOKEN_PLUGIN
            else:
                config['quantization'][
                    'quant_algo'] = QuantAlgo.W8A8_SQ_PER_TENSOR_PLUGIN

    if args.int8_kv_cache:
        config['quantization']['kv_cache_quant_algo'] = QuantAlgo.INT8

    if args.weight_only_precision == 'int4_gptq':
        config['quantization'].update({
            "group_size": args.group_size,
            "has_zero_point": True,
            "pre_quant_scale": False,
            'quant_algo': QuantAlgo.W4A16_GPTQ
        })

    with open(os.path.join(args.output_dir, 'config.json'), 'w') as f:
        json.dump(config, f, indent=4)

    if args.weight_only_precision == 'int8':
        plugin_weight_only_quant_type = torch.int8
    elif args.weight_only_precision == 'int4':
        plugin_weight_only_quant_type = torch.quint4x2

    act_range = {}
    llama_qkv_para = {}
    # smoother for inputs of self_attn.o_proj and mlp.down_proj
    llama_smoother = {}
    model = None
    if args.model_dir is not None:
        hf_model = LlamaForCausalLM if args.model_type != "mixtral" else MixtralForCausalLM

        model = hf_model.from_pretrained(args.model_dir,
                                         torch_dtype='auto',
                                         device_map="auto",
                                         trust_remote_code=True)

        if args.smoothquant is not None or args.int8_kv_cache:
            os.environ["TOKENIZERS_PARALLELISM"] = os.environ.get(
                "TOKENIZERS_PARALLELISM", "false")
            if args.load_model_on_cpu:
                logger.warning(
                    "Note that running capture_activation_range on cpu would be very small."
                )
            tokenizer = LlamaTokenizer.from_pretrained(args.model_dir,
                                                       padding_side='left')
            dataset = load_calib_dataset(args.calib_dataset,
                                         cache_dir=args.dataset_cache_dir)

            act_range = capture_activation_range(model, tokenizer, dataset)
            if args.smoothquant is not None:
                smooth_llama_model(model, act_range, args.smoothquant,
                                   llama_qkv_para, llama_smoother)
    convert_args = {
        'hf_model': model,
        'act_range': act_range,
        'llama_qkv_para': llama_qkv_para,
        'llama_smoother': llama_smoother
    }

    def covert_and_save(rank, convert_args):
        mapping = Mapping(world_size=world_size,
                          rank=rank,
                          tp_size=args.tp_size,
                          pp_size=args.pp_size)

        if args.use_weight_only and args.weight_only_precision == 'int4_gptq':
            assert False, "Never supported"
        else:
            if args.load_by_shard:
                weights = load_weights_from_hf_by_shard(
                    args.model_dir, PretrainedConfig.from_dict(config))

            else:
                weights = convert_hf_llama(
                    convert_args['hf_model'],
                    mapping,
                    rank,
                    dtype=args.dtype,
                    use_weight_only=args.use_weight_only,
                    plugin_weight_only_quant_type=plugin_weight_only_quant_type,
                    use_parallel_embedding=args.use_parallel_embedding,
                    sharding_dim=args.embedding_sharding_dim,
                    share_embedding_table=args.use_embedding_sharing,
                    use_smooth_quant=args.smoothquant,
                    per_channel=args.per_channel,
                    per_token=args.per_token,
                    int8_kv_cache=args.int8_kv_cache,
                    act_range=convert_args['act_range'],
                    qkv_para=convert_args['llama_qkv_para'],
                    smoother=convert_args['llama_smoother'])

                def load_medusa_hf(medusa_path: str,
                                   mapping=Mapping(),
                                   dtype='float32'):
                    logger.info("Loading Medusa heads' weights ...")
                    is_ckpt_safetensors = False

                    ckpt_file = Path(medusa_path) / "medusa_lm_head.pt"
                    if not ckpt_file.exists():
                        ckpt_file = Path(
                            medusa_path) / "medusa_lm_head.safetensors"
                        is_ckpt_safetensors = True

                    if is_ckpt_safetensors:
                        logger.info("Safetensors Found ...")
                        from safetensors.torch import load_file
                        state_dict = load_file(ckpt_file)
                    else:
                        state_dict = torch.load(ckpt_file, map_location="cpu")

                    torch_dtype = str_dtype_to_torch(dtype)
                    weights = {}

                    for h in range(args.num_medusa_heads):
                        for l in range(args.num_medusa_layers):
                            w = state_dict[f"{h}.{l}.linear.weight"].clone().to(
                                torch_dtype)

                            split_v = split(w, mapping.tp_size, mapping.tp_rank)
                            weights.update(
                                get_tllm_linear_weight(
                                    split_v,
                                    f'medusa_heads.{h}.medusa_layers.{l}.linear.',
                                    None, args.use_weight_only,
                                    plugin_weight_only_quant_type))

                            b = state_dict[f"{h}.{l}.linear.bias"].clone().to(
                                torch_dtype)

                            weights[
                                'medusa_heads.{}.medusa_layers.{}.linear.bias'.
                                format(h, l)] = split(b, mapping.tp_size,
                                                      mapping.tp_rank)

                        lm = state_dict[
                            f"{h}.{args.num_medusa_layers}.weight"].clone().to(
                                torch_dtype)  # LM Head

                        weights['medusa_heads.{}.lm_head.weight'.format(
                            h)] = split(lm, mapping.tp_size, mapping.tp_rank)

                    return weights

                if args.medusa_model_dir is not None:
                    config_file = Path(args.medusa_model_dir) / "config.json"
                    with open(config_file) as fp:
                        config = json.load(fp)
                    args.num_medusa_heads = config.get('medusa_num_heads',
                                                       args.num_medusa_heads)
                    args.num_medusa_layers = config.get('medusa_num_layers',
                                                        args.num_medusa_layers)
                    if args.fixed_num_medusa_heads is not None and args.fixed_num_medusa_heads != args.num_medusa_heads:
                        logger.info(
                            f"fixing num_medusa_heads from {args.num_medusa_heads} to {args.fixed_num_medusa_heads}"
                        )
                        args.num_medusa_heads = args.fixed_num_medusa_heads

                    assert args.max_medusa_token_len > 0, "should have max_medusa_token_len > 0"

                    medusa_weights = load_medusa_hf(args.medusa_model_dir,
                                                    mapping,
                                                    dtype=args.dtype)
                    weights.update(medusa_weights)

        safetensors.torch.save_file(
            weights, os.path.join(args.output_dir, f'rank{rank}.safetensors'))

    if args.workers == 1:
        for rank in range(world_size):
            covert_and_save(rank, convert_args)
    else:
        with ThreadPoolExecutor(max_workers=args.workers) as p:
            futures = [
                p.submit(covert_and_save, rank, convert_args)
                for rank in range(world_size)
            ]
            exceptions = []
            for future in as_completed(futures):
                try:
                    future.result()
                except Exception as e:
                    traceback.print_exc()
                    exceptions.append(e)
            assert len(
                exceptions
            ) == 0, "Checkpoint conversion failed, please check error log."

    tok = time.time()
    t = time.strftime('%H:%M:%S', time.gmtime(tok - tik))
    print(f'Total time of converting checkpoints: {t}')