convert_checkpoint.py

# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# 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.

import argparse
import copy
import json
import os
import traceback
from argparse import Namespace
from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path
from typing import Dict, Optional

import safetensors
import torch
from transformers.models.auto import AutoModel
from transformers.models.auto.configuration_auto import AutoConfig
from transformers.models.llama.configuration_llama import LlamaConfig
from transformers.models.llama.modeling_llama import LlamaForCausalLM

import tensorrt_llm.models.modeling_utils
from examples.medusa.convert_checkpoint import convert_hf_llama
from tensorrt_llm._utils import str_dtype_to_torch
from tensorrt_llm.mapping import Mapping
from tensorrt_llm.models.llama.convert import (dup_kv_weight,
                                               get_tllm_linear_weight,
                                               get_weight, get_weight_and_bias,
                                               split)

BASE_MODEL_TLLM_WEIGHT_PREFIX = "base_model."
DRAFTER_TLLM_WEIGHT_PREFIX = "drafter."


def parse_arguments():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model_dir", type=str, default=None, required=True)
    parser.add_argument("--drafter_model_dir",
                        type=str,
                        default=None,
                        required=True)

    parser.add_argument("--tp_size",
                        type=int,
                        default=1,
                        help="N-way tensor parallelism size")
    parser.add_argument("--dtype",
                        type=str,
                        default="float16",
                        choices=["float32", "bfloat16", "float16"])

    parser.add_argument("--storage-type",
                        "-t",
                        type=str,
                        default="fp32",
                        choices=["fp32", "fp16"])
    parser.add_argument("--load_model_on_cpu", action="store_true")
    parser.add_argument(
        "--use_parallel_embedding",
        action="store_true",
        default=False,
        help="By default embedding parallelism is disabled.",
    )
    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 (=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(
        "--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(
        "--dense_context_fmha",
        default=False,
        action="store_true",
        help=
        "Enable dense fmha in context phase, otherwise sliding window attention."
        "If dense_context_fmha=False, the sliding window size is the max attention window size.",
    )

    parser.add_argument(
        "--redrafter_draft_len_per_beam",
        type=int,
        default=5,
        help=
        "Number of times that the Recurrent Drafter runs the beam search to generate draft"
        "candidates. Note that this draft_len does not include the first true/guaranteed token.",
    )
    parser.add_argument(
        "--redrafter_num_beams",
        type=int,
        default=5,
        help="Number of beam search candidates to keep during the Recurrent"
        "Drafter beam search iterations.",
    )
    parser.add_argument(
        "--redrafter_no_greedy_search",
        action="store_false",
        default=True,
        dest="redrafter_greedy_search",
        help=
        "Whether Redrafter will use the token with the highest probability from lm_head"
        "output or randomly sampled from the probability distribution.",
    )

    return parser.parse_args()


def hf_llama_model(
    hf_model: LlamaForCausalLM,
    mapping: Mapping,
    dtype: torch.dtype = torch.float32,
    use_parallel_embedding: bool = False,
    sharding_dim: int = 0,
    additional_tllm_prefix: str = "",
) -> Dict[str, torch.Tensor]:
    weights = {}
    model_params = dict(hf_model.named_parameters())
    num_attention_heads = hf_model.config.num_attention_heads
    hidden_size = hf_model.config.hidden_size
    head_size = hidden_size // num_attention_heads
    num_key_value_heads = hf_model.config.num_key_value_heads

    for layer_idx in range(hf_model.config.num_hidden_layers):
        hf_prefix = f"model.layers.{layer_idx}."
        tllm_prefix = f"{additional_tllm_prefix}transformer.layers.{layer_idx}."

        # load qkv
        q_weight = get_weight(model_params, hf_prefix + "self_attn.q_proj",
                              dtype)
        k_weight = get_weight(model_params, hf_prefix + "self_attn.k_proj",
                              dtype)
        v_weight = get_weight(model_params, hf_prefix + "self_attn.v_proj",
                              dtype)

        if num_key_value_heads < mapping.tp_size:
            # duplicate the KV heads up to mapping.tp_size
            k_weight = dup_kv_weight(k_weight, num_key_value_heads,
                                     mapping.tp_size)
            v_weight = dup_kv_weight(v_weight, num_key_value_heads,
                                     mapping.tp_size)
        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_qkv = torch.concat((wq, wk, wv))

        weights.update(
            get_tllm_linear_weight(split_qkv, tllm_prefix + "attention.qkv."))

        # load dense
        attn_dense_weight = get_weight(model_params,
                                       hf_prefix + "self_attn.o_proj", dtype)
        split_attn_dense_weight = split(attn_dense_weight,
                                        mapping.tp_size,
                                        mapping.tp_rank,
                                        dim=1)
        weights.update(
            get_tllm_linear_weight(split_attn_dense_weight,
                                   tllm_prefix + "attention.dense."))

        # load mlp (merge gate + fc = gate_fc, and then proj)
        mlp_up_weight = get_weight(model_params, hf_prefix + "mlp.up_proj",
                                   dtype)
        mlp_up_weight = split(mlp_up_weight,
                              mapping.tp_size,
                              mapping.tp_rank,
                              dim=0)

        mlp_gate_weight = get_weight(model_params, hf_prefix + "mlp.gate_proj",
                                     dtype)
        mlp_gate_weight = split(mlp_gate_weight,
                                mapping.tp_size,
                                mapping.tp_rank,
                                dim=0)

        mlp_gate_fc_weight = torch.concat((mlp_up_weight, mlp_gate_weight))

        weights.update(
            get_tllm_linear_weight(mlp_gate_fc_weight,
                                   tllm_prefix + "mlp.gate_fc."))

        mlp_proj_weight = get_weight(model_params, hf_prefix + "mlp.down_proj",
                                     dtype)
        split_mlp_proj_weight = split(mlp_proj_weight,
                                      mapping.tp_size,
                                      mapping.tp_rank,
                                      dim=1)
        weights.update(
            get_tllm_linear_weight(split_mlp_proj_weight,
                                   tllm_prefix + "mlp.proj."))

        # Layer norms do not use tensor parallelism
        input_ln_weight = get_weight(model_params,
                                     hf_prefix + "input_layernorm", dtype)
        weights[tllm_prefix + "input_layernorm.weight"] = input_ln_weight

        post_ln_weight = get_weight(model_params,
                                    hf_prefix + "post_attention_layernorm",
                                    dtype)
        weights[tllm_prefix + "post_layernorm.weight"] = post_ln_weight

    embed_tokens = get_weight(model_params, "model.embed_tokens", dtype)

    weights[f"{additional_tllm_prefix}lm_head.weight"] = split(
        embed_tokens.clone(), mapping.tp_size, mapping.tp_rank, dim=0)

    if use_parallel_embedding:
        embed_tokens = split(embed_tokens,
                             mapping.tp_size,
                             mapping.tp_rank,
                             dim=sharding_dim)

    weights[
        f"{additional_tllm_prefix}transformer.vocab_embedding.weight"] = embed_tokens

    ln_f_w = get_weight(model_params, "model.norm", dtype)
    weights[f"{additional_tllm_prefix}transformer.ln_f.weight"] = ln_f_w

    return weights


def hf_drafter(
    hf_model: Namespace,  #DrafterModel, # TODO:
    mapping: Mapping,
    dtype: torch.dtype = torch.float32,
    additional_tllm_prefix: str = "",
) -> Dict[str, torch.Tensor]:
    """
    Possible tensor names for Drafter checkpoints:
        input_proj.weight
        input_proj.bias
        lm_head.0.linear.weight
        lm_head.0.linear.bias
        lm_head.1.linear.weight
        lm_head.1.linear.bias
        lm_head.2.weight
        rnn_u.weight
        rnn_u.bias
        rnn_w.weight

        OR

        input_projs.weight
        input_projs.bias
        lm_heads.0.linear.weight
        lm_heads.0.linear.bias
        lm_heads.1.linear.weight
        lm_heads.1.linear.bias
        lm_heads.2.weight

        OR

        0.0.linear.weight
        0.0.linear.bias
        0.1.linear.weight
        0.1.linear.bias
        0.2.weight

    """

    def get_weight_and_bias_with_multiple_possible_names(
            model_params, dtype, names_to_try, bias=True):
        w, b = None, None
        for name in names_to_try:
            try:
                if bias:
                    w, b = get_weight_and_bias(model_params, name, dtype)
                else:
                    w = get_weight(model_params, name, dtype)
                break
            except:
                pass
        if not bias:
            return w
        return w, b

    weights = {}
    # TODO: When ReDrafter is added to Transformers
    # model_params = dict(hf_model.named_parameters())
    model_params = dict(hf_model.named_parameters)

    if hf_model.config.hidden_size * 2 != hf_model.config.exit_dim:
        input_proj_weight, input_proj_bias = get_weight_and_bias_with_multiple_possible_names(
            model_params, dtype, ["input_proj", "input_projs"])
        weights[f"{additional_tllm_prefix}input_proj.weight"] = split(
            input_proj_weight, mapping.tp_size, mapping.tp_rank, dim=0)
        weights[f"{additional_tllm_prefix}input_proj.bias"] = split(
            input_proj_bias, mapping.tp_size, mapping.tp_rank, dim=0)

    for layer_idx in range(hf_model.config.num_draft_layers):
        layer_weight, layer_bias = get_weight_and_bias_with_multiple_possible_names(
            model_params, dtype, [
                f"lm_head.{layer_idx}.linear", f"lm_heads.{layer_idx}.linear",
                f"0.{layer_idx}.linear"
            ])
        weights[
            f"{additional_tllm_prefix}layers.{layer_idx}.linear.weight"] = split(
                layer_weight, mapping.tp_size, mapping.tp_rank, dim=0)
        weights[
            f"{additional_tllm_prefix}layers.{layer_idx}.linear.bias"] = split(
                layer_bias, mapping.tp_size, mapping.tp_rank, dim=0)

    last_layer_weight = get_weight_and_bias_with_multiple_possible_names(
        model_params,
        dtype, [
            f"lm_head.{hf_model.config.num_draft_layers}",
            f"lm_heads.{hf_model.config.num_draft_layers}",
            f"0.{hf_model.config.num_draft_layers}"
        ],
        bias=False)
    weights[f"{additional_tllm_prefix}lm_head.weight"] = split(
        last_layer_weight, mapping.tp_size, mapping.tp_rank, dim=0)

    if hf_model.config.rnn:
        # rnn_u has both weight and bias
        rnn_u_weight, rnn_u_bias = get_weight_and_bias(model_params, "rnn_u",
                                                       dtype)
        weights[f"{additional_tllm_prefix}rnn_u.weight"] = split(
            rnn_u_weight, mapping.tp_size, mapping.tp_rank, dim=0)
        weights[f"{additional_tllm_prefix}rnn_u.bias"] = split(rnn_u_bias,
                                                               mapping.tp_size,
                                                               mapping.tp_rank,
                                                               dim=0)

        # rnn_w only has weight
        rnn_w_weight = get_weight(model_params, "rnn_w", dtype)
        weights[f"{additional_tllm_prefix}rnn_w.weight"] = split(
            rnn_w_weight, mapping.tp_size, mapping.tp_rank, dim=0)

    return weights


def hf_llama_config(
    hf_config: LlamaConfig,
    dtype: str = "float32",
    logits_dtype: str = "float32",
    mapping: Mapping = Mapping(1),
) -> tensorrt_llm.models.modeling_utils.PretrainedConfig:
    return tensorrt_llm.models.modeling_utils.PretrainedConfig(
        architecture="LlamaForCausalLM",
        dtype=dtype,
        logits_dtype=logits_dtype,
        vocab_size=hf_config.vocab_size,
        max_position_embeddings=hf_config.max_position_embeddings,
        hidden_size=hf_config.hidden_size,
        num_hidden_layers=hf_config.num_hidden_layers,
        num_attention_heads=hf_config.num_attention_heads,
        num_key_value_heads=hf_config.num_key_value_heads,
        hidden_act=hf_config.hidden_act,
        intermediate_size=hf_config.intermediate_size,
        norm_epsilon=hf_config.rms_norm_eps,
        position_embedding_type="rope_gpt_neox",
        mapping=mapping,
        quantization=tensorrt_llm.models.modeling_utils.QuantConfig(),
        rotary_base=getattr(hf_config, "rope_theta", 500000.0),
        rotary_scaling=getattr(hf_config, "rotary_scaling", None),
    )


def hf_redrafter_config(
    tllm_base_model_config: tensorrt_llm.models.modeling_utils.PretrainedConfig,
    redrafter_config: Namespace,  # DrafterConfig
    redrafter_num_beams: int,
    redrafter_draft_len_per_beam: int,
    redrafter_greedy_search: bool,
) -> tensorrt_llm.models.modeling_utils.PretrainedConfig:
    tllm_config = copy.deepcopy(tllm_base_model_config)

    tllm_config.base_model_architecture = tllm_config.architecture
    tllm_config.architecture = "ReDrafterForCausalLM"
    setattr(tllm_config, "redrafter_num_layers",
            redrafter_config.num_draft_layers)
    setattr(tllm_config, "redrafter_hidden_size", redrafter_config.hidden_size)
    setattr(tllm_config, "redrafter_exit_dim", redrafter_config.exit_dim)
    setattr(tllm_config, "redrafter_is_rnn", redrafter_config.rnn)

    # These three configs look like runtime parameters. But for TensorRT-LLM
    # implementation, they are required to be provided at engine build time and
    # TensorRT needs to unroll loops with set number of loop iterations.
    setattr(tllm_config, "redrafter_num_beams", redrafter_num_beams)
    setattr(tllm_config, "redrafter_draft_len_per_beam",
            redrafter_draft_len_per_beam)
    setattr(tllm_config, "redrafter_greedy_search", redrafter_greedy_search)

    return tllm_config


def convert_and_save(
    rank: int,
    tp_size: int,
    hf_base_model: LlamaForCausalLM,
    hf_drafter_model: Optional[AutoModel],
    dtype: str,
    use_parallel_embedding: bool,
    embedding_sharding_dim: int,
    output_dir: str,
) -> None:
    mapping = Mapping(
        world_size=tp_size,
        rank=rank,
        tp_size=tp_size,
    )
    weights = convert_hf_llama(
        hf_base_model,
        mapping,
        rank,
        dtype=dtype,
        use_parallel_embedding=use_parallel_embedding,
        sharding_dim=embedding_sharding_dim,
        # use_weight_only=args.use_weight_only,
        # plugin_weight_only_quant_type=plugin_weight_only_quant_type,
        # 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']
    )

    if hf_drafter_model is not None:
        drafter_weights = hf_drafter(
            hf_drafter_model,
            mapping,
            dtype=str_dtype_to_torch(dtype),
            additional_tllm_prefix=(DRAFTER_TLLM_WEIGHT_PREFIX
                                    if hf_drafter_model is not None else ""),
        )
        weights.update(drafter_weights)

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


def multi_worker_convert_and_save(
    workers: int,
    tp_size: int,
    hf_base_model: LlamaForCausalLM,
    hf_drafter_model: Optional[AutoModel],
    dtype: str,
    use_parallel_embedding: bool,
    embedding_sharding_dim: int,
    output_dir: str,
) -> None:
    with ThreadPoolExecutor(max_workers=workers) as p:
        futures = [
            p.submit(
                convert_and_save,
                rank,
                tp_size,
                hf_base_model,
                hf_drafter_model,
                dtype,
                use_parallel_embedding,
                embedding_sharding_dim,
                output_dir,
            ) for rank in range(tp_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."


def create_and_save_config(args):
    mapping = Mapping(
        world_size=args.tp_size,
        tp_size=args.tp_size,
        pp_size=1,
    )
    base_model_hf_config = AutoConfig.from_pretrained(args.model_dir)
    tllm_model_config = hf_llama_config(
        base_model_hf_config,
        dtype=args.dtype,
        mapping=mapping,
    )

    if args.drafter_model_dir:
        # TODO: When ReDrafter is added to Transformers
        # drafter_hf_config = AutoConfig.from_pretrained(args.drafter_model_dir)
        with open(Path(args.drafter_model_dir, "config.json")) as fp:
            drafter_hf_config = Namespace(**json.load(fp))
        tllm_model_config = hf_redrafter_config(
            tllm_model_config,
            drafter_hf_config,
            args.redrafter_draft_len_per_beam,
            args.redrafter_num_beams,
            args.redrafter_greedy_search,
        )

    with open(os.path.join(args.output_dir, "config.json"), "w") as f:
        json.dump(tllm_model_config.to_dict(), f, indent=4)
    return drafter_hf_config


def main():
    args = parse_arguments()
    if not os.path.exists(args.output_dir):
        os.makedirs(args.output_dir)
    drafter_hf_config = create_and_save_config(args)

    hf_base_model = LlamaForCausalLM.from_pretrained(
        args.model_dir,
        torch_dtype="auto",
    )

    hf_drafter_model: Optional[AutoModel] = None
    if args.drafter_model_dir:
        # TODO: When ReDrafter is added to Transformers
        # hf_drafter_model = AutoModel.from_pretrained(
        #     args.drafter_model_dir,
        #     torch_dtype="auto",
        # )
        ckpt_file = Path(args.drafter_model_dir, "model.safetensors")
        if not Path.exists(ckpt_file):
            ckpt_file = Path(args.drafter_model_dir, "model.pt")
        print(f"Loading drafter from {ckpt_file}")
        if str(ckpt_file).endswith(".safetensors"):
            drafter_ckpt = {}
            with safetensors.safe_open(ckpt_file, framework="pt",
                                       device="cpu") as f:
                key: str = None
                for key in f.keys():
                    drafter_ckpt[key] = f.get_tensor(key)
        else:
            drafter_ckpt = torch.load(ckpt_file, map_location='cpu')
        hf_drafter_model = Namespace(**{
            "named_parameters": drafter_ckpt,
            "config": drafter_hf_config
        })

    multi_worker_convert_and_save(
        args.workers,
        args.tp_size,
        hf_base_model,
        hf_drafter_model,
        args.dtype,
        args.use_parallel_embedding,
        args.embedding_sharding_dim,
        args.output_dir,
    )


if __name__ == "__main__":
    main()