build.py

# SPDX-FileCopyrightText: Copyright (c) 2022-2023 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 os
import time
from pathlib import Path

import tensorrt as trt
import torch
import torch.multiprocessing as mp
from transformers import BloomConfig, BloomForCausalLM

import tensorrt_llm
from tensorrt_llm._utils import str_dtype_to_trt
from tensorrt_llm.builder import Builder
from tensorrt_llm.logger import logger
from tensorrt_llm.mapping import Mapping
from tensorrt_llm.models import smooth_quantize, weight_only_quantize
from tensorrt_llm.network import net_guard
from tensorrt_llm.plugin.plugin import ContextFMHAType
from tensorrt_llm.quantization import QuantMode

from weight import load_from_hf_bloom, load_from_bin, parse_config, check_embedding_share  # isort:skip

MODEL_NAME = "bloom"

import onnx
import tensorrt as trt
from onnx import TensorProto, helper


def trt_dtype_to_onnx(dtype):
    if dtype == trt.float16:
        return TensorProto.DataType.FLOAT16
    elif dtype == trt.float32:
        return TensorProto.DataType.FLOAT
    elif dtype == trt.int32:
        return TensorProto.DataType.INT32
    else:
        raise TypeError("%s is not supported" % dtype)


def to_onnx(network, path):
    inputs = []
    for i in range(network.num_inputs):
        network_input = network.get_input(i)
        inputs.append(
            helper.make_tensor_value_info(
                network_input.name, trt_dtype_to_onnx(network_input.dtype),
                list(network_input.shape)))

    outputs = []
    for i in range(network.num_outputs):
        network_output = network.get_output(i)
        outputs.append(
            helper.make_tensor_value_info(
                network_output.name, trt_dtype_to_onnx(network_output.dtype),
                list(network_output.shape)))

    nodes = []
    for i in range(network.num_layers):
        layer = network.get_layer(i)
        layer_inputs = []
        for j in range(layer.num_inputs):
            ipt = layer.get_input(j)
            if ipt is not None:
                layer_inputs.append(layer.get_input(j).name)
        layer_outputs = [
            layer.get_output(j).name for j in range(layer.num_outputs)
        ]
        nodes.append(
            helper.make_node(str(layer.type),
                             name=layer.name,
                             inputs=layer_inputs,
                             outputs=layer_outputs,
                             domain="com.nvidia"))

    onnx_model = helper.make_model(helper.make_graph(nodes,
                                                     'attention',
                                                     inputs,
                                                     outputs,
                                                     initializer=None),
                                   producer_name='NVIDIA')
    onnx.save(onnx_model, path)


def get_engine_name(model, dtype, tp_size, rank):
    return '{}_{}_tp{}_rank{}.engine'.format(model, dtype, tp_size, rank)


def serialize_engine(engine, path):
    logger.info(f'Serializing engine to {path}...')
    tik = time.time()
    with open(path, 'wb') as f:
        f.write(bytearray(engine))
    tok = time.time()
    t = time.strftime('%H:%M:%S', time.gmtime(tok - tik))
    logger.info(f'Engine serialized. Total time: {t}')


def parse_arguments():
    parser = argparse.ArgumentParser()
    parser.add_argument('--world_size',
                        type=int,
                        default=1,
                        help='world size, only support tensor parallelism now')
    parser.add_argument('--model_dir', type=str, default=None)
    parser.add_argument('--bin_model_dir', type=str, default=None)
    parser.add_argument('--dtype',
                        type=str,
                        default='float16',
                        choices=['float32', 'float16'])
    parser.add_argument(
        '--timing_cache',
        type=str,
        default='model.cache',
        help=
        'The path of to read timing cache from, will be ignored if the file does not exist'
    )
    parser.add_argument('--log_level', type=str, default='info')
    parser.add_argument('--vocab_size', type=int, default=250680)
    parser.add_argument('--n_layer', type=int, default=32)
    parser.add_argument('--n_positions', type=int, default=2048)
    parser.add_argument('--n_embd', type=int, default=4096)
    parser.add_argument('--n_head', type=int, default=32)
    parser.add_argument('--mlp_hidden_size', type=int, default=None)
    parser.add_argument('--max_batch_size', type=int, default=8)
    parser.add_argument('--max_input_len', type=int, default=1024)
    parser.add_argument('--max_output_len', type=int, default=1024)
    parser.add_argument('--max_beam_width', type=int, default=1)
    parser.add_argument('--use_gpt_attention_plugin',
                        nargs='?',
                        const='float16',
                        type=str,
                        default=False,
                        choices=['float16', 'float32'])
    parser.add_argument('--use_gemm_plugin',
                        nargs='?',
                        const='float16',
                        type=str,
                        default=False,
                        choices=['float16', 'float32'])
    parser.add_argument('--enable_context_fmha',
                        default=False,
                        action='store_true')
    parser.add_argument('--enable_context_fmha_fp32_acc',
                        default=False,
                        action='store_true')
    parser.add_argument(
        '--use_layernorm_plugin',
        nargs='?',
        const='float16',
        type=str,
        default=False,
        choices=['float16', 'float32'],
        help=
        "Activates layernorm plugin. You can specify the plugin dtype or leave blank to use the model dtype."
    )
    parser.add_argument('--parallel_build', default=False, action='store_true')
    parser.add_argument('--visualize', default=False, action='store_true')
    parser.add_argument('--enable_debug_output',
                        default=False,
                        action='store_true')
    parser.add_argument('--gpus_per_node', type=int, default=8)
    parser.add_argument(
        '--output_dir',
        type=str,
        default='bloom_outputs',
        help=
        'The path to save the serialized engine files, timing cache file and model configs'
    )
    # Arguments related to the quantization of the model.
    parser.add_argument(
        '--use_smooth_quant',
        default=False,
        action="store_true",
        help=
        'Use the SmoothQuant method to quantize activations and weights for the various GEMMs.'
        'See --per_channel and --per_token for finer-grained quantization options.'
    )
    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'],
        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(
        '--per_channel',
        default=False,
        action="store_true",
        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',
        default=False,
        action="store_true",
        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(
        '--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(
        '--use_lookup_plugin',
        nargs='?',
        const=None,
        default=False,
        choices=['float16', 'float32', 'bfloat16'],
        help="Activates the lookup plugin which enables embedding sharing.")

    args = parser.parse_args()
    logger.set_level(args.log_level)

    if args.model_dir is not None:
        hf_config = BloomConfig.from_pretrained(args.model_dir)
        args.n_embd = hf_config.hidden_size
        args.n_head = hf_config.num_attention_heads
        args.n_layer = hf_config.num_hidden_layers
        args.vocab_size = hf_config.vocab_size
    elif args.bin_model_dir is not None:
        logger.info(f"Setting model configuration from {args.bin_model_dir}.")
        n_embd, n_head, n_layer, vocab_size, _, rotary_pct, bias, inter_size, multi_query_mode, dtype, prompt_num_tasks, prompt_max_vocab_size = parse_config(
            Path(args.bin_model_dir) / "config.ini")
        args.n_embd = n_embd
        args.n_head = n_head
        args.n_layer = n_layer
        args.vocab_size = vocab_size

    assert not (
        args.use_smooth_quant and args.use_weight_only
    ), "You cannot enable both SmoothQuant and INT8 weight-only together."

    if args.use_smooth_quant:
        args.quant_mode = QuantMode.use_smooth_quant(args.per_token,
                                                     args.per_channel)
    elif args.use_weight_only:
        args.quant_mode = QuantMode.use_weight_only(
            args.weight_only_precision == 'int4')
    else:
        args.quant_mode = QuantMode(0)

    if args.int8_kv_cache:
        args.quant_mode = args.quant_mode.set_int8_kv_cache()

    return args


def build_rank_engine(builder: Builder,
                      builder_config: tensorrt_llm.builder.BuilderConfig,
                      engine_name, rank, args):
    '''
       @brief: Build the engine on the given rank.
       @param rank: The rank to build the engine.
       @param args: The cmd line arguments.
       @return: The built engine.
    '''
    kv_dtype = str_dtype_to_trt(args.dtype)

    # Share_embedding_table can be set True only when:
    # 1) the weight for lm_head() does not exist while other weights exist
    # 2) For multiple-processes, use_parallel_embedding=True and embedding_sharding_dim == 0.
    # Besides, for TensorRT 9.0, we can observe the engine size reduction when the lookup and gemm plugin are enabled.
    share_embedding_table = False
    if args.use_embedding_sharing:
        if args.world_size > 1:
            if args.model_dir is not None and args.embedding_sharding_dim == 0 and args.use_parallel_embedding:
                share_embedding_table = check_embedding_share(args.model_dir)
        else:
            if args.model_dir is not None:
                share_embedding_table = check_embedding_share(args.model_dir)

        if not share_embedding_table:
            logger.warning(f'Cannot share the embedding lookup table.')

    if share_embedding_table:
        logger.info(
            'Engine will share embedding and language modeling weights.')

    # Initialize Module
    tensorrt_llm_bloom = tensorrt_llm.models.BloomForCausalLM(
        num_layers=args.n_layer,
        num_heads=args.n_head,
        hidden_size=args.n_embd,
        vocab_size=args.vocab_size,
        max_position_embeddings=args.n_positions,
        dtype=kv_dtype,
        mapping=Mapping(world_size=args.world_size,
                        rank=rank,
                        tp_size=args.world_size),  # TP only
        use_parallel_embedding=args.use_parallel_embedding,
        embedding_sharding_dim=args.embedding_sharding_dim,
        share_embedding_table=share_embedding_table,
        quant_mode=args.quant_mode)
    if args.use_smooth_quant:
        tensorrt_llm_bloom = smooth_quantize(tensorrt_llm_bloom,
                                             args.quant_mode)
    elif args.use_weight_only:
        tensorrt_llm_bloom = weight_only_quantize(tensorrt_llm_bloom,
                                                  args.quant_mode)

    if args.model_dir is not None:
        logger.info(f'Loading HF BLOOM ... from {args.model_dir}')
        tik = time.time()
        hf_bloom = BloomForCausalLM.from_pretrained(args.model_dir,
                                                    torch_dtype="auto")
        tok = time.time()
        t = time.strftime('%H:%M:%S', time.gmtime(tok - tik))
        logger.info(f'HF BLOOM loaded. Total time: {t}')
        print(hf_bloom)
        load_from_hf_bloom(tensorrt_llm_bloom,
                           hf_bloom,
                           rank,
                           args.world_size,
                           fp16=(args.dtype == 'float16'),
                           use_parallel_embedding=args.use_parallel_embedding,
                           sharding_dim=args.embedding_sharding_dim,
                           share_embedding_table=share_embedding_table)
    elif args.bin_model_dir is not None:
        load_from_bin(tensorrt_llm_bloom,
                      args.bin_model_dir,
                      rank,
                      args.world_size,
                      args.dtype,
                      use_parallel_embedding=args.use_parallel_embedding,
                      sharding_dim=args.embedding_sharding_dim,
                      share_embedding_table=share_embedding_table)

    # Module -> Network
    network = builder.create_network()
    network.trt_network.name = engine_name
    if args.use_gpt_attention_plugin:
        network.plugin_config.set_gpt_attention_plugin(
            dtype=args.use_gpt_attention_plugin)
    if args.use_gemm_plugin:
        network.plugin_config.set_gemm_plugin(dtype=args.use_gemm_plugin)
    if args.use_layernorm_plugin:
        network.plugin_config.set_layernorm_plugin(
            dtype=args.use_layernorm_plugin)
    if args.use_lookup_plugin:
        # Use the plugin for the embedding parallelism
        network.plugin_config.set_lookup_plugin(dtype=args.dtype)
    assert not (args.enable_context_fmha and args.enable_context_fmha_fp32_acc)
    if args.enable_context_fmha:
        network.plugin_config.set_context_fmha(ContextFMHAType.enabled)
    if args.enable_context_fmha_fp32_acc:
        network.plugin_config.set_context_fmha(
            ContextFMHAType.enabled_with_fp32_acc)
    # Quantization plugins.
    if args.use_smooth_quant:
        network.plugin_config.set_smooth_quant_gemm_plugin(dtype=args.dtype)
        network.plugin_config.set_layernorm_quantization_plugin(
            dtype=args.dtype)

        network.plugin_config.set_quantize_tensor_plugin()
        network.plugin_config.set_quantize_per_token_plugin()
    elif args.use_weight_only:
        network.plugin_config.set_weight_only_quant_matmul_plugin(
            dtype=args.dtype)
    if args.world_size > 1:
        network.plugin_config.set_nccl_plugin(args.dtype)
    with net_guard(network):
        # Prepare
        network.set_named_parameters(tensorrt_llm_bloom.named_parameters())

        # Forward
        inputs = tensorrt_llm_bloom.prepare_inputs(args.max_batch_size,
                                                   args.max_input_len,
                                                   args.max_output_len, True,
                                                   args.max_beam_width)
        tensorrt_llm_bloom(*inputs)
        if args.enable_debug_output:
            # mark intermediate nodes' outputs
            for k, v in tensorrt_llm_bloom.named_network_outputs():
                v = v.trt_tensor
                v.name = k
                network.trt_network.mark_output(v)
                v.dtype = kv_dtype
        if args.visualize:
            model_path = os.path.join(args.output_dir, 'test.onnx')
            to_onnx(network.trt_network, model_path)

    tensorrt_llm.graph_rewriting.optimize(network)

    engine = None

    # Network -> Engine
    engine = builder.build_engine(network, builder_config)
    if rank == 0:
        config_path = os.path.join(args.output_dir, 'config.json')
        builder.save_config(builder_config, config_path)
    return engine


def build(rank, args):
    torch.cuda.set_device(rank % args.gpus_per_node)
    tensorrt_llm.logger.set_level(args.log_level)
    if not os.path.exists(args.output_dir):
        os.makedirs(args.output_dir)

    # when doing serializing build, all ranks share one engine
    builder = Builder()

    cache = None
    for cur_rank in range(args.world_size):
        # skip other ranks if parallel_build is enabled
        if args.parallel_build and cur_rank != rank:
            continue
        # NOTE: when only int8 kv cache is used together with paged kv cache no int8 tensors are exposed to TRT
        int8_trt_flag = args.quant_mode.has_act_and_weight_quant(
        ) or args.quant_mode.has_int8_kv_cache()
        builder_config = builder.create_builder_config(
            name=MODEL_NAME,
            precision=args.dtype,
            timing_cache=args.timing_cache if cache is None else cache,
            tensor_parallel=args.world_size,  # TP only
            parallel_build=args.parallel_build,
            num_layers=args.n_layer,
            num_heads=args.n_head,
            hidden_size=args.n_embd,
            vocab_size=args.vocab_size,
            max_position_embeddings=args.n_positions,
            max_batch_size=args.max_batch_size,
            max_input_len=args.max_input_len,
            max_output_len=args.max_output_len,
            int8=(args.quant_mode.has_act_and_weight_quant()
                  or args.quant_mode.has_int8_kv_cache()),
            quant_mode=args.quant_mode)
        builder_config.trt_builder_config.builder_optimization_level = 1
        engine_name = get_engine_name(MODEL_NAME, args.dtype, args.world_size,
                                      cur_rank)
        engine = build_rank_engine(builder, builder_config, engine_name,
                                   cur_rank, args)
        assert engine is not None, f'Failed to build engine for rank {cur_rank}'

        if cur_rank == 0:
            # Use in-memory timing cache for multiple builder passes.
            if not args.parallel_build:
                cache = builder_config.trt_builder_config.get_timing_cache()

        serialize_engine(engine, os.path.join(args.output_dir, engine_name))

    if rank == 0:
        ok = builder.save_timing_cache(
            builder_config, os.path.join(args.output_dir, "model.cache"))
        assert ok, "Failed to save timing cache."


if __name__ == '__main__':
    args = parse_arguments()
    logger.set_level(args.log_level)
    tik = time.time()
    if args.parallel_build and args.world_size > 1 and \
            torch.cuda.device_count() >= args.world_size:
        logger.warning(
            f'Parallelly build TensorRT engines. Please make sure that all of the {args.world_size} GPUs are totally free.'
        )
        mp.spawn(build, nprocs=args.world_size, args=(args, ))
    else:
        args.parallel_build = False
        logger.info('Serially build TensorRT engines.')
        build(0, args)

    tok = time.time()
    t = time.strftime('%H:%M:%S', time.gmtime(tok - tik))
    logger.info(f'Total time of building all {args.world_size} engines: {t}')