Skip to content

Latest commit

 

History

History
612 lines (519 loc) · 27.5 KB

perf-overview.md

File metadata and controls

612 lines (519 loc) · 27.5 KB
Raw

(perf-overview)=

Important

As of TensorRT-LLM v0.10, these performance benchmarks have changed methodology to utilize in-flight batching and no longer utilize static benchmarking. These numbers are initial measurements and are expected to improve in future releases.

Overview

This document summarizes performance measurements of TensorRT-LLM on H100 (Hopper), L40S (Ada) and A100 (Ampere) GPUs for a few key models.

The data in the following tables is provided as a reference point to help users validate observed performance. It should not be considered as the peak performance that can be delivered by TensorRT-LLM.

Known Issues

The following issues are being addressed to improve the efficiency of TensorRT-LLM.

Unexpected extra GPU memory allocation when enabling --multiple_profiles

We observed that enabling multiple profiles can lead to extra unexpected GPU memory usage on some cases starting from v0.11. The issue will be addressed in future releases.

Fused Matmul + Gated-SiLU (LLaMA)

The current implementation combines two Matmul operations into one Matmul followed by a separate SwiGLU kernel (when --use_fused_mlp is enabled). There is also a more efficient implementation that runs single Matmul + SwiGLU fused kernel for FP8 on Hopper (when --use_fused_mlp --gemm_swiglu_plugin fp8 is enabled). The gemm_swiglu_plugin will support more data types and GPU architectures in the future release.

Throughput Measurements

The below table shows performance data where a local inference client is fed requests at an infinite rate (no delay between messages), and shows the throughput client-server scenario under maximum load.

The performance numbers below were collected using the steps described in this document.

All data in the table below was generated using version 0.11.0 and presents token throughput in tokens/second.
GPU H200 141GB HBM3 GH200 120GB H100 80GB HBM3 H100 80GB HBM3 A100-SXM4-80GB L40S
Precision FP8 FP8 FP8 FP16 FP16 FP8
Model Input/Output Lengths TP
GPTJ 6B 128/128 1 25116.54 24998.09 24456.84 13328.96 6168.8 7737.44
2048/128 1 2845.4 2840.46 2781.11 1410.81 662 83.46
128/2048 1 8165.91 7936.16 7643.02 3503.41 2213.44 1927.91
2048/2048 1 3560.37 3197.21 3081.26 1326.79 893.43
LLaMA v2 7B 128/128 1 19695.41 19509.49 17684.88 11605.69 5286.1 6655.52
2048/128 1 2471.89 2401.29 2342.71 1173.81 558.56 644.72
128/2048 1 7867.28 6689.51 6814.72 3074.4 1813.79 1591.51
2048/2048 1 3215.63 3015.84 2820.31 1289.87 716.55 653.19
LLaMA v3 8B 128/128 1 29084.05 29197.48 27781.28 15225.75 6450.88 8929.6
2048/128 1 3699.64 3780.47 3555.57 1844.38 775.18 1052.3
128/2048 1 23723.81 22055.94 17894.85 8415.67 4837.47 4497.21
2048/2048 1 11193.29 8877.13 8398.71 3996.93 2271.65 1911.63
Mistral 7B 128/128 1 31618.59 31868.45 30400.21 16108.11 6749.91 10237.23
2048/128 1 3791.1 3795.27 3618.11 1896.76 783.94 1126.08
128/2048 1 25646.02 20491.88 20518.75 10018.54 5358.28 5441.98
2048/2048 1 12068.11 9462.96 9504.59 4383.42 2465.77 2213.69
LLaMA v2 70B 128/128 2 6652.29 5619.41 6502.44
4 10921.65 11043 10448.46 6219.11 2487.78 1549.09
8 15878.34 14781.66 10093.27 4233.24 1497.68
2048/128 2 766.38 647.73 747.14
4 1296.75 1298.94 1231.26 714.07 285.9 179.19
8 1930.16 1808.02 1230.66 494.29 176.24
128/2048 2 7014.86 4844.17 5267.56
4 13365.86 11596.55 9202.42 3787.24 2267.02 1772.45
8 18861.53 17085.82 7846.64 5096.52 2290.99
2048/2048 2 3554.71 2843.31 2457.73
4 6604.37 5969.11 4586.99 1994.1 1137.22 890.83
8 10034.12 7647.54 4347.09 2152.35 1130.36
LLaMA v3 70B 128/128 4 9872.81
8 15255 13853.05 4033.42
2048/128 4 1284.88
8 1918.47 1738.94 476.42
128/2048 4 9996.88
8 19071.39 10887.34 3373.71
2048/2048 4 4985.31
8 9387.81 6029.39 1824.06
Mixtral 8x7B 128/128 2 26317.73 21768.19 24770.44 11821.14 5522.43
2048/128 2 3181.76 2545.52 2973.11 1391.28 636.77
128/2048 2 30105.61 23643.33 22120.85 6337.02 3698.23
2048/2048 2 15002.42 11683.11 11486.66 3024.95 1710.53
Falcon 180B 128/128 4 5647.01 5568.91
8 9304.06 8885.39 2171.78
2048/128 4 670.99 693.82 667.8
8 1103.18 1065.16 238.61
128/2048 4 8358.01 6655.38 6376.89
8 14514.24 12447.25 2657.9
2048/2048 4 4169.39 3415.05 3412.09
8 7524.11 6326.46 1392.31

TP stands for Tensor Parallelism

Reproducing Benchmarked Results

Building the TensorRT-LLM Container

In order to benchmark TensorRT-LLM, you will need to follow the Quick Start build process to create a baseline container for building a wheel. Additionally, the development container needs a copy of the source code to build the wheel and the benchmarking script. Create the right build environment, use the following :

git clone https://github.com/NVIDIA/TensorRT-LLM.git
cd TensorRT-LLM
git submodule update --init --recursive
git lfs install
git lfs pull
make -C docker build
make -C docker run LOCAL_USER=1

Warning

If you have elevated privileges on your system, then skip the make -C docker run LOCAL_USER=1 command above as it may make it so that you cannot access some required system libraries within the container because the build forces your UID and GID to match those that are set for your non-elevated user. There are cases where the container will be booted as root (i.e. on some SLURM systems with the pyxis plugin) which will cause libraries to be missing.

If you are benchmarking in a shared environment, you need to specify the GPU indices that you would like the container to use, otherwise the Makefile defaults to loading the container with all GPUs on the system. For example, if you only have the 4 higher indices of GPUs on your system you can configure it using the following example:

NV_GPU=0,1,2,3
make -C docker run LOCAL_USER=1 GPU_OPTS='--gpus \"device=${NV_GPU}\"'

Additionally, if you'd like to mount external storage to access persistent storage, or previously built engines, you can mount directories as follows (simply replace source and destination with the appropriate paths):

make -C docker run LOCAL_USER=1 DOCKER_RUN_ARGS="-v /source:/destination"

Once the container starts, you'll need to build the wheel and the benchmarking scripts. From the code root (the default directory when the container is loaded), the following commands will build the TensorRT-LLM wheel, install dependencies, and build the benchmark scripts:

python3 ./scripts/build_wheel.py --benchmarks --trt_root /usr/local/tensorrt
pip install ./build/tensorrt_llm*.whl

Methodology

The following tables are references for commands that are used as part of the benchmarking process.

Commands

Stage Description Command
Build Build a TensorRT-LLM engine trtllm-build --model_config $model_cfg --use_fused_mlp --gpt_attention_plugin float16 --output_dir $engine_dir --max_batch_size $max_batch_size --max_input_len 2048 --max_seq_len 2048 --reduce_fusion disable --workers $tp_size --max_num_tokens $max_num_tokens --use_paged_context_fmha enable --multiple_profiles enable
Dataset Create a synthetic dataset benchmarks/cpp/prepare_dataset.py --output=$dataset_file --tokenizer=$model_name token-norm-dist --num-requests=2000 --input-mean=$isl --output-mean=$osl --input-stdev=0 --output-stdev=0
Run Run a benchmark with a dataset mpirun -n $tp_size --allow-run-as-root --oversubscribe cpp/build/benchmarks/gptManagerBenchmark --engine_dir $engine_dir --type IFB --dataset $dataset_file --eos_id -1 --scheduler_policy guaranteed_no_evict --kv_cache_free_gpu_mem_fraction 0.99 --output_csv result.csv --request_rate -1.0 --enable_chunked_context --warm_up 0

Variables

Name Description
$isl Benchmark input sequence length.
$osl Benchmark output sequence length.
$tp_size Number of GPUs to run the benchmark with
$engine_dir Location to store built engine file (can be deleted after running benchmarks).
$model_cfg Name of the model configuration JSON file to use for building.
$model_name HuggingFace model name eg. meta-llama/Llama-2-7b-hf or use the path to a local weights directory
$dataset_file Location of the dataset file generated by prepare_dataset.py
$results_csv Path to store end results to.
$max_batch_size Absolute maximum number of concurrent requests an engine can handle during one iteration.
$max_num_tokens Maximum number of total tokens an engine can handle during one iteration.

Engine Building

All benchmarks were run using a single engine with a configuration that is capable of handling the maximum sequence lengths encountered during benchmarking. For each benchmark, regardless of input/output sequence length, you can reuse the single engine to run all tests. Each engine will be built with a paged KV cache and in-flight batching enabled. For more information see the documentation about in-flight batching.

In order to build an engine you will need to run the following command by specifying a configuration file for the model that you would like to build (see below). The general build command is as follows:

trtllm-build --model_config $model_cfg --use_fused_mlp --gpt_attention_plugin float16 --output_dir $engine_dir --max_batch_size $max_batch_size --max_input_len 2048 --max_seq_len 2048 --reduce_fusion disable --workers $tp_size --max_num_tokens $max_num_tokens --use_paged_context_fmha enable --multiple_profiles enable

Some notes about the command:

  • --workers affects the number of threads that build the engine file and does not necessarily need to match the TP size. Make sure to set the tensor parallelism in the $model_cfg JSON file. See below
  • You can run benchmarks for datasets that fit within the bounds of the max_input_len and max_seq_len parameters.

Engine Configuration Files

In order to configure the TensorRT-LLM build process for benchmarking, you need to provide trtllm-build a configuration file that specifies the following the network configuration, parallelism mapping, and quantization options.

Below we document how to benchmark each model on an H100-HBM3-80GB system and reproduce the throughput numbers we document on our [Performance section](#performance of-tensorrt-llm).

Important

In order to change the parallelism for a build, you need to modify the mapping dictionary in your configuration file. The settings must conform to the following condition: world_size == tp_size * pp_size.

Note

All configurations below are set to run utilizing FP8 by default. If you would like to run on an A100 system, see our notes about disabling FP8 quantization.

Network Configuration Files and Settings

Each network has its own configuration file. All networks are configured to run using FP8 quantization by default. Additionally, each network has a specific tuning for the $max_batch_size and $max_num_tokens parameters -- at times varying for some input and output sequence legnths within the same model.

![Note] General settings are specified by "General" in the "ISL/OSL" column. For special cases, specific input and output sequence lengths will be specified.

Model Configuration File (FP8)
EleutherAI/gpt-j-6b 
{
    "architecture": "GPTJForCausalLM",
    "dtype": "float16",
    "num_hidden_layers": 28,
    "num_attention_heads": 16,
    "hidden_size": 4096,
    "norm_epsilon": 1e-05,
    "vocab_size": 50400,
    "position_embedding_type": "rope_gptj",
    "max_position_embeddings": 2048,
    "hidden_act": "gelu_new",
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    },
    "rotary_dim": 64,
    "kv_dtype": "float16"
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
1 2048 128 General
1 2048 2048 128, 128
tiiuae/falcon-180B 
{
    "architecture": "FalconForCausalLM",
    "dtype": "bfloat16",
    "num_hidden_layers": 80,
    "num_attention_heads": 232,
    "num_key_value_heads": 8,
    "hidden_size": 14848,
    "norm_epsilon": 1e-05,
    "vocab_size": 65024,
    "position_embedding_type": "rope_gpt_neox",
    "max_position_embeddings": 2048,
    "hidden_act": "gelu",
    "use_parallel_embedding": false,
    "embedding_sharding_dim": 0,
    "share_embedding_table": false,
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    },
    "mapping": {
        "world_size": 8,
        "tp_size": 8,
        "pp_size": 1
    },
    "bias": false,
    "parallel_attention": true,
    "new_decoder_architecture": true,
    "kv_dtype": "float16"
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
4 8192 4096 General
8 8192 2048 General
meta-llama/Llama-2-7b-hf 
{
    "architecture": "LlamaForCausalLM",
    "dtype": "float16",
    "num_hidden_layers": 32,
    "num_attention_heads": 32,
    "hidden_size": 4096,
    "intermediate_size": 11008,
    "num_key_value_heads": 32,
    "vocab_size": 32000,
    "position_embedding_type": "rope_gpt_neox",
    "max_position_embeddings": 4096,
    "hidden_act": "silu",
    "rotary_base": 10000.0,
    "rotary_scaling": null,
    "norm_epsilon": 1e-05,
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    },
    "kv_dtype": "float16"
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
1 8192 4096 General
meta-llama/Llama-2-70b-hf 
{
    "architecture": "LlamaForCausalLM",
    "dtype": "float16",
    "num_hidden_layers": 80,
    "num_attention_heads": 64,
    "hidden_size": 8192,
    "intermediate_size": 28672,
    "num_key_value_heads": 8,
    "vocab_size": 32000,
    "position_embedding_type": "rope_gpt_neox",
    "max_position_embeddings": 4096,
    "hidden_act": "silu",
    "rotary_base": 10000.0,
    "rotary_scaling": null,
    "norm_epsilon": 1e-05,
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    },
    "mapping": {
        "world_size": 4,
        "tp_size": 4,
        "pp_size": 1
    },
    "kv_dtype": "float16"
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
2 2048 2048 General
4 8192 4096 General
4 8192 256 128, 4096
8 16384 8192 General
8 16384 1024 128, 2048
meta-llama/Meta-Llama-3-8B 
{
    "architecture": "LlamaForCausalLM",
    "num_hidden_layers": 32,
    "num_attention_heads": 32,
    "num_key_value_heads": 8,
    "hidden_size": 4096,
    "vocab_size": 128256,
    "max_position_embeddings": 8192,
    "hidden_act": "silu",
    "norm_epsilon": 1e-05,
    "dtype": "float16",
    "position_embedding_type": "rope_gpt_neox",
    "intermediate_size": 14336,
    "rotary_base": 500000.0,
    "rope_theta": 500000.0,
    "rotary_scaling": null,
    "mapping": {
        "world_size": 1,
        "tp_size": 1,
        "pp_size": 1
    },
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    },
    "kv_dtype": "float16"
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
1 8192 2048 General
meta-llama/Meta-Llama-3-70B 
{
    "architecture": "LlamaForCausalLM",
    "num_hidden_layers": 80,
    "num_attention_heads": 64,
    "num_key_value_heads": 8,
    "hidden_size": 8192,
    "vocab_size": 128256,
    "max_position_embeddings": 8192,
    "hidden_act": "silu",
    "dtype": "float16",
    "norm_epsilon": 1e-05,
    "position_embedding_type": "rope_gpt_neox",
    "intermediate_size": 28672,
    "rotary_base": 500000.0,
    "rope_theta": 500000.0,
    "rotary_scaling": null,
    "mapping": {
        "world_size": 4,
        "tp_size": 4,
        "pp_size": 1
    },
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    },
    "kv_dtype": "float16"
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
4 1024 2048 General
8 16384 8192 General
mistralai/Mixtral-8x7B-v0.1 
{
    "architecture": "MixtralForCausalLM",
    "num_hidden_layers": 32,
    "num_attention_heads": 32,
    "num_key_value_heads": 8,
    "hidden_size": 4096,
    "norm_epsilon": 1e-05,
    "vocab_size": 32000,
    "max_position_embeddings": 32768,
    "head_size": 128,
    "hidden_act": "swiglu",
    "dtype": "float16",
    "position_embedding_type": "rope_gpt_neox",
    "intermediate_size": 14336,
    "moe_num_experts": 8,
    "moe_top_k": 2,
    "rotary_base": 1000000.0,
    "rope_theta": 1000000.0,
    "mapping": {
        "world_size": 1,
        "tp_size": 1,
        "pp_size": 1
    },
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    },
    "kv_dtype": "float16"
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
2 3072 2048 General
4 8192 8192 General
mistralai/Mistral-7B-v0.1 
{
    "architecture": "MistralForCausalLM",
    "num_hidden_layers": 32,
    "num_attention_heads": 32,
    "num_key_value_heads": 8,
    "hidden_size": 4096,
    "norm_epsilon": 1e-05,
    "vocab_size": 32000,
    "max_position_embeddings": 32768,
    "hidden_act": "silu",
    "dtype": "float16",
    "logits_dtype": "float32",
    "position_embedding_type": "rope_gpt_neox",
    "use_parallel_embedding": false,
    "embedding_sharding_dim": 0,
    "share_embedding_table": false,
    "intermediate_size": 14336,
    "use_prompt_tuning": false,
    "mapping": {
        "world_size": 1,
        "tp_size": 1,
        "pp_size": 1
    },
    "quantization": {
        "quant_algo": "FP8",
        "kv_cache_quant_algo": "FP8"
    }
}
$tp_size $max_num_tokens $max_batch_size ISL/OSL
1 8192 4098 General

Running on A100

To run the benchmarks on A100, you will need to undefine or remove the following quantization fields from each config json file, because FP8 computation is a feature in H100 and newer GPUs.

"quantization": {
	"quant_algo": null,
	"kv_cache_quant_algo": null,
}

Preparing a Dataset

In order to prepare a dataset, you can use the provided script. To generate a synthetic dataset, run the following command:

benchmarks/cpp/prepare_dataset.py --output=$dataset_file --tokenizer=$model_name token-norm-dist --num-requests=$num_requests --input-mean=$isl --output-mean=$osl --input-stdev=0 --output-stdev=0

The command will generate a JSON file located at the path specified $dataset_file where all requests are of the same input/output sequence length combinations. The script works by using the tokenizer to retrieve the vocabulary size and randomly sample token IDs from it to create entirely random sequences. In the command above, all requests will be uniform because the standard deviations for both input and output sequences are set to 0.

For each input and output sequence length combination, the table below details the $num_requests that were used. For shorter input and output lengths, a larger number of messages were used to guarantee that the system hit a steady state because requests enter and exit the system at a much faster rate. For longer input/output sequence lengths, requests remain in the system longer and therefore require less requests to achieve steady state.

Input Length Output Length $num_requests
128 128 30000
128 2048 3000
128 4096 1500
2048 128 3000
2048 2048 1500

Running the Benchmark

To run the benchmark with the generated data set, simply run the following command from the root of the TensorRT-LLM repository. See the variables section for reference on variable values.

mpirun -n $tp_size --allow-run-as-root --oversubscribe cpp/build/benchmarks/gptManagerBenchmark --engine_dir $engine_dir --type IFB --dataset $dataset_file --eos_id -1 --scheduler_policy guaranteed_no_evict --kv_cache_free_gpu_mem_fraction 0.99 --output_csv result.csv --request_rate -1.0 --enable_chunked_context --warm_up 0

[!Warning] GH200 benchmarks For GH200 benchmarks, the command above must be modified to use --kv_cache_free_gpu_mem_fraction 0.95 to avoid an out of memory scenario.

The command will run the gptManagerBenchmark binary that will report the throughput and other metrics as part of its output that can be compared with the table in the Performance section of this README.