Overview

This recipe contains information and scripts to produce performance results for the Llama 3.1 training workload. The scripts help perform environment setup, dataset setup, and launch benchmark jobs. This variant of the workload is best-suited for GPU clusters with

• At least 576 GPUs with at least 80 GB memory each. Training of this 405-billion parameter variant of the workload will not fit on fewer GPUs with less memory.
  • 32 GPUs with at least 80GB memory is the minimum when running proxy configs: <576 GPUs.
• H100 GPUs. This workload runs with BF16 or FP8, which are both supported by H100 GPUs.

Expected Performance

Performance for Llama 3.1 training is measured by seconds per iteration, or in other words seconds per training step. This metric is logged for every training step in a .out file which is generated inside of the $STAGE_PATH/results/$GSW_VERSION/$DTYPE/405b/$JOB_TOTAL_GPUS folder.

Since the performance fluctuates significantly at the beginning, we are using the last training step timing to obtain throughput value.

grep train_step_timing results/*.out
Epoch 0: : 100%|██████████| 50/50 [16:26<00:00, v_num=gjbq, reduced_train_loss=11.70, global_step=49.00, consumed_samples=12600.0, train_step_timing in s=12.80]

To obtain throughput as a tokens per second measurement, follow this formula:

(sequence length) * (global batch size) / (training_step_timing) = (throughput in tokens per second)

E.g. 8192 * 252 / 12.84 = 160778

To calculate time to train estimate:

(total tokens) / (throughput in tokens per second) / (number of seconds in a day) = (time to train in days)

E.g. 1e12 / 160778 / 86400 = 71.99 days

To calculate the model flops utilization (MFU):

MFU = (global batch size) * (model flops) / (training step time) / (number of GPUs) / (peak GPU FLOPS)

The peak theoretical throughput for H100 FP8 is 1979 TFLOPS and for H100 BF16 is 989 TFLOPS.

The model flops for Llama 3.1 405b for GBS=1 is 2.17E+16. Calculation shown here.

E.g. Llama 3.1 405b FP8 on 576x H100 GPUs (GBS=252)

peak FP8 FLOPS for H100 = 1979 TFLOPS
training step time = 11.24
model flops = 2.17E+16
MFU = 252 * 2.17E+16 / 11.24 / 576 / 1979E+12 = 42.71%

For proxy configs (<576 GPUs scales) we don't provide time to train estimates to avoid misleading conclusions. Proxy configs are not realistic and were created to allow fit of Llama model to smaller number of GPUs than intended.

Prerequisites

This recipe requires access to Llama 3.1. Instructions are below if needed.

Request Access

A HuggingFace account is required and you will need to create a HuggingFace access token in order to run the training script. Add the generated token to your environment via export HF_TOKEN=<your token>.

Access to Llama 3.1 must be requested through Meta's website then requested on the HuggingFace Llama page. The approval process is not automatic and could take a day or more.

Download benchmark

Note, that this step requires that you have NGC CLI tooling installed on your slurm login node. If you don't have it installed follow these instructions:

1. Click the top-right "Download" button and select "CLI". This will copy the download command into clipboard.
2. From your terminal on login node, paste the download command and execute it.
3. Extract the downloaded benchmark zip containing required scripts and configuration files.

Prepare Environment

Create a staging area by running the attached setup.sh. The script converts the docker image from nvcr.io/nvidia/nemo:24.09 to the nvidia+nemo+24.09.sqsh file under the $STAGE_PATH folder and copies NeMo Launcher code from the container.

# Set the path where all artifacts will be downloaded
export STAGE_PATH=<path to your shared file system folder> (e.g. /lustre/myproject/nemo)
# Set the Slurm partition to launch against
export SLURM_PARTITION=<partition>
# Set the Slurm account to launch against
export SLURM_ACCOUNT=<account>
# Set the number of GPUs per node according to Slurm's gres, this is usually 8 or "" - https://slurm.schedmd.com/gres.html
export SLURM_GPUS_PER_NODE=""
# Set HuggingFace token
export HF_TOKEN=<your token>

# Run the setup
bash ./setup.sh

Note:

• Slurm parameters might not be applicable to all environments. Please consult with your system administrator and update or remove parameters as needed.
• Setup script expects your $PYTHONUSERBASE/bin folder (typically ~/.local/bin) to be in your PATH. You can check this with python3 -m site --user-base. If you encounter error messages such as:

  WARNING: The script huggingface-cli is installed in '$PYTHONUSERBASE/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
  ..snip..
  /cm/local/apps/slurm/var/spool/job1490931/slurm_script: line xx: huggingface-cli: command not found
  

  Please add $PYTHONUSERBASE/bin to your PATH, and retry.

Prepare Dataset

Llama 3.1 405B uses synthetic data for training. A dataset does not need to be prepared. Note that Llama3.1 405B uses the GPT2BPETokenizer as a proxy.

Run Training

NeMo Launcher is using the Hydra framework to process command line arguments and pass them down as hyperparameters to a multi-node job performing the training.

The training will run for the first 50 steps and will stop afterwards. Log files and results will be located under the $STAGE_PATH/results/$GSW_VERSION/$DTYPE/405b/$JOB_TOTAL_GPUS folder.

Below is a command template for launching Llama 3.1 405b model training.

DTYPE=<fp8/bf16> MODEL_SIZE=405b sbatch -A ${SLURM_ACCOUNT} -p ${SLURM_PARTITION} -N ${NUM_NODES} ./launch.sh

Where:

• DTYPE and MODEL_SIZE are required environment variables.
  • DTYPE can be either fp8 or bf16.
  • MODEL_SIZE should be 405b in this case.
• NUM_NODES can be calculate by N_GPUS / N_GPUS_PER_NODE, N_GPUS_PER_NODE is 8 for DGX H100, therefore for 576 GPUs scale, NUM_NODES should be 576 / 8 = 72.
• -A and -p parameters are not applicable to all environments. Please consult with your system administrator and update or remove parameters as needed.

Note: it might be necessary to pass --gres=gpu:8 to sbatch for certain clusters on encountering errors like GPU not found. See https://slurm.schedmd.com/gres.html

The following applies only to the full model scales: 576, 1152, 2304 GPUs. Configurations and Global batch size changes for proxy configs <576 GPUs.

It is important to maintain these values for model parallelism settings in order to accurately assess performance results for completed jobs against expected baseline for the non-proxy 405b configurations:

• training.model.tensor_model_parallel_size=8
• training.model.pipeline_model_parallel_size=9
• training.model.virtual_pipeline_model_parallel_size=7
• training.model.context_parallel_size=2

Global batch size (training.model.global_batch_size) value should scale with total number GPUs. The starting global batch size for 576 GPUs is 252, therefore it should set to <number of total gpus> * 252 / 576.

Notes

model flops = (sequence length) * ((attention flops) + (mlp flops) + (embedding flops))

model flops breakdown:
    attention flops = 12 * (number of layers) * (hidden size)^2 * (1 + (number of query groups)/(number of attention heads) + (sequence length)/(hidden size))
    mlp flops = 18 * (number of layers) * (FFN size) * (hidden size)
    embedding flops = 6 * (vocab size) * (hidden size)

Llama 3.1 405b calculation:
    sequence length = 8192
    attention flops = 12 * 126 * 16384^2 * (1 + 16/128 + 8192/16384) = 659,545,915,392
    mlp flops = 18 * 126 * 53248 * 16384 = 1,978,637,746,176
    embedding flops = 6 * 128256 * 16384 = 12,608,077,824

    model flops = 8129 * (659,545,915,392 + 1,978,637,746,176 + 12,608,077,824) = 2.17E16