Overview

This recipe contains information and scripts to produce performance results for the Maxtext Llama2 70B training workload. The scripts help perform environment setup and launch benchmark jobs.

This variant of the workload is best-suited for GPU clusters with:

• At least 64 GPUs with at least 80 GB memory each. Training of this 70-billion parameter variant of the workload will not fit on fewer GPUs with less memory.
• This workload runs with BF16 or FP8 precision. FP8 is only supported by H100 GPUs. BF16 recipes are suitable for both A100 and H100 GPUs.

Expected Slurm Performance

Performance for Maxtext Llama2 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/$DTYPE/70b/$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 'completed step:' results/$GSW_VERSION/bf16/70b/64/log-maxtext-64_70b_1407993.out | tail -1
completed step: 49, seconds: 7.908, TFLOP/s/device: 460.480, Tokens/s/device: 1035.970, total_weights: 524288, loss: 0.000

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. 4096 * 128 / 7.908 = 66298.43

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 / 66298.43 / 86400 = 174.57 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 2 70b for GBS=1 is 1.82E+15. Calculation shown here.

E.g. Llama 2 70b BF16 on 64x H100 GPUs (GBS=128)

peak FLOPS for H100 = 989 TFLOPS 
training step time = 7.908 s 
model flops = 1.82E+15
MFU = 128 * 1.82E+15 / 7.908 / 64 / 989E+12 = 46.54% 

Prepare Slurm Environment

Create a staging area by running the attached setup.sh. The script converts the docker image from ghcr.io/nvidia/jax:maxtext-2024-12-09 to the nvidia+jax+maxtext-2024.12.09.sqsh file under the $STAGE_PATH folder.

# 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="batch"
# Set the Slurm account to launch against
export SLURM_ACCOUNT="account_name"
# Set the number of GPUs per node according to Slurm's gres, this is usually 8 or null - https://slurm.schedmd.com/gres.html
export SLURM_GPUS_PER_NODE=null

# Run the setup
bash ./setup.sh

Prepare Dataset

Since Maxtext Llama training only uses synthetic datasets, this step is omitted.

Run Slurm Training

Once the environment has been prepared, it is time to train a model.

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/$DTYPE/70b/$JOB_TOTAL_GPUS folder.

Below is a slurm command template for launching Maxtext Llama2 70b model training.

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

Where:

• DTYPE is a required environment variable.
  • DTYPE can be either fp8 or bf16.
• NUM_NODES can be calculated by N_GPUS / N_GPUS_PER_NODE, N_GPUS_PER_NODE is 8 for DGX H100, therefore for 256 GPUs scale, NUM_NODES should be 256 / 8 = 32.

Note: that 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

Global batch size ( training.model.global_batch_size) value should be set to <number of nodes> * <gpus per node> * <mbs> . E.g., 8 * 8 * 2 = 128 (in the example above) since we are using FSDP+DP.

Notes

model flops = (sequence length) * ((attention flops) + (mlp flops) + (normalization 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 2 70b calculation: 
    sequence length = 4096
    attention flops = 12 * 80 * 8192^2 * (1 + 8/64 + 4096/8192) = 104,689,827,840
    mlp flops = 18 * 80 * 28672 * 8192 = 338,228,674,560
    embedding flops = 6 * 32000 * 8192 = 1,572,864,000

    model flops = 4096 * (104,689,827,840 + 338,228,674,560 + 1,572,864,000) = 1.82E+15