This recipe contains information and scripts to produce performance results for the Maxtext Llama3 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:
Performance for Maxtext Llama3 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/128/log-maxtext_llama3_70b_128_2102379.out | tail -1
completed step: 49, seconds: 15.570, TFLOP/s/device: 506.597, Tokens/s/device: 1052.266, total_weights: 2097152, 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. 8192 * 256 / 15.570 = 134691.84
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 / 134691.84 / 86400 = 85.93 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 Llama3 70b for GBS=1 is 3.94E+15. Calculation shown here.
E.g. Llama3 70b BF16 on 128x H100 GPUs (GBS=256)
peak FLOPS for H100 = 989 TFLOPS
training step time = 15.570 s
model flops = 3.94E+15
MFU = 256 * 3.94E+15 / 15.570 / 128 / 989E+12 = 51.17%
| Maxtext Llama3 70b BF16 (FSDP+DP MBS=2) | Throughput on 128x H100 GPUs (FSDP=128 GBS=256) | Throughput on 256x H100 GPUs (FSDP=256 GBS=512) | Throughput on 512x H100 GPUs (FSDP=128 GBS=1024) | Throughput on 1024x H100 GPUs (FSDP=128 GBS=2048) | Throughput on 2048x H100 GPUs (FSDP=128 GBS=4096) |
|---|---|---|---|---|---|
| Training step time (seconds per step) | 15.563 | 15.444 | 15.519 | 15.603 | 15.898 |
| Throughput in tokens per second | 134752.43 | 271581.46 | 540537.92 | 1075255.78 | 2110607.12 |
| Model flops utilization | 51.20% | 51.59% | 51.34% | 51.06% | 50.12% |
| Time to train 1T tokens in days | 85.89 | 42.62 | 21.41 | 10.76 | 5.48 |
| Maxtext Llama3 70b FP8 (FSDP+DP MBS=2) | Throughput on 128x H100 GPUs (FSDP=128 GBS=256) | Throughput on 256x H100 GPUs (FSDP=256 GBS=512) | Throughput on 512x H100 GPUs (FSDP=128 GBS=1024) | Throughput on 1024x H100 GPUs (FSDP=128 GBS=2048) | Throughput on 2048x H100 GPUs (FSDP=128 GBS=4096) |
|---|---|---|---|---|---|
| Training step time (seconds per step) | 9.074 | 9.488 | 9.244 | 9.483 | 9.339 |
| Throughput in tokens per second | 231116.60 | 442064.08 | 907465.17 | 1769188.65 | 3592936.29 |
| Model flops utilization | 43.88% | 41.97% | 43.07% | 41.99% | 42.64% |
| Time to train 1T tokens in days | 50.08 | 26.18 | 12.75 | 6.54 | 3.22 |
No special pre-requisites or access required for this recipe.
We reference a number of Slurm commands and parameters in this document. A brief summary is included below. It's important to note these are a guide and might not be applicable to all environments. Please consult with your system administrator for the parameters that are specific to your system.
Common parameters:
SBATCH_PARTITION or -p - Partition (or queue) to use.SBATCH_ACCOUNT or -A - Slurm account to associate with your job, different from your user. Meant for accounting purposes.SBATCH_GPUS_PER_NODE or --gres=gpu:<num gpus> - If your cluster is configured with GRES this should be set to all GPUs in a node. Ignore if not configured.These parameters can be set either by exporting the environment variable or using the corresponding sbatch flag.
Create a staging area by running the attached setup.sh. The script converts the docker image from the registry to a 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/maxtext)
# Run the setup
sbatch -A ${SBATCH_ACCOUNT} -p ${SBATCH_PARTITION} -N 1 ./setup.sh
Check the corresponding slurm-<job_id>.out file for status information.
Important: STAGE_PATH used in this step must be used when running the workload.
Since Maxtext Llama training only uses synthetic datasets, this step is omitted.
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 Llama3 70b model training.
DTYPE=<fp8/bf16> sbatch -A ${SBATCH_ACCOUNT} -p ${SBATCH_PARTITION} -N ${NUM_NODES} ./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.Global batch size ( training.model.global_batch_size) value should be set to <number of nodes> * <gpus per node> * <mbs> . E.g., 16 * 8 * 2 = 256 (in the example above) since we are using FSDP+DP.
Nsight profiling is used as part of this benchmark. See Run Nsight Profiling for more info.
The workload uses [PGLE](See https://github.com/google/paxml?tab=readme-ov-file#run-pgle-workflow-on-gpu) and turns on Nsight profiling by default. The profiles are located at $STAGE_PATH/results/$GSW_VERSION/${DTYPE}/70b/$JOB_TOTAL_GPUS/nsys.
Each job generates two nsys-rep files:
*-prof.nsys-rep is the PGLE profiling run at the start of the benchmark.*-perf.nsys-rep is the profile taken during the performace run. Use this profile when debugging performance issues.Options to customize profiling behaviour are not currently available for this workload. GPU device metrics capture is not currently supported for this workload.
In order to view the profile traces (*.nsys-rep files) interactively:
Since most of the benchmarking jobs run on multiple GPUs, there will be multiple .nsys-rep files generated for each run. Multi-Report Analysis Guide will be very helpful to automate the analysis and get to results quicker by using Nsight recipes.
See these tutorials to get a quick start if you are new to Nsight profiling.
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)
Llama3 70b calculation:
sequence length = 8192
attention flops = 12 * 80 * 8192^2 * (1 + 8/64 + 8192/8192) = 136,902,082,560
mlp flops = 18 * 80 * 28672 * 8192 = 338,228,674,560
embedding flops = 6 * 128256 * 8192 = 6,304,038,912
model flops = 8192 * (136,902,082,560 + 338,228,674,560 + 6,304,038,912) = 3.94E+15