sbhavani / h100-performance-tests

h100 testing scripts

This repository tested the T-FLOPS on an 8xH100 node, gratiously provided by CoreWeave.

To run:

1. Clone this repository
2. Install requirements pip install -r h100-stuff/requirements.txt (and ensure you have git-lfs installed. See here for directions)
3. wandb login to track with wandb (Optional)
4. accelerate launch --config_file bf16_fsdp.yml run_summarization_no_trainer.py --model_name_or_path t5-small --dataset_name cnn_dailymail --dataset_config "3.0.0" --source_prefix "summarize: " --output_dir tst-summarization --per_device_train_batch_size=1 --per_device_eval_batch_size=1 --max_train_steps 3

Report

For the full Weights and Biases log, please check out the workspace here.

Overall we saw a 25% increase in terms of FLOPS comparing bf16 to fp8 on the H100's. Below are two graphs comparing the T-FLOPS at different batch sizes, one at 8 per GPU (64 actual) and the other at 4 per GPU (32 actual). The model utilized was the "T5-Large" varient of the T5 models, so as to use a large model to properly test the capabilities of the cluster.

To see how T-FLOPS were calculated, please see the notebook here. Note that you need to download the chart data for "batch_time" from the Weights and Biases logs, grouped by the "xxx-steps-bs-y" key.

Footnote: The numbers reported also show an issue somewhere on the architecture or code that potentially wasn't allowing for the full TFLOPS performance gain. There was a negligble difference between training on BF16 and full precision on the node, leading to this conclusion. (Review the W&B logs to see this).

fp8 or bf16 on multi-node

Change each yml to be (maintaining the mixed_precision already stored there):

And pass it to accelerate launch under the --config_file param

compute_environment: LOCAL_MACHINE
deepspeed_config: {}
distributed_type: 'MULTI_GPU'
downcast_bf16: 'no'
fsdp_config: {}
machine_rank: 0
main_process_ip: null
main_process_port: null
main_training_function: main
-mixed_precision: 'bf16'
+mixed_precision: 'fp8'
num_machines: 1
num_processes: 1
num_processes: 8
use_cpu: false