This recipe contains information and scripts to produce performance results for the Mistral Hugging Face fine-tuning training workload using PEFT and FSDP. The scripts help perform environment setup, dataset setup, and launch benchmark jobs. This variant of the workload is best-suited for GPU clusters with:
Performance for HF Mistral fine tuning is measured by train samples per second, which is logged in the .out file associated with the job.
grep train_samples_per_second log-hf_mistral_7b_32_656947.out
{'train_runtime': 2950.1412, 'train_samples_per_second': 555.363, 'train_steps_per_second': 0.034, 'train_loss': 1.0721950674057006, 'epoch': 6.25}
| Mistral 7b BF16 | Train samples per second on 8x H100 GPUs | Train samples per second on 16x H100 GPUs | Train samples per second on 32x H100 GPUs | Train samples per second on 64x H100 GPUs | Train samples per second on 128x H100 GPUs | Train samples per second on 256x H100 GPUs |
|---|---|---|---|---|---|---|
| Training samples per second | 16.287 | 37.895 | 81.626 | 161.273 | 308.95 | 555.363 |
This recipe requires access to Hugging Face Mistral. Instructions are below if needed.
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 setup.sh script. The script converts the docker image from nvcr.io/nvidia/pytorch:24.02-py3 to the nvidia+pytorch+24.02.sqsh file under the $STAGE_PATH folder and downloads DHS-LLM workshop source code.
# Set the path where all artifacts will be downloaded
export STAGE_PATH=<path to your shared file system folder> (e.g. /lustre/myproject/nemo)
# 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.
Access to Mistral 7B must be requested on Hugging Face Mistral 7B.
To download the model and dataset you will need to create a Hugging Face access token with READ privileges. You will use your HF user name and access token as the user/password for the git clones. For more information see: https://huggingface.co/docs/hub/en/security-tokens
Note: Cloning the model can take well over an hour, and you will be prompted twice for user/password. After the second prompt it'll appear as if it's hung.
cd $STAGE_PATH
# Only needs to be performed once
git lfs install
git clone https://huggingface.co/mistralai/Mistral-7B-v0.1
If the model download step was successful there should these files in the $STAGE_PATH/Mistral-7B-v0.1 folder.
README.md config.json generation_config.json model-00001-of-00002.safetensors model-00002-of-00002.safetensors model.safetensors.index.json pytorch_model-00001-of-00002.bin pytorch_model-00002-of-00002.bin pytorch_model.bin.index.json special_tokens_map.json tokenizer.json tokenizer.model tokenizer_config.json
git clone https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k
If the dataset clone step was successful there should these files in the $STAGE_PATH/ultrachat_200k/data folder
test_gen-00000-of-00001-3d4cd8309148a71f.parquet test_sft-00000-of-00001-f7dfac4afe5b93f4.parquet train_gen-00000-of-00003-a6c9fb894be3e50b.parquet train_gen-00001-of-00003-d6a0402e417f35ca.parquet train_gen-00002-of-00003-c0db75b92a2f48fd.parquet train_sft-00000-of-00003-a3ecf92756993583.parquet train_sft-00001-of-00003-0a1804bcb6ae68c6.parquet train_sft-00002-of-00003-ee46ed25cfae92c6.parquet
Once the environment has been prepared, it is time to train a model. Run the launch.sh script with sbatch for launching Hugging Face MISTRAL 7b model training on 1 to 32 nodes with BF16 precision.
Log files will be located under ${STAGE_PATH}/results/$GSW_VERSION/bf16/7b/$JOB_TOTAL_GPUS.
sbatch -A ${SBATCH_ACCOUNT} -p ${SBATCH_PARTITION} -N ${NUM_NODES} ./launch.sh
Where:
NUM_NODES can be calculate 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.We do not expose profiling options for this workload at this time.
accelerate launches on every node and pip install requirements.txt is run as part of srun command to ensure compute nodes have same environment. PYTHONPATH is set for this.