Use FSDP to fine-tune Mixtral-8x7B on an A4 Slurm cluster

Access Mixtral by using Hugging Face

To use Hugging Face to access Mixtral, do the following:

1. Create a Hugging Face read access token.
2. Copy and save the read access token value. You use it later in this tutorial.

Prepare your environment

You do the following steps on your local machine to prepare for deployment of the cluster.

1. Clone the Google Cloud Cluster Toolkit repository:

   git clone https://github.com/GoogleCloudPlatform/cluster-toolkit.git
   

2. Create a Cloud Storage bucket:

   export BUCKET_NAME="your-unique-bucket-name"
   gcloud storage buckets create gs://${BUCKET_NAME}
   

Create an A4 Slurm cluster

To create an A4 Slurm cluster, do the following:

1. Go to the cloned cluster-toolkit directory:

   cd cluster-toolkit
   

2. If it's your first time using Cluster Toolkit, then build the gcluster binary:

   make
   

3. Go to the examples/machine-learning/a4-highgpu-8g directory.

   Open the a4high-slurm-deployment.yaml file and edit it as follows:

   terraform_backend_defaults:
     type: gcs
     configuration:
       bucket: BUCKET_NAME
   
   vars:
     deployment_name: DEPLOYMENT_NAME
     project_id: PROJECT_ID
     region: REGION
     zone: ZONE
     a4h_cluster_size: 2
     a4h_reservation_name: RESERVATION_NAME
   

   Replace the following:

   • BUCKET_NAME: : the name of the Cloud Storage bucket that you created in the previous section.
   • PROJECT_ID: the ID of the Google Cloud project where your Cloud Storage exists and where you want to create your Slurm cluster.
   • REGION: the region where your reservation exists.
   • ZONE: the zone where your reservation exists.
   • A4h_reservation_name: use the name of your A4 reservation.

4. Open the a4high-slurm-blueprint.yaml file and edit it as follows:

   • Remove the filestore_homefs module.
   • Enable the lustrefs and private-service-access modules.
   • In the vars block, configure the following:
     1. Find slurm_vars and set install_managed_lustre to true.
     2. Set the per_unit_storage_throughput parameter to 500.
     3. Set the size_gib parameter to 36000.

5. Deploy the cluster:

   ./gcluster deploy -d examples/machine-learning/a4-highgpu-8g/a4high-slurm-deployment.yaml \
     examples/machine-learning/a4-highgpu-8g/a4high-slurm-blueprint.yaml \
     --auto-approve
   

   The ./gcluster deploy command starts a two-phase process, which is as follows:

   • The first phase builds a custom image with all software pre-installed, which can take up to 35 minutes to complete.
   • The second phase deploys the cluster by using that custom image. This process should complete more quickly than the first phase.

Prepare your workload

To prepare your workload, follow these steps:

1. Create workload scripts.

2. Upload scripts to the Slurm cluster.

3. Connect to the Slurm cluster .

4. Install frameworks and tools.

Create workload scripts

To create the scripts that your fine-tuning workload will use, follow these steps:

1. To set up the Python virtual environment, create the install_environment.sh file with the following content:

   #!/bin/bash
   # This script sets a reliable environment for FSDP training.
   # It is meant to be run on a compute node.
   set -e
   
   # --- 1. Create the Python virtual environment ---
   VENV_PATH="$HOME/.venv/venv-fsdp"
   if [ ! -d "$VENV_PATH" ]; then
     echo "--- Creating Python virtual environment at $VENV_PATH ---"
     python3 -m venv $VENV_PATH
   else
     echo "--- Virtual environment already exists at $VENV_PATH ---"
   fi
   
   source $VENV_PATH/bin/activate
   
   # --- 2. Install Dependencies ---
   echo "--- [STEP 2.1] Upgrading build toolchain ---"
   pip install --upgrade pip wheel packaging
   
   echo "--- [STEP 2.2] Installing PyTorch Nightly ---"
   pip install --force-reinstall --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
   
   echo "--- [STEP 2.3] Installing application dependencies ---"
   if [ -f "requirements-fsdp.txt" ]; then
       pip install -r requirements-fsdp.txt
   else
       echo "ERROR: requirements-fsdp.txt not found!"
       exit 1
   fi
   
   # --- [STEP 2.4] Build Flash Attention from Source ---
   echo "--- Building flash-attn from source... This will take a while. ---"
   # Use all available CPU cores to speed up the build
   MAX_JOBS=$(nproc) pip install flash-attn --no-build-isolation
   
   # --- 3. Download the Model ---
   echo "--- [STEP 2.5] Downloading Mixtral model ---"
   if [ -z "$HF_TOKEN" ]; then
     echo "ERROR: The HF_TOKEN environment variable is not set."; exit 1;
   fi
   pip install huggingface_hub[cli]
   
   # Execute the CLI using its full, explicit path
   $VENV_PATH/bin/huggingface-cli download mistralai/Mixtral-8x7B-v0.1 --local-dir ~/Mixtral-8x7B-v0.1 --token $HF_TOKEN
   
   echo "--- Environment setup complete. ---"
   

2. To specify the Python dependencies for the training script, create a requirements-fsdp.txt file with the following content:

   transformers==4.55.0
   datasets==4.0.0
   peft==0.16.0
   accelerate==1.9.0
   trl==0.21.0
   
   # Other dependencies
   sentencepiece==0.2.0
   protobuf==6.31.1
   

3. Specify train-mixtral.py as the main training script:

   import torch
   from torch.distributed.fsdp import MixedPrecision
   from datasets import load_dataset
   import shutil
   import os
   import torch.distributed as dist
   
   from peft import LoraConfig, PeftModel, get_peft_model
   from transformers import (
       AutoModelForCausalLM,
       AutoTokenizer,
       TrainingArguments,
       HfArgumentParser,
   )
   
   from torch.distributed import get_rank, get_world_size
   
   from transformers.models.mixtral.modeling_mixtral import MixtralDecoderLayer
   from trl import SFTTrainer
   from dataclasses import dataclass, field
   from typing import Optional
   
   @dataclass
   class ScriptArguments:
       model_id: str = field(default="mistralai/Mixtral-8x7B-v0.1", metadata={"help": "Hugging Face model ID from the Hub"})
       dataset_name: str = field(default="philschmid/gretel-synthetic-text-to-sql", metadata={"help": "Dataset from the Hub"})
       run_inference_after_training: bool = field(default=False, metadata={"help": "Run sample inference on rank 0 after training"})
       dataset_subset_size: Optional[int] = field(default=None, metadata={"help": "Number of samples to use from the dataset for training. If None, uses the full dataset."})
   
   @dataclass
   class PeftArguments:
       lora_r: int = field(default=16, metadata={"help": "LoRA attention dimension"})
       lora_alpha: int = field(default=32, metadata={"help": "LoRA alpha scaling factor"})
       lora_dropout: float = field(default=0.05, metadata={"help": "LoRA dropout probability"})
   
   @dataclass
   class SftTrainingArguments(TrainingArguments):
       max_length: Optional[int] = field(default=2048, metadata={"help": "The maximum sequence length for SFTTrainer"})
       packing: Optional[bool] = field(default=False, metadata={"help": "Enable packing for SFTTrainer"})
       ddp_find_unused_parameters: Optional[bool] = field(default=False, metadata={"help": "When using FSDP activation checkpointing, this must be set to False for Mixtral"})
   
   def formatting_prompts_func(example):
       system_message = "You are a text to SQL query translator. Users will ask you questions in English and you will generate a SQL query based on the provided SCHEMA."
       user_prompt = f"### SCHEMA:\n{example['sql_context']}\n\n### USER QUERY:\n{example['sql_prompt']}"
       response = f"\n\n### SQL QUERY:\n{example['sql']}"
       return f"{system_message}\n\n{user_prompt}{response}"
   
   def main():
       parser = HfArgumentParser((ScriptArguments, PeftArguments, SftTrainingArguments))
       script_args, peft_args, training_args = parser.parse_args_into_dataclasses()
   
       training_args.gradient_checkpointing = True
       training_args.gradient_checkpointing_kwargs = {"use_reentrant": True}
   
       training_args.optim = "adamw_torch_fused"
   
       bf16_policy = MixedPrecision(
           param_dtype=torch.bfloat16,
           reduce_dtype=torch.bfloat16,
           buffer_dtype=torch.bfloat16,
       )
   
       training_args.fsdp = "full_shard"
       training_args.fsdp_config = {
           "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
           "fsdp_transformer_layer_cls_to_wrap": [MixtralDecoderLayer],
           "fsdp_state_dict_type": "SHARDED_STATE_DICT",
           "fsdp_offload_params": False,
           "fsdp_forward_prefetch": True,
           "fsdp_mixed_precision_policy": bf16_policy
       }
   
       tokenizer = AutoTokenizer.from_pretrained(script_args.model_id, trust_remote_code=True)
   
       tokenizer.pad_token = tokenizer.eos_token
       tokenizer.padding_side = "right"
   
       model = AutoModelForCausalLM.from_pretrained(
           script_args.model_id,
           torch_dtype=torch.bfloat16,
           trust_remote_code=True,
           attn_implementation="flash_attention_2",
       )
   
       peft_config = LoraConfig(
           r=peft_args.lora_r,
           lora_alpha=peft_args.lora_alpha,
           lora_dropout=peft_args.lora_dropout,
           bias="none",
           task_type="CAUSAL_LM",
           target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
       )
   
       model = get_peft_model(model, peft_config)
   
       data_splits = load_dataset(script_args.dataset_name)
   
       dataset = data_splits["train"]
       eval_dataset = data_splits["test"]
   
       if script_args.dataset_subset_size is not None:
           dataset = dataset.select(range(script_args.dataset_subset_size))
   
       dataset = dataset.shuffle(seed=training_args.seed)
   
       trainer = SFTTrainer(
           model=model,
           args=training_args,
           train_dataset=dataset,
           eval_dataset=eval_dataset,
           formatting_func=formatting_prompts_func,
           processing_class=tokenizer,
       )
   
       trainer.train()
   
       dist.barrier()
       if trainer.is_world_process_zero():
           best_model_path = trainer.state.best_model_checkpoint
   
           final_model_dir = os.path.join(training_args.output_dir, "final_best_model")
           print(f"Copying best model to: {final_model_dir}")
   
           if os.path.exists(final_model_dir):
               shutil.rmtree(final_model_dir)
           shutil.copytree(best_model_path, final_model_dir)
   
           if script_args.run_inference_after_training:
               del model, trainer
               torch.cuda.empty_cache()
               run_post_training_inference(script_args, final_model_dir, tokenizer)
   
   def run_post_training_inference(script_args, best_model_path, tokenizer):
       print("\n" + "="*50)
       print("=== RUNNING POST-TRAINING INFERENCE TEST ===")
       print("="*50 + "\n")
   
       base_model = AutoModelForCausalLM.from_pretrained(
           script_args.model_id,
           torch_dtype=torch.bfloat16,
           trust_remote_code=True,
           device_map="auto"
       )
       model = PeftModel.from_pretrained(base_model, best_model_path)
       model = model.merge_and_unload()
       model.eval()
   
       # Define the test case
       schema = "CREATE TABLE artists (Name TEXT, Country TEXT, Genre TEXT)"
       system_message = "You are a text to SQL query translator. Users will ask you questions in English and you will generate a SQL query based on the provided SCHEMA."
       question = "Show me all artists from the Country just north of the USA."
   
       prompt = f"{system_message}\n\n### SCHEMA:\n{schema}\n\n### USER QUERY:\n{question}\n\n### SQL QUERY:\n"
   
       print(f"Test Prompt:\n{prompt}")
   
       inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
   
       print("\n--- Generating SQL... ---")
       outputs = model.generate(
           **inputs,
           max_new_tokens=100,
           pad_token_id=tokenizer.eos_token_id,
           do_sample=False,
           temperature=None,
           top_p=None,
       )
   
       generated_sql = tokenizer.decode(outputs[0], skip_special_tokens=True)[len(prompt):].strip()
   
       print(f"\n--- Generated SQL Query ---")
       print(generated_sql)
       print("\n" + "="*50)
       print("=== INFERENCE TEST COMPLETE ===")
       print("="*50 + "\n")
   
   if __name__ == "__main__":
       main()
   

4. To specify the tasks for the jobs to run on your Slurm cluster, create the train-mixtral.sh file with the following content:

   #!/bin/bash
   #SBATCH --job-name=mixtral-fsdp
   #SBATCH --nodes=2
   #SBATCH --ntasks-per-node=8
   #SBATCH --gpus-per-node=8
   #SBATCH --partition=a4high
   #SBATCH --output=mixtral-%j.out
   #SBATCH --error=mixtral-%j.err
   
   set -e
   set -x
   
   echo "--- Slurm Job Started ---"
   
   # --- Define Paths ---
   LOCAL_SSD_PATH="/mnt/localssd/job_${SLURM_JOB_ID}"
   VENV_PATH="${HOME}/.venv/venv-fsdp"
   MODEL_PATH="${HOME}/Mixtral-8x7B-v0.1"
   
   # --- STAGE 1: Stage Data to Local SSD on Each Node ---
   srun --ntasks=$SLURM_NNODES --ntasks-per-node=1 bash -c "
   echo '--- Staging on node: $(hostname) ---'
   mkdir -p ${LOCAL_SSD_PATH}
   
   echo 'Copying virtual environment...'
   rsync -a -q ${VENV_PATH}/ ${LOCAL_SSD_PATH}/venv/
   
   echo 'Copying model weights...'
   rsync -a ${MODEL_PATH}/ ${LOCAL_SSD_PATH}/model/
   
   mkdir -p ${LOCAL_SSD_PATH}/hf_cache
   
   echo '--- Staging on $(hostname) complete ---'
   "
   echo "--- Staging complete on all nodes ---"
   
   # --- STAGE 2: Run the Training Job ---
   echo "--- Launching Distributed Training with GIB NCCL Plugin ---"
   nodes=( $( scontrol show hostnames "$SLURM_JOB_NODELIST" ) )
   head_node=${nodes[0]}
   head_node_ip=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address)
   
   export MASTER_ADDR=$head_node_ip
   export MASTER_PORT=29500
   
   export NCCL_SOCKET_IFNAME=enp0s19
   
   export NCCL_NET=gIB
   
   # export NCCL_DEBUG=INFO # Un-comment to diagnose NCCL issues if needed
   
   srun --cpu-bind=none --accel-bind=g bash -c '
   # Activate the environment from the local copy
   source '${LOCAL_SSD_PATH}'/venv/bin/activate
   
   # Point Hugging Face cache to the local SSD
   export HF_HOME='${LOCAL_SSD_PATH}'/hf_cache
   
   export RANK=$SLURM_PROCID
   export WORLD_SIZE=$SLURM_NTASKS
   export LOCAL_RANK=$SLURM_LOCALID
   
   export LD_LIBRARY_PATH=/usr/local/gib/lib64:$LD_LIBRARY_PATH
   source /usr/local/gib/scripts/set_nccl_env.sh
   
   # --- Launch the training ---
   python \
       '${SLURM_SUBMIT_DIR}'/train-mixtral.py \
       --model_id="'${LOCAL_SSD_PATH}'/model/" \
       --output_dir="${HOME}/outputs/mixtral_job_${SLURM_JOB_ID}" \
       --dataset_name="philschmid/gretel-synthetic-text-to-sql" \
       --seed=900913 \
       --bf16=True \
       --num_train_epochs=3 \
       --per_device_train_batch_size=32 \
       --gradient_accumulation_steps=4 \
       --learning_rate=4e-5 \
       --logging_steps=3 \
       --lora_r=32 \
       --lora_alpha=32 \
       --lora_dropout=0.05 \
       --eval_strategy=steps \
       --eval_steps=10 \
       --save_strategy=steps \
       --save_steps=10 \
       --load_best_model_at_end=False \
       --metric_for_best_model=eval_loss \
       --run_inference_after_training \
       --dataset_subset_size=67000
   '
   
   # --- STAGE 3: Cleanup ---
   echo "--- Cleaning up local SSD on all nodes ---"
   srun --ntasks=$SLURM_NNODES --ntasks-per-node=1 bash -c "rm -rf ${LOCAL_SSD_PATH}"
   
   echo "--- Slurm Job Finished ---"
   

Upload Scripts to the Slurm cluster

To upload the scripts that you created in the previous section to the Slurm cluster, do the following:

1. To identify your login node, list all of the VMs in your project:

   gcloud compute instances list
   

   The name of the login node is similar to a4-high-login-001.

2. Upload your scripts to the login node's home directory:

   # Run this from your local machine where you created the files
   LOGIN_NODE_NAME="your-login-node-name" # e.g., a4high-login-001
   PROJECT_ID="your-gcp-project-id"
   ZONE="your-cluster-zone" # e.g., us-west4-a
   
   gcloud compute scp --project="$PROJECT_ID" --zone="$ZONE" --tunnel-through-iap \
     ./install_environment.sh \
     ./requirements-fsdp.txt \
     ./train-mixtral.py \
     ./train-mixtral.sh \
     "${LOGIN_NODE_NAME}":~/
   

Connect to the Slurm cluster

Connect to the Slurm cluster by connecting to the login node through SSH:

gcloud compute ssh $LOGIN_NODE_NAME \
    --project=$PROJECT_ID \
    --tunnel-through-iap \
    --zone=$ZONE

Install frameworks and tools

After you connect to the login node, install frameworks and tools.

1. Export your Hugging Face token:

   # On the login node
   export HF_TOKEN="hf_..." # Replace with your token
   

2. Run the installation script on a compute node.

   # On the login node
   srun \
     --job-name=env-setup \
     --nodes=1 \
     --ntasks=1 \
     --gpus-per-node=1 \
     --partition=a4high \
     bash ./install_environment.sh
   

   This command sets up the virtual environment, installs all dependencies, and downloads the Mixtral model weights into ~/Mixtral-8x7B-v0.1. This process can take over 30 minutes to complete.

Start your fine-tuning workload

To start your training your workload, do the following:

1. Submit the job to the Slurm scheduler:

   # On the login node
   sbatch train-mixtral.sh
   

2. On the login node in your Slurm cluster, you can monitor the job's progress by checking the output files created in your home directory:

   # On the login node
   tail -f mixtral-*.out
   

   If your job successfully starts, then the .err file shows a progress bar that updates as your job progresses.

   The job has two main phases:

   • Copying the large base model to the local SSD of each compute node.
   • The training job, which begins once the copying of the model is complete.

   The whole job takes about 40 minutes to run.