Use FSDP to fine-tune Llama 4 on an A4 Slurm cluster

Access Llama 4 by using Hugging Face

To use Hugging Face to access Llama 4, do the following:

1. Sign the consent agreement to use Llama 4.

2. Create a Hugging Face read access token.

   Click Your Profile > Settings > Access tokens > +Create new token

3. Copy and save the read access token value. You use it later in this tutorial.

Prepare your environment

To prepare your environment, follow these steps:

1. Clone the Cluster Toolkit GitHub repository:

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

2. Create a Cloud Storage bucket:

   gcloud storage buckets create gs://BUCKET_NAME \
       --project=PROJECT_ID
   

   Replace the following:

   • BUCKET_NAME: a name for your Cloud Storage bucket that follows bucket naming requirements.

   • PROJECT_ID: the ID of the Google Cloud project where you want to create your Cloud Storage bucket.

Create an A4 Slurm cluster

To create an A4 Slurm cluster, follow these steps:

1. Go to the 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:

   cd examples/machine-learning/a4-highgpu-8g/
   

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

   terraform_backend_defaults:
     type: gcs
     configuration:
       bucket: BUCKET_NAME
   
   vars:
     deployment_name: a4-high
     project_id: PROJECT_ID
     region: REGION
     zone: ZONE
     a4h_cluster_size: 2
     a4h_reservation_name: RESERVATION_URL
   

   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.

   • RESERVATION_URL: the URL of the reservation that you want to use to create your Slurm cluster. Based on the project in which the reservation exists, specify one of the following values:

     • The reservation exists in your project: RESERVATION_NAME

     • The reservation exists in a different project, and your project can use the reservation: projects/RESERVATION_PROJECT_ID/reservations/RESERVATION_NAME

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 is 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.

   If the first phase succeeds but the second phase fails, then you can try to deploy the Slurm cluster again by skipping the first phase:

   ./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 --skip "image" -w
   

Prepare your workload

To prepare your workload, you do the following:

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 up a consistent environment for FSDP training.
   # It is meant to be run once on the login node of your Slurm cluster
   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
   
   # --- 3. Download the Model ---
   echo "--- [STEP 2.4] Downloading Llama4 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 meta-llama/Llama-4-Scout-17B-16E-Instruct --local-dir ~/Llama-4-Scout-17B-16E-Instruct --token $HF_TOKEN
   
   echo "--- Environment setup complete. ---"
   

   This script sets up a reliable Python virtual environment, installs a PyTorch nightly build, and downloads the Llama 4 model.

2. To specify the Python dependencies for the training script, create a requirements-fsdp.txtfile 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
   

3. Specify llama4-train-distributed.py as the main training script:

   import torch
   from datasets import load_dataset
   from peft import LoraConfig, PeftModel
   from transformers import (
       AutoModelForCausalLM,
       AutoTokenizer,
       TrainingArguments,
       HfArgumentParser,
   )
   
   from torch.distributed import get_rank, get_world_size
   
   from transformers.models.llama4.modeling_llama4 import Llama4TextDecoderLayer
   from trl import SFTTrainer
   from dataclasses import dataclass, field
   from typing import Optional
   
   @dataclass
   class ScriptArguments:
       model_id: str = field(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=True, metadata={"help": "When using FSDP activation checkpointing, this must be set to True"})
   
   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": False}
   
       training_args.optim = "adamw_torch_fused"
   
       training_args.fsdp = "full_shard"
       training_args.fsdp_config = {
           "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
           "fsdp_transformer_layer_cls_to_wrap": [Llama4TextDecoderLayer],
           "fsdp_state_dict_type": "FULL_STATE_DICT",
           "fsdp_offload_params": False,
           "fsdp_forward_prefetch": True,
       }
   
       tokenizer = AutoTokenizer.from_pretrained(script_args.model_id, trust_remote_code=True)
   
       model = AutoModelForCausalLM.from_pretrained(
           script_args.model_id,
           torch_dtype=torch.bfloat16,
           trust_remote_code=True,
           attn_implementation="sdpa",
       )
   
       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"],
       )
       rank = get_rank()
       world_size = get_world_size()
   
       dataset = load_dataset(script_args.dataset_name, split="train")
   
       if script_args.dataset_subset_size is not None:
           dataset = dataset.select(range(script_args.dataset_subset_size))
       else:
           print(f"Using the full dataset with {len(dataset)} samples.")
   
       dataset = dataset.shuffle(seed=training_args.seed)
       print(f"Dataset shuffled with seed: {training_args.seed}.")
   
       if world_size > 1:
           print(f"Sharding dataset for Rank {rank} of {world_size}.")
           dataset = dataset.shard(num_shards=world_size, index=rank)
   
       print("Initializing SFTTrainer...")
       trainer = SFTTrainer(
           model=model,
           args=training_args,
           train_dataset=dataset,
           peft_config=peft_config,
           formatting_func=formatting_prompts_func,
           processing_class=tokenizer,
       )
   
       trainer.train()
   
       trainer.save_model(training_args.output_dir)
   
       if script_args.run_inference_after_training and trainer.is_world_process_zero():
           del model
           del trainer
           torch.cuda.empty_cache()
           run_post_training_inference(script_args, training_args, tokenizer)
   
   def run_post_training_inference(script_args, training_args, tokenizer):
       """
       Loads the fine-tuned PEFT adapter from the local output directory and runs inference.
       This should only be called on rank 0 after training is complete.
       """
       print("\n" + "="*50)
       print("=== RUNNING POST-TRAINING INFERENCE TEST ===")
       print("="*50 + "\n")
   
       # Load the base model and merge the adapter.
       base_model = AutoModelForCausalLM.from_pretrained(
           script_args.model_id,
           torch_dtype=torch.bfloat16,
           trust_remote_code=True,
           device_map="auto"
       )
       # Load the PEFT adapter and merge it into the base model
       model = PeftModel.from_pretrained(base_model, training_args.output_dir)
       model = model.merge_and_unload() # Merge weights for faster inference
       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."
   
       # This must match the formatting_func exactly
       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()
   

   This script utilizes the TRL Supervised Fine-Tuning (SFT) Trainer to manage FSDP training loops, Low-Rank Adaptation (LoRA) configuration, and data formatting.

4. To specify the tasks for the jobs to run on your Slurm cluster, create the submit.slurm file with the following content:

   #!/bin/bash
   #SBATCH --job-name=llama4-fsdp-fixed
   #SBATCH --nodes=2
   #SBATCH --ntasks-per-node=8
   #SBATCH --gpus-per-node=8
   #SBATCH --partition=a4high
   #SBATCH --output=llama4-%j.out
   #SBATCH --error=llama4-%j.err
   
   set -e
   set -x
   
   echo "--- Slurm Job Started ---"
   echo "Job ID: $SLURM_JOB_ID"
   echo "Node List: $SLURM_JOB_NODELIST"
   
   # --- Define Paths ---
   LOCAL_SSD_PATH="/mnt/localssd/job_${SLURM_JOB_ID}"
   VENV_PATH="${HOME}/.venv/venv-fsdp"
   MODEL_PATH="${HOME}/Llama-4-Scout-17B-16E-Instruct"
   
   # --- 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 --info=progress2 ${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 \
       '${SLURM_SUBMIT_DIR}'/llama4-train-distributed.py \
         --model_id="'${LOCAL_SSD_PATH}'/model/" \
         --output_dir="'${LOCAL_SSD_PATH}'/outputs/" \
         --dataset_name="philschmid/gretel-synthetic-text-to-sql" \
         --seed=900913 \
         --bf16=True \
         --num_train_epochs=1 \
         --per_device_train_batch_size=2 \
         --gradient_accumulation_steps=4 \
         --learning_rate=2e-5 \
         --logging_steps=10 \
         --lora_r=16 \
         --lora_alpha=32 \
         --lora_dropout=0.05 \
         --run_inference_after_training
   '
   
   # --- STAGE 3: Copy Final Results Back to Persistent Storage ---
   echo "--- Copying final results from local SSD to shared storage ---"
   PERSISTENT_OUTPUT_DIR="${HOME}/outputs/llama4_job_${SLURM_JOB_ID}"
   mkdir -p "$PERSISTENT_OUTPUT_DIR"
   
   # Only copy from the head node where trl has combined the results
   srun --nodes=1 --ntasks=1 -w "$head_node" \
     rsync -a --info=progress2 "${LOCAL_SSD_PATH}/outputs/" "${PERSISTENT_OUTPUT_DIR}/"
   
   # --- STAGE 4: 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, follow these steps:

1. To identify your login node, list all A4 VMs in your project:

   gcloud compute instances list --filter="machineType:a4-highgpu-8g"
   

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

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

   gcloud compute scp --project="$PROJECT_ID" --zone="$ZONE" --tunnel-through-iap \
     ./install_environment.sh \
     ./requirements-fsdp.txt \
     ./llama4-train-distributed.py \
     ./submit.slurm \
     "${LOGIN_NODE_NAME}":~/
   

   Replace LOGIN_NODE_NAME with the name of the login node.

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 by doing the following:

1. Export your Hugging Face token:

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

2. Run the installation script:

   # On the login node
   chmod +x install_environment.sh
   ./install_environment.sh
   

   This command sets up a virtual environment with all the required dependencies, and downloads the model weights into the ~/Llama-4-Scout-17B-16E-Instruct file.

   Because the model download is very large (~200 GB), this process takes around 30 minutes, depending on network conditions.

Start your fine-tuning workload

To start training your workload, do the following:

1. Submit the job to the Slurm scheduler:

   sbatch submit.slurm
   

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 llama4-*.out
   

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

   This job should take a bit over an hour to complete on the Slurm Cluster. 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. This job takes about 35 minutes to run.