|
| 1 | +#!/bin/bash |
| 2 | +# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved. |
| 3 | +# |
| 4 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | +# you may not use this file except in compliance with the License. |
| 6 | +# You may obtain a copy of the License at |
| 7 | +# |
| 8 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | +# |
| 10 | +# Unless required by applicable law or agreed to in writing, software |
| 11 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | +# See the License for the specific language governing permissions and |
| 14 | +# limitations under the License. |
| 15 | + |
| 16 | +# ============================================================================== |
| 17 | +# Nemotron 3 Nano Parameter-Efficient Fine-Tuning (PEFT) with LoRA |
| 18 | +# |
| 19 | +# Nemotron 3 Nano is a 30B parameter model with A3B (Active 3 Billion) architecture |
| 20 | +# LoRA/DoRA significantly reduces memory requirements |
| 21 | +# Supports multiple parallelism configs: each "TP,PP,EP,CP,SP" runs sequentially. |
| 22 | +# |
| 23 | +# Usage: |
| 24 | +# 1. Modify the #SBATCH directives below for your cluster |
| 25 | +# 2. Set CONTAINER_IMAGE to your container path |
| 26 | +# 3. Set PARALLELISM_CONFIGS (TP,PP,EP,CP,SP per entry; CP = context parallel size, 1 = disabled) |
| 27 | +# 4. Submit: sbatch slurm_peft.sh |
| 28 | +# ============================================================================== |
| 29 | + |
| 30 | +#SBATCH --job-name=nemotron3-lora |
| 31 | +#SBATCH --nodes=2 |
| 32 | +#SBATCH --ntasks-per-node=8 |
| 33 | +#SBATCH --gpus-per-node=8 |
| 34 | +#SBATCH --time=08:00:00 |
| 35 | +#SBATCH --partition=gpu |
| 36 | +#SBATCH --account=my_account |
| 37 | +#SBATCH --output=logs/nemotron3_lora_%j.out |
| 38 | +#SBATCH --error=logs/nemotron3_lora_%j.err |
| 39 | +#SBATCH --exclusive |
| 40 | + |
| 41 | +# ============================================================================== |
| 42 | +# CONFIGURATION |
| 43 | +# ============================================================================== |
| 44 | + |
| 45 | +# Workspace directory for checkpoints and results |
| 46 | +WORKSPACE=${WORKSPACE:-/workspace} |
| 47 | + |
| 48 | +# Model and training configurations |
| 49 | +PRETRAINED_CHECKPOINT=${WORKSPACE}/models/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 |
| 50 | +MODEL_NAME=nemotron_3_nano |
| 51 | +DATASET_NAME=squad |
| 52 | +SEQ_LENGTH=2048 |
| 53 | +TRAIN_ITERS=50 |
| 54 | +GLOBAL_BATCH_SIZE=8 |
| 55 | +MICRO_BATCH_SIZE=1 |
| 56 | +EVAL_ITERS=10 |
| 57 | +LR_WARMUP_ITERS=5 |
| 58 | +LOG_INTERVAL=1 |
| 59 | +WANDB_PROJECT=megatron-bridge-${DATASET_NAME} |
| 60 | +GLOBAL_BATCH_SIZE=16 |
| 61 | + |
| 62 | +# Parallelism configs: "TP,PP,EP,CP,SP" per entry |
| 63 | +PARALLELISM_CONFIGS=("4,1,8,1,True" "2,2,8,1,True" "2,1,8,2,True") |
| 64 | + |
| 65 | +# Container image (required) |
| 66 | +CONTAINER_IMAGE="" |
| 67 | +# CONTAINER_IMAGE="/path/to/container.sqsh" |
| 68 | + |
| 69 | +# Container mounts (optional, space-separated) |
| 70 | +CONTAINER_MOUNTS="" |
| 71 | +# CONTAINER_MOUNTS="/data:/data /workspace:/workspace" |
| 72 | + |
| 73 | +# ============================================================================== |
| 74 | +# Environment Setup |
| 75 | +# ============================================================================== |
| 76 | + |
| 77 | +# NCCL optimizations for large-scale training |
| 78 | +export TORCH_NCCL_AVOID_RECORD_STREAMS=1 |
| 79 | +export NCCL_NVLS_ENABLE=0 |
| 80 | + |
| 81 | +# UV cache on shared filesystem (recommended for multi-node setups) |
| 82 | +# Pre-sync once before submitting jobs: UV_CACHE_DIR=/path/to/cache uv sync |
| 83 | +# export UV_CACHE_DIR="/path/to/shared/uv_cache" |
| 84 | + |
| 85 | +# HuggingFace cache directory (recommended for shared filesystem) |
| 86 | +# export HF_HOME="/path/to/shared/HF_HOME" |
| 87 | + |
| 88 | +# Authentication tokens (set these for your environment) |
| 89 | +# export HF_TOKEN="hf_your_token_here" |
| 90 | +# export WANDB_API_KEY="your_wandb_key_here" |
| 91 | + |
| 92 | +# ============================================================================== |
| 93 | +# Job Execution |
| 94 | +# ============================================================================== |
| 95 | + |
| 96 | +echo "======================================" |
| 97 | +echo "Nemotron 3 Nano LoRA Fine-Tuning Job" |
| 98 | +echo "======================================" |
| 99 | +echo "Job ID: $SLURM_JOB_ID" |
| 100 | +echo "Nodes: $SLURM_JOB_NUM_NODES" |
| 101 | +echo "GPUs per node: $SLURM_GPUS_PER_NODE" |
| 102 | +echo "Model: $MODEL_NAME" |
| 103 | +echo "Parallelism configs: ${PARALLELISM_CONFIGS[*]}" |
| 104 | +echo "PEFT: LoRA" |
| 105 | +echo "======================================" |
| 106 | + |
| 107 | +# Create logs directory if it doesn't exist |
| 108 | +mkdir -p logs |
| 109 | + |
| 110 | +# Require container image |
| 111 | +if [ -z "$CONTAINER_IMAGE" ]; then |
| 112 | + echo "ERROR: CONTAINER_IMAGE must be set. Please specify a valid container image." |
| 113 | + exit 1 |
| 114 | +fi |
| 115 | + |
| 116 | +# Build srun command (shared across configs) |
| 117 | +SRUN_CMD="srun --mpi=pmix --container-image=$CONTAINER_IMAGE" |
| 118 | +if [ -n "$CONTAINER_MOUNTS" ]; then |
| 119 | + SRUN_CMD="$SRUN_CMD --container-mounts=$CONTAINER_MOUNTS" |
| 120 | +fi |
| 121 | +echo "SRUN base: $SRUN_CMD" |
| 122 | +echo "======================================" |
| 123 | + |
| 124 | +# Run each parallelism config in sequence |
| 125 | +CONFIG_INDEX=0 |
| 126 | +for CONFIG in "${PARALLELISM_CONFIGS[@]}"; do |
| 127 | + IFS=',' read -r TP PP EP CP SP <<< "$CONFIG" |
| 128 | + CONFIG_INDEX=$((CONFIG_INDEX + 1)) |
| 129 | + echo "" |
| 130 | + echo "======================================" |
| 131 | + echo "Config $CONFIG_INDEX/${#PARALLELISM_CONFIGS[@]}: TP=$TP, PP=$PP, EP=$EP, SP=$SP, CP=$CP" |
| 132 | + echo "======================================" |
| 133 | + |
| 134 | + # Build CLI overrides for this config |
| 135 | + CLI_OVERRIDES=" \ |
| 136 | + checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \ |
| 137 | + train.train_iters=$TRAIN_ITERS \ |
| 138 | + train.global_batch_size=$GLOBAL_BATCH_SIZE \ |
| 139 | + train.micro_batch_size=$MICRO_BATCH_SIZE \ |
| 140 | + train.eval_iters=$EVAL_ITERS \ |
| 141 | + scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \ |
| 142 | + checkpoint.save=${WORKSPACE}/results/${MODEL_NAME}_lora_tp${TP}_pp${PP}_ep${EP}_sp${SP}_cp${CP} \ |
| 143 | + logger.log_interval=$LOG_INTERVAL \ |
| 144 | + logger.wandb_project=$WANDB_PROJECT \ |
| 145 | + logger.wandb_exp_name=${MODEL_NAME}_${DATASET_NAME}_lora_tp${TP}_pp${PP}_ep${EP}_sp${SP}_cp${CP} \ |
| 146 | + model.tensor_model_parallel_size=$TP \ |
| 147 | + model.pipeline_model_parallel_size=$PP \ |
| 148 | + model.expert_model_parallel_size=$EP \ |
| 149 | + model.sequence_parallel=$SP \ |
| 150 | + model.context_parallel_size=$CP \ |
| 151 | + model.calculate_per_token_loss=True \ |
| 152 | + train.global_batch_size=$GLOBAL_BATCH_SIZE \ |
| 153 | + dataset.packed_sequence_specs.pad_seq_to_mult=$((CP * 2)) \ |
| 154 | + dataset.packed_sequence_specs.packed_sequence_size=$SEQ_LENGTH \ |
| 155 | + dataset.seq_length=$SEQ_LENGTH \ |
| 156 | + model.seq_length=$SEQ_LENGTH |
| 157 | + " |
| 158 | + |
| 159 | + CMD="python /opt/Megatron-Bridge/scripts/training/run_recipe.py" |
| 160 | + CMD="$CMD --recipe ${MODEL_NAME}_finetune_config" |
| 161 | + CMD="$CMD --peft_scheme lora" |
| 162 | + # Collapse newlines so bash -c receives a single command |
| 163 | + CMD="$CMD $(echo "$CLI_OVERRIDES" | tr '\n' ' ' | sed 's/ \+/ /g')" |
| 164 | + |
| 165 | + echo "Executing command..." |
| 166 | + echo $CMD |
| 167 | + echo "======================================" |
| 168 | + |
| 169 | + |
| 170 | + |
| 171 | + $SRUN_CMD bash -c "$CMD" |
| 172 | + RUN_EXIT=$? |
| 173 | + if [ $RUN_EXIT -ne 0 ]; then |
| 174 | + echo "ERROR: Config TP=$TP, PP=$PP, EP=$EP, SP=$SP, CP=$CP failed with exit code $RUN_EXIT" |
| 175 | + exit $RUN_EXIT |
| 176 | + fi |
| 177 | +done |
| 178 | + |
| 179 | +echo "======================================" |
| 180 | +echo "Job completed (all ${#PARALLELISM_CONFIGS[@]} configs)" |
| 181 | +echo "======================================" |
0 commit comments