|
| 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 | +# MiniMax-M2 Conversion Round-Trip Verification (Multi-Node via Slurm) |
| 18 | +# |
| 19 | +# MiniMax-M2 (MoE: 256 experts, top-8, ~230GB fp8) |
| 20 | +# The full model OOMs on a single 8-GPU node — minimum 2 nodes (16 GPUs) |
| 21 | +# with EP >= 8. |
| 22 | +# |
| 23 | +# Sweeps multiple parallelism configs (TP,PP,EP) to verify HF <-> Megatron |
| 24 | +# round-trip conversion. Each config runs sequentially. |
| 25 | +# |
| 26 | +# Usage: |
| 27 | +# 1. Fill in CONTAINER_IMAGE, CONTAINER_MOUNTS, and token exports |
| 28 | +# 2. Adjust PARALLELISM_CONFIGS if needed |
| 29 | +# 3. Submit: sbatch slurm_conversion.sh |
| 30 | +# ============================================================================== |
| 31 | + |
| 32 | +#SBATCH --job-name=minimax-m2-roundtrip |
| 33 | +#SBATCH --nodes=2 |
| 34 | +#SBATCH --ntasks-per-node=8 |
| 35 | +#SBATCH --gpus-per-node=8 |
| 36 | +#SBATCH --time=4:00:00 |
| 37 | +#SBATCH --account=<your-account> |
| 38 | +#SBATCH --partition=batch |
| 39 | +#SBATCH --output=logs/minimax_m2_roundtrip_%j.log |
| 40 | +#SBATCH --exclusive |
| 41 | + |
| 42 | +# ── Container ──────────────────────────────────────────────────────────── |
| 43 | +CONTAINER_IMAGE="" |
| 44 | +# CONTAINER_IMAGE="/path/to/container.sqsh" |
| 45 | +CONTAINER_MOUNTS="" |
| 46 | +# CONTAINER_MOUNTS="/path/to/shared/storage:/mnt/storage,/path/to/project:/opt/Megatron-Bridge" |
| 47 | +WORKDIR="/opt/Megatron-Bridge" |
| 48 | + |
| 49 | +# ── Tokens / Caches ────────────────────────────────────────────────────── |
| 50 | +# export HF_TOKEN="hf_your_token_here" |
| 51 | +# export HF_HOME="/path/to/shared/HF_HOME" |
| 52 | +# export UV_CACHE_DIR="/path/to/shared/uv_cache" |
| 53 | + |
| 54 | +# ── Parallelism configs: "TP,PP,EP" per entry ──────────────────────────── |
| 55 | +# TP*PP*EP must equal total GPUs (NODES * GPUS_PER_NODE). |
| 56 | +# EP must divide 256 (number of experts). |
| 57 | +# Note: TP=2,EP=4 on 8 GPUs (1 node) OOMs — use EP >= 8 with 2+ nodes. |
| 58 | +PARALLELISM_CONFIGS=("2,1,8" "1,2,8" "2,2,4") |
| 59 | + |
| 60 | +# ── Model ───────────────────────────────────────────────────────────────── |
| 61 | +MODEL_NAME=MiniMax-M2 |
| 62 | +HF_MODEL_ID=MiniMaxAI/$MODEL_NAME |
| 63 | + |
| 64 | +# ── Environment ─────────────────────────────────────────────────────────── |
| 65 | +export TORCH_NCCL_AVOID_RECORD_STREAMS=1 |
| 66 | +export NCCL_NVLS_ENABLE=0 |
| 67 | + |
| 68 | +# ============================================================================== |
| 69 | +# Job Execution |
| 70 | +# ============================================================================== |
| 71 | + |
| 72 | +echo "======================================" |
| 73 | +echo "MiniMax-M2 Round-Trip Conversion Sweep" |
| 74 | +echo "Job: $SLURM_JOB_ID | Nodes: $SLURM_JOB_NUM_NODES" |
| 75 | +echo "Parallelism configs: ${PARALLELISM_CONFIGS[*]}" |
| 76 | +echo "======================================" |
| 77 | + |
| 78 | +mkdir -p logs |
| 79 | + |
| 80 | +if [ -z "$CONTAINER_IMAGE" ]; then |
| 81 | + echo "ERROR: CONTAINER_IMAGE must be set." |
| 82 | + exit 1 |
| 83 | +fi |
| 84 | + |
| 85 | +SRUN_CMD="srun --mpi=pmix --container-image=$CONTAINER_IMAGE" |
| 86 | +if [ -n "$CONTAINER_MOUNTS" ]; then |
| 87 | + SRUN_CMD="$SRUN_CMD --container-mounts=$CONTAINER_MOUNTS" |
| 88 | +fi |
| 89 | + |
| 90 | +CONFIG_INDEX=0 |
| 91 | +for CONFIG in "${PARALLELISM_CONFIGS[@]}"; do |
| 92 | + IFS=',' read -r TP PP EP <<< "$CONFIG" |
| 93 | + CONFIG_INDEX=$((CONFIG_INDEX + 1)) |
| 94 | + |
| 95 | + echo "" |
| 96 | + echo "======================================" |
| 97 | + echo "Config $CONFIG_INDEX/${#PARALLELISM_CONFIGS[@]}: TP=$TP, PP=$PP, EP=$EP" |
| 98 | + echo "======================================" |
| 99 | + |
| 100 | + # Sync dependencies once per node, then run the roundtrip |
| 101 | + CMD="if [ \"\$SLURM_LOCALID\" -eq 0 ]; then uv sync; else sleep 10; fi && " |
| 102 | + CMD="${CMD}uv run --no-sync python examples/conversion/hf_megatron_roundtrip_multi_gpu.py" |
| 103 | + CMD="$CMD --hf-model-id $HF_MODEL_ID" |
| 104 | + CMD="$CMD --tp $TP --pp $PP --ep $EP" |
| 105 | + CMD="$CMD --trust-remote-code" |
| 106 | + |
| 107 | + echo "Executing: $CMD" |
| 108 | + |
| 109 | + $SRUN_CMD bash -c "cd $WORKDIR && $CMD" |
| 110 | + RUN_EXIT=$? |
| 111 | + if [ $RUN_EXIT -ne 0 ]; then |
| 112 | + echo "ERROR: Config TP=$TP, PP=$PP, EP=$EP failed (exit $RUN_EXIT)" |
| 113 | + exit $RUN_EXIT |
| 114 | + fi |
| 115 | + echo "[OK] Config $CONFIG_INDEX: TP=$TP, PP=$PP, EP=$EP passed" |
| 116 | +done |
| 117 | + |
| 118 | +echo "" |
| 119 | +echo "======================================" |
| 120 | +echo "All ${#PARALLELISM_CONFIGS[@]} configs passed" |
| 121 | +echo "======================================" |
0 commit comments