Megatron-DeepSpeed¶

Megatron-DeepSpeed is a scalable, highly performant library for training large language models on any GPU¹.

In particular, it retains the core 4D parallelism² functionality of the NVIDIA / Megatron-LM library, while leveraging the microsoft / DeepSpeed library for efficient scaling and 🍋 saforem2 / ezpz for automated device + backend selection.

Getting Started¶

1. Clone the argonne-lcf / Megatron-DeepSpeed repository:

   git clone https://github.com/argonne-lcf/Megatron-DeepSpeed
   cd Megatron-DeepSpeed
   

2. Set up your environment:

   export PBS_O_WORKDIR=$(pwd)
   source <(curl -s https://raw.githubusercontent.com/saforem2/ezpz/refs/heads/main/src/ezpz/bin/utils.sh)
   ezpz_setup_env
   

   [Optional] Setup WandB

   To enable Weights & Biases (WandB) logging, we need to install and login:

   python3 -m pip install wandb --upgrade
   wandb login
   

   NOTE: WandB can be disabled by setting export WANDB_DISABLED=1

   See wandb: Quickstart for additional information

3. Install dependencies:

   1. 🍋 saforem2 / ezpz:

   python3 -m pip install -e "git+https://github.com/saforem2/ezpz#egg=ezpz" --require-virtualenv
   

   1. microsoft / DeepSpeed:

   python3 -m pip install deepspeed
   

4. Launch training:

   # Before launching, `PBS_O_WORKDIR` should be set to Megatron-DeepSpeed's PATH
   # and venv inside Megatron-DeepSpeed/venv should be activated.
   TP=2 NLAYERS=10 DATA_FILE_LIST=ALCF/data-lists/aurora/books.txt bash train_aGPT_7B.sh
   

   This will launch a distributed pre-training run with:

   • NLAYERS=10: Llama style model consisting of 10 layers

   • TP=2: Split across 2 Tensor Parallel groups

   • DATA_FILE_LIST: Using the Books corpus of the Dolma dataset

   Overridable Options

   This is a simple subset of the overridable options.

   The full list (as well as their default values) can be found in ALCF / helpers.sh

   • DTYPE: Data type
   • DATA_FILE_LIST: Data file list
   • FFN_HIDDEN_SIZE: Feedforward Neural Network projection size
   • GRAD_ACC_STEPS: Gradient accumulation steps
   • HEADS: Number of attention heads
   • HIDDEN: Hidden size
   • MICRO_BATCH: Micro batch size
   • NO_FLASH_ATTN: No Flash Attention
   • NLAYERS: Number of layers
   • NUM_KV_HEAD: Number of key-value heads
   • OPT: Optimizer
     • adam
     • adam8bit
     • adamw
     • adamwschedulefree
     • apex.adam
     • apex.sgd
     • ds.fusedlamb
     • ds.onebitlamb
     • galoreadamw
     • galoreadamw8bit
     • galoreadamw8bitperlayer
     • ipex.fusedlamb
     • ipex.lamb
     • shampoo
     • sgd
     • sgdschedulefree
     • sophiag
   • PP: Pipeline parallelism degree
   • SEQ: Sequence length
   • SP: Sequence parallelism (Ulysses) degree
   • TP: Tensor parallelism degree
   • TRAIN_TOKENS: Number of training tokens
   • TRAIN_ITERS: Number of training iterations
   • USE_ACTIVATION_CHECKPOINTING: Use activation checkpointing
   • WEIGHT_DECAY: Weight decay
   • ZERO_STAGE: Zero stage