accmt-cli 1.2.0.2

CLI for AcceleratorModule library (accmt).

AcceleratorModule CLI

CLI for AcceleratorModule library (accmt).

This is a command-line tool wrapper around Accelerate's command-line utilities ('accelerate').

Installation

accmt-cli is automatically installed when installing accmt library. You also install it via pip:

pip install accmt-cli

Launch

You can launch any distributed training process with the following command:

accmt launch [-N][-n, --gpus][-O1][--strat] <your_python_script> [...]

Where:

• -N (optional): Corresponds to the number of processes, or a Python-like slice to take GPUs from a certain index (e.g. '-N=2:', to take GPUs from index 2).
• -n or --gpus (optional): Corresponds to a list of CUDA devices (e.g. '-n=1,3,5,6', to take GPUs indices 1, 3, 5 and 6).
• -O1 (optional): Corresponds to the optimization of type 1, which calculates the efficient number for 'OMP_NUM_THREADS', depending on how many processes you will run you training script.
• --strat (optional): Corresponds to the specific strategy to implement, or a configuration file path from Accelerate ('accelerate config --config_file=your-config.yaml'). See 'accmt strats' for specific strategies. ... (optional): You can add here any additional arguments that your Python script might have.

Get model from checkpoint

You can get a model from any checkpoint using the following command:

accmt get <checkpoint> --out=<output-model-directory> [--dtype]

Where:

• --out or -O (REQUIRED): Output model directory name where to save the model.
• --dtype (Optional): PyTorch data type of model parameters. Default is 'float32'.

Strats

You can check the specific strats included with the following command:

accmt strats [--ddp][--fsdp][--deepspeed]

Where:

• --ddp: To only filter for DDP strategies.
• --fsdp: To only filter for FSDP strategies.
• --deepspeed: To only filter for DeepSpeed strategies.

Example

Generate an example HPS file config with the following command:

accmt example

This will generate a file on your current directory called 'hps_example.yaml'.