---

name: configuration-and-cli description: "Use this sub-skill when working with Hugging Face Accelerate configuration files and CLI commands, including accelerate config/default/update/env/launch/test/estimate-memory/merge-weights/to-fsdp2, launch command construction, multi-node and SLURM planning, config validation, and CLI troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Accelerate Configuration and CLI

Use this sub-skill to create, validate, update, and reason about Accelerate config files and command-line invocations.

When to Use

• Build or debug accelerate launch commands for scripts, modules, non-Python executables, multi-node jobs, or SLURM wrappers.
• Create or adapt default_config.yaml or custom --config_file YAML for local CPU, single accelerator, multi-GPU, multi-XPU, or multi-node runs.
• Explain or safely run accelerate config, config default, config update, env, test, estimate-memory, merge-weights, or to-fsdp2.
• Diagnose missing configs, invalid YAML, unknown keys, conflicting launcher flags, rendezvous mistakes, subprocess tracebacks, and training-script argument separation.

Routing Boundaries

• For training-loop code changes such as Accelerator.prepare, dataloaders, gradient accumulation, trackers, and checkpoint hooks, use ../training-loop-integration/.
• For DeepSpeed, FSDP, Megatron-LM, TPU, and backend-specific semantics beyond CLI flag placement, use ../distributed-training-backends/.
• For estimate-memory interpretation and big-model device-map loading workflows, use ../big-model-inference/.

Start Here

1. Read references/cli-reference.md for command inventory, safe command patterns, and which commands mutate files.
2. Read references/configuration.md before editing YAML or choosing top-level config keys.
3. Read references/launching.md before composing accelerate launch, multi-node, module, no-Python, or SLURM commands.
4. Read references/troubleshooting.md when a config or launch command fails.

Bundled Helpers

• scripts/validate_accelerate_config.py: validates YAML syntax, known top-level keys, required fields, and common local/multi-node conflicts without importing Accelerate.
• scripts/print_accelerate_help_summary.py: invokes installed accelerate --help subcommands when available and prints a compact availability summary.

Run helpers with python path/to/script.py --help for usage.