[ENHANCEMENT] Add a minimal "quickstart" example for single-GPU / CPU local runs

[nathon-lee]

Background

Megatron-LM and Megatron Core are geared toward large-scale, multi-GPU and multi-node training. While this is essential for production and research at scale, it raises the entry barrier for new contributors and developers who don't have access to large GPU clusters. A minimal, local quickstart that runs on a single GPU (and provides a CPU fallback) would make the project much more accessible: it enables quicker development cycles, easier PR testing for external contributors, and a fast sanity-check path in CI.

Proposed change

• Add an examples/quickstart/ directory containing:
  • train_small_gpt.py: a simplified training wrapper that runs a tiny GPT-like model (few layers, small hidden size) for a configurable number of steps.
  • configs/quickstart_small.yaml: minimal config for model size, dataset (synthetic/random data), optimizer, and runtime settings.
• Add a doc page docs/get-started/local_quickstart.md with step-by-step instructions for:
  • setting up a Python virtual environment,
  • installing only the minimal dependencies required for the quickstart (torch, numpy, etc.),
  • running the example on a single GPU and on CPU (fallback).
• Ensure the quickstart does not require TransformerEngine, FP8, or other optional binary dependencies. Use synthetic/random data so no dataset downloads are needed.
• Optionally, add a lightweight script or Makefile target to run the quickstart easily (e.g., tools/quickstart/run_quickstart.sh).

Acceptance criteria

• A user can follow docs/get-started/local_quickstart.md and run a full training loop locally on:
  • a single GPU (if available), completing in a short time (e.g., <10 minutes for default steps),
  • or on CPU (with a longer but still reasonable runtime for verification).
• The example uses synthetic/random input data and does not require external datasets or heavy optional dependencies.
• The new files are placed under examples/quickstart/ and the docs are linked from the main README.md under "Getting Started".
• Example is minimal, easy to read, and well-documented; it includes suggested command lines and expected lightweight output.

Implementation notes / suggestions

• train_small_gpt.py can be a thin wrapper that constructs a minimal model from transformer or a small custom torch.nn.Module that mimics the model shape used by Megatron modules. It should:
  • accept a --device flag (cuda or cpu) and a --steps flag for number of optimization steps,
  • simulate datasets using random tensors to avoid downloads,
  • log a couple of metrics (loss) to stdout so users can verify training progress.
• Keep the dependency list in docs minimal: torch, numpy, and pyyaml for configs. Mark transformer-engine and other heavy extras as optional for this quickstart.
• Add a short CI sanity job (optional) that runs the example on CPU to verify the quickstart remains functional. This job can be lightweight and time-limited.

Suggested branch name
doc/quickstart-single-gpu

Labels
documentation, enhancement, good-first-issue

Estimated difficulty
Low

I am willing to submit a PR
I’m happy to implement this and submit a PR that adds examples/quickstart/ and docs/get-started/local_quickstart.md. The PR will include the minimal script(s), config, and documentation. If maintainers prefer a slightly different layout, I can adapt the changes to match repo conventions.

[sbhavani]

@nathon-lee could you run the existing examples/run_simple_mcore_train_loop.py and see if anything needs to be improved there? We also have quickstart documentation here: https://docs.nvidia.com/megatron-core/developer-guide/latest/get-started/quickstart.html.

[nathon-lee]

thanks. I'll give it a try