recipe-recommender

Choosing a Megatron Bridge training recipe involves matching your model size, GPU topology, and training goal to a best-match indexed entry point. The recommender selects the appropriate recipe and outputs the run_recipe or run_script invocation structure so you can start training immediately.

Configuring long-context training beyond 8K tokens requires selecting specific CP variants and adjusting TP/PP/CP/EP constraints. The recommender points you to the correct long-context recipe entry points and provides the necessary adjustment constraints for your parallelism tuning.

Yes, you can use perf recipes for throughput benchmarking on mock data instead of library recipes for actual training. The recommender helps you choose between library and perf recipes, including key limitations and validation warnings for benchmarking scenarios.

Yes, you need to provide your model name or size, GPU count and type, training goal (pretrain, SFT, or PEFT), and sequence length. The recommender uses these inputs to map to appropriate recipe entry points and generate the exact entry-point command to run.

Recommended recipes provide a best-match starting configuration to avoid trial-and-error, but they include key limitations and validation warnings. You must still verify adjustment constraints for TP/PP/CP/EP parallelism tuning against your specific hardware and throughput benchmarking requirements.

Yes, recipe recommendation supports pretraining, SFT, and PEFT workflows. It maps your selected training mode and model size to appropriate recipe entry points, providing the correct run_recipe or run_script invocation structure for your specific fine-tuning goal.