Configuration

CLI Configuration Reference

Flag	Default	Description
--model	tiny	tiny (from scratch) or qwen (pretrained LoRA)
--cpu	false	Force CPU training
--corpus-file	built-in examples	Path to UTF-8, one example per line
--batch-size	2	Per-rank batch size
--seq-len	64	Maximum sequence length (at least 4)
--epochs	1	Number of full training epochs
--max-steps	4	Maximum optimizer steps per rank per epoch; 0 means all batches
--lr	tiny 3e-4, Qwen 2e-4	Optional learning-rate override
--repeat	4	Corpus repetition multiplier
--save-dir	ddp_example_output	Folder for resulting artifacts

Distributed environment variables normally supplied by torchrun: WORLD_SIZE, RANK, and LOCAL_RANK. Use --standalone --nproc_per_node=N for a local single-node demonstration.

Not implemented: gradient accumulation, an LR scheduler, mixed-precision autocast configuration, DeepSpeed, model-parallel tensor partitioning, or evaluation flags.