Summary

Adds a lightweight validation layer and a configuration summary printed at startup, inspired by GPT-NeoX, resolving Issue #124.

Key features

• megatron/arguments.py
  • _validate_and_summarize_args(args) — runs sanity checks:
    • hidden_size % num_attention_heads == 0
    • global_batch_size % data_parallel_size == 0
    • pad_vocab_size_to (if set) divisible by TP size
    • fp16 / bf16 mutual-exclusion echoed
  • Builds a rank-0 console table summarising world-size layout, model dims, batch sizes, precision, and passes.
  • Raises ValueError if any rule fails, aborting early before costly init.

Why it matters

Early mis-configs (e.g., mismatched hidden/head sizes or bad batch divisibility) now surface instantly, saving hours of debugging and wasted GPU time.

Testing

• pytest -q tests — all existing tests pass.
• Launched pretrain_gpt_tiny.sh on 1 GPU and 4 GPU runs; summary appears once on rank 0.
• Introduced an invalid hidden_size (not divisible by heads) — run aborts immediately with clear error.

Backward compatibility

Purely additive logging/validation. No impact on training logic or performance.

Fixes #124