Model Converter Test Harness

Day-0 model coverage (DeepSeek-V4/V4.1, GLM5-Next, Kimi-K3, Qwen3.8-Next) adds new HF↔Megatron converters. Every converter must pass the shared harness in tests/converters/harness.py:

1. Bitwise round trip — for each Megatron parameter, export to HF names and reload: loader(exporter(w)) == w exactly (assert_roundtrip_bitwise).

2. Logits preservation — a randomly initialized 2-layer HF model (tiny config from tests/converters/tiny_configs/<family>.json) is mapped HF→Megatron→HF and the reconstructed model must produce allclose logits (fp32, rtol=atol=1e-5; assert_hf_logits_preserved).

Each family ships a tiny 2-layer config so the checks run on CPU CI in seconds.

Run:

python -m pytest tests/test_hf_to_megatron.py -k tiny_harness -x -q