Non-record: Entropy-aware Int5-odd BTT-MLP with materialized eval

[LucasErcolano]

Summary

This PR adds a non-record submission package for an entropy-aware int5_odd export objective combined with a TT/BTT-inspired structured MLP and an evaluation-time materialization path.

Main pieces included in the package:

• entropy-aware training on a 5-bin odd grid {-2,-1,0,1,2} aligned to the exported int5_odd + zlib artifact
• structured MLP matrices via a 2-core TT/BTT-style StructuredLinear in the MLP while keeping attention dense
• evaluation-time materialization of the structured MLP into dense weights so validation runs through standard F.linear
• local and cloud helper scripts, plus the logs referenced by the README

Main Result

Recovered 1x H100 80GB HBM3 result (non-record, VAL_TOKEN_LIMIT=1048576):

• params: 25,727,104
• pre-quant val_bpb: 5.3457
• roundtrip val_bpb: 5.88802157
• quantized size: 5,184,543 bytes
• total package size: 5,267,667 bytes
• peak memory: 50290 MiB allocated / 78558 MiB reserved
• optimizer-step training time: 603.522 s for 31 steps

Systems Note

The key engineering win here is the eval-time materialization hook for the structured MLP. On the 1,048,576-token validation cap, the materialized path reduced eval time from roughly 96.99 s to 0.851 s, which is what made this structured stack operationally usable on H100 for evaluation.

Validation

• py -3.11 -m py_compile records/track_non_record_16mb/2026-03-31_EntropyAware_Int5Odd_BTTMLP_3060/train_gpt.py
• package includes the logs referenced by the README for compile, init, lambda, scout, local smoke, and H100 runs

Scope

This is intentionally a non-record exploratory PR. The current package is anchored to the best fully recovered 1xH100 run and not a full-validation leaderboard submission.