Add mirror recurrence non-record submission

[cschubiner]

Summary

• add a non-record Apple Silicon MLX submission exploring mirrored depth recurrence
• reuse 9 unique transformer blocks across 18 logical layers while keeping the int8+zlib artifact under the 16 MB cap
• include the completed train log, submission metadata, and the serialization fix needed to exclude non-tensor schedule state from export

Run details

• hardware: Apple M5 Max
• tokenizer/data: fineweb10B_sp1024, full validation split, 1 training shard
• final int8+zlib roundtrip: val_loss 3.77618886, val_bpb 2.23647175
• bytes_total: 8040848

Notes

• this is a non-record experiment, not a record-track claim
• the architectural idea is mirrored recurrent depth: encoder uses blocks 0..8 and decoder reuses them in reverse order 8..0

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6853b49028

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Reject UNIQUE_LAYERS values the mirror schedule cannot reach

If a caller sets UNIQUE_LAYERS above the encoder depth but below NUM_LAYERS, this schedule never visits the extra blocks. For example, with NUM_LAYERS=18, UNIQUE_LAYERS=10, encoder_schedule is 0..8 and decoder_schedule is its reverse, so block 9 is allocated, counted, and serialized but never used in GPT.__call__. That silently changes the requested architecture for any UNIQUE_LAYERS > num_layers // 2 (except the unique_layers == num_layers fallback), so this should either be validated away or the schedule should be built to actually consume all unique blocks.

Useful? React with 👍 / 👎.

[cschubiner]

Closing this one as superseded by #56, which is the stronger Apple Silicon MLX local submission on the same account (val_bpb=1.8440 vs 2.2365). I explored additional local recurrent and width-scaling variants from this workspace, but none beat #56.