[WIP] Compression-aware fixed-step research

[JusticeShultz]

Summary

This is a WIP local research branch for Parameter Golf focused on improving post-roundtrip val_bpb under the actual artifact constraint, using 1x RTX 3090 for local search before moving to the target 8xH100 environment.

The biggest shift in this branch is methodological: local ranking moved away from noisy short wallclock runs and onto a fixed-step exact roundtrip track. That made the local loop much more trustworthy and changed the research conclusions substantially.

The current branch focus is now:

• compression-aware training
• export-aware regularization
• dense near-cap scaling
• fixed-step post-roundtrip evaluation

As of March 19, 2026, train_gpt.py remains under the repo hard cap at 1492 lines.

What This Branch Adds

Compression-aware / export-aware training knobs

• COMPRESSION_REG_WEIGHT
• COMPRESSION_GRID_REG_WEIGHT
• COMPRESSION_SCALE_REG_WEIGHT
• COMPRESSION_RANK1_REG_WEIGHT
• TERNARY_REG_WEIGHT
• OUTLIER_REG_WEIGHT

Architecture / export knobs

• NUM_UNIQUE_BLOCKS
• WINDOW_SIZE
• EMBED_DIM
• INT8_AXIS_MODE
• INT8_RESIDUAL_RANK
• INT8_RESIDUAL_BUDGET_BYTES

Hybrid eval-time knobs

• EVAL_CACHE_MIX_WEIGHT
• EVAL_BIGRAM_MIX_WEIGHT
• EVAL_CACHE_SIZE

Local evaluation / search controls

• VAL_MAX_TOKENS
• ROUNDTRIP_VAL_MAX_TOKENS
• FINAL_ROUNDTRIP_EVAL
• fixed-step local runs via ITERATIONS
• dedicated local sweep launchers for:
  • fixed-step roundtrip comparisons
  • export-aware sweeps
  • iso-byte dense sweeps
  • high-cap dense width/depth sweeps

What Has Been Tried

1. Matched local roundtrip baseline

• Run: baselinert3090_20260318_181344
• Exact final roundtrip: val_bpb=2.11089617
• Artifact: 6,705,058 bytes

2. Compression-aware baseline

• Run: compressrt3090_20260318_175828
• Knobs: COMPRESSION_REG_WEIGHT=0.005
• Exact final roundtrip: val_bpb=2.06085837
• Artifact: 6,839,798 bytes

This was the first clear local win over the matched baseline.

3. Fixed-step methodology pivot

The local wallclock track turned out to be too noisy to trust for small deltas, so ranking moved to a fixed-step exact roundtrip track.

Dense fixed-step control:

• Run: fixedsteprtsweep_20260318_221632_base_a
• Exact final roundtrip: val_bpb=2.04299145

This became the new local control.

4. Export-aware compression regularization

Best export-aware probe:

• Run: exportaware_fixedstep_20260318_223456_g010_r000
• Knobs:
  • COMPRESSION_REG_WEIGHT=0.005
  • COMPRESSION_GRID_REG_WEIGHT=0.10
• Exact final roundtrip: val_bpb=2.04288777
• Artifact: 6,663,470 bytes

This slightly but repeatably improved the fixed-step control.

Follow-up export-aware checks:

• scale-aware regularization: regressed
• nearby grid weights (0.08, 0.12): regressed
• tiny outlier suppression: regressed

5. Branches that did not currently win locally

These were tested and are currently parked:

• hybrid eval-time sidecar
• sparse attention
• recurrent/shared-block variants
• ternary/low-bit shaping
• residual-budget tuning
• factorized/reused-block variants in the tested regime

Important nuance: these negative results were gathered before or outside the stronger near-cap dense regime, so they are not being treated as globally dead ideas.

6. Iso-byte dense frontier sweep

This changed the branch direction significantly.

Results:

• b10 -> 2.02814871 at 9,683,932 bytes
• b12 -> 2.05262920 at 11,334,608 bytes
• b14 -> 2.03768242 at 13,094,288 bytes
• b155 -> 2.00290272 at 13,741,308 bytes

This showed that simply spending more of the byte budget on a dense compression-aware model mattered much more than most under-cap micro-ideas.

7. High-cap dense width/depth frontier

Recovered / rerun near-cap results:

• w608_l12 -> 2.00551677 at 14,371,393 bytes
• w624_l12 -> 2.01128088 at 15,024,114 bytes
• d576_l14 -> 1.99806297 at 15,222,128 bytes
• w640_l12 -> 2.00505534 at 15,658,993 bytes

Current Best Local Result

Current local leader:

• Run: highcapdense_rerun_20260319_d576_l14
• Shape: 14 layers / 576 dim / 8 heads / 4 KV heads
• Knobs:
  • COMPRESSION_REG_WEIGHT=0.005
  • COMPRESSION_GRID_REG_WEIGHT=0.10
• Exact final roundtrip: val_bpb=1.99806297
• Total artifact: 15,222,128 bytes

Current Findings

• Optimizing for the roundtripped artifact matters more than pre-quant proxy quality.
• The fixed-step exact roundtrip local track is much more trustworthy than short local wallclock ranking.
• Compression-aware training is the only clearly first-order win so far.
• Export-aware grid alignment is promising, but currently a small gain, not a complete solution.
• Dense scaling near the byte cap dominates most earlier under-cap micro-ideas.
• In the near-cap local regime tested so far, depth currently looks better than width.
• Many earlier negative results from recurrence / sparsity / ternary likely reflect the wrong byte regime, not necessarily globally bad ideas.

Current / Next Direction

The branch has now shifted from “small-model micro-tuning” to “near-cap dense frontier + export-side improvements.”

The next likely high-upside directions are:

• export-side symmetry-aware permutation for compression
• tensor sensitivity mapping / heterogeneous export allocation
• building new export ideas on top of the deeper dense 14x576 control instead of the older 6.6 MB model

Caveats

• These are local experiments on 1x RTX 3090 under Windows.
• They are not leaderboard claims and are not directly comparable to the official 8xH100 / 10 minute challenge runs.
• Local evaluation is still a proxy, even though it is now much more reliable than the earlier wallclock track.
• No tokenizer work has been explored yet.
• No dataset/accounting tricks or external-compute loopholes are being used.

Why This Branch Exists

The goal of this branch is to build a reproducible local search loop that ranks ideas against a closer approximation of the real challenge objective, so the strongest branch can be taken to target compute once grant access is available.