AGILLM-4.3-EGGROLL-Experimental

AGILLM 4.3 EGGROLL Experimental

This repository is an experimental, default-off extension of AGILLM 4.3. It adds a PyTorch EGGROLL-style evolution-strategy sidecar for the model's discrete top-1 Mixture-of-Experts routers while preserving the existing backpropagation and DiffusionBlock training path.

The implementation is inspired by Evolution Strategies at the Hyperscale, which introduces Evolution Guided GeneRal Optimisation via Low-rank Learning (EGGROLL): https://eshyperscale.github.io/imgs/paper.pdf

Published locations

• GitHub source: Marxist-Leninist/AGILLM4.3-EGGROLL-Experimental
• Canonical Hugging Face repository: MarxistLeninist/AGILLM-4.3-EGGROLL-Experimental
• Legacy Hugging Face code/documentation mirror: OpenTransformer/AGILLM-4.3-EGGROLL-Experimental

MarxistLeninist is the main Hugging Face account and contains the code, model card, metadata, and full frozen checkpoint. OpenTransformer has reached its public-storage allowance and is retained only as a legacy code/documentation mirror; it is not the canonical repository for this experiment.

Status

• Production AGILLM 4.3 v22 remains untouched and continues training.
• EGGROLL is disabled unless --eggroll_every_steps is greater than zero.
• The first implementation targets only top-1 MoE router matrices. It does not replace AdamW, AR/SAT/NAT training, or the existing full-stack anchors.
• The frozen starting checkpoint is step 1,777,105, with 87,348,264,960 tokens seen.
• Experimental code SHA-256: f487cf2fbb38b69bf9f19322759176c3e4b66636a6375fc30759434be22aab80.
• Base checkpoint manifest SHA-256: 72ee3cf3ae4e3dedb22f16d8c2d21c80c66e0bdcd2001f606f030886cc98e2af.

What was added

The sidecar uses antithetic low-rank perturbations, candidate-major population batching, deterministic counter-derived seeds, full-stack autoregressive cross-entropy scoring, separate search and guard crops, update-RMS trust-region clipping, exact rollback on regression, no-signal rejection, JSONL receipts, dry-run mode, and checkpointed experiment state.

The disabled path was tested against v22 and produced bitwise-identical output. CPU tests also verified perturbation symmetry, descent on a known quadratic objective, accepted router mutation, and exact restoration after forced rejection.

Why this is optional

EGGROLL can make population evaluation vastly more efficient than naïve evolution strategies. That does not automatically make it more sample-efficient than token-level backpropagation. In the paper's from-scratch integer-language-model experiment, EGGROLL demonstrated a capability that ordinary backpropagation could not directly provide for that integer model, but the largest run still used far more total GPU-hours than its floating-point backpropagation comparison. The correct AGILLM test is therefore an equal-GPU-hour A/B experiment, not a victory announcement generated by enthusiasm and a command-line flag.

It is most promising here for discrete router choices and, later, verifiable outcome objectives where ordinary gradients are absent, noisy, or misaligned with the actual target.

Smoke test

python agillm43_singlefile_intelligence_v23_eggroll_experimental.py eggroll-smoke

Recommended first experiment

Use the included launch_eggroll_experiment.sh. It resumes from the frozen checkpoint and starts with --eggroll_dry_run, so candidate populations are measured and logged without changing model weights. Remove only that flag after the receipts show usable pair signal and acceptable overhead.

Conservative starting settings are population 8, rank 1, one router per event, 128-token search and guard crops, two guard crops, and one event every 5,000 committed training steps.

Important limitations

1. This version directly updates selected router weights after an ordinary AdamW commit. AdamW's stored moments for those few parameters are therefore slightly stale. The update is deliberately tiny and clipped, but a future implementation should either project the ES update through a dedicated optimiser or reconcile the main optimiser state.
2. Guard crops are independent regions of the current training batch, not a held-out benchmark.
3. Only AR cross-entropy is used as EGGROLL fitness in v1. SAT, NAT, reasoning rewards, code execution, and tool-use rewards remain future experiments.
4. This is a single-GPU research path. The paper's strongest throughput result relies on large batched populations and specialised multi-device inference infrastructure.

Files

• agillm43_singlefile_intelligence_v23_eggroll_experimental.py: complete single-file trainer and inference runtime.
• EXPERIMENT.md: design, estimator, safety boundaries, and A/B protocol.
• EXPERIMENT_MANIFEST.json: hashes, checkpoint lineage, hardlink snapshot receipt, and test record.
• Canonical Hugging Face base_checkpoint/: exact block-sharded package used as the experiment base.
• launch_eggroll_experiment.sh: dry-run-first launch generated from the live production command.