Non-record: 12L Low-Rank Q + QAT (1xH100, pre-quant 1.2035)#316

Open

SkywardSyntax wants to merge 6 commits intoopenai:mainfrom

SkywardSyntax:pr/non-record-research-pipeline

Commits on Mar 21, 2026

[1xH100] Initial implementation: 12L + Low-Rank Q + QAT + FTLE + Stride-OGD

SkywardSyntax
and
claude
committed
[1xH100] Fix GQA compat for PyTorch 2.4 + fix quantization bit search order

SkywardSyntax
and
claude
committed
[1xH100] 2000-step test: 1.2517 bpb — beats baseline at only 2000 steps

SkywardSyntax
and
claude
committed
[1xH100] 7900-step full run: pre-quant val_bpb=1.2035

SkywardSyntax
and
claude
committed
[1xH100] FTLE ablation: uniform quantization strictly better, update docs

SkywardSyntax
and
claude
committed
Non-record: 12L Low-Rank Q + QAT research pipeline (1xH100)
SkywardSyntax
committed