LR-AttnRes-converted-full-rms-20260919

Full-width-RMS conversions of standard Attention Residuals

Six original standard checkpoints, converted by a shared orthogonal residual-basis change followed by exact query-span restriction. No additional training. The source checkpoints and SHA256 hashes are pinned in protocol.json. Respect the source checkpoint licenses.

Each completed subdirectory contains the full BF16 model, reduced query parameters, rotation, numerical qualification, and paired 100M-target plain cross-entropy results. The RMS denominator still uses all residual coordinates. Source values retain full width. The exported configuration requires the provided model implementation and Fast-AttnRes 2.0.1. This is a numerical-fidelity experiment, not a measured inference-speed claim.

Load with the code/ directory on sys.path, reconstruct ModelConfig from checkpoint['model_args'], strict-load checkpoint['model'], move to CUDA BF16, and call require_fast_attnres() before forward. See execution.json and protocol.json for the exact environment, code hashes and data identity. The private experiment source is archived on branch conversion-audit-20260919; raw paired batch results are archived on conversion-results-20260919.