CR4 RCM Rollout 256: RoPE ablations at 5k
Model-only CR4 compressor checkpoints trained for 5,000 optimizer iterations on the 256-record SF-DMD c1-1 rollout dataset.
Shared training setup: Wan2.1 1.3B SF-DMD c1-1 initialization, future-only cache-aware self-forcing, CR4 (memory_ratio=0.25), batch size 1, learning rate 5e-5, prefix lengths {1,4,8,12,16,20}, and four-step rollouts.
| Folder | memory_rope_mode |
Description |
|---|---|---|
source_aligned |
source_aligned |
Source-time-aligned 15x26 memory grid stretched over the native spatial range. |
jit_identity |
jit_identity |
Generator memory RoPE phases are zero. |
compact_grid |
compact_grid |
Dense 15x26 memory coordinates without spatial stretching. |
compact_grid_time_cr4 |
compact_grid_time_cr4 |
Compact spatial grid with memory time scaled by 1/4; asymmetric diagnostic. |
packed_temporal_grid |
packed_temporal_grid |
Primary temporal model: all CR4 slots are packed row-major into native 30x52 grids and future time follows the packed memory timeline. |
Use branch exp/packed-temporal-memory-grid-future-sf-20260720 at a87a28eba3edc42d362f8eed0f9bddd569955a86. The packed-temporal implementation commit is 945550f2b0e463aac5217940ec4aa753f4e75e12. Pass the folder's exact memory_rope_mode during both compression and generation.
Each folder is an Imaginaire/PyTorch distributed checkpoint containing only iter_000005000/model. Optimizer state is intentionally omitted. See checkpoint_manifest.json for provenance and hashes.