…vas cap Three audit-flagged fixes: #2 Temporal consistency loss was DEAD at the trainer call site. losses.py supports out_prev_for_consistency + motion_lr_prev_to_curr but sr_train_v7.py never passed them. motion_n_to_np1 was built by the dataset, RRM-zeroed when applicable, then dropped on the floor. Fix: pass out_main_n as the prev frame + motion_n_to_np1 + scale_for_warp. Temp loss now actually fires when curriculum stage 3 enables it. #3 Parent-child drift only saw the LAST renders gradient. The model stashed only one positions tensor in _last_positions_for_grad, overwriting it on every render_canvas call. With the trainer doing 3 renders per sample (t=0, t=2, t=1) and drift firing once after backward, only t=1s gradient counted. In curriculum stage 1 with lambda_fg=0, that means ZERO drift signal until step 20K. Fix: _retained_positions_for_grad is now a list; drift sums grads across all retained renders, then clears. New test test_drift_accumulates_gradient_across_multiple_renders verifies. #5 Smoke launcher validated canvas count against canvasCap=4096 but the trainer default is canvas_capacity=16384 (post-2026-05-13 spawner-config-rationale change). Bumped to match. 132/132 v7 tests pass. The currently-running pico-005 trainer (pid 21440, started 13:15:44 with the rasterizer-grad fix in cd73cc5) is correct for curriculum stage 1 -- temp loss is 0 and FG loss is 0 through step 20K. These fixes start mattering at step 20K when stage 2 kicks in; safe to pick up at the next natural restart. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

v6.2: Custom counting-sort tile bin (replace torch.sort) #3

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions