plan(perf): K-loop block-per-batch parallelization implementation plan

Browse Source

Bite-sized 22-task plan implementing the design at
docs/superpowers/specs/2026-05-17-kloop-parallelization-design.md.

Four atomic commits per the spec's Rollout section:
  Commit 1: reduce_axis0 kernel + cfc_step refactor (Tasks 1-10)
  Commit 2: GRN bwd refactor (Tasks 11-14)
  Commit 3: VSN bwd refactor (Tasks 15-17)
  Commit 4: attention_pool bwd refactor (Tasks 18-20)

Each commit covers kernel rewrite + per-batch scratch buffers +
reducer launches + memset_zeros wiring + smoke validation.

Tests added across commits:
  - cfc_bwd_b1_oracle.rs (Task 6): B=1 oracle vs single-sample helper
    within relative_eq 1e-5 (FP-tolerant, not bit-exact)
  - stacked_trainer_loss_shrinks_at_batch_32 (Task 7): first test
    that exercises the cross-batch reducer path
  - cfc_bwd_scratch_clears_between_steps (Task 8): scratch-zero
    regression guard

Acceptance gates 1-8 from the spec are mapped to Tasks 6, 7, 8, 9,
21, 22 (local + cluster A/B perf). Reference baseline for gate #8
is t6z89's per-epoch AUC trajectory.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

...

This commit is contained in:

jgrusewski

2026-05-17 23:27:20 +02:00

parent ea1d8703b7

commit c508d101c1

1 changed files with 1861 additions and 0 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

1861

docs/superpowers/plans/2026-05-17-kloop-parallelization.md Normal file

File diff suppressed because it is too large Load Diff