jgrusewski 9eea7f2977 feat(moe): moe_mixture_backward + moe_dgate_reduce kernels + FD test ...

Two-stage backward: moe_mixture_backward computes de_k = g · dh_s2 in
one kernel; moe_dgate_reduce computes dg via shmem reduction over c.

FD test verifies analytic backward matches numerical for B=2, K=4, C=32
within 1e-3 tolerance (perturbation 1e-3).

All test data flows via mapped pinned buffers (no HtoD/HtoH).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-04-27 18:42:29 +02:00

..

infra

feat(infra): GPU training launcher with local/cloud routing

2026-02-23 11:19:58 +01:00

plans

feat(dqn): enable Branching DQN with 45 factored actions (5×3×3)

2026-03-11 22:00:13 +01:00

superpowers

spec+plan(moe): correct gate input dim 42 → STATE_DIM=128

2026-04-27 18:22:40 +02:00

dqn-gpu-hot-path-audit.md

feat(dqn): Plan 1 Task 9 — atoms_update GPU kernel + AtomsMonitor + delete CPU paths

2026-04-24 18:28:19 +02:00

dqn-named-dims.md

feat(dqn-v2): D.6 plan_isv[6] = remaining_fraction

2026-04-24 20:36:46 +02:00

dqn-wire-up-audit.md

feat(moe): moe_mixture_backward + moe_dgate_reduce kernels + FD test

2026-04-27 18:42:29 +02:00

h_s2_consumers_audit.md

audit(dqn-v2): h_s2 audit anchor corrections (post-2c.3+4 reconnaissance)

2026-04-25 12:29:47 +02:00

isv-slots.md

refactor(dqn-v2): delete legacy CPU-DtoH per_branch_vsn_mean / per_branch_target_drift

2026-04-25 11:02:23 +02:00

ml-supervised-to-dqn-concept-audit.md

plan4(task7): Part E audit close-out — every supervised concept landed or OUT

2026-04-26 10:36:03 +02:00