8000 Noprop fm architecture 14903760323221249793 by LokiMetaSmith · Pull Request #5 · volotat/mini-AGI · GitHub
[go: up one dir, main page]

Skip to content

Noprop fm architecture 14903760323221249793 - #5

Closed
LokiMetaSmith wants to merge 2 commits into
volotat:mainfrom
LokiMetaSmith:noprop-fm-architecture-14903760323221249793
Closed

LokiMetaSmith wants to merge 2 commits into
volotat:mainfrom
LokiMetaSmith:noprop-fm-architecture-14903760323221249793

Conversation

@LokiMetaSmith
Copy link
Copy Markdown

No description provided.

LokiMetaSmith and others added 2 commits September 22, 2026 05:51
This refactor pivots the mini-AGI architecture from an autoregressive
next-byte PonderNet setup to a NoProp-FM continuous time dynamics
architecture. The changes introduce an ODE vector field approach using
Euler integration for inference and Flow Matching loss for training,
entirely eliminating BPTT. The recurrent blocks now act as point-wise
MoE networks, while causal attention is maintained purely in the prelude.

Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com>
This refactor pivots the mini-AGI architecture from an autoregressive
next-byte PonderNet setup to a NoProp-FM continuous time dynamics
architecture. The changes introduce an ODE vector field approach using
Euler integration for inference and Flow Matching loss for training,
entirely eliminating BPTT. The recurrent blocks now act as point-wise
MoE networks, while causal attention is maintained purely in the prelude.

Also updated README.md to reflect the new capabilities and structural changes.

Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com>
Maxritz pushed a commit to Maxritz/mini-AGI-cpp that referenced this pull request Sep 25, 2026
…weep (case 2b)

The d_lam path was implemented but missing the 1/M batch mean, making
every halting gradient M times too large (8x at production M=8).
Verified via a cached-lam finite difference in double precision that
sweeps M={1,2,3}: rel error was exactly 1/M with the defect and
passed with the fix.

- src/backward.cpp: fold inv_M into step_ce so d_lam and dc_prev scale correctly
- src/backward.hpp: publish d_lam_per_step via optional out-param
- tests/test_backward.cpp: case 2b check_halting_scaling + halting_identity;
  also fixed a latent exp-for-log typo in ce_of that had masked an entire defect turn
- tests/test_pool_backward.cpp: covers pooled MLP in isolation
- src/paged.*: assert_slots_valid on mutable_w* to prevent pre-begin_segment writes
- src/recur.hpp: begin_segment() marked REQUIRED
- STATUS.md: gap volotat#11 CLOSED, gap volotat#12 CLOSED, ce_of typo recorded as bug volotat#5

Test: 8/8 ctest pass; defect reintroduction fails 6 checks, fix passes all.
Train smoke: 3 steps, loss 5.5458 (forward unchanged).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet
3AFB

Development

Successfully merging this pull request may close these issues.

1 participant

0