Skip to content

[WIP] Matmul infrastructure - #98

Merged
dance858 merged 6 commits into
mainfrom
matmul-infrastructure
Sep 1, 2026
Merged

[WIP] Matmul infrastructure#98
dance858 merged 6 commits into
mainfrom
matmul-infrastructure

Conversation

@dance858

@dance858 dance858 commented Jun 8, 2026

Copy link
Copy Markdown
Collaborator

I'm happy with this code, but should wrap it into quad form and add tests to python side before we merge it

dance858 and others added 6 commits August 31, 2026 19:13
The dense composition path formed term1 = 2w J_f^T Q J_f by filling a
length-n all-2w vector every evaluation, purely because only the
B^T diag(d) A dispatcher form was reachable. Use the new
BTA_matrices_fill_values and carry the weight with one dscal on the
result instead, mirroring term2. The diag_w field, its per-init
allocation, and its free all go away.

Also bump hess_term1's values_version after the direct ->x writes in
both backends (the matrix.h contract for writers; previously only
hess_term2 was bumped), and add a param-Q + nonlinear-child regression
test: the refreshed Hessian must match a freshly built engine, both
anchored against numerical differentiation. Verified load-bearing by
mutation (dropping the dscal and staling QJf_dense both fail it).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HRnTPZsCXK6x2PSGUzanmg
@dance858
dance858 force-pushed the matmul-infrastructure branch from c15275b to 7354349 Compare September 1, 2026 02:41
@dance858
dance858 merged commit 942df2d into main Sep 1, 2026
11 of 12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant