Fix GDN/KDA bugs, require CUDA kernels, add cache-aware tests#451

Merged

tscholak merged 9 commits into

oo/apriel_modeling_bugfrom

fix/require-cuda-kernels-no-fallbacks

Jan 19, 2026

Commits on Jan 17, 2026

Require CUDA kernels with no silent fallbacks

tscholak
and
claude
committed
Fix GDN chunk mode to use initial_state from cache

tscholak
and
claude
committed
Add extended cache tests for GDN and KDA equivalence

tscholak
and
claude
committed
Merge cache and non-cache equivalence tests with proper Apriel2Cache

tscholak
and
claude
committed
Refactor mixer tests and fix KDA mode selection

tscholak
and
claude
committed
Remove redundant test_causal_conv1d.py

tscholak
and
claude
committed
Consolidate cache.py into modeling_apriel2.py

tscholak
and
claude
committed
Enable bf16 tests and fix dtype handling

tscholak
and
claude
committed
Refactor test_mixer_equivalence.py: extract config fixtures and helpers

tscholak
and
claude
committed