Commit Graph

2 Commits

Author SHA1 Message Date
Dan Sedano b5b5c40fd7 Add xformers-only estimate test 2026-08-13 13:15:52 -07:00
Dan Sedano 2bc2e8d73e Add unit tests for the attention-dependent memory estimate
Covers which attention checks select the efficient formula in
BaseModel.memory_required, including the new flash_attention_enabled()
path, and pins the 7.5x bf16 ratio between the conservative and efficient
estimates so a retune of either formula fails the test and gets updated
consciously.
2026-08-13 12:49:01 -07:00