Files
automatic/modules/lora
CalamitousFelicitousness 4c85a60c3e perf(lora): skip the discarded init when building lora modules
create_module built each up/down module with the default constructor,
which kaiming-initializes the parameter, then copied the stored weight
over the whole thing. The init is thrown away every time and costs about
four times the copy: 22.1ms per module against 2.5ms, or 5.8s against
0.7s over a 264-module lora, on every load.

skip_init constructs on meta and materializes uninitialized, so the copy
still fully defines the parameter. Dtype, device and values are
unchanged, including the fp32 upcast of bf16 files that the copy performs.
2026-08-10 02:42:05 +01:00
..
2026-07-03 22:05:33 +02:00
2026-07-24 18:20:41 +02:00
2026-02-19 09:15:37 +01:00
2026-02-19 11:09:13 +01:00
2026-02-19 09:15:37 +01:00
2026-06-04 12:36:10 +02:00
2024-11-29 08:01:26 -05:00
2026-07-08 09:50:26 +02:00