JWLHS
|
16def1ff18
|
chore: remove FIX comments, simplify cd dtype, remove __del__
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
16a70f05e8
|
fix: register_buffer for int4 tensors, kron-based LoKr path with ceiling-div padding, remove __doc__ overrides
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
d574e8242d
|
fix: register_buffer for quant tensors, kron-based LoKr path, remove __doc__ overrides
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
4ee3d16d84
|
fix: preserve linear bias during INT4 reconstruction
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
335eb29ca5
|
resolve rebase conflict
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
c7408ee3f9
|
resolve rebase conflict
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
ed64dd8f08
|
fix: filter_fn returns True to quantize, not skip
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
8631934ed7
|
fix: add docstrings for coverage threshold
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
c44d799d2a
|
fix: add _lora_prerotated flag — avoid double-rotation with pre-rotated LoRA
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
f1666813e8
|
fix: address CodeRabbit feedback — QuaRot error surface, LoRA shape validation, Sylvester bug fix, cache cap
|
2026-07-19 01:56:23 +08:00 |
|
JWLHS
|
5abbdb8661
|
feat: add torchao INT4 weight-only quantization backend
|
2026-07-19 01:56:23 +08:00 |