new awq kernels paths (#2572)

* new awq kernels paths

bump 3.5 (#2565)

v3.4.3 (#2499)

Speeder (#2494)

* put preds on cpu directly. Less important but align GS on beam_search

3.4.1 (#2480)

v3.4 (#2466)

fix readme layout (#2421)

prevent from merge/unmerge LoRA weights with quantized weights (#2399)

bump 3.1.3 (#2382)

bump version 3.1.2 (#2368)