AlpinDale
|
ca7028d5ca
sampler: simplify logits resort in _apply_top_k_top_p (#1067)
|
6 days ago |
AlpinDale
|
4593a3b306
chore: remove dead code from triton sampling kernels (#1049)
|
1 week ago |
AlpinDale
|
2261a0e8dd
cpu: fix issue with sampling kernels (#1016)
|
2 weeks ago |
AlpinDale
|
0dfa6b60ec
core: support logprobs with multi-step scheduling (#963)
|
2 weeks ago |
AlpinDale
|
22a4cd4595
core: fix spec decode metrics and envs circular import (#889)
|
3 weeks ago |
AlpinDale
|
901900854e
chore: consolidate environment variables within one file (#882)
|
4 weeks ago |
AlpinDale
|
ef99a567b6
fix: temp_last warning being repeated for every output token (#869)
|
1 month ago |
50h100a
|
fc3c1cd5a5
this is getting its own commit because lint failures like that are exactly why people stop using linters
|
1 month ago |
50h100a
|
60a5d0fb80
rewrite DRY to be a lot faster
|
1 month ago |
AlpinDale
|
0035dc42ed
sampler: optimize DRY performance using z-algorithm (#856)
|
1 month ago |
AlpinDale
|
2150bb5019
sampler: add range parameter for DRY (#855)
|
1 month ago |
AlpinDale
|
72c505ad84
sampler: fix dry concurrency issue (#852)
|
1 month ago |
Selali
|
14ac216498
sampler: add output_tokens to DRY sampler (#849)
|
1 month ago |
AlpinDale
|
dfa34d1b24
feat: add sampler_priorty (#837)
|
1 month ago |
AlpinDale
|
60f7b828d5
feat: add skew sampling (#834)
|
1 month ago |
AlpinDale
|
ba9d8f631a
feat: add no_repeat_ngram sampler (#832)
|
1 month ago |
Selali
|
4c4a365f77
feat: Add DRY (Don't Repeat Yourself) sampling (#827)
|
1 month ago |
AlpinDale
|
bfc8988116
feat: add cuda sampling kernels for top_k and top_p (#828)
|
1 month ago |
AlpinDale
|
22427602eb
feat: add top-nsigma sampling method
|
1 month ago |
50h100a
|
273c61d406
guard against nan temperature from dynatemp (or anywhere else).
|
1 month ago |
50h100a
|
a22e887319
why we don't use the github website editor to make changes
|
1 month ago |
50h100a
|
54a8320638
logit shenanigans to prevent even worse shenanigans
|
1 month ago |
50h100a
|
b6a897d2a1
fix temperature, and address those pernicious dynatemp NaNs
|
1 month ago |
50h100a
|
83040c6389
Mask dynatemp using min/max, rather than exp
|
1 month ago |
Pyroserenus
|
99fc6f4697
fix: crash in token bans (#764)
|
3 months ago |
AlpinDale
|
ad181e3fef
feat: bring back dynatemp (#754)
|
3 months ago |
AlpinDale
|
6329c2d53f
chore: re-enable custom token bans (#751)
|
3 months ago |
AlpinDale
|
abd9d5799a
feat: add XTC Sampling (#740)
|
3 months ago |
AlpinDale
|
198029295c
fix: empty sampler output when temperature is too low (#709)
|
4 months ago |
AlpinDale
|
1394008421
chore: decouple `should_modify_greedy_probs_inplace (#671)
|
4 months ago |