Block a user
788dadeb88
a number of upgrades to SFT script to bring it up to date w.r.t. pretraining and tuning some of its kwargs based on sweeps
tacit
synced commits to refs/pull/498/merge at tacit/nanochat from mirror
2026-02-16 16:40:25 +00:00
788dadeb88
a number of upgrades to SFT script to bring it up to date w.r.t. pretraining and tuning some of its kwargs based on sweeps
tacit
synced commits to refs/pull/520/merge at tacit/nanochat from mirror
2026-02-16 16:40:25 +00:00
788dadeb88
a number of upgrades to SFT script to bring it up to date w.r.t. pretraining and tuning some of its kwargs based on sweeps
tacit
synced commits to refs/pull/516/merge at tacit/nanochat from mirror
2026-02-16 16:40:25 +00:00
788dadeb88
a number of upgrades to SFT script to bring it up to date w.r.t. pretraining and tuning some of its kwargs based on sweeps
tacit
synced commits to refs/pull/526/merge at tacit/nanochat from mirror
2026-02-16 16:40:25 +00:00
788dadeb88
a number of upgrades to SFT script to bring it up to date w.r.t. pretraining and tuning some of its kwargs based on sweeps
788dadeb88
a number of upgrades to SFT script to bring it up to date w.r.t. pretraining and tuning some of its kwargs based on sweeps
tacit
synced and deleted reference 2026-02-16 16:40:24 +00:00
refs/tags/refs/pull/151/merge at tacit/nanochat from mirror
tacit
synced and deleted reference 2026-02-16 00:20:29 +00:00
refs/tags/refs/pull/512/merge at tacit/nanochat from mirror
tacit
synced and deleted reference 2026-02-16 00:20:29 +00:00
refs/tags/refs/pull/477/merge at tacit/nanochat from mirror
tacit
synced and deleted reference 2026-02-16 00:20:29 +00:00
refs/tags/refs/pull/492/merge at tacit/nanochat from mirror
tacit
synced and deleted reference 2026-02-16 00:20:29 +00:00
refs/tags/refs/pull/515/merge at tacit/nanochat from mirror
e569b59f92
delete torchao dependency, create our own exact API-matched version of Float8Linear, document it very well. for some poorly understood reason, the performance is not only ~identical but actually runs 3% faster. despite of it being significantly simpler and much less code. i don't fully understand why/how atm
ff46300720
tune miniseries just a bit, fairly cosmetic, keep to even depths where the math works out nicely in model sizing
tacit
synced commits to refs/pull/447/merge at tacit/nanochat from mirror
2026-02-13 14:09:54 +00:00
e569b59f92
delete torchao dependency, create our own exact API-matched version of Float8Linear, document it very well. for some poorly understood reason, the performance is not only ~identical but actually runs 3% faster. despite of it being significantly simpler and much less code. i don't fully understand why/how atm
tacit
synced commits to refs/pull/455/merge at tacit/nanochat from mirror
2026-02-13 14:09:54 +00:00
e569b59f92
delete torchao dependency, create our own exact API-matched version of Float8Linear, document it very well. for some poorly understood reason, the performance is not only ~identical but actually runs 3% faster. despite of it being significantly simpler and much less code. i don't fully understand why/how atm
tacit
synced commits to refs/pull/485/merge at tacit/nanochat from mirror
2026-02-13 14:09:54 +00:00
e569b59f92
delete torchao dependency, create our own exact API-matched version of Float8Linear, document it very well. for some poorly understood reason, the performance is not only ~identical but actually runs 3% faster. despite of it being significantly simpler and much less code. i don't fully understand why/how atm
e569b59f92
delete torchao dependency, create our own exact API-matched version of Float8Linear, document it very well. for some poorly understood reason, the performance is not only ~identical but actually runs 3% faster. despite of it being significantly simpler and much less code. i don't fully understand why/how atm
tacit
synced commits to refs/pull/510/merge at tacit/nanochat from mirror
2026-02-13 14:09:54 +00:00
e569b59f92
delete torchao dependency, create our own exact API-matched version of Float8Linear, document it very well. for some poorly understood reason, the performance is not only ~identical but actually runs 3% faster. despite of it being significantly simpler and much less code. i don't fully understand why/how atm
ff46300720
tune miniseries just a bit, fairly cosmetic, keep to even depths where the math works out nicely in model sizing