Robert McMenemy comments

Results 12 comments of


                                            Robert McMenemy

Offset not working

This works as a fix for cursor ` let allNFTs = []; const timer = (ms) => new Promise((res) => setTimeout(res, ms)); let cursor = null; do { const NFTs...

i can't get cuda working

> not sure if I wanna screw anything up, at the moment I run it with certain parameters, which my 8gb seems to handle. Problem is I can't seem to...

Cleanup fetch_file_cached() operations in the download model

My bad good spot graham will undo with a new commit

Rework forward pass to remove old gradients

This may be linked to issue #27

Rework forward pass to remove old gradients

Good spot yet again @dancergraham have switched over to nvidia apex for fusedlayernorm on GPU can you try this now ?

Rework forward pass to remove old gradients

NB this requires the external lib for apex to work should speed up rendering once it fires up. Any issues let me know and I'll rework @dancergraham (have added to...

Rework forward pass to remove old gradients

Looking into other alternatives to layernorm and it seems instance or group normalisation may help speed things up here ! Will ping another refactored PR soon

Rework forward pass to remove old gradients

Shall remove the apex lib but keep the forward pass change this should still preserve performance without the lib dependency

Rework forward pass to remove old gradients

@dancergraham try this now, textvec rendering should still be significantly faster whilst I find a native way of speeding up layer norm on gpu / cuda

Rework forward pass to remove old gradients

My bad @dancergraham forgot I added as a param, changed back so .to() uses torch.deice directly rather than by reference from param list