-
https://git.ecker.tech/ aims to provide a place to share my efforts while maintaining true ownership of my code, as I do not trust GitHub.
XMR: 4B9TQdkAkBFYrbj5ztvTx89e5LpucPeTSPzemCihdDi9EBnx7btn8RDNZTBz2zihWsjMnDkzn5As1LU6gLv3KQy8BLsZ8SG
- Joined on
Oct 10, 2022
mrq deleted branch main from mrq/bitsandbytes-rocm
mrq deleted branch master from mrq/bitsandbytes-rocm
mrq pushed to main at mrq/bitsandbytes-rocm
- c88f97a9c8 drop support for gfx903 because depending on hipblaslt gums up too many things
- e38b9e91b7 Revert get_cuda_version ROCM version change
- c97c78bd66 Update README rocm quickstart
- 0b481bfcc2 Use workaround for ROCm wave32 recognition
- 615d47583f README: Add quickstart and info section
- Compare 10 commits »
mrq commented on issue mrq/ai-voice-cloning#411
Cant get training to work? (windows 10 nvida GPU)Well, from a glance your training settings are whack. Your dataset size is 93, yet your batch size is set to that (which is already an uneven number). Your gradient accumulation factor is also…
mrq commented on issue mrq/ai-voice-cloning#410
Learning state restart question> the learning rate will restart only 4 times in the entire training Correct.
mrq pushed to master at mrq/vall-e
- 08bae355eb actually use langs from the dataloader
mrq pushed to master at mrq/vall-e
- 8193d6b938 actually use langs from the dataloader
mrq pushed to master at mrq/vall-e
- 8740cdefc6 added initial support for languages (still testing, marked as model version 3), added experimental 'context extend by limiting the resp context' (untested)
mrq pushed to master at mrq/ai-voice-cloning
- 2830d1fa96 fixed setup scripts and Dockerfile to NOT use extra-index-url and instead use index-url (how this happened I don't know, since pytorch instructions use index-url), '''fixed''' phonemizing japanese for VALL-E with pykakasi
mrq commented on issue mrq/ai-voice-cloning#405
Does not run in DockerAh, I see now. The original Dockerfile was using `extra-index-url`, which will be ignored if the packages are found elsewhere (such as in the base index), especially without an explicit version…
mrq pushed to master at mrq/vall-e
- 6045cbce94 added experimental option to append utterances for training target (emphasis on experimental)
mrq pushed to master at mrq/vall-e
- 7facacf7c9 separated samplers into its own file, don't bother copying the logits back to the GPU after sampling, it's not necessary
mrq pushed to master at mrq/vall-e
- 100dd164e6 apply phoneme cleanup in inferencing as well
mrq pushed to master at mrq/vall-e
- b4405c98ea remove double spaces in the text phonemes (might have caused problems.........)