emcodem
emcodem
@Const-me for me the lz4 stuff made loading the model fail (windows). To make it work again, I needed to revert to the commit before "[Switched shader compression algorithm from...
@philipag that did the trick, damn i thought about it but didnt actually do it :-(
fixed in this patch. Issue is that expComputeTokenLevelTimestamps assumes "unkown timestamps" are
you can try -mc 0 option
Thanks for trying, please share minimal example file
A few tests with the workaround mentioned are done now, it looks like we still get repeated text output but it seems to be a lot better than before, e.g....
thanks @VRCWizard that opened a lot of insights for me. I believe setting the max-content (-mc) option to 0 is the equivalent of the mentioned condition_on_previous_text of the python implementation...
Just adding some relevant informations. Just found that openai says " The model will only consider the final 224 tokens of the prompt and ignore anything earlier." https://platform.openai.com/docs/guides/speech-to-text/longer-inputs That seconds...
As most readers here might not be able to add the changes i mentioned above in the code for testing, i thought it might be a good idea to share...
@eagleftw023 > -mc 223 or 224? Honestly very good question. I wrote 223 on purpose because in my last test it made a difference to 224 so i thought we...