Ability to set ComputeType (int8, float16 etc) for transcription in peertube-runner #6354

New issue

Open

opened 2026-02-22 12:47:37 -05:00 by deekerman · 0 comments

deekerman commented

2026-02-22 12:47:37 -05:00

Owner

Originally created by @lachlan-stevens on GitHub (Dec 6, 2025).

Describe the problem to be solved

I would like to be able to configure my peertube-runner instances to use different compute type based on their hardware config.

whisper-ctranslate2 accepts the following:

default
auto
int8
int8_float32
int8_float16
int8_bfloat16
int16
float16
bfloat16
float32

openai-whisper accepts the following:

float16 (default)
float32

I have forked the repo and created a reasonably simple modifications that accepts a new computeType key in config.toml and uses this at runtime (including validating it matches the expected values on both whisper-ctranslate2 and openai-whisper, falling back to the default value if it does not match the expected format, or if it is not specified (backward compatibility).

Creating this issue to discuss prior to submitting PR as per requirements in PR template.

Describe the solution you would like

Accept a PR which implements this functionality and passes CI.

Originally created by @lachlan-stevens on GitHub (Dec 6, 2025). ### Describe the problem to be solved I would like to be able to configure my peertube-runner instances to use different compute type based on their hardware config. whisper-ctranslate2 accepts the following: - default - auto - int8 - int8_float32 - int8_float16 - int8_bfloat16 - int16 - float16 - bfloat16 - float32 openai-whisper accepts the following: - float16 (default) - float32 I have forked the repo and created a reasonably simple modifications that accepts a new `computeType` key in config.toml and uses this at runtime (including validating it matches the expected values on both whisper-ctranslate2 and openai-whisper, falling back to the default value if it does not match the expected format, or if it is not specified (backward compatibility). Creating this issue to discuss prior to submitting PR as per requirements in PR template. ### Describe the solution you would like Accept a PR which implements this functionality and passes CI.