amd and/or apple metal/mps acceleration

[BBC-Esq]

Does faster-whisper support either AMD or metal/mps acceleration out of the box? I noticed on the ctranslate2 instruction page it states that it can be built with Apple's Acceleration, but faster-whisper is a separate repository and so it's not clear from the readme whether faster-whisper supports apple's metal/mps and/or "Accelerate?" I'm presuming from the ctranslate2 instructions that faster-whisper also doesn't support AMD GPU acceleration?

Editing this response to also ask about oneDNN, Openblas, and Ruy, which are all indicated as "off" in the instructions to compile Ctranslate2, but not mentioned on the faster-whisper respository.

Thanks!

[phineas-pta]

"off" is default value but u can change to your need when compiling

[BBC-Esq]

"off" pertains to which of my multiple questions??

[BBC-Esq]

I might be misunderstanding, but...The only instructions I'm aware of regarding "compiling" pertain to Ctranslate2, not faster-whisper. If I use pip install faster-whisper does it come with the certain backends enabled or not? Faster-whisper uses Ctranslate2, but it is itself a separate library. So my questions are whether faster-whisper has the backends enabled, and if not, how can i enable them?

[phineas-pta]

default backend is here https://opennmt.net/CTranslate2/hardware_support.html

to enable certain backend u compile ctranslate2 from source before install faster-whisper

[BBC-Esq]

Thank you! That answers my questions. Now, faster-whisper automatically installs ctranslate2 as part of its dependencies in requirements.txt. Do I need to modify installation requirements.txt to NOT install ctranslate2 since I already compiled it manually? In other words, I don't want my pip install faster-whisper to accidentally override my manually-compiled version of ctranslate2?

[phineas-pta]

if u already got latest ctranslate2 compiled and installed as python package then no overwrite to worry about

[BBC-Esq]

Thanks again sir! Will close this issue out.