Does faster whisper support AMD GPU?

[apollos]

If yes, is there a installation or build steps basing on ROCm?

[PinW]

CTranslate2 just merged support for AMD cards via HIP a few days ago: OpenNMT/CTranslate2#1989

Release with wheels here: https://github.com/OpenNMT/CTranslate2/releases/tag/v4.7.1

With some elbow grease, I even managed to get faster-whisper working with an unsupported 5700 XT card

[urec56]

Hello!
@PinW How did you do that?

I'm trying to start it on 6700 XT card and rocm 7.1.1 and it just shows me the message libamdhip64.so.7: undefined symbol: hsa_amd_enable_logging, version ROCR_1

Also I'm trying to use it in develop nix shell, and I don't know what additional packages it needs

[PinW]

Hello! @PinW How did you do that?

I'm trying to start it on 6700 XT card and rocm 7.1.1 and it just shows me the message libamdhip64.so.7: undefined symbol: hsa_amd_enable_logging, version ROCR_1

Also I'm trying to use it in develop nix shell, and I don't know what additional packages it needs

I believe the CT2 wheels are built on rocm 7.2, you could try that.

[urec56]

They are. But I have the 6.0 rocm in my system and it can't work with 7+ version in docker or nix-env.
I tried to build wheels on 7.1.1 version and it doesn't work too. Also, I can't build wheels on 6+ version
So, now I'm trying to build wheels of side project https://github.com/arlo-phoenix/CTranslate2-rocm with 6+ version support

[PinW]

@urec56 https://github.com/PinW/ctranslate2-rocm-rdna1

Maybe you can use my project as a reference. Was a pain but I got it working eventually.

[urec56]

Thank you!
Now I've realized that the main problem in my setup is 6.* drivers instead of 7.* in my system. I can update to 7.*, but I don't want
I've tried several methods to run different builds of the lib, but I gave up. It wastes too much time
Also, I'm on linux and your method is for windows. I hope it will help someone

[nabe2030]

Yes, as of CTranslate2 v4.7.1 (released Feb 2026, see #1989), faster-whisper runs on AMD GPUs with ROCm. The setup is reasonably simple on Linux now:

Linux setup:

1. Install ROCm 7.x — sudo apt install rocm works if you're on Ubuntu with AMD's apt repo configured
2. Get the official CTranslate2 ROCm wheel from the v4.7.1 release page: rocm-python-wheels-Linux.zip (contains wheels for Python 3.10–3.14)
3. pip install the wheel for your Python version, then pip install faster-whisper
4. Use as normal with device="cuda" — CTranslate2 keeps that name even on ROCm

The wheel ships GPU kernels for: gfx803, gfx900, gfx906, gfx908, gfx90a, gfx942, gfx950, gfx1030, gfx1100, gfx1101, gfx1102, gfx1150, gfx1151, gfx1200, gfx1201. That covers most ROCm-supported AMD GPUs from Polaris/Vega through Instinct MI355 and RDNA4. If your card isn't on this list (e.g. gfx1032 / RX 6600), see OpenNMT/CTranslate2#2012 for an env-var workaround that worked for some RDNA2 cards.

Confirmed configurations from the community:

• gfx1151 (Strix Halo / Ryzen AI Max+ 395) — full recipe at https://github.com/nabe2030/faster-whisper-rocm-strix-halo, ~11.5x realtime English / 6.56x realtime 30-min Japanese long-form
• gfx1032 (RX 6600, RDNA2) — works with CT2_CUDA_ALLOCATOR=cub_caching per OpenNMT/CTranslate2#2012
• Instinct line (gfx906/gfx942) — sssshhhhhh's primary test target in #1989

For Windows, install instructions are in sssshhhhhh's Jan 31 comment on #1989 — slightly more involved due to ROCm SDK packaging on Windows.

For non-English long-form audio (Japanese, possibly Mandarin / Korean / etc.), vad_filter=False and condition_on_previous_text=False produced both better quality and ~2x throughput in my testing versus the library defaults. May not generalize to all languages, but worth trying if the defaults give you hallucinations or dropped segments.