ROCm (gfx1201/RX 9070 XT): faster-whisper crash in CTranslate2 4.7.1

Description

faster-whisper crashes on ROCm with CTranslate2 on AMD RX 9070 XT (gfx1201).

I can reproduce this on stock Fedora kernel (6.18.14-200.fc43) with a minimal script (device="cuda").

Expected behavior: transcription completes normally, or returns a handled exception.
Actual behavior: process aborts with a GPU memory access fault and core dump.

Environment:

• OS: Fedora 43
• GPU: AMD Radeon RX 9070 XT (gfx1201)
• ROCm: 7.2.0
• Python: 3.12.12
• CTranslate2: 4.7.1 (ROCm wheel)
• faster-whisper: 1.2.1
• Runtime env: LD_LIBRARY_PATH=/opt/rocm/lib
• Kernel: 6.18.14-200.fc43.x86_64 (stock Fedora updates-testing)

Reproducibility

Minimal script:

from faster_whisper import WhisperModel

model = WhisperModel("small", device="cuda", compute_type="float16")
segments, info = model.transcribe("input_media.mp3", vad_filter=False)
for s in segments:
    print(f"[{s.start:.2f} -> {s.end:.2f}] {s.text}")

Run command:

LD_LIBRARY_PATH=/opt/rocm/lib python3.12 min_repro.py input_media.mp3 --model small --compute-type float16

Observed result on stock kernel (6.18.14-200.fc43):
small:

• float16 (no language hint): Memory access fault ... Page not present, then core dump / abort (EXIT_CODE=134)

Input tested:

• A short English Apollo 13 MP3 clip
• Additional local checks on a longer Japanese MP4 with large-v3 also failed (int8: GPU memory fault; float32: OOM followed by abnormal termination), so this is not limited to one clip/model setting.

Representative crash output (stock kernel, small + float16):

Memory access fault by GPU node-1 ... Reason: Page not present or supervisor privilege.
Failed to write segment data to pipe: Bad address
GPU coredump: handler exited with error (status: 1)
GPU core dump failed
timeout: the monitored command dumped core

repro_run.log

coredump_gdb_bt_13360.txt

rocminfo.txt

Start by running the minimal repro in min_repro.py with the stated ROCm, CTranslate2, Fedora kernel, and RX 9070 XT environment. Read repro_run.log, coredump_gdb_bt_13360.txt, and rocminfo.txt alongside the ROCm execution path to identify the fault. Done means transcription completes or returns a handled exception without a GPU memory fault, abort, or core dump.