4.8.0 regression: WhisperModel.align() fails with parallel_for failed: cudaErrorInvalidDevice on Tesla T4 (works on 4.7.0)

[christopher5106]

Description

After upgrading from CTranslate2 4.7.0 to 4.8.0, Whisper word-timestamp alignment (WhisperModel.align(), called by faster-whisper's add_word_timestamps → find_alignment) crashes on a Tesla T4 (sm_75) with:

RuntimeError: parallel_for failed: cudaErrorInvalidDevice: invalid device ordinal

Plain transcription (encode/generate) is unaffected — only the alignment path fails. Pinning back to 4.7.0 resolves it completely, with no other change. I suspect the Thrust 1.12.0 → CCCL 2.7.0 swap in 4.8.0 (the only GPU-infra change in that release per the changelog), since parallel_for is the failing Thrust primitive and align() is the op that uses it.

Environment

• CTranslate2: 4.8.0 (broken) vs 4.7.0 (works) — single-variable change
• faster-whisper: 1.2.0 (constraint ctranslate2>=4.0,<5, so it floats to 4.8.0)
• GPU: NVIDIA Tesla T4 (compute capability 7.5), single GPU, no MIG, default device_index=0
• Base image: nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04, Python 3.11
• Model: large-v3, compute_type="float16", device="cuda"

Reproduction

from faster_whisper import WhisperModel
model = WhisperModel("large-v3", device="cuda", compute_type="float16")
# ~30s speech clip; word_timestamps triggers WhisperModel.align()
segments, _ = model.transcribe("audio.wav", word_timestamps=True, vad_filter=False)
for s in segments:        # crash occurs while consuming the generator
    print(s.words)

Traceback

File ".../faster_whisper/transcribe.py", line 1253, in generate_segments
    self.add_word_timestamps(...)
File ".../faster_whisper/transcribe.py", line 1564, in add_word_timestamps
    alignments = self.find_alignment(...)
File ".../faster_whisper/transcribe.py", line 1683, in find_alignment
    results = self.model.align(...)
RuntimeError: parallel_for failed: cudaErrorInvalidDevice: invalid device ordinal

Workaround

Pin ctranslate2==4.7.0.

Notes

Reproduced deterministically across container restarts and model sizes (medium, large-v3). Likely related to but distinct from #2063 (also a 4.8.0 regression, but CPU-thread oversubscription on macOS).

[jordimas]

Hello. I tried to reproduce on a Tesla T4 with CTranslate2 4.8.0, faster-whisper 1.2.0, large-v3, Python 3.11, and the same nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04 base image, but the 33s word-timestamp alignment path completed successfully. I was unable to reproduce it.

If you can provide a docker image where the issue can be produce it, it will be great. Also if you can test it in other GPUS, this will be great input.

Thanks

Note: #2063 is unrelated to this issue.