python - Torch model call works fine on Github Codespaces, but crashes on Render.com - Stack Overflow

The following code:

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import torch

audio_proc = Wav2Vec2Processor.from_pretrained(
    "vitouphy/wav2vec2-xls-r-300m-timit-phoneme"
)
audio_model = Wav2Vec2ForCTC.from_pretrained(
    "vitouphy/wav2vec2-xls-r-300m-timit-phoneme"
)

def compute_phonemes(audio_content: bytes) -> list[dict[str, Any]]:
    try:
        audio_file = io.BytesIO(audio_content)
        speech, sr = librosa.load(audio_file, sr=16000)
    except Exception as e:
        print("Error loading speech", e)
        return []

    model_inputs = audio_proc(
        speech, sampling_rate=sr, return_tensors="pt", padding=True
    )

    with torch.no_grad():
        logits = audio_model(**model_inputs).logits

works completely fine on Github Codespaces 2CPU instance and runs in about .4 seconds

However when the exact same code, the exact same API, is deployed on Render — which usually works like a charm — then I got the following error logs, systematically:

[2025-01-09 18:25:30 +0000] [98] [CRITICAL] WORKER TIMEOUT (pid:141)
[2025-01-09 18:25:31 +0000] [98] [ERROR] Worker (pid:141) was sent code 134!

I traced down the issue and this is exactly this line that triggers the SIGABRT signal (code 134):

logits = audio_model(**model_inputs).logits

The same code will run "fine" on Railway but take 10s instead of 0.4s on Github Codespaces.

I can't see any difference in configuration between the two environments. Same Python version (3.12) and same PyTorch version (2.2.0)

Any idea or pointers to a solution (or an alternative) would be welcome.