[Crash] Colab Instantly Crashes with Whisper + unsloth — Small Dataset, CPU Only, No Traceback

[C0deXG]

1. Did you update?

Yes. Installed fresh in clean runtime:

pip install --upgrade unsloth unsloth_zoo

---

2. Colab or Kaggle or local / cloud?

Colab Pro (paid user)

---

3. Number of GPUs used (nvidia-smi)?

None — crash happens even with CPU-only runtime. No GPU involved.

---

4. Which notebook?

Unsloth Official Whisper Notebook

---

5. Paste Unsloth printout with 🦥

Colab runtime crashes instantly before this can print. I never see the :sloth: printout. The notebook dies with:

0.00s - Note: Debugging will proceed. Set PYDEVD_DISABLE_FILE_VALIDATION=1 to disable this validation.

I tried:

import os
os.environ['PYDEVD_DISABLE_FILE_VALIDATION'] = '1'

But the crash still occurs.

---

6. Which trainer? SFTTrainer, GRPOTrainer etc?

No training starts. Crash happens before any trainer is invoked.

---

7. Minimal code to reproduce (no HF token):

from datasets import load_dataset, Audio
import tqdm

# Dataset is small (~15k examples)
dataset = load_dataset("Private dataset", split="train+test")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
dataset = dataset.train_test_split(test_size=0.06)

model.generation_config.language = "<||>"
model.generation_config.task = "transcribe"
model.config.suppress_tokens = []
model.generation_config.forced_decoder_ids = None

def formatting_prompts_func(example):
    audio_arrays = example['path']['array']
    sampling_rate = example["path"]["sampling_rate"]
    features = tokenizer.feature_extractor(audio_arrays, sampling_rate=sampling_rate)
    tokenized_text = tokenizer.tokenizer(example["text"])
    return {
        "input_features": features.input_features[0],
        "labels": tokenized_text.input_ids,
    }

train_dataset = [formatting_prompts_func(example) for example in tqdm.tqdm(dataset['train'], desc='Train split')]
test_dataset = [formatting_prompts_func(example) for example in tqdm.tqdm(dataset['test'], desc='Test split')]

---

✅ What I Already Tried

• Restarted runtime and used fresh notebook
• Set PYDEVD_DISABLE_FILE_VALIDATION = 1 to bypass debug validation
• Tried using .map() instead of list comprehension (same crash)
• Dataset is ~15k examples, not large
• No GPU / CUDA involved
• Crash is instant and consistent before training starts
• Suspect cause might be Unsloth + Whisper generation config or tokenizer/feature interaction

---

Request

Please confirm if:

• unsloth fully supports Whisper + audio datasets
• This crash is known or related to tokenizer/feature_extractor + patched model
• Any fix or stable workaround is available

Thanks for the great work — happy to help debug or test!

[C0deXG]

I would appreciate any help

[rolandtannous]

@C0deXG is this output all you get

0.00s - Note: Debugging will proceed. Set PYDEVD_DISABLE_FILE_VALIDATION=1 to disable this validation.

no other error stack trace?
what is the GPU you initially tried? T4? A100?

[C0deXG]

@C0deXG is this output all you get

0.00s - Note: Debugging will proceed. Set PYDEVD_DISABLE_FILE_VALIDATION=1 to disable this validation.

no other error stack trace? what is the GPU you initially tried? T4? A100?

I tried A100 before it worked fine but A100 was eating too much so I changed GPU to L4 and other GPU then that’s when this issues starting happening.

[rolandtannous]

what is the error trace stack you got ? can you please share?

[github-actions]

Is this issue still important to you? Apologies in advance we might have missed this issue as well. For faster response times, please post on our Reddit server - https://www.reddit.com/r/unsloth or our Discord - https://discord.com/invite/unsloth