Diagnosing Kaggle Notebook Accelerator Failures: Quota, Session Limits, and OOM Errors

Recognizable condition

When you try to start a GPU or TPU accelerator in a Kaggle Notebook you may see one of the following:

• A banner that says your weekly accelerator quota has been exceeded.
• The notebook disconnects or stops responding after a few hours of training.
• The session ends abruptly after a period of inactivity.
• Training stops with a CUDA out‑of‑memory (OOM) error.

These symptoms point to common failure modes on Kaggle’s free accelerator environment.

Cause / diagnostic table

Symptom	Likely cause family
Quota‑exceeded banner	Weekly accelerator‑hour quota exhausted
Session dies after ~9 h (GPU/TPU) or ~12 h (CPU)	Maximum session length reached
Notebook disconnects after idle period	Idle timeout enforced by the platform
CUDA OOM error	Model/batch size exceeds GPU memory
Accelerator fails to attach	Missing phone verification or competition‑specific internet‑off rule

Ordered checks

1. Read the exact error banner – Note the wording (quota, session limit, idle, OOM, verification). This tells you which check to perform next.
2. Open the notebook Settings → Accelerator panel – Look at the quota meter that shows used vs. available accelerator hours for the current week. The meter updates in near‑real time.
3. Confirm phone verification – Go to Account → Settings → Phone Verification. Accelerator attachment and external internet are blocked until verification is complete.
4. Check competition rules – If you are in a competition, open the competition page and look for “Internet off” or “External data bans”. These rules override the default notebook behavior.
5. Compare planned run length to session caps – Estimate how many hours your training will need. If it exceeds the typical session length (≈9 h for GPU/TPU, ≈12 h for CPU), you will need to checkpoint and restart.
6. Inspect output size – In the notebook sidebar, check the size of /kaggle/working. Kaggle imposes a size cap on this directory (historically ~20 GB). Exceeding it can cause silent failures or forced termination.

Fixes tied to findings

Quota exhausted

• Wait for the weekly reset (UTC midnight Sunday/Monday).
• Switch to the other accelerator type (e.g., use TPU if GPU quota is spent, or vice‑versa) if your code supports it.
• Run the workload on CPU only (no accelerator) for testing or lightweight experiments.

Idle or session‑length disconnects

• Replace long interactive sessions with a committed headless run: click Save & Run All and close the notebook. The run continues until completion or a hard limit.
• Insert periodic checkpoint saves to /kaggle/working and design your script to resume from the latest checkpoint after a restart.

CUDA OOM

• Reduce batch size or sequence length.
• Enable mixed‑precision training (torch.cuda.amp or TensorFlow tf.keras.mixed_precision).
• Use gradient accumulation to simulate larger batches without exceeding memory.

Attach blocked (verification or competition rule)

• Complete phone verification in account settings.
• If the competition forbids internet, pre‑package all required pip libraries and datasets as Kaggle Datasets or Model attachments, then disable external calls.

Escalation criteria

Escalate when:

• The quota meter shows remaining hours but the accelerator still fails to attach.
• The same error appears across fresh sessions, different browsers, or after clearing cache.
• Your workload genuinely exceeds the free limits (e.g., needs >30 h GPU/week or >20 GB persistent output).

In these cases, post a detailed description (including the exact banner text, quota meter reading, and notebook ID) to the Kaggle forums or contact support. For sustained heavy training, consider moving to a local GPU or a paid cloud instance.