Accelerator Quota Tracking
Kaggle provides standard weekly quotas for hardware accelerators:
- GPU Quota: 30 hours per rolling 7-day window (shared between T4 and P100 instances)
- TPU Quota: 20 hours per rolling 7-day window (for TPU v3-8 instances)
Because Kaggle's public API does not expose an endpoint to query live quota balances, kagglex computes usage locally from run records stored in ~/.kagglex/runs.json.
Viewing Quota Dashboard
Inspect your rolling accelerator usage with kagglex quota:
Sample output:
Kaggle Accelerator Quota Usage (Past 7 Days):
ACCELERATOR USED HOURS QUOTA LIMIT REMAINING UTILIZATION RUNS
GPU 12.45 hrs 30.00 hrs 17.55 hrs 41.5% 4
TPU 2.30 hrs 20.00 hrs 17.70 hrs 11.5% 1
Recent Runs in Quota Window:
- 2026-09-18 14:20:10 UTC [GPU: t4-2x] username/exp-bert-finetune (duration: 3.50 hrs, status: complete)
- 2026-09-17 09:12:44 UTC [GPU: t4-2x] username/exp-lora-training (duration: 8.95 hrs, status: complete)
Customizing Rolling Windows and Limits
You can adjust the time window or custom quota thresholds using CLI flags:
# Check past 14 days with custom 40-hour limit
kagglex quota --days 14 --gpu-limit 40.0 --tpu-limit 30.0
Low Quota Pre-Flight Warnings
When submitting a new experiment via kagglex run, kagglex automatically estimates remaining quota. If your available balance falls below 2 hours, a warning is logged before execution:
Centralized Storage
All run records are stored in ~/.kagglex/runs.json. This ensures that runs submitted from different repositories on your machine are aggregated under your account's single global quota.