TorchedUp
LearnBetaProblemsSystem DesignSoonPremium
TorchedUp
LearnBetaProblemsSystem DesignSoonPremium
Learn/Debug Gauntlet
🔍

Debug Gauntlet

Production ML code breaks for the same handful of reasons over and over: shape mismatches, wrong reduction axes, forgotten train/eval mode switches, broken residual connections, missed numerical stability tricks. This track gives you working code that doesn't, and asks you to find the why. No bug-count hints — you finish when the tests pass.

11 problems · suggested order

  1. ○1#59Debug: Attention Scaleeasy
  2. ○2#60Debug: BatchNorm at Inferenceeasy
  3. ○3#63Debug: Loss Reduction (sum vs mean)easy
  4. ○4#131Dataloader Bottleneck Gapmedium
  5. ○5#128Detect Gradient Explosionmedium
  6. ○6#62Debug: Gradient Accumulationmedium
  7. ○7#61Debug: RNN Vanishing Gradientmedium
  8. ○8#129Detect Dead Networkmedium
  9. ○9#64Debug: Broken Transformer Encoder (5 Bugs)hard
  10. ○10#100Debug Decoder I (NumPy)easy
  11. ○11#104Debug Decoder I (PyTorch)easy
Tracks are curated by hand. The order above is the suggested learning progression — feel free to skip around if you already know a topic.

© 2026 TorchedUp. All rights reserved.

ChangelogContact UsTerms of ServicePrivacy PolicyRefund Policy