This almost always indicates an environment or preprocessing mismatch.

Training pipelines often include steps—normalization, tokenization, feature encoding—that are not replicated exactly in production. Even small differences in default parameters can cause large output changes.

Verify that the same preprocessing code runs in both environments, ideally by packaging it with the model artifact. Also confirm that model weights, framework versions, and inference settings match training.

Another subtle issue is switching from GPU to CPU inference without testing numerical stability.

Common mistakes:

• Reimplementing preprocessing instead of reusing it

• Different library versions in production

• Using training-time batch behavior during inference

Treat preprocessing as part of the model, not an external dependency.