Avoid double-application in layered preprocessing and loss paths

Summary#

When adjacent layers both perform the same normalization or shift, the bug often survives basic smoke tests because tensor shapes still look valid. Treat label transforms, decoder input preparation, and loss helpers as a single contract, and make one layer solely responsible while testing the exact semantic alignment of inputs and targets.

Problem#

Multiple Hugging Face fixes showed model forward() code manually shifting labels or tokens before calling a generic loss helper that also shifts internally, producing semantically wrong training targets without obvious crashes.

Solution#

Define one canonical owner for sequence shifting, pass arguments by explicit name instead of relying on positional binding, and add regression tests that assert token-at-position t predicts the intended target rather than a double-shifted one.

Failure Modes#

Manual preprocessing duplicates logic already embedded in shared loss utilities
Positional arguments bind to the wrong parameter and bypass intended safeguards
Tests validate loss execution or shape compatibility but not label/target alignment

Sources#

https://github.com/tensorflow/tensorflow/pull/123726
https://github.com/tensorflow/tensorflow/pull/123676
https://github.com/tensorflow/tensorflow/pull/123339
https://github.com/huggingface/transformers/pull/47395
https://github.com/huggingface/transformers/pull/46895
https://github.com/huggingface/transformers/pull/47379
https://github.com/huggingface/transformers/pull/46607
https://github.com/huggingface/transformers/pull/46952
https://github.com/huggingface/transformers/pull/47170
https://github.com/huggingface/transformers/pull/47420
https://github.com/huggingface/transformers/pull/47454
https://github.com/huggingface/transformers/pull/47385
https://github.com/huggingface/transformers/pull/46935
https://github.com/microsoft/ML-For-Beginners/pull/1002
https://github.com/microsoft/ML-For-Beginners/pull/1001
https://github.com/microsoft/ML-For-Beginners/pull/1000
https://github.com/microsoft/ML-For-Beginners/pull/994
https://github.com/microsoft/ML-For-Beginners/pull/991
https://github.com/microsoft/ML-For-Beginners/pull/990
https://github.com/microsoft/ML-For-Beginners/pull/989
https://github.com/ClickHouse/ClickHouse/pull/107586
https://github.com/ClickHouse/ClickHouse/pull/111255
https://github.com/ClickHouse/ClickHouse/pull/111136
https://github.com/ClickHouse/ClickHouse/pull/110706
https://github.com/ClickHouse/ClickHouse/pull/110797
https://github.com/ClickHouse/ClickHouse/pull/108735
https://github.com/ClickHouse/ClickHouse/pull/108678
https://github.com/ClickHouse/ClickHouse/pull/111162
https://github.com/ClickHouse/ClickHouse/pull/111187
https://github.com/ClickHouse/ClickHouse/pull/111039
https://github.com/ClickHouse/ClickHouse/pull/111285
https://github.com/ClickHouse/ClickHouse/pull/110710
https://github.com/ClickHouse/ClickHouse/pull/108991
https://github.com/vercel/turborepo/pull/13424
https://github.com/vercel/turborepo/pull/13423
https://github.com/vercel/turborepo/pull/13419
https://github.com/vercel/turborepo/pull/13418
https://github.com/vercel/turborepo/pull/13420
https://github.com/vercel/turborepo/pull/13408
https://github.com/vercel/turborepo/pull/13401
mined_at: 2026-07-22T00:41:10Z

Sagwan Revalidation 2026-07-22T01:23:32Z#

verdict: ok
note: 중복 시프트 방지 원칙과 명시 인자 권장은 여전히 최신 관행이다