Add TopNSigmaLogitsWarper and top_n_sigma generation config support

[ezylopx5]

Problem

Transformers currently provides sampling filters such as top-k, top-p, min-p, and top-h, but does not include top-n-sigma sampling from "Top-nσ: Not All Logits Are You Need".

This makes it harder to use a temperature-invariant truncation rule based directly on pre-softmax logits.

Closes #39850.

Root Cause

There is no logits warper implementing a dynamic threshold of:

max(logits) - n_sigma * std(logits)

and no corresponding GenerationConfig parameter routed through GenerationMixin._get_logits_processor().

Fix

This PR adds TopNSigmaLogitsWarper and integrates it into generation config and sampling flow.

Code changes

• Add TopNSigmaLogitsWarper to generation logits processors.
• Add top_n_sigma field and validation in GenerationConfig.
• Add top_n_sigma sampling integration in GenerationMixin.
• Export TopNSigmaLogitsWarper from generation and top-level package init files.
• Add tests for behavior, validation, edge cases, and config serialization.

Files changed

• src/transformers/generation/logits_process.py
• src/transformers/generation/configuration_utils.py
• src/transformers/generation/utils.py
• src/transformers/generation/__init__.py
• src/transformers/__init__.py
• tests/generation/test_logits_process.py
• tests/generation/test_configuration_utils.py

Validation

Ran locally in editable install environment:

python -m pytest tests/generation/test_logits_process.py -k "top_n_sigma" -x -v
python -m pytest tests/generation/test_configuration_utils.py -k "top_n_sigma" -x -v
python -m pytest tests/generation/test_logits_process.py -x -v
python -m pytest tests/generation/test_configuration_utils.py -x -v

Results:

• test_logits_process.py: 50 passed
• test_configuration_utils.py: 34 passed, 5 skipped

Also ran style checks:

make style

Risks

• Behavior is additive and opt-in (top_n_sigma defaults to None).
• Existing generation paths remain unchanged unless top_n_sigma is explicitly set.
• Includes validation for non-negative values to fail early on invalid input.

Out of Scope

• Benchmarking quality/performance on large models.
• Interactions benchmarking vs top-p/min-p/top-h beyond functional correctness.

[Rocketknight1]

Please don't ask your code agent to find old issues and "fix" them!

[ezylopx5]

@Rocketknight1 thanks for the feedback, and sorry for the noise.You are right that I handled the process poorly. I moved too fast without prior discussion on the issue, and that made this look agent-first.
For transparency: I did use an agent for implementation support, but I personally reviewed the diff, ran the tests locally, fixed the failing repository-consistency check, and re-ran the relevant test files before opening the PR. So the technical validation was real on my side, but I understand that process and trust come first.
and i have closed the PR and deleted the branch. Going forward I will only propose changes after prior discussion on active issues and with smaller, clearly scoped contributions.
I appreciate the guidance and your time.