Rename Kimi resize helper for NaViT patch-grid semantics

[aswanth-07]

Summary

• rename resize_to_patch_budget to resize_to_navit_patch_grid
• preserve Kimi's existing NaViT resize and padding geometry
• describe max_patches as a pre-padding patch budget
• update all model, numerical-test, and unit-test call sites
• add a focused test showing that merge-grid padding can raise the final patch count above the pre-padding budget

Rationale

The helper follows Kimi's navit_resize_image behavior. It applies the total and per-side limits while calculating the resize scale, then pads the result to the merge grid. That padding can make the final raw-patch count exceed max_patches.

The previous helper name implied a strict final budget. The new name describes the operation without changing the image geometry used by Titan's Kimi path.

Related: #4093

Validation

• rebased onto current main at f81fc0afe
• py -3.13 -m pytest tests/unit_tests/test_mm_dataset_preprocessing.py -q with plugin autoload disabled: 10 passed
• Torchtitan pre-commit hooks for the five changed files: whitespace, AST, merge-conflict, large-file, EOF, license, flake8, ufmt, pydoclint, and codespell checks passed
• py -3.13 -m compileall -q for the renamed model and test call sites: passed
• git diff --check: passed

The optional grain dependency is unavailable on this host, so test_qwen_multimodal_data.py is left to dependency-complete CI. Local Pyrefly still expects the configured ../pytorch source checkout, so CI provides type validation.

AI assistance disclosure

OpenAI Codex assisted with repository auditing, implementation, test execution, and drafting this pull request. The updated code has not yet received a human code review.

[pytorch-bot]

Workflows were awaiting approval. CI has now been triggered for the ciflow labels on this PR.

[shuhuayu]

hi @aswanth-07, the current approach is implemented to match the kimi's navit_resize_image (https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/media_utils.py#L85-L133), which does not compress the number of patches as this pr did. In this case, i would suggest renaming resize_to_patch_budget to resize_to_navit_patch_grid, and describe max_patches as a “pre-padding patch budget” rather than a strict maximum number of patches. if we implement the new function as in this pr, it is faithful to the name but not used in titan yet. what do you think?

[aswanth-07]

Thanks, that makes sense. I checked the behavior against Kimi's navit_resize_image and updated the PR accordingly. The post-padding compression is gone, so Titan keeps the existing Kimi geometry.

I renamed the helper to resize_to_navit_patch_grid across all call sites, documented max_patches as a pre-padding budget, and added a focused test showing that padding may take the final grid above that budget. I also closed #4093 since the reported behavior is expected for this path.

[shuhuayu]

looks good to me, can you also refactor the kimi k3 model?