fix(image/ocr): import torchvision lazily in deepseek_ocr (#5189)

[Anai-Guo]

Problem

xinference/model/__init__._install() imports every model family at actor-creation
time. That chain reaches model/image/__init__.py → .ocr → ocr/__init__.py →
deepseek_ocr.py, which does a module-level from torchvision import transforms.

Net effect: torchvision is imported into every model subprocess — audio models,
LLMs, anything — regardless of whether the model has anything to do with OCR.

Why this is the odd one out

Every other torchvision consumer in the tree already imports it lazily, inside the
function that needs it:

file	import site
model/llm/transformers/multimodal/intern_vl.py	inside method (line ~113)
model/llm/transformers/multimodal/ovis2.py	inside method (line ~99)
model/video/diffusers.py	inside method (line ~393)
model/image/ocr/deepseek_ocr.py	module level ← this PR

So this looks like a plain oversight rather than a deliberate choice.

Change

Move the import into the only two scopes that use transforms:

• normalize_transform()
• BasicImageTransform.__init__()

That is the whole diff — 5 lines, no behavior change for OCR users (the import
just happens on first use instead of at module import).

Relation to #5189

This is the "additionally" item called out at the end of #5189, filed as its own
minimal change. It does not claim to fix the root cause of #5189 (the
_create_subpool / _prepare_virtual_env ordering — that reorder touches the
launch flow and deserves a separate, maintainer-directed discussion).

What it does do is remove one eager host torchvision import from subprocesses
whose virtualenv installs its own pinned torchvision, which the issue identifies as
what aggravates the operator torchvision::nms does not exist failure variant for
audio models that never use torchvision at all.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request refactors the imports in deepseek_ocr.py by moving torchvision to local scopes for lazy loading. The reviewer suggested a minor optimization to place the lazy import after the initial None checks in normalize_transform to prevent unnecessary import overhead and potential ImportErrors when the library is not installed.

[qinxuye]

LGTM