Fix/model cache Qwen/CogView4 cancel repair

[JPPhoto]

Summary

Fixes a canceled-run regression where the next Qwen-based or CogView4-based text encoder invocation could fail with a CPU/CUDA device mismatch.

The change adds a repair hook to partially loaded cached models so required non-autocast tensors are restored to the compute device before inference, while keeping model-cache VRAM accounting consistent. The invocation paths now use that repair flow, and a focused cached-model regression test was added.

Related Issues / Discussions

https://discord.com/channels/1020123559063990373/1049495067846524939/1481290964587778273

QA Instructions

1. Start a generation that uses the Qwen-based text encoder path on a low-VRAM system.
2. Cancel it while the text encoder is running.
3. Start another generation with the same model.
4. Verify the next run no longer fails with Expected all tensors to be on the same device.

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[lstein]

Works. I don't see any obvious code problems.

[Pfannkuchensack]

Code review

Found 1 issue:

1. Invocation node version strings are not bumped in any of the three modified nodes. PR perf(flux2): optimize cache locking in Klein encoder to fix #7513 #8863 (which modified the same flux2_klein_text_encoder.py node) required a version bump before merging — reviewer @Pfannkuchensack requested it and @lstein blocked merge until it was done. PR Fix/model cache Qwen/CogView4 cancel repair #8959 makes behavioral changes (adds repair hooks with runtime side-effects and warning logs) to all three nodes but leaves their version strings unchanged:

InvokeAI/invokeai/app/invocations/cogview4_text_encoder.py

Lines 23 to 25 in 385462f

	category="conditioning",
	version="1.0.0",
	classification=Classification.Prototype,

InvokeAI/invokeai/app/invocations/flux2_klein_text_encoder.py

Lines 48 to 50 in 385462f

	category="conditioning",
	version="1.1.1",
	classification=Classification.Prototype,

InvokeAI/invokeai/app/invocations/z_image_text_encoder.py

Lines 37 to 39 in 385462f

	category="conditioning",
	version="1.1.0",
	classification=Classification.Prototype,

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[JPPhoto]

Code review

Found 1 issue:

1. Invocation node version strings are not bumped in any of the three modified nodes.

Internal changes to the node that don't alter its functionality do not require a node version bump, in general. If you are fundamentally changing how a node works, a version bump is a good idea to alert users. If you're altering its external interface (inputs/outputs), it's essential.

[Pfannkuchensack]

Works good.