[stanleys12]

Summary

A standalone LTX-2.5 video or audio VAE file (for example ltx-2.5-22b_video_vae_bf16.safetensors) got imported as an SD1.x VAE. Its causal convs store weights as decoder.conv_in.conv.weight, so the probe couldn't read a latent width and fell back to the filename. vae maps to SD1 there.

The checkpoint VAE probe now rejects state dicts without a 2-D decoder.conv_in.weight, so those files stay Unknown. I also removed the name-only fallback in _get_base_or_raise since nothing can reach it anymore.

Related Issues / Discussions

Related to #9777
Related to #9780

QA Instructions

• uv run --no-sync pytest tests/backend/model_manager/configs -n 4 --ignore tests/backend/model_manager/configs/test_ministral3_encoder_config.py: 894 passed, 1 skipped.
• uv run --no-sync pytest tests/backend/model_manager/configs/test_ministral3_encoder_config.py: 11 passed. I ran it on its own because its big temp safetensors files filled my nearly full disk under -n 4 (ENOSPC).
• uv run --no-sync pytest tests/backend/model_manager/configs/test_vae_base_identification.py -k decoder_conv_in with vae.py from main: 2 failed (both cases resolve to SD1). They pass with the fix.
• ruff check and ruff format --check are clean on vae.py and test_vae_base_identification.py.
• Not run locally: tests/model_identification (no LFS fixtures here).

Review

No material findings.

Compatibility / Rollout

No migration. Models already registered as SD1 VAEs from earlier imports keep their entry until removed and re-scanned. No API or persisted-schema changes.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)

AI use

I drafted this with an AI coding assistant and checked it by running the tests listed above.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}