component_quantization replaced the global flag instead of refining it, so
naming ONE component switched load_in_4bit off for every other one. Measured
2026-10-29 on LTX-2.5 (load_in_4bit=True, map naming transformer and
text_encoder): ten of its twelve components loaded at full precision, among
them connectors at 11.82 GB and prompt_enhancer at 9.51 GB -- 21.3 GB of a
25.3 GB card before the quantized transformer arrived, which is why the
resident attempt OOMed on exactly those components.

The candidate list was the second half of it. Written for unet/SD-era
pipelines, it matched only transformer*/unet/text_encoder*, so connectors and
prompt_enhancer could not be reached by the global flag either. A modern video
pipeline carries several multi-GB Linear-heavy parts under other names.

Both builders are fixed together -- build_pipeline_quant_config and the
incremental component cache -- because they had the same shape, and a cached
build that quantizes differently from a direct load is its own bug. An
explicit map entry still wins, including "none"/"full", which is how a
component is opted out.

Wan2.2-TI2V-5B shows the size of this: load_in_4bit=True with a map naming
only text_encoder and vae, so its transformer was never quantized at all --
a 5B model measuring 40.0 GB on a 25.3 GB card.

Separately, the message printed when eviction cannot make room claimed
"Remaining models are busy or VRAM is held elsewhere" without having
established either, and it was usually false. Over one production day, 10 of
12 of these were a model asking for more than the whole card (25.4 to 40.0 GB
against 25.3 GB), four of them printing "freed 0.0 GB (now 25.3 GB free)" --
the card was already empty. "Skipping eviction of ... still busy" never
appeared once: eviction was never blocked by a busy model.

That wording sent an investigation at the scheduler when the causes were a
model larger than the GPU and components that were never quantized. It now
names which of the three cases applies, and for the common one points at the
two things that actually help.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014T3rZozi1Ehk96NG1WesgC