QwenImage21Cache: cache placement is not weight quantization

Understand cache placement and storage precision as separate controls from model-file quantization.

Updated 2026-09-24Published revision: 4Source baseline: not recorded

Source and status

QwenImage21Cache is an experimental core node in the reviewed v0.37.0 implementation. It accepts a model, clones its configuration and returns a model with cache options. Q04

Placement

The device choices are auto, gpu, cpu and off. Disabling the cache recomputes the prefix at each step and can help isolate a cache-related problem; it is not an automatic speed optimization. CPU cache placement does not convert the entire generation pipeline into CPU inference.

Storage precision

The dtype choices are default, int8 and int4. They describe KV-cache storage, not a conversion of the model checkpoint into GGUF. Keep diffusion weight precision, encoder precision and cache precision as separate fields in a test record. Q04

Controlled comparison

Keep the references, prompt, dimensions, seed and software unchanged. Record auto/default first, then change one option. Compare peak device memory, host memory, latency and output differences independently. A failed run is evidence about that combination, not every GPU. This site has not executed the benchmark and does not attach a predicted speedup percentage.

Sources and verification

Source and code review on 2026-09-22. No GPU run, weight download or quality benchmark. Procedures are editorial guidance, not a confirmed diagnosis of your environment.

Q04 · ComfyUI 0.37.0 Qwen nodes