Systems

Cross-cutting subsystems that span multiple packages or warrant a dedicated explanation. The Packages view is "what's in each directory"; this view is "how do the moving parts work."

Page	Topic
Prompt execution	The path from POST /prompt to running PyTorch — queue, validate, DAG, cache, run
Sampling pipeline	What KSampler actually does: schedule → sigmas → denoise loop → CFG → output
Conditioning and CLIP	Text encoders, tokenization, the CLIP type, conditioning lists
Model management	VRAM scheduling, dtype/attention probes, ModelPatcher, async offload, DynamicVRAM
LoRA and hooks	Low-rank adapters, scoped hooks, weight adapters, patcher extensions
Quantization	QuantizedTensor, MixedPrecisionOps, FP8 layers, scaling parameters
Asset system	Content-hashed asset DB, scanner, seeder, ingest pipeline, REST API
Custom-nodes system	How custom nodes load, the V1/V3 schemas, ComfyUI-Manager interaction

For deployable units (apps, services), there is only one runtime — the ComfyUI process — so we use the systems lens rather than apps. See Architecture for the bird's-eye view.