Technical
W4A4 and W4A8: The New Quantization Toolkit for Squeezing More Out of a Checkpoint
ยท RenderBob team
ComfyUI's Quantization Toolkit, formerly the INT8 Toolkit, adds native W4A4, W4A8 and W8A8 quantisation, including LoRA support. Here is what the notation means and how to pick a mode.
ComfyUI's Quantization Toolkit, formerly published as the INT8 Toolkit, brings native support for quantising diffusion models to W4A4, W4A8 or W8A8, loading pre-quantised checkpoints, applying quantisation-aware LoRAs, and tuning low-VRAM inference. It is a more systematic, mode-based update to the quantisation stack this blog has covered piecemeal with FP8 and NVFP4.
Decoding the notation
W4A4 means 4-bit weights and 4-bit activations. W8A8 means 8-bit weights and 8-bit activations. W4A8 is an asymmetric middle ground: 4-bit weights paired with 8-bit activations. Weights are the model's learned parameters, static and quantised once. Activations are the intermediate values flowing through the network during a forward pass, quantised on the fly. Lower bit-widths save more VRAM and run faster, and they push harder on quality. Activations are typically more sensitive to aggressive quantisation than weights, which is why W4A8 exists as a deliberate middle option.
How the toolkit tells you to choose
Start with int8, the simplest, most broadly compatible W8A8 path, for general use. Move to int4_mixed, an architecture-aware mix of W4A4 and W8A8 that applies the more aggressive format only where the model can tolerate it, when memory pressure justifies it. Treat w4a8 as experimental: it is a distinct kernel format (asymmetric 4-bit weights with ConvRot INT8 activations, falling back to W8A8 on incompatible shapes), not a preset inside int4_mixed. Test it deliberately rather than assuming it slots in as a drop-in upgrade.
LoRA compatibility is the studio-critical detail. The toolkit supports applying LoRAs to an already-quantised model, with standard, stochastic-requantised, and dynamic-runtime modes for how that interaction is handled. A studio that has invested in character or brand LoRAs needs those adapters to keep working when the base model is running quantised for VRAM reasons. A toolkit that treats this as a first-class path removes a real risk from adopting aggressive quantisation on production checkpoints.
Quantisation mode is a quality-versus-VRAM trade-off that should be made once, benchmarked against real production shots, and applied consistently. Pin the tested mode in the studio toolkit rather than rediscovering it on every new project.
More from the blog
- Block Swap, Explained: The Technique Behind Almost Every Low-VRAM Video Workflow
Scroll through almost any low-VRAM video workflow in ComfyUI and you will find block swap doing the heavy lifting under a different node name each time.
- The FP8 File That Silently Runs at FP16: A VRAM Gotcha Worth Knowing
A checkpoint saved in FP8 does not guarantee it runs at FP8. Some loaders and backends upcast the weights, so the file is small and the memory it consumes is not.