W4A4 (4-bit weights and 4-bit activations, general term)

GROUP SIZE LEFT NULL 2026-10-04: this is the GENERAL TERM W4A4, a class spanning MXFP4 (32), NVFP4 (16), INT4 grouped (64/128) and Atom's 128-group variant. The four concrete records in this repo already carry their own verified values. Putting a number on the umbrella term would give one of those widths to all of them. A category record, not a single scheme: 'W4A4' in the wild usually means one of three very different things, and conflating them is the single most common error in quantization blog posts. (1) Integer W4A4 with INT4 tensor cores - does not exist on shipping NVIDIA or AMD parts; Hopper and older have no INT4 path at all and Ampere-era 'INT4' marketing was sparsity, not 4-bit multiply. (2) FP4 W4A4 - this IS native, on Blackwell, and that is exactly what nvfp4 is. (3) Simulated/fake W4A4 in research papers - SpinQuant, QuaRot, QuIP# - which measure quality but buy no speed without custom kernels. Note TensorRT-LLM's own naming: it ships W4A16 and W4A8 variants of AWQ and GPTQ and no W4A4 row at all (quant-trtllm-quantization), which is the honest state of the art. The distinction to carry away: '4-bit weights with 16-bit compute' (AWQ, GPTQ, NF4, EXL2, all GGUF k-quants) is a memory optimization; '4-bit weights and 4-bit activations with hardware support' currently exists only for FP4 on Blackwell.