vLLM quantization docs - Supported Hardware table (Implementation x architecture)

The Implementation x Hardware compatibility table, extracted verbatim. Columns Volta(Turing etc.) rows: AWQ = no Volta, yes Turing/Ampere/Ada/Hopper, yes Intel GPU, yes x86 CPU, NO AMD GPU, NO Arm CPU. GPTQ = yes everywhere including Volta. Marlin (GPTQ/AWQ/FP8/FP4) = NO Volta, Turing marked with a footnote star, yes Ampere/Ada/Hopper, no AMD/Intel/CPU. llm-compressor INT8 W8A8 = no Volta, yes Turing onward, yes x86 CPU. llm-compressor INT8 W4A8 = NO NVIDIA GPU AT ALL, yes Arm CPU only. llm-compressor FP8 W8A8 = Ada and Hopper only, nowhere else. bitsandbytes = Volta through Hopper. DeepSpeedFP and GGUF = Volta through Hopper, no CPU/AMD/Intel. Footnotes: Volta=SM7.0, Turing=SM7.5, Ampere=SM8.0/8.6, Ada=SM8.9, Hopper=SM9.0; the Turing star on Marlin means 'Turing does not support Marlin MXFP4'. Accessed 2026-10-04.