Feature matrix

	CPU (AVX/AVX2)	CPU (ARM NEON)	Metal	CUDA	ROCm	SYCL	Vulkan	Kompute
K-quants	✅	✅	✅	✅	✅	✅	✅ 🐢⁵	🚫
I-quants	✅ 🐢⁴	✅ 🐢⁴	✅ 🐢⁴	✅	✅	Partial¹	✅ 🐢⁴	🚫
Parallel Multi-GPU⁶	N/A	N/A	N/A	✅	✅	Sequential only	Sequential only	❓
K cache quants	✅	✅	✅	✅	✅	❓	✅	🚫
MoE architecture	✅	✅	✅	✅	✅	❓	✅	🚫
Flash Attention	✅	✅	✅	✅	✅	❓	Partial⁷	🚫

• ✅: feature works
• 🚫: feature does not work
• ❓: unknown, please contribute if you can test it yourself
• 🐢: feature is slow
• ¹: IQ3_S and IQ1_S, see #5886
• ²: Only with -ngl 0
• ³: Inference is 50% slower
• ⁴: Slower than K-quants of comparable size
• ⁵: Generally the CUDA or ROCM backends are faster, though there are cases where Vulkan has faster text generation. See #10879 for benchmarks.
• ⁶: By default, all GPU backends can utilize multiple devices by running them sequentially. The CUDA code (which is also used for ROCm via HIP) also has code for running GPUs in parallel via --split-mode row. However, this is optimized relatively poorly and is only faster if the interconnect speed is fast vs. the speed of a single GPU.
• ⁷: This is only implemented for coopmat2. Otherwise the Flash Attention ops will run slowly on CPU instead.