Back to feed

b8690

Apr 7, 2026
Meta/llama.cppCLIvb8690

vulkan: add FA dequant for q4_1, q5_0, q5_1, iq4_nl (#21029)

Add dequantize4() implementations for Q4_1, Q5_0, Q5_1, and IQ4_NL in the flash attention base shader. Register them in the shader generator, pipeline creation, and enable in the scalar/coopmat1 FA support check.

macOS/iOS:

Linux:

Windows:

openEuler: