mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-09-11 04:56:56 +02:00
Julian Pscheid
3ff67eb43d
The Imagination proprietary Vulkan compiler returns VK_ERROR_UNKNOWN from vkCreateComputePipelines for every dequant mul_mat_vec shader built with the subgroup-only reduction that requires a subgroup size >= 16. That covers the k-quants, the i-quants, TQ2_0, MXFP4 and NVFP4. ggml rethrows, so the first generated token of any such model kills the process. Reproduced on a Pixel 11 Pro (PowerVR C-Series CXTP-48-1536 MC1, driver 1.662.3024, subgroup size 128, min 32, max 128). The failure is independent of subgroup size: 32, 64 and 128 all fail, as does dropping the full-subgroups flag and the required-subgroup-size pNext. The legacy quants, which use the plain subgroup reduction, compile and run fine. The shared-memory reduction variant compiles and matches the CPU reference for q2_K, q3_K, q4_K, q5_K and q6_K. The hybrid variant also compiles but costs 27% of token throughput (3.78 vs 5.20 t/s on Qwen3.5-2B-Q4_K_M).