497ea2832 Docs: state the batch-invariance guarantee as 1 to 4 columns 2578fdfa3 Fix: keep GGML_CUDA_RESTRICT off the PTQ1_0 mat-vec kernel signature 588318660 Fix: use the mat-vec kernel for bf16 matrices under 64 rows at 2 to 8 columns a55badae9 Add: GGML_CUDA_BATCH_INVARIANT for batch-invariant small-batch kernels a4344b545 Test: add Bonsai 2 projection shapes to the mul_mat perf cases c7551241a Add: dedicated PTQ1_0 mat-vec kernel with full lane utilization 98ea41080 Add: planar-transposed activation layout for the PTQ1_0 mat-vec path