ci : fix CUDA builds

.github/workflows/build.yml

@@ -459,7 +459,7 @@ jobs:
           path: build/bin/${{ matrix.build }}
 
   windows-cublas:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     strategy:
       matrix:

ggml-cuda/mmvq.cu

@@ -75,7 +75,7 @@ static __global__ void mul_mat_vec_q(
             tmp[j][i] = warp_reduce_sum(tmp[j][i]);
         }
 
-        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) {
+        if (threadIdx.x < rows_per_cuda_block) {
             dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
         }
     }