CUDA: replace GGML_CUDA_F16 with CUDA arch checks (#15433)

This commit is contained in:
Johannes Gäßler
2025-08-20 16:58:49 +02:00
committed by GitHub
parent fec9519802
commit 7a6e91ad26
12 changed files with 32 additions and 86 deletions

View File

@@ -32,7 +32,7 @@ static __global__ void k_get_rows(
const int y_offset = qr == 1 ? 1 : qk/2;
// dequantize
dfloat2 v;
float2 v;
dequantize_kernel(src0_row, ib, iqs, v);
dst_row[iybs + iqs + 0] = ggml_cuda_cast<dst_t>(v.x);