diff options
Diffstat (limited to 'ggml/src/ggml-cuda')
| -rw-r--r-- | ggml/src/ggml-cuda/convert.cu | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu index 7089a6df..28b2415b 100644 --- a/ggml/src/ggml-cuda/convert.cu +++ b/ggml/src/ggml-cuda/convert.cu @@ -705,7 +705,7 @@ static __global__ void dequantize_block_iq2_k(const void * __restrict__ vx, dst_ int ib128 = tid/16; // 0 or 1 int il = tid%16; // 0...15 dst_t * y = yy + i*QK_K + 128*ib128 + 2*il; - const float d = (float)x[i].d * 1.025f; //1.0325f; + const float d = (float)x[i].d; const float dl1 = d * (2*((x[i].scales[4*ib128+0] >> 4*(il/8)) & 0xf) - 15); const float dl2 = d * (2*((x[i].scales[4*ib128+1] >> 4*(il/8)) & 0xf) - 15); const float dl3 = d * (2*((x[i].scales[4*ib128+2] >> 4*(il/8)) & 0xf) - 15); @@ -730,7 +730,7 @@ static __global__ void dequantize_block_iq3_k(const void * __restrict__ vx, dst_ int ib128 = tid/16; // 0 or 1 int il = tid%16; // 0...15 dst_t * y = yy + i*QK_K + 128*ib128 + 2*il; - const float d = (float)x[i].d * 1.01f; //1.0125f; + const float d = (float)x[i].d; const uint16_t sh = x[i].scales_h >> (8*ib128 + (il/8)); const float dl1 = d * ((2*((x[i].scales_l[4*ib128+0] >> 4*(il/8)) & 0xf) + 1) * ((sh & 0x01) ? -1 : 1)); const float dl2 = d * ((2*((x[i].scales_l[4*ib128+1] >> 4*(il/8)) & 0xf) + 1) * ((sh & 0x04) ? -1 : 1)); |