@@ -12134,16 +12134,16 @@ static float ggml_rope_ntkv2_corr_factor(const int n_dims, const float n_rot, co
12134
12134
void ggml_rope_ntkv2_corr_factors(int n_dims, const float freq_base, float factors[4]) {
12135
12135
// Interpolation constants found experimentally for LLaMA (might not be totally optimal though)
12136
12136
// Do not change unless there is a good reason for doing so!
12137
- static const float BETA_0 = 1.75f;
12138
- static const float BETA_1 = 1.25f;
12137
+ static const float BETA_0 = 1.75f;
12138
+ static const float BETA_1 = 1.25f;
12139
12139
static const float GAMMA_0 = 16.0f;
12140
12140
static const float GAMMA_1 = 2.0f;
12141
12141
12142
12142
// start and end correction factors
12143
- factors[0] = maxf(0.0f, floorf(ggml_rope_ntkv2_corr_factor(n_dims, BETA_0, freq_base)));
12144
- factors[1] = minf(n_dims - 1.0f , ceilf(ggml_rope_ntkv2_corr_factor(n_dims, BETA_1, freq_base)));
12145
- factors[2] = maxf(0.0f, floorf(ggml_rope_ntkv2_corr_factor(n_dims, GAMMA_0, freq_base)));
12146
- factors[3] = minf(n_dims - 1.0f , ceilf(ggml_rope_ntkv2_corr_factor(n_dims, GAMMA_1, freq_base)));
12143
+ factors[0] = maxf(0, floorf(ggml_rope_ntkv2_corr_factor(n_dims, BETA_0, freq_base)));
12144
+ factors[1] = minf(n_dims - 1, ceilf(ggml_rope_ntkv2_corr_factor(n_dims, BETA_1, freq_base)));
12145
+ factors[2] = maxf(0, floorf(ggml_rope_ntkv2_corr_factor(n_dims, GAMMA_0, freq_base)));
12146
+ factors[3] = minf(n_dims - 1, ceilf(ggml_rope_ntkv2_corr_factor(n_dims, GAMMA_1, freq_base)));
12147
12147
}
12148
12148
12149
12149
static void ggml_compute_forward_rope_f32(
0 commit comments