huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-18 19:47:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

neon optimized q8k multiplication. (#1021)

Browse Source 
* neon optimized q8k multiplication.

* Bugfixes.

* simdification.
This commit is contained in:
Laurent Mazare
2023-10-02 23:26:34 +01:00
committed by GitHub
parent cddfc3944c
commit 7670fe7d1f
2 changed files with 36 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
candle-core/src/quantized/k_quants.rs
View File

@ -1756,14 +1756,18 @@ impl GgmlType for BlockQ8K {
const BLCK_SIZE: usize = QK_K;
type VecDotType = BlockQ8K;
#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "neon")]
return super::neon::vec_dot_q8k_q8k(n, xs, ys);
Self::vec_dot_unopt(n, xs, ys)
}
fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
let qk = QK8_0;
if n % QK8_0 != 0 {
crate::bail!("vec_dot_q8_0_q8_0: {n} is not divisible by {qk}")
let qk = QK_K;
if n % QK_K != 0 {
crate::bail!("vec_dot_q8k_q8k: {n} is not divisible by {qk}")
}
// Generic implementation.