huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 11:56:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

Potential bug on q4k.

Browse Source
This commit is contained in:
Nicolas Patry
2024-01-05 14:13:22 +01:00
parent fa3ea98ba9
commit 504d0b9ac7
1 changed files with 11 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
candle-core/tests/quantized_tests.rs
View File

@ -632,7 +632,17 @@ fn quantized_matmul_q4k() -> Result<()> {
let rhs = quantized::QTensor::quantize::<BlockQ4K>(&rhs)?; let rhs = quantized::QTensor::quantize::<BlockQ4K>(&rhs)?;
let rhs = quantized::QMatMul::from_qtensor(rhs)?; let rhs = quantized::QMatMul::from_qtensor(rhs)?;
let mm = rhs.forward(&lhs)?; let qmm = rhs.forward(&lhs)?;
let error: f32 = ((&mm - &qmm)?.abs()? / &mm.abs()?)?
.sum_all()?
.to_scalar()?;
let error = error / (m * n) as f32;
assert!(
error < 0.01,
"{error} is too big, shouldn't exceed a few percent. \nGot:{qmm}\nExpected:\n{mm} "
);
assert_eq!(mm.dims(), [m, n]); assert_eq!(mm.dims(), [m, n]);
let dst = mm.flatten_all()?.to_vec1::<f32>()?; let dst = mm.flatten_all()?.to_vec1::<f32>()?;