huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-14 09:57:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Cuda quantization padding fix.

Browse Source
This commit is contained in:
laurent
2024-09-25 23:35:16 +02:00
parent fd3b53f48b
commit 5221146cfa
2 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
candle-core/src/quantized/cuda.rs
View File

@ -34,7 +34,10 @@ fn ceil_div(p: usize, q: usize) -> usize {
}
fn pad(p: usize, q: usize) -> usize {
ceil_div(p, q) * q
// Overallocate by q rather than just padding by q as this should pad the last row
// and we don't have enough information here to know how many elements to add :(
// ceil_div(p, q) * q
p + q
}
fn quantize_q8_1(
@ -439,7 +442,7 @@ impl QCudaStorage {
}
_ => crate::bail!("only f32 can be quantized"),
};
let src_len = src.len();
let src_len = pad(src.len(), MATRIX_ROW_PADDING);
let src = crate::Storage::Cpu(crate::CpuStorage::F32(src));
let mut qcpu_storage = crate::Device::Cpu.qzeros(src_len, self.dtype)?;
qcpu_storage.quantize(&src)?;

2
candle-core/src/quantized/utils.rs
View File

@ -18,7 +18,7 @@ pub(super) fn group_for_quantization<'a, 'b, T: super::k_quants::GgmlType>(
let actual_blocks = ys.len();
// Validate that the input is the right size
if expected_blocks != actual_blocks {
if actual_blocks < expected_blocks {
crate::bail!("quantize {dtype:?}: expected {expected_blocks} blocks but only {actual_blocks} were provided!")
}