huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-15 10:26:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

Clippy 1.86 fixes for cuda. (#2868)

Browse Source
This commit is contained in:
Laurent Mazare
2025-04-05 15:45:35 +02:00
committed by GitHub
parent bc33df77e1
commit 338f6a102e
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
candle-core/src/quantized/cuda.rs
View File

@ -73,7 +73,7 @@ fn dequantize_f32(
elem_count: usize,
dev: &CudaDevice,
) -> Result<CudaStorage> {
let nb = (elem_count + 255) / 256;
let nb = elem_count.div_ceil(256);
let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
GgmlDType::Q4_0 => ("dequantize_block_q4_0_f32", false, 32, nb),
GgmlDType::Q4_1 => ("dequantize_block_q4_1_f32", false, 32, nb),
@ -133,7 +133,7 @@ fn dequantize_f16(
elem_count: usize,
dev: &CudaDevice,
) -> Result<CudaStorage> {
let nb = (elem_count + 255) / 256;
let nb = elem_count.div_ceil(256);
let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
GgmlDType::Q4_0 => ("dequantize_block_q4_0_f16", false, 32, nb),
GgmlDType::Q4_1 => ("dequantize_block_q4_1_f16", false, 32, nb),
@ -278,8 +278,8 @@ fn mul_mat_vec_via_q8_1(
// https://github.com/ggerganov/llama.cpp/blob/facb8b56f8fd3bb10a693bf0943ae9d69d0828ef/ggml-cuda/mmvq.cu#L98
let (nblocks, nwarps) = match b_size {
1 => (nrows as u32, 4),
2..=4 => ((nrows as u32 + 1) / 2, 4),
5..=8 => ((nrows as u32 + 1) / 2, 2),
2..=4 => ((nrows as u32).div_ceil(2), 4),
5..=8 => ((nrows as u32).div_ceil(2), 2),
_ => crate::bail!("unexpected bsize {b_size}"),
};
let cfg = cudarc::driver::LaunchConfig {