huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 19:58:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Avoid some overflows on wasm32. (#968)

Browse Source
This commit is contained in:
Laurent Mazare
2023-09-26 11:15:38 +01:00
committed by GitHub
parent 2dd43d6cdd
commit 4abc1ea34d
2 changed files with 14 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@ -135,7 +135,13 @@ pub fn qtensor_from_ggml(
dims: Vec<usize>, dims: Vec<usize>,
) -> Result<super::QTensor> { ) -> Result<super::QTensor> {
let tensor_elems = dims.iter().product::<usize>(); let tensor_elems = dims.iter().product::<usize>();
let size_in_bytes = tensor_elems * ggml_dtype.type_size() / ggml_dtype.blck_size(); let blck_size = ggml_dtype.blck_size();
if tensor_elems % blck_size != 0 {
crate::bail!(
"the number of elements {tensor_elems} is not divisible by the block size {blck_size}"
)
}
let size_in_bytes = tensor_elems / blck_size * ggml_dtype.type_size();
match ggml_dtype { match ggml_dtype {
GgmlDType::F32 => from_raw_data::<f32>(raw_data, size_in_bytes, dims), GgmlDType::F32 => from_raw_data::<f32>(raw_data, size_in_bytes, dims),

9
candle-core/src/quantized/gguf_file.rs
View File

@ -59,8 +59,13 @@ impl TensorInfo {
tensor_data_offset: u64, tensor_data_offset: u64,
) -> Result<QTensor> { ) -> Result<QTensor> {
let tensor_elems = self.shape.elem_count(); let tensor_elems = self.shape.elem_count();
let size_in_bytes = let blck_size = self.ggml_dtype.blck_size();
tensor_elems * self.ggml_dtype.type_size() / self.ggml_dtype.blck_size(); if tensor_elems % blck_size != 0 {
crate::bail!(
"the number of elements {tensor_elems} is not divisible by the block size {blck_size}"
)
}
let size_in_bytes = tensor_elems / blck_size * self.ggml_dtype.type_size();
let mut raw_data = vec![0u8; size_in_bytes]; let mut raw_data = vec![0u8; size_in_bytes];
reader.seek(std::io::SeekFrom::Start(tensor_data_offset + self.offset))?; reader.seek(std::io::SeekFrom::Start(tensor_data_offset + self.offset))?;
reader.read_exact(&mut raw_data)?; reader.read_exact(&mut raw_data)?;