huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-16 18:48:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Typos and format and CD only when PR lands.

Browse Source
This commit is contained in:
Nicolas Patry
2023-08-02 19:18:43 +02:00
parent 1b2b32e58d
commit dba31473d4
3 changed files with 6 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
candle-examples/src/lib.rs
View File

@ -73,8 +73,8 @@ let mmap = unsafe { Mmap::map(&file).unwrap() };
// Use safetensors directly
let tensors = SafeTensors::deserialize(&mmap[..]).unwrap();
let view = tensors
.tensor("bert.encoder.layer.0.attention.self.query.weight")
.unwrap();
.tensor("bert.encoder.layer.0.attention.self.query.weight")
.unwrap();
// We're going to load shard with rank 1, within a world_size of 4
// We're going to split along dimension 0 doing VIEW[start..stop, :]
@ -86,7 +86,7 @@ let mut tp_shape = view.shape().to_vec();
let size = tp_shape[0];
if size % world_size != 0 {
panic!("The dimension is not divisble by `world_size`");
panic!("The dimension is not divisble by `world_size`");
}
let block_size = size / world_size;
let start = rank * block_size;
@ -102,7 +102,7 @@ tp_shape[dim] = block_size;
// Convert safetensors Dtype to candle DType
let dtype: DType = dtype.try_into().unwrap();
// TODO: Implement from_buffer_iterator to we can skip the extra CPU alloc.
// TODO: Implement from_buffer_iterator so we can skip the extra CPU alloc.
let raw: Vec<u8> = iterator.into_iter().flatten().cloned().collect();
let tp_tensor = Tensor::from_raw_buffer(&raw, dtype, &tp_shape, &Device::Cpu).unwrap();
// ANCHOR_END: book_hub_3