candle::Tensor vs candle_core::Tensor

[vishpat]

What is the difference between

candle_core::tensor::Tensor vs candle_core::Tensor

[cirocavani]

As I can see, they are the same.

candle/candle-core/src/lib.rs

Line 84 in 4fd00b8

pub use tensor::{Tensor, TensorId};

Also, I think candle:: instead of candle_core:: is just an artifact of how the workspace is configured.

For example, in candle-examples:

candle/candle-examples/Cargo.toml

Line 14 in 4fd00b8

candle = { workspace = true }

candle/Cargo.toml

Line 34 in 4fd00b8

candle = { path = "./candle-core", package = "candle-core", version = "0.4.1" }

So, when using candle published crates, you will have candle_core::.

For example, to convert an example from candle-examples to use the crates, you have to rewrite the "uses" of candle.

Something like:

(your project)
cargo add candle-core@0.4.0
cargo add candle-nn@0.4.0
cargo add candle-transformers@0.4.0

curl -kLo src/main.rs https://github.com/huggingface/candle/raw/6400e1b0a08d594e1448d522a41bddc98c584313/candle-examples/examples/phi/main.rs
sed -i '' 's/candle::/candle_core::/' src/main.rs

I did this yesterday to test only the CPU and Metal devices.

Full "decoupling" of the Phi example:

https://github.com/cirocavani/seventeen-candle/blob/8909d6cc03025261b424d6a342845baf978a13b3/phi/README.md