Enable native sign operation for Candle backend (#1647)

* Enable native sign operation for Candle backend

* Use fixed revision

Cargo.toml

@@ -25,7 +25,8 @@ license = "MIT OR Apache-2.0"
 [workspace.dependencies]
 async-trait = "0.1.80"
 bytemuck = "1.14"
-candle-core = { version = "0.4.1" }
+# candle-core = { version = "0.4.1" }
+candle-core = { git = "https://github.com/huggingface/candle.git", rev = "2817643" }
 clap = { version = "4.5.4", features = ["derive"] }
 colored = "2.1.0"
 console_error_panic_hook = "0.1.7"

crates/burn-candle/src/ops/base.rs

@@ -149,3 +149,7 @@ pub fn expand<E: CandleElement, const D1: usize, const D2: usize>(
 ) -> CandleTensor<E, D2> {
     CandleTensor::new(tensor.tensor.broadcast_as(&shape.dims).unwrap())
 }
+
+pub fn sign<E: CandleElement, const D: usize>(tensor: CandleTensor<E, D>) -> CandleTensor<E, D> {
+    CandleTensor::new(tensor.tensor.sign().unwrap())
+}

crates/burn-candle/src/ops/int_tensor.rs

@@ -8,7 +8,7 @@ use crate::{
     Candle, CandleTensor,
 };
 
-use super::base::{expand, permute};
+use super::base::{expand, permute, sign};
 
 impl<F: FloatCandleElement, I: IntCandleElement> IntTensorOps<Self> for Candle<F, I> {
     fn int_empty<const D: usize>(shape: Shape<D>, device: &Device<Self>) -> IntTensor<Self, D> {
@@ -447,6 +447,7 @@ impl<F: FloatCandleElement, I: IntCandleElement> IntTensorOps<Self> for Candle<F
         expand(tensor, shape)
     }
 
-    // TODO add sign operator once Candle supports it:
-    // https://github.com/huggingface/candle/issues/1827
+    fn int_sign<const D: usize>(tensor: IntTensor<Self, D>) -> IntTensor<Self, D> {
+        sign(tensor)
+    }
 }

crates/burn-candle/src/ops/tensor.rs

@@ -11,7 +11,7 @@ use crate::{
     Candle, CandleTensor,
 };
 
-use super::base::{expand, permute};
+use super::base::{expand, permute, sign};
 
 impl<F: FloatCandleElement, I: IntCandleElement> FloatTensorOps<Self> for Candle<F, I> {
     fn float_from_data<const D: usize>(
@@ -539,6 +539,7 @@ impl<F: FloatCandleElement, I: IntCandleElement> FloatTensorOps<Self> for Candle
         expand(tensor, shape)
     }
 
-    // TODO add sign operator once Candle supports it:
-    // https://github.com/huggingface/candle/issues/1827
+    fn float_sign<const D: usize>(tensor: FloatTensor<Self, D>) -> FloatTensor<Self, D> {
+        sign(tensor)
+    }
 }