Get the cpu tests to run.

candle-core/src/cpu_backend.rs

@@ -229,10 +229,6 @@ impl CpuStorage {
         D::cpu_storage_as_slice(self)
     }
 
-    pub fn as_mut_slice<D: crate::WithDType>(&mut self) -> Result<&mut [D]> {
-        D::cpu_storage_as_mut_slice(self)
-    }
-
     pub(crate) fn to_dtype(&self, layout: &Layout, dtype: DType) -> Result<Self> {
         // TODO: find a way around the quadratic number of cases below.
         match (self, dtype) {
@@ -581,6 +577,7 @@ impl CpuStorage {
         layout_f: &Layout,
     ) -> Result<Self> {
         // TODO: Support types that could be casted to a boolean.
+        // TODO: this should use the layout.
         let pred = self.as_slice::<u32>()?;
         match (t, f) {
             (Self::BF16(t), Self::BF16(f)) => {
@@ -618,6 +615,7 @@ impl CpuStorage {
         hidden_size: usize,
         vocab_size: usize,
     ) -> Result<Self> {
+        // TODO: this should use the layout.
         let ids = self.as_slice::<u32>()?;
         map1!(vs, take_impl1, ids, layout, vocab_size, hidden_size)
     }

candle-core/src/dtype.rs

@@ -41,7 +41,6 @@ pub trait WithDType: Sized + Copy {
     }
 
     fn cpu_storage_as_slice(s: &CpuStorage) -> Result<&[Self]>;
-    fn cpu_storage_as_mut_slice(s: &mut CpuStorage) -> Result<&mut [Self]>;
     fn cpu_storage_data(s: CpuStorage) -> Result<Vec<Self>>;
 }
 
@@ -75,17 +74,6 @@ macro_rules! with_dtype {
                     }),
                 }
             }
-
-            fn cpu_storage_as_mut_slice(s: &mut CpuStorage) -> Result<&mut [Self]> {
-                match s {
-                    CpuStorage::$dtype(data) => Ok(data),
-                    _ => Err(Error::UnexpectedDType {
-                        expected: DType::$dtype,
-                        got: s.dtype(),
-                        msg: "unexpected dtype",
-                    }),
-                }
-            }
         }
     };
 }

candle-core/src/strided_index.rs

@@ -17,7 +17,7 @@ impl<'a> StridedIndex<'a> {
             None
         } else {
             // This applies to the scalar case.
-            Some(0)
+            Some(layout.start_offset())
         };
         StridedIndex {
             next_storage_index,

candle-core/src/tensor.rs

@@ -317,6 +317,7 @@ impl Tensor {
     unary_op!(sqrt, Sqrt);
     unary_op!(gelu, Gelu);
     unary_op!(relu, Relu);
+
     pub fn to_scalar<S: crate::WithDType>(&self) -> Result<S> {
         if self.rank() != 0 {
             return Err(Error::UnexpectedNumberOfDims {