feat(slice): add feature flag and guards for std/alloc

Author: LVivona

binding/python/Cargo.toml

@@ -15,4 +15,6 @@ memmap2 = "0.9"
 bincode = "2.0.1" 
 
 [dependencies.bintensors]
-path = "../../bintensors"
+path = "../../bintensors"
+default-features = false
+features = ["std", "slice"]

bintensors/Cargo.toml

@@ -9,17 +9,7 @@ readme = "README.md"
 keywords = ["bintensors", "serde", "tensors", "huggingface", "pytorch"]
 repository = "https://github.com/GnosisFoundation/bintensors"
 description = """
-bintensors is a high-performance binary tensor serialization 
-format designed to be faster than safetensors by eliminating 
-JSON metadata overhead. Instead of using JSON for metadata storage, 
-bintensors employs a compact binary header, reducing parsing time and improving I/O efficiency. 
-The format begins with an 8-byte unsigned integer (u64), indicating the size of the binary header. 
-This header contains essential tensor metadata, including the data type (dtype), 
-the shape of the tensor stored as a length-prefixed list, 
-and data offsets specifying where each tensor's raw values are stored within the file. 
-An index map is also included for named tensor lookups. With its efficient structure, 
-bintensors provides a fast and lightweight alternative for serializing and deserializing large models, 
-making it ideal for machine learning, scientific computing, and other high-performance applications.
+Bintensors is a high-performance binary tensor serialization format designed to be faster eliminating use of JSON serialization metadata. 
 """
 exclude = ["target/*", "Cargo.lock"]
 
@@ -43,6 +33,7 @@ sha1 = "0.10.6"
 default = ["std"]
 std = ["bincode/std", "serde/std"]
 alloc = ["bincode/alloc", "bincode/serde", "serde/alloc", "hashbrown"]
+slice = []
 
 [[bench]]
 name = "benchmark"

bintensors/src/lib.rs

@@ -4,6 +4,9 @@
 )]
 #![doc = include_str!("../DOC_README.md")]
 #![cfg_attr(not(feature = "std"), no_std)]
+
+#[cfg(any(feature = "std", feature = "alloc"))]
+#[cfg(feature = "slice")]
 pub mod slice;
 pub mod tensor;
 /// serialize_to_file only valid in std
@@ -20,6 +23,8 @@ extern crate alloc;
 compile_error!("must choose either the `std` or `alloc` feature, but not both.");
 #[cfg(all(not(feature = "std"), not(feature = "alloc")))]
 compile_error!("must choose either the `std` or `alloc` feature");
+#[cfg(all(not(feature = "std"), not(feature = "alloc"), feature = "slice"))]
+compile_error!("must choose either the `std` or `alloc` feature to use the slice feature.");
 
 /// A facade around all the types we need from the `std`, `core`, and `alloc`
 /// crates. This avoids elaborate import wrangling having to happen in every

bintensors/src/tensor.rs

@@ -1,5 +1,6 @@
 //! Module Containing the most important structures
 use crate::lib::{BTreeMap, Cow, HashMap, String, ToString, Vec};
+#[cfg(feature = "slice")]
 use crate::slice::{InvalidSlice, SliceIterator, TensorIndexer};
 use bincode::{Decode, Encode};
 use digest::Digest;
@@ -731,6 +732,7 @@ impl<'data> TensorView<'data> {
     }
 
     /// The various pieces of the data buffer according to the asked slice
+    #[cfg(feature = "slice")]
     pub fn sliced_data(
         &'data self,
         slices: &[TensorIndexer],
@@ -818,7 +820,7 @@ impl Dtype {
 mod tests {
 
     use super::*;
-    use crate::slice::IndexOp;
+
     use proptest::prelude::*;
     #[cfg(not(debug_assertions))]
     use sha1::Sha1;
@@ -829,6 +831,9 @@ mod tests {
     #[cfg(not(feature = "std"))]
     extern crate std;
 
+    #[cfg(feature = "slice")]
+    use crate::slice::IndexOp;
+
     const MAX_DIMENSION: usize = 8;
     const MAX_SIZE: usize = 8;
     const MAX_TENSORS: usize = 8;
@@ -995,6 +1000,7 @@ mod tests {
         assert_eq!(tensor.data().as_ptr() as usize % tensor.dtype().size(), 0);
     }
 
+    #[cfg(feature = "slice")]
     #[test]
     fn test_slicing() {
         let data: Vec<u8> = vec![0.0f32, 1.0, 2.0, 3.0, 4.0, 5.0]