burn-store 0.22.1

//! Bridge between burn-core tensors and the tensor-agnostic burnpack format entries.
//!
//! [`burn_pack::Tensor`] carries the format-level facts about a tensor (name, dtype, shape,
//! optional param id) and a source for its raw little-endian bytes, which may not exist yet.
//! burn-store uses it as its single tensor-transport type: what a [`Collector`](crate::Collector)
//! produces, what an [`Applier`](crate::Applier) consumes, and what the stores read and write.
//!
//! This module is the seam where burn-core's [`TensorData`] meets those raw bytes. The lazy
//! paths are the load-bearing ones: a tensor built by [`from_tensor`] holds the
//! (reference-counted) device tensor and reads it back only when the bytes are finally drawn,
//! [`map_data`] composes a transform onto a byte source without materializing it, and
//! `from_pytorch` (with the `pytorch` feature) leaves a checkpoint tensor's bytes unread.

use alloc::format;
use alloc::string::{String, ToString};

use burn_pack::{Error as PackError, Tensor as PackTensor};

use burn_core::tensor::kind::Basic;
use burn_core::tensor::quantization::quantized_data_len;
use burn_core::tensor::{Bytes, DType, Shape, Tensor, TensorData};

/// Number of bytes a tensor of this dtype and shape serializes to.
///
/// [`Writer`](burn_pack::Writer) lays out the whole container from these lengths before any
/// data is produced, which is what lets a save's peak memory be bounded by the largest single
/// tensor rather than by the model. Quantized data packs its values and appends scales inline,
/// so its length is not a product of the shape and dtype and gets its own accounting.
pub fn data_len(dtype: DType, shape: &Shape) -> usize {
    match dtype {
        DType::QFloat(scheme) => quantized_data_len(&scheme, shape),
        _ => shape.iter().product::<usize>() * dtype.size(),
    }
}

/// Run a data-producing closure, turning a panic into an error.
///
/// A backend can panic reading a tensor back from its device, and that panic would otherwise
/// unwind through [`Writer`](burn_pack::Writer) partway through a container. burn-pack cannot
/// guard against it: `catch_unwind` needs `std`, which that crate deliberately does not
/// require. So the guard lives here, wrapped around every provider burn-store builds.
///
/// Classified as a validation failure rather than an I/O one on purpose. It arrives mid-write,
/// alongside the writer's own file errors, and calling it I/O would send the reader looking at
/// their disk.
#[cfg(feature = "std")]
fn guarded(f: impl Fn() -> Result<TensorData, PackError>) -> Result<TensorData, PackError> {
    // AssertUnwindSafe: the shared closure is not UnwindSafe, and the only state that could be
    // observed after a panic is the tensor being read, which is dropped either way.
    std::panic::catch_unwind(std::panic::AssertUnwindSafe(f)).unwrap_or_else(|payload| {
        // Carry the panic's own message through. Everything from a device OOM to a bug in an
        // adapter's transform arrives here, and without the payload they are indistinguishable
        // to anything handling the error rather than reading stderr.
        let cause = payload
            .downcast_ref::<&str>()
            .map(|s| (*s).to_string())
            .or_else(|| payload.downcast_ref::<String>().cloned())
            .unwrap_or_else(|| "unknown panic payload".to_string());

        Err(PackError::ValidationError(format!(
            "panic while producing tensor data: {cause}"
        )))
    })
}

/// Run a data-producing closure.
///
/// See the `std` variant. `catch_unwind` is unavailable here, so a panic propagates.
#[cfg(not(feature = "std"))]
fn guarded(f: impl Fn() -> Result<TensorData, PackError>) -> Result<TensorData, PackError> {
    f()
}

/// The `Send + Sync` bound a data closure carries on a target with pointer-width atomics, and
/// nothing on one without.
///
/// [`PackTensor::deferred`] asks for `Send + Sync` providers only where `alloc::sync` exists so
/// deferred tensors preserve the thread-safety of resident ones. Without pointer-width atomics,
/// it uses an `Rc`-backed provider and cannot itself be `Send` or `Sync`, regardless of the
/// closure's bounds. Naming that difference once keeps every constructor below from having to
/// be written twice.
///
/// Implemented for every type that satisfies it; the blanket impl makes a manual
/// implementation impossible.
#[cfg(target_has_atomic = "ptr")]
pub trait MaybeSendSync: Send + Sync {}
#[cfg(target_has_atomic = "ptr")]
impl<T: Send + Sync> MaybeSendSync for T {}

/// Empty on a target without pointer-width atomics, where an `Rc`-backed [`PackTensor`] cannot
/// be `Send` or `Sync` regardless of the provider's own bounds.
#[cfg(not(target_has_atomic = "ptr"))]
pub trait MaybeSendSync {}
#[cfg(not(target_has_atomic = "ptr"))]
impl<T> MaybeSendSync for T {}

/// Build a tensor whose data is produced on demand by `data_fn`.
///
/// `dtype` and `shape` describe what `data_fn` will return, and fix the byte length the writer
/// reserves before the closure ever runs, so a transform that changes either must be declared
/// here rather than left for the closure to reveal. A provider that hands back something else
/// is rejected when the data is drawn, so the failure is an error rather than a misplaced or
/// misread tensor.
pub fn deferred(
    name: String,
    dtype: DType,
    shape: Shape,
    param_id: Option<u64>,
    data_fn: impl Fn() -> Result<TensorData, PackError> + MaybeSendSync + 'static,
) -> PackTensor {
    let byte_len = data_len(dtype, &shape);
    let declared = shape.clone();

    PackTensor::deferred(name, dtype, shape, param_id, byte_len, move || {
        let data = guarded(&data_fn)?;

        // The byte length is checked downstream, but two dtypes of the same width (F32 and
        // I32, say) produce identical lengths, and the declared one is what a reader will
        // reinterpret the bytes as. Nothing else catches that, so check it here.
        if data.dtype() != dtype || *data.shape() != declared {
            return Err(PackError::ValidationError(format!(
                "provider produced {:?} {:?}, but the tensor declared {:?} {:?}",
                data.dtype(),
                data.shape(),
                dtype,
                declared
            )));
        }

        Ok(data.into_bytes())
    })
}

/// Snapshot a module parameter without reading it back from its device.
///
/// The tensor is cloned (cheap, reference-counted) and read only when the bytes are drawn, so
/// a whole module can be traversed for its metadata alone. One function covers float, int and
/// bool parameters: [`Basic`] is what supplies `dtype`, `shape` and `to_data` for all three.
pub fn from_tensor<const D: usize, K: Basic + 'static>(
    tensor: &Tensor<D, K>,
    name: String,
    param_id: Option<u64>,
) -> PackTensor {
    let (dtype, shape) = (tensor.dtype(), tensor.shape());
    let tensor = tensor.clone();

    deferred(name, dtype, shape, param_id, move || Ok(tensor.to_data()))
}

/// Build a tensor from data already in hand.
pub fn from_data(data: TensorData, name: String, param_id: Option<u64>) -> PackTensor {
    let (bytes, shape, dtype) = data.into_parts();
    PackTensor::new(name, dtype, shape, param_id, bytes)
}

/// Read a tensor's bytes back as [`TensorData`], leaving it intact.
///
/// Costs whatever the source costs: a deferred tensor re-runs its provider, since nothing is
/// cached, and a resident one copies its bytes unless they can be shared. Prefer [`into_data`]
/// where the tensor is not needed afterwards.
pub fn to_data(tensor: &PackTensor) -> Result<TensorData, PackError> {
    tensor_data(tensor.to_bytes()?, tensor.shape.clone(), tensor.dtype)
}

/// Take a tensor's bytes as [`TensorData`], producing them if deferred.
pub fn into_data(tensor: PackTensor) -> Result<TensorData, PackError> {
    let (_, dtype, shape, _, bytes) = tensor.into_parts()?;

    tensor_data(bytes, shape, dtype)
}

/// Build [`TensorData`] from raw bytes, reporting a length that disagrees with the shape and
/// dtype as an error rather than a panic.
pub(crate) fn tensor_data(
    bytes: Bytes,
    shape: Shape,
    dtype: DType,
) -> Result<TensorData, PackError> {
    TensorData::try_from_bytes(bytes, shape, dtype)
        .map_err(|err| PackError::TensorBytesSizeMismatch(err.to_string()))
}

/// Wrap a tensor's data in a transform, keeping it deferred.
///
/// How adapters chain. Each declares the `dtype` and `shape` its transform produces, because
/// the writer needs the resulting byte length up front, while the transform itself runs only
/// once the bytes are finally drawn. `name` is passed explicitly because some adapters rename
/// as they go.
///
/// Takes the tensor by value: the closure holds it for the adapted tensor's whole lifetime, so
/// borrowing here would only mean cloning it into the closure anyway.
pub fn map_data(
    tensor: PackTensor,
    name: String,
    dtype: DType,
    shape: Shape,
    f: impl Fn(TensorData) -> TensorData + MaybeSendSync + 'static,
) -> PackTensor {
    let param_id = tensor.param_id;

    deferred(name, dtype, shape, param_id, move || {
        to_data(&tensor).map(&f)
    })
}

/// Wrap a PyTorch checkpoint tensor, leaving its bytes unread until they are drawn.
///
/// PyTorch carries no parameter identity, so `param_id` stays `None`. Takes the tensor by
/// value for the same reason [`map_data`] does.
#[cfg(feature = "pytorch")]
pub fn from_pytorch(tensor: pytorch_reader::Tensor) -> PackTensor {
    use std::io::ErrorKind;

    let (name, dtype, shape, read) = tensor.into_parts();
    let dtype = pytorch_dtype(dtype);
    let shape = Shape::from(shape);
    let declared = shape.clone();

    deferred(name, dtype, shape, None, move || {
        // `Tensor::read` documents the split: these two kinds mean the file disagrees with
        // itself, any other that the operating system could not read it.
        let bytes = read().map_err(|err| match err.kind() {
            ErrorKind::InvalidData | ErrorKind::UnexpectedEof => {
                PackError::ValidationError(err.to_string())
            }
            _ => PackError::IoError(err.to_string()),
        })?;
        tensor_data(Bytes::from_bytes_vec(bytes), declared.clone(), dtype)
    })
}

#[cfg(feature = "pytorch")]
fn pytorch_dtype(dtype: pytorch_reader::DType) -> DType {
    use burn_core::tensor::BoolStore;
    use pytorch_reader::DType as Pt;

    match dtype {
        Pt::F64 => DType::F64,
        Pt::F32 => DType::F32,
        Pt::F16 => DType::F16,
        Pt::BF16 => DType::BF16,
        Pt::I64 => DType::I64,
        Pt::I32 => DType::I32,
        Pt::I16 => DType::I16,
        Pt::I8 => DType::I8,
        Pt::U64 => DType::U64,
        Pt::U32 => DType::U32,
        Pt::U16 => DType::U16,
        Pt::U8 => DType::U8,
        Pt::Bool => DType::Bool(BoolStore::Native),
    }
}

#[cfg(all(test, feature = "std"))]
mod tests {
    use super::*;
    use alloc::format;
    use alloc::string::ToString;
    use alloc::vec;
    use burn_core::tensor::quantization::{QuantScheme, QuantStore, QuantValue, ScaleDtype};
    use burn_core::tensor::{Bool, Device, Distribution, Int, shape};

    /// `byte_len` is computed from the declared shape and dtype rather than observed, and the
    /// burnpack writer reserves exactly that many bytes before ever running the provider. A
    /// disagreement would silently misplace everything written after this tensor.
    ///
    /// The length is also checked inside the materialization path, so a mismatch surfaces as
    /// an error rather than a wrong number; report both the same way.
    fn assert_byte_len_matches(tensor: PackTensor) {
        let (declared, name) = (tensor.byte_len(), tensor.name.clone());

        match to_data(&tensor) {
            Ok(data) => assert_eq!(
                declared,
                data.bytes().len(),
                "byte_len disagrees with the materialized bytes for {name}"
            ),
            Err(e) => panic!("byte_len disagrees with the materialized bytes for {name}: {e}"),
        }
    }

    #[test]
    fn data_len_covers_every_dtype_family() {
        let device = Device::default();

        let floats = Tensor::<2>::from_data([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], &device);
        let ints = Tensor::<2, Int>::from_data([[1, 2], [3, 4]], &device);
        let bools = Tensor::<2, Bool>::from_data([[true, false], [false, true]], &device);

        assert_byte_len_matches(from_tensor(&floats, "float".to_string(), None));
        assert_byte_len_matches(from_tensor(&ints, "int".to_string(), None));
        assert_byte_len_matches(from_tensor(&bools, "bool".to_string(), None));
    }

    /// Quantized is the one family where [`data_len`] reconstructs the layout instead of
    /// deriving it from an element count, so it is where the two drift apart.
    ///
    /// The shapes and schemes here hit the two ways that reconstruction has been wrong: a
    /// value-byte count that is not a multiple of the 4-byte scale alignment (which a spurious
    /// round-up inflates), and a sub-byte `QuantValue` under `QuantStore::Native` (whose stored
    /// width rounds down to zero bytes if divided rather than div_ceil'd). A 32x32 Q8 tensor
    /// passes either way, so it cannot stand in for these.
    ///
    /// Both axes rely on the test backend storing quantized values natively, one `i8` per
    /// value: `quantize_dynamic` rewrites the scheme's store to `QuantStore::Native` even
    /// though `QuantScheme::default()` starts as `PackedU32`. A backend honoring `PackedU32`
    /// would produce 4-byte-exact value counts and exercise neither axis.
    #[test]
    fn data_len_matches_materialized_bytes_when_quantized() {
        let device = Device::default();

        let schemes = [
            ("q8", QuantScheme::default()),
            ("q4", QuantScheme::default().with_value(QuantValue::Q4S)),
            ("q2", QuantScheme::default().with_value(QuantValue::Q2S)),
        ];
        // Element counts that are and are not multiples of the 4-byte scale alignment.
        let shapes = [shape![32, 32], shape![3, 3], shape![5, 5], shape![2, 3]];

        for (name, scheme) in schemes {
            for shape in &shapes {
                let tensor = Tensor::<2>::random(shape.clone(), Distribution::Default, &device)
                    .quantize_dynamic(&scheme);
                assert_byte_len_matches(from_tensor(&tensor, format!("{name}-{shape:?}"), None));
            }
        }
    }

    /// The same agreement against what `TensorData::quantized` writes, over block sizes rather
    /// than through a backend, so the per-block and per-tensor scale arithmetic is pinned
    /// independently of what any backend happens to produce.
    #[test]
    fn data_len_matches_quantized_bytes() {
        let base = QuantScheme::default()
            .with_value(QuantValue::Q8S)
            .with_store(QuantStore::Native);

        // 8 values in blocks of 4 lands on `QPARAM_ALIGN`, which would hide a padding
        // assumption; 6 in blocks of 3 and 10 in blocks of 5 do not.
        for (values, block) in [(8usize, 4usize), (6, 3), (10, 5)] {
            let scales = vec![0.5f32; values / block];
            let one_level = base.per_block([block as u8], ScaleDtype::UE4M3);
            let two_level = base
                .per_block([block as u8], ScaleDtype::UE4M3)
                .per_tensor(ScaleDtype::F32);

            for (scheme, global) in [(one_level, None), (two_level, Some(3.0f32))] {
                let data =
                    TensorData::quantized(vec![0i8; values], [values], scheme, &scales, global);

                assert_eq!(
                    data_len(data.dtype(), data.shape()),
                    data.bytes().len(),
                    "predicted size disagrees with the written bytes for {values} values \
                     in blocks of {block}, {scheme:?}"
                );
            }
        }
    }

    /// Packed quantized storage divides only the packed dimension, so a non-divisible extent
    /// pads once per line rather than once over the flattened tensor. No current backend can
    /// materialize such a tensor, so this pins the formula against the storage shape the
    /// allocation would use (`CubeTensor::quantized_storage`) rather than against real bytes.
    #[test]
    fn data_len_packs_per_line_for_packed_stores() {
        // Q4 in u32 words: 8 values per storage element, packed along the last dimension.
        let scheme = QuantScheme::default().with_value(QuantValue::Q4S);
        let packed = |shape: Shape| data_len(DType::QFloat(scheme), &shape);

        // 3 lines x ceil(3 / 8) = 3 u32 words = 12 value bytes, plus one tensor-level f32
        // scale. Flattening first would say ceil(9 / 8) * 4 = 8.
        assert_eq!(packed(shape![3, 3]), 12 + 4);

        // The same over more than two dimensions: 2 * 5 lines x ceil(9 / 8) = 20 words.
        assert_eq!(packed(shape![2, 5, 9]), 20 * 4 + 4);

        // A divisible extent is unaffected, and agrees with the flattened count.
        assert_eq!(packed(shape![4, 8]), 4 * 4 + 4);
    }

    /// A backend can panic reading a tensor back from its device. That panic must not unwind
    /// through the burnpack writer partway through a container, so every provider built here
    /// is wrapped: the panic becomes an ordinary error the writer can report and abort on.
    #[test]
    fn a_panicking_provider_becomes_an_error() {
        let tensor = deferred("weight".to_string(), DType::F32, shape![2, 2], None, || {
            panic!("device readback panicked")
        });

        let err = to_data(&tensor).expect_err("a panicking provider must not unwind");
        assert!(
            matches!(&err, PackError::ValidationError(m) if m.contains("panic")),
            "expected a validation error naming the panic, got {err:?}"
        );
    }

    /// A materialization failure must reach the caller through the writer with its class and
    /// its tensor intact: a panic is a validation failure, not an I/O one, and it arrives
    /// mid-write alongside the writer's real file errors. Reverting the classification, or
    /// dropping the writer's `in_tensor` annotation, fails here.
    #[test]
    fn a_write_time_failure_keeps_its_class_and_names_its_tensor() {
        let tensor = deferred(
            "encoder.weight".to_string(),
            DType::F32,
            shape![1],
            None,
            || panic!("device readback panicked"),
        );

        let err = burn_pack::Writer::new(vec![tensor])
            .into_bytes()
            .expect_err("a panicking provider must fail the write");

        assert!(
            matches!(&err, PackError::ValidationError(m)
                if m.contains("tensor 'encoder.weight'") && m.contains("device readback panicked")),
            "expected a named ValidationError carrying the panic message, got {err:?}"
        );
    }

    /// A provider's own error is passed through untouched; only panics are translated.
    #[test]
    fn a_provider_error_passes_through() {
        let tensor = deferred("weight".to_string(), DType::F32, shape![2, 2], None, || {
            Err(PackError::IoError("simulated IO error".to_string()))
        });

        let err = to_data(&tensor).unwrap_err();
        assert!(
            matches!(&err, PackError::IoError(m) if m == "simulated IO error"),
            "expected the provider's own error, got {err:?}"
        );
    }

    /// A deferred tensor must not run its provider until the bytes are asked for. This is what
    /// bounds a save's peak host memory by the largest single tensor rather than by the whole
    /// model, and it is invisible in the metadata, so pin it on the call count.
    #[test]
    fn deferred_runs_its_provider_only_when_asked() {
        use core::sync::atomic::{AtomicUsize, Ordering};
        // Tests only build for std targets, so the atomic pointer is always available here.
        use alloc::sync::Arc;

        let calls = Arc::new(AtomicUsize::new(0));
        let counter = calls.clone();
        let data = TensorData::from([1.0f32, 2.0, 3.0, 4.0]);

        let tensor = deferred(
            "weight".to_string(),
            DType::F32,
            shape![4],
            None,
            move || {
                counter.fetch_add(1, Ordering::Relaxed);
                Ok(data.clone())
            },
        );

        // Metadata alone costs nothing.
        assert_eq!(tensor.byte_len(), 16);
        assert_eq!(tensor.shape, shape![4]);
        assert_eq!(calls.load(Ordering::Relaxed), 0);

        to_data(&tensor).unwrap();
        assert_eq!(calls.load(Ordering::Relaxed), 1);
    }

    /// One generic constructor has to cover all three parameter kinds, since `Basic` is what
    /// supplies `dtype`, `shape` and `to_data` for each.
    #[test]
    fn from_tensor_covers_every_kind() {
        let device = Device::default();

        let floats = from_tensor(
            &Tensor::<2>::from_data([[1.0, 2.0], [3.0, 4.0]], &device),
            "float".to_string(),
            Some(7),
        );
        assert_eq!(floats.name, "float");
        assert_eq!(floats.shape, shape![2, 2]);
        assert_eq!(floats.param_id, Some(7));
        assert_eq!(*to_data(&floats).unwrap().shape(), shape![2, 2]);

        let ints = from_tensor(
            &Tensor::<2, Int>::from_data([[1, 2], [3, 4]], &device),
            "int".to_string(),
            None,
        );
        assert_eq!(ints.dtype, device.settings().int_dtype.into());

        let bools = from_tensor(
            &Tensor::<2, Bool>::from_data([[true, false], [false, true]], &device),
            "bool".to_string(),
            None,
        );
        assert_eq!(*to_data(&bools).unwrap().shape(), shape![2, 2]);
    }

    /// A provider that returns a different dtype than the tensor declared must be refused.
    /// Two dtypes of the same width give identical byte lengths, so the length check cannot
    /// see this, and a reader would reinterpret the bytes as the declared type.
    #[test]
    fn a_provider_contradicting_its_declared_dtype_is_refused() {
        let tensor = deferred("weight".to_string(), DType::F32, shape![4], None, || {
            Ok(TensorData::from([1i32, 2, 3, 4]))
        });

        let err = to_data(&tensor).unwrap_err();
        assert!(
            matches!(&err, PackError::ValidationError(m) if m.contains("I32") && m.contains("F32")),
            "expected a validation error naming both dtypes, got {err:?}"
        );
    }

    /// The same for a shape that permutes to the same element count.
    #[test]
    fn a_provider_contradicting_its_declared_shape_is_refused() {
        let tensor = deferred("weight".to_string(), DType::F32, shape![2, 3], None, || {
            Ok(TensorData::new(vec![0.0f32; 6], shape![3, 2]))
        });

        assert!(matches!(
            to_data(&tensor).unwrap_err(),
            PackError::ValidationError(_)
        ));
    }

    /// `into_data` yields what `to_data` does, taking the tensor rather than borrowing it.
    #[test]
    fn into_data_matches_to_data() {
        let device = Device::default();
        let tensor = from_tensor(
            &Tensor::<2>::from_data([[1.0, 2.0], [3.0, 4.0]], &device),
            "weight".to_string(),
            None,
        );

        let borrowed = to_data(&tensor).unwrap();
        let taken = into_data(tensor).unwrap();
        assert_eq!(borrowed.shape(), taken.shape());
        assert_eq!(borrowed.dtype(), taken.dtype());
        assert_eq!(borrowed.as_bytes(), taken.as_bytes());
    }

    /// `map_data` has to declare the length its transform will produce, not the one it
    /// consumed. A cast that shrinks the data while still claiming the source's length would
    /// be caught only once the writer had already committed its offsets.
    #[test]
    fn map_data_declares_the_transformed_length() {
        let device = Device::default();
        let source = from_tensor(
            &Tensor::<2>::from_data([[1.0, 2.0], [3.0, 4.0]], &device),
            "weight".to_string(),
            None,
        );
        assert_eq!(source.byte_len(), 16);

        let shape = source.shape.clone();
        let cast = map_data(source, "weight".to_string(), DType::F16, shape, |data| {
            data.convert_dtype(DType::F16)
        });

        assert_eq!(cast.byte_len(), 8);
        assert_byte_len_matches(cast);
    }

    /// A checkpoint tensor crosses over with its identity intact and its bytes reinterpreted
    /// under the mapped dtype. Bool is the variant with no direct counterpart, and the one
    /// whose bytes must arrive as valid `bool`s.
    #[cfg(feature = "pytorch")]
    #[test]
    fn from_pytorch_maps_dtype_and_reads_bytes() {
        use burn_core::tensor::BoolStore;

        let path = concat!(
            env!("CARGO_MANIFEST_DIR"),
            "/pytorch-tests/tests/boolean/boolean.pt"
        );
        let mut tensors = pytorch_reader::PytorchReader::new(path)
            .unwrap()
            .into_tensors();
        let tensor = from_pytorch(tensors.remove("buffer").unwrap());

        assert_eq!(tensor.name, "buffer");
        assert_eq!(tensor.dtype, DType::Bool(BoolStore::Native));
        assert_eq!(tensor.shape, shape![3]);
        assert_byte_len_matches(tensor.clone());

        let data = into_data(tensor).unwrap();
        assert_eq!(data.as_slice::<bool>().unwrap(), &[true, false, true]);
    }

    /// Every checkpoint dtype maps to the burn dtype of the same name and width. The two
    /// sides of a swapped arm (`I16` to `U16`, say) would agree with each other everywhere
    /// downstream, so only a check against the names catches it.
    #[cfg(feature = "pytorch")]
    #[test]
    fn pytorch_dtype_maps_every_variant_by_name() {
        use pytorch_reader::DType as Pt;

        let all = [
            Pt::F64,
            Pt::F32,
            Pt::F16,
            Pt::BF16,
            Pt::I64,
            Pt::I32,
            Pt::I16,
            Pt::I8,
            Pt::U64,
            Pt::U32,
            Pt::U16,
            Pt::U8,
            Pt::Bool,
        ];
        for pt in all {
            let mapped = pytorch_dtype(pt);
            assert_eq!(mapped.size(), pt.size(), "{pt:?}");
            let expected = match pt {
                Pt::Bool => "Bool(Native)".to_string(),
                other => format!("{other:?}"),
            };
            assert_eq!(format!("{mapped:?}"), expected);
        }
    }
}