cjc_runtime/

tensor.rs

//! N-dimensional tensor with element-wise, reduction, linalg, and NN operations.
//!
//! [`Tensor`] is the primary numerical type in CJC. It is backed by a
//! [`Buffer<f64>`] with COW (copy-on-write) semantics, so cloning a tensor
//! is O(1) and mutation triggers a deep copy only when the buffer is shared.
//!
//! # Determinism Guarantees
//!
//! - **Reductions** (`sum`, `mean`, `sum_axis`) use [`BinnedAccumulatorF64`]
//!   for order-invariant, bit-identical results.
//! - **Matmul** uses Kahan-compensated accumulation (sequential path) or
//!   tiled + parallel strategies (large matrices) with deterministic per-row
//!   thread assignment.
//! - **SIMD** kernels (via [`tensor_simd`]) avoid hardware FMA to preserve
//!   cross-platform bit-identity.
//! - **No** `HashMap`/`HashSet` anywhere -- all ordering is deterministic.
//!
//! # Layout
//!
//! Tensors use row-major (C-order) layout with explicit strides. Non-contiguous
//! views (from `slice`, `transpose`, `broadcast_to`) share the underlying
//! buffer with adjusted strides and offset. Call [`Tensor::to_contiguous`] to
//! materialize a contiguous copy when needed.
//!
//! # Operation Categories
//!
//! | Category | Methods |
//! |----------|---------|
//! | Construction | `zeros`, `ones`, `randn`, `from_vec`, `from_bytes` |
//! | Shape | `shape`, `ndim`, `len`, `reshape`, `transpose`, `slice`, `unsqueeze`, `squeeze`, `flatten` |
//! | Element-wise | `add`, `sub`, `mul_elem`, `div_elem`, `elem_pow`, `map`, `map_simd` |
//! | Reductions | `sum`, `mean`, `sum_axis`, `mean_axis`, `var_axis`, `std_axis` |
//! | Linalg | `matmul`, `bmm`, `linear`, `einsum` |
//! | NN | `softmax`, `layer_norm`, `relu`, `sigmoid`, `gelu`, `conv1d`, `conv2d`, `maxpool2d` |
//! | Indexing | `get`, `set`, `gather`, `scatter`, `index_select`, `argsort` |
//! | Attention | `scaled_dot_product_attention`, `split_heads`, `merge_heads` |
//!
//! [`Buffer<f64>`]: crate::buffer::Buffer
//! [`BinnedAccumulatorF64`]: crate::accumulator::BinnedAccumulatorF64
//! [`tensor_simd`]: crate::tensor_simd

use cjc_repro::Rng;

use crate::accumulator::{binned_sum_f64, BinnedAccumulatorF64};
use cjc_repro::KahanAccumulatorF64;

use crate::accumulator;
use crate::buffer::Buffer;
use crate::dispatch;
use crate::error::RuntimeError;
use crate::kernel as kernel_fns;
use crate::tensor_simd::{self, BinOp, UnaryOp};
use crate::tensor_tiled::TiledMatmul;

// ---------------------------------------------------------------------------
// 2. Tensor Runtime
// ---------------------------------------------------------------------------

/// An N-dimensional tensor backed by a [`Buffer<f64>`](crate::buffer::Buffer).
///
/// Supports element-wise arithmetic (SIMD-accelerated), matrix multiplication
/// (tiled + parallel), numerically-stable reductions via
/// [`BinnedAccumulatorF64`](crate::accumulator::BinnedAccumulatorF64),
/// and neural network operations (softmax, layer norm, attention).
///
/// # Memory
///
/// The underlying [`Buffer<f64>`](crate::buffer::Buffer) uses copy-on-write
/// semantics. Cloning a `Tensor` is O(1). Operations like `reshape` and
/// `transpose` return zero-copy views when possible. Mutation via `set`
/// triggers a deep copy only when the buffer is shared.
#[derive(Debug, Clone)]
pub struct Tensor {
    /// The underlying COW data buffer.
    pub buffer: Buffer<f64>,
    /// Dimensions of the tensor (e.g., `[3, 4]` for a 3x4 matrix).
    pub(crate) shape: Vec<usize>,
    /// Row-major strides for each dimension. Non-contiguous views may have
    /// strides that differ from the standard row-major pattern (e.g., stride=0
    /// for broadcast dimensions).
    pub(crate) strides: Vec<usize>,
    /// Byte offset into the buffer where this tensor's data begins.
    /// Non-zero for slice views.
    pub(crate) offset: usize,
}

impl Tensor {
    // -- Construction -------------------------------------------------------

    /// Compute row-major strides for a given shape.
    pub(crate) fn compute_strides(shape: &[usize]) -> Vec<usize> {
        let mut strides = vec![1usize; shape.len()];
        for i in (0..shape.len().saturating_sub(1)).rev() {
            strides[i] = strides[i + 1] * shape[i + 1];
        }
        strides
    }

    /// Total number of elements implied by `shape`.
    fn shape_numel(shape: &[usize]) -> usize {
        shape.iter().product()
    }

    /// Create a tensor filled with zeros.
    pub fn zeros(shape: &[usize]) -> Self {
        let numel = Self::shape_numel(shape);
        Tensor {
            buffer: Buffer::alloc(numel, 0.0),
            shape: shape.to_vec(),
            strides: Self::compute_strides(shape),
            offset: 0,
        }
    }

    /// Create a tensor filled with ones.
    pub fn ones(shape: &[usize]) -> Self {
        let numel = Self::shape_numel(shape);
        Tensor {
            buffer: Buffer::alloc(numel, 1.0),
            shape: shape.to_vec(),
            strides: Self::compute_strides(shape),
            offset: 0,
        }
    }

    /// Create a tensor filled with samples from the standard normal
    /// distribution, drawn deterministically from `rng`.
    pub fn randn(shape: &[usize], rng: &mut Rng) -> Self {
        let numel = Self::shape_numel(shape);
        let data: Vec<f64> = (0..numel).map(|_| rng.next_normal_f64()).collect();
        Tensor {
            buffer: Buffer::from_vec(data),
            shape: shape.to_vec(),
            strides: Self::compute_strides(shape),
            offset: 0,
        }
    }

    /// Create a tensor from raw data and a shape. Returns an error if the
    /// number of elements does not match the shape.
    pub fn from_vec(data: Vec<f64>, shape: &[usize]) -> Result<Self, RuntimeError> {
        let numel = Self::shape_numel(shape);
        if data.len() != numel {
            return Err(RuntimeError::ShapeMismatch {
                expected: numel,
                got: data.len(),
            });
        }
        Ok(Tensor {
            buffer: Buffer::from_vec(data),
            shape: shape.to_vec(),
            strides: Self::compute_strides(shape),
            offset: 0,
        })
    }

    // -- Accessors ----------------------------------------------------------

    /// The shape of this tensor.
    pub fn shape(&self) -> &[usize] {
        &self.shape
    }

    /// Number of dimensions.
    pub fn ndim(&self) -> usize {
        self.shape.len()
    }

    /// Total number of elements.
    pub fn len(&self) -> usize {
        Self::shape_numel(&self.shape)
    }

    /// Whether the tensor has zero elements.
    pub fn is_empty(&self) -> bool {
        self.len() == 0
    }

    /// Flatten a multi-dimensional index to a linear offset in the buffer.
    fn linear_index(&self, indices: &[usize]) -> Result<usize, RuntimeError> {
        if indices.len() != self.shape.len() {
            return Err(RuntimeError::DimensionMismatch {
                expected: self.shape.len(),
                got: indices.len(),
            });
        }
        let mut off = self.offset;
        for (i, &idx) in indices.iter().enumerate() {
            if idx >= self.shape[i] {
                return Err(RuntimeError::IndexOutOfBounds {
                    index: idx,
                    length: self.shape[i],
                });
            }
            off += idx * self.strides[i];
        }
        Ok(off)
    }

    /// Whether this tensor is contiguous in memory (row-major, no offset).
    pub fn is_contiguous(&self) -> bool {
        if self.offset != 0 {
            return false;
        }
        let expected = Self::compute_strides(&self.shape);
        self.strides == expected
    }

    /// Create a zero-copy slice (view) of this tensor.
    /// `ranges` contains `(start, end)` for each dimension.
    pub fn slice(&self, ranges: &[(usize, usize)]) -> Result<Tensor, RuntimeError> {
        if ranges.len() != self.shape.len() {
            return Err(RuntimeError::DimensionMismatch {
                expected: self.shape.len(),
                got: ranges.len(),
            });
        }
        let mut new_offset = self.offset;
        let mut new_shape = Vec::with_capacity(ranges.len());
        for (i, &(start, end)) in ranges.iter().enumerate() {
            if end > self.shape[i] || start > end {
                return Err(RuntimeError::IndexOutOfBounds {
                    index: end,
                    length: self.shape[i],
                });
            }
            new_offset += start * self.strides[i];
            new_shape.push(end - start);
        }
        Ok(Tensor {
            buffer: self.buffer.clone(), // shared — zero copy
            shape: new_shape,
            strides: self.strides.clone(),
            offset: new_offset,
        })
    }

    /// Materialize a contiguous copy if this tensor is non-contiguous.
    pub fn to_contiguous(&self) -> Tensor {
        if self.is_contiguous() {
            return self.clone();
        }
        let data = self.to_vec();
        Tensor {
            buffer: Buffer::from_vec(data),
            shape: self.shape.clone(),
            strides: Self::compute_strides(&self.shape),
            offset: 0,
        }
    }

    /// Create a broadcast view of this tensor to `target_shape`.
    /// Uses stride=0 for dimensions that need broadcasting (size 1 -> target size).
    pub fn broadcast_to(&self, target_shape: &[usize]) -> Result<Tensor, RuntimeError> {
        let src_ndim = self.shape.len();
        let tgt_ndim = target_shape.len();
        if tgt_ndim < src_ndim {
            return Err(RuntimeError::InvalidOperation(
                "cannot broadcast to a smaller rank".to_string(),
            ));
        }
        let pad = tgt_ndim - src_ndim;
        let mut new_strides = vec![0usize; tgt_ndim];
        for i in 0..tgt_ndim {
            if i < pad {
                // Padded dimension: stride = 0 (broadcast)
                new_strides[i] = 0;
            } else {
                let src_i = i - pad;
                if self.shape[src_i] == target_shape[i] {
                    new_strides[i] = self.strides[src_i];
                } else if self.shape[src_i] == 1 {
                    new_strides[i] = 0; // broadcast
                } else {
                    return Err(RuntimeError::ShapeMismatch {
                        expected: target_shape[i],
                        got: self.shape[src_i],
                    });
                }
            }
        }
        Ok(Tensor {
            buffer: self.buffer.clone(),
            shape: target_shape.to_vec(),
            strides: new_strides,
            offset: self.offset,
        })
    }

    /// Read the element at the given multi-dimensional index.
    pub fn get(&self, indices: &[usize]) -> Result<f64, RuntimeError> {
        let offset = self.linear_index(indices)?;
        self.buffer
            .get(offset)
            .ok_or(RuntimeError::IndexOutOfBounds {
                index: offset,
                length: self.buffer.len(),
            })
    }

    /// Write the element at the given multi-dimensional index.
    pub fn set(&mut self, indices: &[usize], val: f64) -> Result<(), RuntimeError> {
        let offset = self.linear_index(indices)?;
        self.buffer.set(offset, val)
    }

    /// Extract the raw data as a `Vec<f64>`, respecting strides and offset.
    pub fn to_vec(&self) -> Vec<f64> {
        if self.is_contiguous() {
            let full = self.buffer.borrow_data();
            let numel = self.len();
            if full.len() == numel {
                return full.to_vec();
            }
            // Buffer may be larger than the tensor's logical size
            // (e.g. Scratchpad pre-allocates extra capacity)
            return full[..numel].to_vec();
        }
        // Non-contiguous: iterate via strided access
        let numel = self.len();
        let mut result = Vec::with_capacity(numel);
        let ndim = self.shape.len();
        let mut indices = vec![0usize; ndim];
        for _ in 0..numel {
            let mut off = self.offset;
            for d in 0..ndim {
                off += indices[d] * self.strides[d];
            }
            result.push(self.buffer.get(off).unwrap_or(0.0));
            // Increment multi-index (row-major order)
            for d in (0..ndim).rev() {
                indices[d] += 1;
                if indices[d] < self.shape[d] {
                    break;
                }
                indices[d] = 0;
            }
        }
        result
    }

    // -- Reshape ------------------------------------------------------------

    /// Reshape to `new_shape`. The new shape must have the same total number
    /// of elements. The returned tensor **shares** the underlying buffer.
    pub fn reshape(&self, new_shape: &[usize]) -> Result<Tensor, RuntimeError> {
        let new_numel = Self::shape_numel(new_shape);
        if new_numel != self.len() {
            return Err(RuntimeError::ShapeMismatch {
                expected: self.len(),
                got: new_numel,
            });
        }
        // Reshape requires contiguous data; materialize if needed
        let tensor = if self.is_contiguous() { self.clone() } else { self.to_contiguous() };
        Ok(Tensor {
            buffer: tensor.buffer,
            shape: new_shape.to_vec(),
            strides: Self::compute_strides(new_shape),
            offset: 0,
        })
    }

    // -- Element-wise operations --------------------------------------------

    /// Apply a binary operation element-wise with broadcasting support.
    fn elementwise_binop(
        &self,
        other: &Tensor,
        op: impl Fn(f64, f64) -> f64,
    ) -> Result<Tensor, RuntimeError> {
        if self.shape == other.shape && self.is_contiguous() && other.is_contiguous() {
            // Fast path: same shape, both contiguous — borrow without cloning
            let a = self.buffer.borrow_data();
            let b = other.buffer.borrow_data();
            let data: Vec<f64> = a.iter().zip(b.iter()).map(|(&x, &y)| op(x, y)).collect();
            return Ok(Tensor {
                buffer: Buffer::from_vec(data),
                shape: self.shape.clone(),
                strides: Self::compute_strides(&self.shape),
                offset: 0,
            });
        }

        // Broadcasting path: compute result shape
        let result_shape = Self::broadcast_result_shape(&self.shape, &other.shape)?;
        let a_broadcast = self.broadcast_to(&result_shape)?;
        let b_broadcast = other.broadcast_to(&result_shape)?;

        let numel = Self::shape_numel(&result_shape);
        let ndim = result_shape.len();
        let mut data = Vec::with_capacity(numel);
        let mut indices = vec![0usize; ndim];

        for _ in 0..numel {
            let mut off_a = a_broadcast.offset;
            let mut off_b = b_broadcast.offset;
            for d in 0..ndim {
                off_a += indices[d] * a_broadcast.strides[d];
                off_b += indices[d] * b_broadcast.strides[d];
            }
            let va = a_broadcast.buffer.get(off_a).ok_or_else(|| {
                RuntimeError::InvalidOperation(format!(
                    "broadcast binop: left operand index {} out of bounds (buffer len {})",
                    off_a,
                    a_broadcast.buffer.len()
                ))
            })?;
            let vb = b_broadcast.buffer.get(off_b).ok_or_else(|| {
                RuntimeError::InvalidOperation(format!(
                    "broadcast binop: right operand index {} out of bounds (buffer len {})",
                    off_b,
                    b_broadcast.buffer.len()
                ))
            })?;
            data.push(op(va, vb));

            for d in (0..ndim).rev() {
                indices[d] += 1;
                if indices[d] < result_shape[d] {
                    break;
                }
                indices[d] = 0;
            }
        }

        Ok(Tensor {
            buffer: Buffer::from_vec(data),
            shape: result_shape.clone(),
            strides: Self::compute_strides(&result_shape),
            offset: 0,
        })
    }

    /// Compute the broadcast result shape for two shapes (NumPy rules).
    fn broadcast_result_shape(a: &[usize], b: &[usize]) -> Result<Vec<usize>, RuntimeError> {
        let max_ndim = a.len().max(b.len());
        let mut result = Vec::with_capacity(max_ndim);
        for i in 0..max_ndim {
            let da = if i < max_ndim - a.len() { 1 } else { a[i - (max_ndim - a.len())] };
            let db = if i < max_ndim - b.len() { 1 } else { b[i - (max_ndim - b.len())] };
            if da == db {
                result.push(da);
            } else if da == 1 {
                result.push(db);
            } else if db == 1 {
                result.push(da);
            } else {
                return Err(RuntimeError::ShapeMismatch {
                    expected: da,
                    got: db,
                });
            }
        }
        Ok(result)
    }

    /// SIMD-accelerated element-wise binary operation for known ops.
    ///
    /// For same-shape contiguous tensors, uses AVX2 (4-wide f64) when available.
    /// Falls back to the generic closure path for broadcast cases.
    fn elementwise_binop_simd(
        &self,
        other: &Tensor,
        op: BinOp,
        fallback: impl Fn(f64, f64) -> f64,
    ) -> Result<Tensor, RuntimeError> {
        if self.shape == other.shape && self.is_contiguous() && other.is_contiguous() {
            // SIMD fast path: same shape, both contiguous
            let a = self.buffer.borrow_data();
            let b = other.buffer.borrow_data();
            let data = tensor_simd::simd_binop(&a, &b, op);
            return Ok(Tensor {
                buffer: Buffer::from_vec(data),
                shape: self.shape.clone(),
                strides: Self::compute_strides(&self.shape),
                offset: 0,
            });
        }
        // Broadcast path: fall through to generic
        self.elementwise_binop(other, fallback)
    }

    /// Element-wise addition (SIMD-accelerated for contiguous same-shape tensors).
    pub fn add(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop_simd(other, BinOp::Add, |a, b| a + b)
    }

    /// Element-wise subtraction (SIMD-accelerated for contiguous same-shape tensors).
    pub fn sub(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop_simd(other, BinOp::Sub, |a, b| a - b)
    }

    /// Element-wise (Hadamard) multiplication (SIMD-accelerated for contiguous same-shape tensors).
    pub fn mul_elem(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop_simd(other, BinOp::Mul, |a, b| a * b)
    }

    /// Element-wise division (SIMD-accelerated for contiguous same-shape tensors).
    pub fn div_elem(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop_simd(other, BinOp::Div, |a, b| a / b)
    }

    /// Fused multiply-add: `self * b + c` element-wise in a single pass.
    ///
    /// Eliminates the intermediate tensor that separate mul + add would create.
    /// Uses software FMA (`a * b + c` with two roundings, not hardware FMA)
    /// to preserve bit-identity with the non-fused path.
    pub fn fused_mul_add(&self, b: &Tensor, c: &Tensor) -> Result<Tensor, RuntimeError> {
        if self.shape != b.shape || self.shape != c.shape {
            return Err(RuntimeError::InvalidOperation(
                "broadcast_fma: all three tensors must have the same shape".to_string(),
            ));
        }
        if self.is_contiguous() && b.is_contiguous() && c.is_contiguous() {
            let a_data = self.buffer.borrow_data();
            let b_data = b.buffer.borrow_data();
            let c_data = c.buffer.borrow_data();
            let n = a_data.len();
            let mut out = vec![0.0f64; n];
            // Software FMA: a*b + c (two roundings — NOT hardware FMA which uses one rounding).
            // This produces identical results to separate broadcast2("mul") + broadcast2("add").
            for i in 0..n {
                out[i] = a_data[i] * b_data[i] + c_data[i];
            }
            return Ok(Tensor {
                buffer: Buffer::from_vec(out),
                shape: self.shape.clone(),
                strides: Self::compute_strides(&self.shape),
                offset: 0,
            });
        }
        // Non-contiguous fallback: mul then add
        let temp = self.mul_elem(b)?;
        temp.add(c)
    }

    // ── v0.1 Broadcasting: additional element-wise binary ops ──

    /// Element-wise power: `a^b`.
    pub fn elem_pow(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop(other, |a, b| a.powf(b))
    }

    /// Element-wise minimum.
    pub fn elem_min(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop(other, |a, b| a.min(b))
    }

    /// Element-wise maximum.
    pub fn elem_max(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop(other, |a, b| a.max(b))
    }

    /// Element-wise atan2(self, other).
    pub fn elem_atan2(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop(other, |a, b| a.atan2(b))
    }

    /// Element-wise hypot(self, other).
    pub fn elem_hypot(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        self.elementwise_binop(other, |a, b| a.hypot(b))
    }

    /// Apply a unary function to every element, returning a new contiguous tensor.
    pub fn map(&self, f: impl Fn(f64) -> f64) -> Tensor {
        let data: Vec<f64> = self.to_vec().iter().map(|&x| f(x)).collect();
        Tensor {
            buffer: Buffer::from_vec(data),
            shape: self.shape.clone(),
            strides: Self::compute_strides(&self.shape),
            offset: 0,
        }
    }

    /// SIMD-accelerated unary map for known operations (sqrt, abs, neg, relu).
    ///
    /// Uses AVX2 (4-wide f64) when available, scalar fallback otherwise.
    /// Bit-identical to `map(f)` for the supported operations.
    pub fn map_simd(&self, op: UnaryOp) -> Tensor {
        let src = self.to_vec();
        let data = tensor_simd::simd_unary(&src, op);
        Tensor {
            buffer: Buffer::from_vec(data),
            shape: self.shape.clone(),
            strides: Self::compute_strides(&self.shape),
            offset: 0,
        }
    }

    // -- Reductions (using BinnedAccumulator) --------------------------------

    /// Sum of all elements (binned accumulation — order-invariant, deterministic).
    pub fn sum(&self) -> f64 {
        let data = self.buffer.borrow_data();
        binned_sum_f64(&data)
    }

    /// Sum of all elements using BinnedAccumulator (order-invariant, deterministic).
    ///
    /// Bit-identical results regardless of element ordering or reduction schedule.
    pub fn binned_sum(&self) -> f64 {
        let data = self.buffer.borrow_data();
        accumulator::binned_sum_f64(&data)
    }

    /// Sum with dispatched strategy based on execution context.
    ///
    /// Uses Kahan in serial mode, Binned in parallel/@nogc/strict/linalg mode.
    pub fn dispatched_sum(&self, ctx: &dispatch::ReductionContext) -> f64 {
        let data = self.buffer.borrow_data();
        dispatch::dispatch_sum_f64(&data, ctx)
    }

    /// Mean of all elements (binned sum / count).
    pub fn mean(&self) -> f64 {
        let n = self.len();
        if n == 0 {
            return 0.0;
        }
        self.sum() / n as f64
    }

    /// Mean with dispatched strategy based on execution context.
    pub fn dispatched_mean(&self, ctx: &dispatch::ReductionContext) -> f64 {
        let n = self.len();
        if n == 0 {
            return 0.0;
        }
        self.dispatched_sum(ctx) / n as f64
    }

    /// Sum along a specific axis, returning a tensor with that dimension reduced.
    ///
    /// Supports N-D tensors. The reduced axis becomes size 1 in the output.
    /// Uses BinnedAccumulator for order-invariant, deterministic summation.
    ///
    /// Examples:
    /// - 2D [M, N] with axis=0: result [1, N] (sum columns)
    /// - 2D [M, N] with axis=1: result [M, 1] (sum rows)
    /// - 3D [A, B, C] with axis=1: result [A, 1, C]
    pub fn sum_axis(&self, axis: usize) -> Result<Tensor, RuntimeError> {
        let ndim = self.ndim();
        if axis >= ndim {
            return Err(RuntimeError::IndexOutOfBounds {
                index: axis,
                length: ndim,
            });
        }

        // Build output shape: same as input but with axis dimension = 1
        let mut out_shape = self.shape.clone();
        out_shape[axis] = 1;
        let out_numel = Self::shape_numel(&out_shape);
        let out_strides = Self::compute_strides(&out_shape);

        let data = self.to_vec();
        let axis_len = self.shape[axis];
        let mut result = vec![0.0f64; out_numel];

        // For each output position, sum over the reduced axis with binned accumulation.
        let mut indices = vec![0usize; ndim];
        for out_idx in 0..out_numel {
            // Compute the N-D index from flat output index
            {
                let mut remaining = out_idx;
                for d in 0..ndim {
                    indices[d] = remaining / out_strides[d];
                    remaining %= out_strides[d];
                }
            }

            let mut acc = BinnedAccumulatorF64::new();
            for k in 0..axis_len {
                // Compute input flat index with indices[axis] = k
                let mut flat = self.offset;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { indices[d] };
                    flat += idx * self.strides[d];
                }
                acc.add(data[flat]);
            }
            result[out_idx] = acc.finalize();
        }

        Tensor::from_vec(result, &out_shape)
    }

    // -- Matrix multiplication (2-D only) -----------------------------------

    /// Negate every element, returning a new tensor.
    pub fn neg(&self) -> Tensor {
        self.map(|x| -x)
    }

    /// Transpose a tensor. For 2-D: swaps rows and columns (zero-copy view).
    /// For N-D: reverses all axes (zero-copy view).
    pub fn transpose(&self) -> Tensor {
        let ndim = self.ndim();
        if ndim <= 1 {
            return self.clone();
        }
        // Reverse shape and strides — zero-copy view
        let mut new_shape = self.shape.clone();
        let mut new_strides = self.strides.clone();
        new_shape.reverse();
        new_strides.reverse();
        Tensor {
            buffer: self.buffer.clone(), // shared — zero copy
            shape: new_shape,
            strides: new_strides,
            offset: self.offset,
        }
    }

    /// Transpose with explicit axis permutation (N-D). Zero-copy view.
    ///
    /// `axes` must be a permutation of `[0, 1, ..., ndim-1]`.
    pub fn transpose_axes(&self, axes: &[usize]) -> Result<Tensor, RuntimeError> {
        let ndim = self.ndim();
        if axes.len() != ndim {
            return Err(RuntimeError::InvalidOperation(
                format!("transpose_axes: expected {} axes, got {}", ndim, axes.len()),
            ));
        }
        // Validate permutation
        let mut seen = vec![false; ndim];
        for &ax in axes {
            if ax >= ndim {
                return Err(RuntimeError::IndexOutOfBounds { index: ax, length: ndim });
            }
            if seen[ax] {
                return Err(RuntimeError::InvalidOperation(
                    format!("transpose_axes: duplicate axis {ax}"),
                ));
            }
            seen[ax] = true;
        }
        let new_shape: Vec<usize> = axes.iter().map(|&ax| self.shape[ax]).collect();
        let new_strides: Vec<usize> = axes.iter().map(|&ax| self.strides[ax]).collect();
        Ok(Tensor {
            buffer: self.buffer.clone(),
            shape: new_shape,
            strides: new_strides,
            offset: self.offset,
        })
    }

    /// Multiply every element by a scalar, returning a new tensor.
    pub fn scalar_mul(&self, s: f64) -> Tensor {
        self.map(|x| x * s)
    }

    // ── Panicking convenience constructors (used by AD engine) --------

    /// Create a tensor from raw data and shape.
    /// **Panics** if `data.len()` does not match the shape.
    pub fn from_vec_unchecked(data: Vec<f64>, shape: &[usize]) -> Tensor {
        Self::from_vec(data, shape).expect("Tensor::from_vec_unchecked: shape mismatch")
    }

    /// Element-wise addition. **Panics** on shape mismatch.
    pub fn add_unchecked(&self, other: &Tensor) -> Tensor {
        self.add(other).expect("Tensor::add shape mismatch")
    }

    /// In-place element-wise addition. **Panics** on shape mismatch.
    ///
    /// Mutates `self` instead of allocating a new tensor. Uses COW: if the
    /// underlying buffer is shared, `make_unique()` deep-copies first.
    pub fn add_assign_unchecked(&mut self, other: &Tensor) {
        assert_eq!(self.shape, other.shape, "Tensor::add_assign shape mismatch");
        self.buffer.make_unique();
        let other_data = other.buffer.borrow_data();
        let mut self_data = self.buffer.borrow_data_mut();
        for (a, b) in self_data.iter_mut().zip(other_data.iter()) {
            *a += b;
        }
    }

    /// Element-wise subtraction. **Panics** on shape mismatch.
    pub fn sub_unchecked(&self, other: &Tensor) -> Tensor {
        self.sub(other).expect("Tensor::sub shape mismatch")
    }

    /// Element-wise multiplication. **Panics** on shape mismatch.
    pub fn mul_elem_unchecked(&self, other: &Tensor) -> Tensor {
        self.mul_elem(other).expect("Tensor::mul_elem shape mismatch")
    }

    /// Element-wise division. **Panics** on shape mismatch.
    pub fn div_elem_unchecked(&self, other: &Tensor) -> Tensor {
        self.div_elem(other).expect("Tensor::div_elem shape mismatch")
    }

    /// Matrix multiplication. **Panics** on dimension mismatch.
    pub fn matmul_unchecked(&self, other: &Tensor) -> Tensor {
        self.matmul(other).expect("Tensor::matmul dimension mismatch")
    }

    /// Matrix multiplication for 2-D tensors.
    ///
    /// `self` is (M, K), `other` is (K, N) => result is (M, N).
    pub fn matmul(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        if self.ndim() != 2 || other.ndim() != 2 {
            return Err(RuntimeError::InvalidOperation(
                "matmul requires 2-D tensors".to_string(),
            ));
        }
        let m = self.shape[0];
        let k = self.shape[1];
        let k2 = other.shape[0];
        let n = other.shape[1];
        if k != k2 {
            return Err(RuntimeError::DimensionMismatch {
                expected: k,
                got: k2,
            });
        }

        let a = self.to_vec();
        let b = other.to_vec();

        // Parallel path (Mode A): parallelize over output rows when the parallel
        // feature is enabled and the matrix is large enough (>= 256 in any dim).
        #[cfg(feature = "parallel")]
        {
            if m >= 256 || n >= 256 || k >= 256 {
                return Self::matmul_parallel_mode_a(&a, &b, m, n, k);
            }
        }

        // Tiled path: use L2-friendly tiled matmul for medium-to-large matrices.
        // Threshold: any dimension >= 64 (the default tile size).
        // NOTE: tiled path uses naive accumulation (not binned) — different
        // numerical path for large matrices, but better cache locality.
        if m >= 64 || n >= 64 || k >= 64 {
            return Self::matmul_tiled(&a, &b, m, n, k);
        }

        // Sequential path: single-threaded with binned accumulation.
        Self::matmul_sequential(&a, &b, m, n, k)
    }

    /// Sequential matmul (always available, deterministic reference).
    fn matmul_sequential(
        a: &[f64], b: &[f64], m: usize, n: usize, k: usize,
    ) -> Result<Tensor, RuntimeError> {
        let mut result = vec![0.0f64; m * n];
        for i in 0..m {
            for j in 0..n {
                let mut acc = KahanAccumulatorF64::new();
                for p in 0..k {
                    acc.add(a[i * k + p] * b[p * n + j]);
                }
                result[i * n + j] = acc.finalize();
            }
        }
        Tensor::from_vec(result, &[m, n])
    }

    /// Tiled matmul: delegates to `TiledMatmul` for L2-cache-friendly tiling.
    ///
    /// Used for medium matrices (any dimension >= 64) where cache locality
    /// matters but parallel overhead isn't justified. The tiled path uses
    /// naive accumulation (not binned accumulation), trading a small amount of
    /// floating-point precision for better cache behavior.
    fn matmul_tiled(
        a: &[f64], b: &[f64], m: usize, n: usize, k: usize,
    ) -> Result<Tensor, RuntimeError> {
        let engine = TiledMatmul::new();
        let result = engine.matmul(a, m, k, b, n);
        Tensor::from_vec(result, &[m, n])
    }

    /// Parallel matmul Mode A: parallelize over output rows using tiled
    /// micro-kernels for cache locality.
    ///
    /// Deterministic because:
    /// - Each output row is computed by exactly one thread.
    /// - Within each row, accumulation uses tiled AXPY (deterministic order).
    /// - No cross-thread reduction or merge of partial sums.
    ///
    /// Uses KahanAccumulatorF64 (lightweight) instead of BinnedAccumulatorF64
    /// (32KB per accumulator) to avoid massive stack pressure in parallel mode.
    #[cfg(feature = "parallel")]
    fn matmul_parallel_mode_a(
        a: &[f64], b: &[f64], m: usize, n: usize, k: usize,
    ) -> Result<Tensor, RuntimeError> {
        use rayon::prelude::*;
        use cjc_repro::KahanAccumulatorF64;

        // For large matrices, use tiled matmul (sequential but cache-friendly).
        // The tiling provides far better cache locality than parallel row-wise
        // with column-strided B access, and avoids the 32KB-per-element
        // BinnedAccumulator overhead that caused the 128→256 regression.
        if m >= 512 && n >= 512 {
            // Only use rayon for very large matrices where thread overhead
            // is amortized. Split into row-bands, each processed with tiled matmul.
            let band_size = (m + rayon::current_num_threads() - 1) / rayon::current_num_threads();
            let band_size = band_size.max(64); // At least 64 rows per band
            let mut result = vec![0.0f64; m * n];

            result
                .par_chunks_mut(band_size * n)
                .enumerate()
                .for_each(|(band_idx, band)| {
                    let i_start = band_idx * band_size;
                    let i_end = (i_start + band_size).min(m);
                    let band_m = i_end - i_start;
                    let a_band = &a[i_start * k .. i_end * k];
                    let engine = crate::tensor_tiled::TiledMatmul::new();
                    let tiled_result = engine.matmul(a_band, band_m, k, b, n);
                    band[..band_m * n].copy_from_slice(&tiled_result);
                });

            return Tensor::from_vec(result, &[m, n]);
        }

        // For medium matrices (256-511), use Kahan accumulator (16 bytes, not 32KB).
        let mut result = vec![0.0f64; m * n];
        result
            .par_chunks_mut(n)
            .enumerate()
            .for_each(|(i, row)| {
                for j in 0..n {
                    let mut acc = KahanAccumulatorF64::new();
                    for p in 0..k {
                        acc.add(a[i * k + p] * b[p * n + j]);
                    }
                    row[j] = acc.finalize();
                }
            });

        Tensor::from_vec(result, &[m, n])
    }

    // -- Transformer Kernels ------------------------------------------------

    /// Batched matrix multiplication.
    ///
    /// `self` is `[..., M, K]`, `other` is `[..., K, N]` => result is `[..., M, N]`.
    /// The batch dimensions must be identical (no broadcast).
    /// For 2-D inputs, delegates to `matmul`.
    pub fn bmm(&self, other: &Tensor) -> Result<Tensor, RuntimeError> {
        if self.ndim() < 2 || other.ndim() < 2 {
            return Err(RuntimeError::InvalidOperation(
                "bmm requires at least 2-D tensors".to_string(),
            ));
        }
        if self.ndim() == 2 && other.ndim() == 2 {
            return self.matmul(other);
        }
        if self.ndim() != other.ndim() {
            return Err(RuntimeError::InvalidOperation(
                format!(
                    "bmm requires same number of dimensions, got {} and {}",
                    self.ndim(),
                    other.ndim()
                ),
            ));
        }
        let nd = self.ndim();
        let batch_dims_a = &self.shape[..nd - 2];
        let batch_dims_b = &other.shape[..nd - 2];
        if batch_dims_a != batch_dims_b {
            return Err(RuntimeError::InvalidOperation(
                format!(
                    "bmm batch dimensions mismatch: {:?} vs {:?}",
                    batch_dims_a, batch_dims_b
                ),
            ));
        }
        let m = self.shape[nd - 2];
        let k = self.shape[nd - 1];
        let k2 = other.shape[nd - 2];
        let n = other.shape[nd - 1];
        if k != k2 {
            return Err(RuntimeError::DimensionMismatch {
                expected: k,
                got: k2,
            });
        }

        let batch_size: usize = batch_dims_a.iter().product();
        let a = self.to_vec();
        let b = other.to_vec();
        let mat_a_stride = m * k;
        let mat_b_stride = k * n;
        let mat_c_stride = m * n;
        let mut result = vec![0.0f64; batch_size * mat_c_stride];

        // Helper closure: compute one batch into c_slice
        let compute_batch = |batch: usize, c_slice: &mut [f64]| {
            let a_slice = &a[batch * mat_a_stride..(batch + 1) * mat_a_stride];
            let b_slice = &b[batch * mat_b_stride..(batch + 1) * mat_b_stride];

            if m >= 64 || n >= 64 || k >= 64 {
                let engine = crate::tensor_tiled::TiledMatmul::new();
                let tiled = engine.matmul(a_slice, m, k, b_slice, n);
                c_slice.copy_from_slice(&tiled);
            } else {
                for i in 0..m {
                    for j in 0..n {
                        let mut acc = KahanAccumulatorF64::new();
                        for p in 0..k {
                            acc.add(a_slice[i * k + p] * b_slice[p * n + j]);
                        }
                        c_slice[i * n + j] = acc.finalize();
                    }
                }
            }
        };

        // Parallel path: parallelize over batches when workload is large enough
        #[cfg(feature = "parallel")]
        {
            if batch_size > 1 && m * k >= 4096 {
                use rayon::prelude::*;
                result
                    .par_chunks_mut(mat_c_stride)
                    .enumerate()
                    .for_each(|(batch, c_slice)| {
                        compute_batch(batch, c_slice);
                    });

                let mut out_shape = batch_dims_a.to_vec();
                out_shape.push(m);
                out_shape.push(n);
                return Tensor::from_vec(result, &out_shape);
            }
        }

        // Sequential fallback
        for batch in 0..batch_size {
            let c_off = batch * mat_c_stride;
            compute_batch(batch, &mut result[c_off..c_off + mat_c_stride]);
        }

        let mut out_shape = batch_dims_a.to_vec();
        out_shape.push(m);
        out_shape.push(n);
        Tensor::from_vec(result, &out_shape)
    }

    /// Softmax along the last dimension (two-pass stable algorithm).
    ///
    /// Pass 1: find max per row (prevents overflow in exp)
    /// Pass 2: compute exp(x - max), accumulate sum, normalize
    ///
    /// For a tensor of shape `[..., N]`, softmax is applied independently
    /// to each length-N slice along the last axis.
    pub fn softmax(&self) -> Result<Tensor, RuntimeError> {
        if self.ndim() == 0 {
            return Err(RuntimeError::InvalidOperation(
                "softmax requires at least 1-D tensor".to_string(),
            ));
        }
        // Avoid allocation when tensor is already contiguous and starts at offset 0
        let data_ref;
        let data_vec;
        let data: &[f64] = if self.is_contiguous() && self.offset == 0 {
            data_ref = self.buffer.borrow_data();
            &data_ref
        } else {
            data_vec = self.to_vec();
            &data_vec
        };
        let n = *self.shape.last().unwrap(); // last dimension size
        let outer: usize = data.len() / n;  // product of all dims except last
        let mut result = vec![0.0f64; data.len()];

        for row in 0..outer {
            let start = row * n;
            let end = start + n;
            let slice = &data[start..end];

            // Pass 1: find max for numerical stability
            let mut max_val = f64::NEG_INFINITY;
            for &v in slice {
                if v > max_val {
                    max_val = v;
                }
            }

            // Pass 2: exp(x - max) and accumulate sum
            let mut exp_vals = vec![0.0f64; n];
            let mut sum = 0.0f64;
            let mut comp = 0.0f64; // Kahan compensation
            for i in 0..n {
                let e = (slice[i] - max_val).exp();
                exp_vals[i] = e;
                // Kahan summation for the denominator
                let y = e - comp;
                let t = sum + y;
                comp = (t - sum) - y;
                sum = t;
            }

            // Normalize
            if sum == 0.0 {
                // Degenerate case: all -inf inputs → uniform
                let uniform = 1.0 / n as f64;
                for i in 0..n {
                    result[start + i] = uniform;
                }
            } else {
                for i in 0..n {
                    result[start + i] = exp_vals[i] / sum;
                }
            }
        }

        Tensor::from_vec(result, &self.shape)
    }

    /// Layer normalization over the last dimension.
    ///
    /// For each length-D slice along the last axis:
    ///   1. mean = Σx / D  (BinnedAccumulator)
    ///   2. var  = Σ(x - mean)² / D  (BinnedAccumulator)
    ///   3. normalized = (x - mean) / √(var + eps)
    ///   4. output = gamma * normalized + beta
    ///
    /// `gamma` and `beta` are 1-D tensors of shape `[D]`.
    /// `eps` is a small constant (typically 1e-5).
    pub fn layer_norm(
        &self,
        gamma: &Tensor,
        beta: &Tensor,
        eps: f64,
    ) -> Result<Tensor, RuntimeError> {
        if self.ndim() == 0 {
            return Err(RuntimeError::InvalidOperation(
                "layer_norm requires at least 1-D tensor".to_string(),
            ));
        }
        let d = *self.shape.last().unwrap();
        if gamma.len() != d || beta.len() != d {
            return Err(RuntimeError::InvalidOperation(
                format!(
                    "layer_norm: gamma/beta length {} must match last dim {}",
                    gamma.len(),
                    d
                ),
            ));
        }

        let data = self.to_vec();
        let gamma_data = gamma.to_vec();
        let beta_data = beta.to_vec();
        let outer = data.len() / d;
        let mut result = vec![0.0f64; data.len()];

        for row in 0..outer {
            let start = row * d;
            let slice = &data[start..start + d];

            // Pass 1: compute mean via BinnedAccumulator
            let mean = binned_sum_f64(slice) / d as f64;

            // Pass 2: compute variance via BinnedAccumulator
            let diffs: Vec<f64> = slice.iter().map(|&x| {
                let diff = x - mean;
                diff * diff
            }).collect();
            let variance = binned_sum_f64(&diffs) / d as f64;

            // Normalize, scale, shift
            let inv_std = 1.0 / (variance + eps).sqrt();
            for i in 0..d {
                let normalized = (slice[i] - mean) * inv_std;
                result[start + i] = gamma_data[i] * normalized + beta_data[i];
            }
        }

        Tensor::from_vec(result, &self.shape)
    }

    /// Apply a function element-wise, reusing the buffer when possible (COW).
    ///
    /// If the tensor is contiguous, starts at offset 0, and has refcount == 1,
    /// the data is mutated in place (zero allocations). Otherwise, allocates
    /// a new buffer.
    fn map_elementwise(&self, f: impl Fn(f64) -> f64) -> Tensor {
        if self.is_contiguous() && self.offset == 0 && self.buffer.refcount() == 1 {
            // Fast path: mutate in place (COW — we're the sole owner)
            let mut data = self.buffer.borrow_data().clone();
            for x in data.iter_mut() {
                *x = f(*x);
            }
            Tensor::from_vec(data, &self.shape).unwrap()
        } else {
            // Fallback: allocate new buffer
            let data = self.to_vec();
            let result: Vec<f64> = data.iter().map(|&x| f(x)).collect();
            Tensor::from_vec(result, &self.shape).unwrap()
        }
    }

    /// ReLU activation: `max(0, x)` element-wise.
    pub fn relu(&self) -> Tensor {
        self.map_elementwise(|x| if x > 0.0 { x } else { 0.0 })
    }

    /// Sigmoid activation: 1 / (1 + exp(-x)) element-wise.
    pub fn sigmoid(&self) -> Tensor {
        self.map_elementwise(|x| 1.0 / (1.0 + (-x).exp()))
    }

    /// Tanh activation element-wise.
    pub fn tanh_activation(&self) -> Tensor {
        self.map_elementwise(|x| x.tanh())
    }

    /// Leaky ReLU activation: max(alpha*x, x) element-wise.
    pub fn leaky_relu(&self, alpha: f64) -> Tensor {
        self.map_elementwise(move |x| if x > 0.0 { x } else { alpha * x })
    }

    /// SiLU (Swish) activation: x * sigmoid(x) element-wise.
    pub fn silu(&self) -> Tensor {
        let data = self.to_vec();
        let result: Vec<f64> = data.iter().map(|&x| x / (1.0 + (-x).exp())).collect();
        Tensor::from_vec(result, &self.shape).unwrap()
    }

    /// Mish activation: x * tanh(softplus(x)) = x * tanh(ln(1 + exp(x))).
    pub fn mish(&self) -> Tensor {
        let data = self.to_vec();
        let result: Vec<f64> = data.iter().map(|&x| {
            let sp = (1.0 + x.exp()).ln();
            x * sp.tanh()
        }).collect();
        Tensor::from_vec(result, &self.shape).unwrap()
    }

    /// Argmax: index of the maximum element (first occurrence, deterministic).
    pub fn argmax(&self) -> usize {
        let data = self.to_vec();
        let mut best_idx = 0;
        let mut best_val = f64::NEG_INFINITY;
        for (i, &v) in data.iter().enumerate() {
            if v > best_val || (v == best_val && i < best_idx) {
                best_val = v;
                best_idx = i;
            }
        }
        best_idx
    }

    /// Argmin: index of the minimum element (first occurrence, deterministic).
    pub fn argmin(&self) -> usize {
        let data = self.to_vec();
        let mut best_idx = 0;
        let mut best_val = f64::INFINITY;
        for (i, &v) in data.iter().enumerate() {
            if v < best_val || (v == best_val && i < best_idx) {
                best_val = v;
                best_idx = i;
            }
        }
        best_idx
    }

    /// Clamp all elements to [min, max].
    pub fn clamp(&self, min: f64, max: f64) -> Tensor {
        let data = self.to_vec();
        let result: Vec<f64> = data.iter().map(|&x| x.max(min).min(max)).collect();
        Tensor::from_vec(result, &self.shape).unwrap()
    }

    /// One-hot encoding: given a 1D tensor of integer indices and a depth,
    /// returns a 2D tensor of shape [len, depth].
    pub fn one_hot(indices: &[usize], depth: usize) -> Result<Tensor, RuntimeError> {
        let n = indices.len();
        let mut data = vec![0.0; n * depth];
        for (i, &idx) in indices.iter().enumerate() {
            if idx >= depth {
                return Err(RuntimeError::InvalidOperation(format!(
                    "one_hot: index {idx} >= depth {depth}"
                )));
            }
            data[i * depth + idx] = 1.0;
        }
        Tensor::from_vec(data, &[n, depth])
    }

    // -----------------------------------------------------------------------
    // Phase B4: Tensor extensions (cat, stack, topk)
    // -----------------------------------------------------------------------

    /// Concatenate tensors along existing axis.
    pub fn cat(tensors: &[&Tensor], axis: usize) -> Result<Tensor, RuntimeError> {
        if tensors.is_empty() {
            return Err(RuntimeError::InvalidOperation("cat: no tensors".to_string()));
        }
        let ndim = tensors[0].ndim();
        if axis >= ndim {
            return Err(RuntimeError::InvalidOperation(
                format!("cat: axis {axis} out of bounds for {ndim}D tensor"),
            ));
        }
        for (i, t) in tensors.iter().enumerate().skip(1) {
            if t.ndim() != ndim {
                return Err(RuntimeError::InvalidOperation(
                    format!("cat: tensor {i} has different ndim"),
                ));
            }
            for d in 0..ndim {
                if d != axis && t.shape[d] != tensors[0].shape[d] {
                    return Err(RuntimeError::InvalidOperation(
                        format!("cat: shape mismatch at dim {d}"),
                    ));
                }
            }
        }
        let mut out_shape = tensors[0].shape.clone();
        for t in tensors.iter().skip(1) {
            out_shape[axis] += t.shape[axis];
        }
        let total = out_shape.iter().product::<usize>();
        let mut result = vec![0.0; total];
        let mut out_strides = vec![1usize; ndim];
        for d in (0..ndim - 1).rev() {
            out_strides[d] = out_strides[d + 1] * out_shape[d + 1];
        }
        let mut offset = 0;
        for t in tensors {
            let t_data = t.to_vec();
            let t_total: usize = t.shape.iter().product();
            let mut t_strides = vec![1usize; ndim];
            for d in (0..ndim - 1).rev() {
                t_strides[d] = t_strides[d + 1] * t.shape[d + 1];
            }
            for idx in 0..t_total {
                let mut remaining = idx;
                let mut out_flat = 0;
                for d in 0..ndim {
                    let coord = remaining / t_strides[d];
                    remaining %= t_strides[d];
                    let out_coord = if d == axis { coord + offset } else { coord };
                    out_flat += out_coord * out_strides[d];
                }
                result[out_flat] = t_data[idx];
            }
            offset += t.shape[axis];
        }
        Tensor::from_vec(result, &out_shape)
    }

    /// Stack tensors along a new axis.
    pub fn stack(tensors: &[&Tensor], axis: usize) -> Result<Tensor, RuntimeError> {
        if tensors.is_empty() {
            return Err(RuntimeError::InvalidOperation("stack: no tensors".to_string()));
        }
        let base_shape = &tensors[0].shape;
        let ndim = base_shape.len();
        if axis > ndim {
            return Err(RuntimeError::InvalidOperation(
                format!("stack: axis {axis} out of bounds"),
            ));
        }
        for (i, t) in tensors.iter().enumerate().skip(1) {
            if &t.shape != base_shape {
                return Err(RuntimeError::InvalidOperation(
                    format!("stack: tensor {i} shape mismatch"),
                ));
            }
        }
        let mut out_shape = Vec::with_capacity(ndim + 1);
        for d in 0..axis { out_shape.push(base_shape[d]); }
        out_shape.push(tensors.len());
        for d in axis..ndim { out_shape.push(base_shape[d]); }
        let total: usize = out_shape.iter().product();
        let mut result = vec![0.0; total];
        let inner_size: usize = base_shape[axis..].iter().product::<usize>().max(1);
        let outer_size: usize = base_shape[..axis].iter().product::<usize>().max(1);
        for (t_idx, t) in tensors.iter().enumerate() {
            let t_data = t.to_vec();
            for outer in 0..outer_size {
                for inner in 0..inner_size {
                    let src = outer * inner_size + inner;
                    let dst = outer * (tensors.len() * inner_size) + t_idx * inner_size + inner;
                    if src < t_data.len() && dst < result.len() {
                        result[dst] = t_data[src];
                    }
                }
            }
        }
        Tensor::from_vec(result, &out_shape)
    }

    /// Top-k values and indices (largest k values from flat data).
    pub fn topk(&self, k: usize) -> Result<(Tensor, Vec<usize>), RuntimeError> {
        let data = self.to_vec();
        let n = data.len();
        if k > n {
            return Err(RuntimeError::InvalidOperation(
                format!("topk: k={k} exceeds data length {n}"),
            ));
        }
        let mut indexed: Vec<(usize, f64)> = data.into_iter().enumerate().collect();
        indexed.sort_by(|a, b| b.1.total_cmp(&a.1).then(a.0.cmp(&b.0)));
        let top_k: Vec<(usize, f64)> = indexed[..k].to_vec();
        let values: Vec<f64> = top_k.iter().map(|&(_, v)| v).collect();
        let indices: Vec<usize> = top_k.iter().map(|&(i, _)| i).collect();
        Ok((Tensor::from_vec(values, &[k])?, indices))
    }

    /// GELU activation (approximate): x * 0.5 * (1 + tanh(√(2/π) * (x + 0.044715 * x³)))
    pub fn gelu(&self) -> Tensor {
        let data = self.to_vec();
        let sqrt_2_over_pi = (2.0_f64 / std::f64::consts::PI).sqrt();
        let result: Vec<f64> = data.iter().map(|&x| {
            let inner = sqrt_2_over_pi * (x + 0.044715 * x * x * x);
            0.5 * x * (1.0 + inner.tanh())
        }).collect();
        Tensor::from_vec(result, &self.shape).unwrap()
    }

    /// Linear layer: output = input @ weight^T + bias
    ///
    /// `self` is `[..., in_features]`, `weight` is `[out_features, in_features]`,
    /// `bias` is `[out_features]`.
    /// Result is `[..., out_features]`.
    pub fn linear(
        &self,
        weight: &Tensor,
        bias: &Tensor,
    ) -> Result<Tensor, RuntimeError> {
        if weight.ndim() != 2 {
            return Err(RuntimeError::InvalidOperation(
                "linear: weight must be 2-D [out_features, in_features]".to_string(),
            ));
        }
        let out_features = weight.shape[0];
        let in_features = weight.shape[1];
        let last_dim = *self.shape.last().ok_or_else(|| {
            RuntimeError::InvalidOperation("linear: input must be at least 1-D".to_string())
        })?;
        if last_dim != in_features {
            return Err(RuntimeError::DimensionMismatch {
                expected: in_features,
                got: last_dim,
            });
        }
        if bias.len() != out_features {
            return Err(RuntimeError::InvalidOperation(
                format!(
                    "linear: bias length {} must match out_features {}",
                    bias.len(),
                    out_features
                ),
            ));
        }

        let data = self.to_vec();
        let w = weight.to_vec();
        let b = bias.to_vec();
        let outer = data.len() / in_features;
        let mut result = vec![0.0f64; outer * out_features];

        for row in 0..outer {
            let x_start = row * in_features;
            let x_slice = &data[x_start..x_start + in_features];
            let y_start = row * out_features;
            for j in 0..out_features {
                let w_start = j * in_features;
                let mut acc = BinnedAccumulatorF64::new();
                for p in 0..in_features {
                    acc.add(x_slice[p] * w[w_start + p]);
                }
                result[y_start + j] = acc.finalize() + b[j];
            }
        }

        let mut out_shape = self.shape[..self.shape.len() - 1].to_vec();
        out_shape.push(out_features);
        Tensor::from_vec(result, &out_shape)
    }

    /// 1D convolution: signal `[signal_len]` * filters `[out_ch, kernel_size]` + bias
    ///
    /// Returns `[out_ch, signal_len - kernel_size + 1]` (valid mode, stride=1).
    pub fn conv1d(
        &self,
        filters: &Tensor,
        bias: &Tensor,
    ) -> Result<Tensor, RuntimeError> {
        if self.ndim() != 1 {
            return Err(RuntimeError::InvalidOperation(
                "conv1d: input must be 1-D [signal_len]".to_string(),
            ));
        }
        if filters.ndim() != 2 {
            return Err(RuntimeError::InvalidOperation(
                "conv1d: filters must be 2-D [out_channels, kernel_size]".to_string(),
            ));
        }
        let signal_len = self.shape[0];
        let out_channels = filters.shape[0];
        let kernel_size = filters.shape[1];
        if signal_len < kernel_size {
            return Err(RuntimeError::InvalidOperation(
                format!(
                    "conv1d: signal_len {} < kernel_size {}",
                    signal_len, kernel_size
                ),
            ));
        }
        if bias.len() != out_channels {
            return Err(RuntimeError::InvalidOperation(
                format!(
                    "conv1d: bias length {} must match out_channels {}",
                    bias.len(), out_channels
                ),
            ));
        }
        let out_len = signal_len - kernel_size + 1;
        let s = self.to_vec();
        let f = filters.to_vec();
        let b = bias.to_vec();
        let mut result = vec![0.0; out_channels * out_len];
        kernel_fns::conv1d_raw(&s, &f, &b, &mut result, signal_len, out_channels, kernel_size);
        Tensor::from_vec(result, &[out_channels, out_len])
    }

    /// 2D convolution — NCHW layout, valid mode, configurable stride.
    ///
    /// # Arguments
    /// - `self`:    `[N, C_in, H, W]` input tensor
    /// - `filters`: `[C_out, C_in, kH, kW]`
    /// - `bias`:    `[C_out]`
    /// - `stride`:  spatial stride (default 1)
    ///
    /// # Returns
    /// `[N, C_out, H_out, W_out]` where `H_out = (H - kH) / stride + 1`.
    ///
    /// Uses `BinnedAccumulatorF64` for every dot product — bit-identical results
    /// across all runs and hardware configurations.
    pub fn conv2d(
        &self,
        filters: &Tensor,
        bias: &Tensor,
        stride: usize,
    ) -> Result<Tensor, RuntimeError> {
        if self.ndim() != 4 {
            return Err(RuntimeError::InvalidOperation(
                "conv2d: input must be 4-D [N, C_in, H, W]".to_string(),
            ));
        }
        if filters.ndim() != 4 {
            return Err(RuntimeError::InvalidOperation(
                "conv2d: filters must be 4-D [C_out, C_in, kH, kW]".to_string(),
            ));
        }
        if stride == 0 {
            return Err(RuntimeError::InvalidOperation(
                "conv2d: stride must be >= 1".to_string(),
            ));
        }

        let n    = self.shape[0];
        let c_in = self.shape[1];
        let h_in = self.shape[2];
        let w_in = self.shape[3];

        let c_out      = filters.shape[0];
        let c_in_check = filters.shape[1];
        let kh         = filters.shape[2];
        let kw         = filters.shape[3];

        if c_in != c_in_check {
            return Err(RuntimeError::InvalidOperation(format!(
                "conv2d: input C_in={} does not match filter C_in={}",
                c_in, c_in_check
            )));
        }
        if h_in < kh || w_in < kw {
            return Err(RuntimeError::InvalidOperation(format!(
                "conv2d: input spatial [{}, {}] is smaller than kernel [{}, {}]",
                h_in, w_in, kh, kw
            )));
        }
        if bias.len() != c_out {
            return Err(RuntimeError::InvalidOperation(format!(
                "conv2d: bias length {} must match C_out={}",
                bias.len(), c_out
            )));
        }

        let h_out = (h_in - kh) / stride + 1;
        let w_out = (w_in - kw) / stride + 1;

        let inp = self.to_vec();
        let flt = filters.to_vec();
        let b   = bias.to_vec();
        let mut result = vec![0.0f64; n * c_out * h_out * w_out];

        kernel_fns::conv2d_raw(&inp, &flt, &b, &mut result,
                           n, c_in, h_in, w_in, c_out, kh, kw, stride);

        Tensor::from_vec(result, &[n, c_out, h_out, w_out])
    }

    /// 2D max-pooling — NCHW layout, non-overlapping windows.
    ///
    /// - `self`: `[N, C, H, W]`
    /// - `ph`, `pw`: pool height/width (stride = window size)
    ///
    /// Returns `[N, C, H/ph, W/pw]`.
    pub fn maxpool2d(&self, ph: usize, pw: usize) -> Result<Tensor, RuntimeError> {
        if self.ndim() != 4 {
            return Err(RuntimeError::InvalidOperation(
                "maxpool2d: input must be 4-D [N, C, H, W]".to_string(),
            ));
        }
        if ph == 0 || pw == 0 {
            return Err(RuntimeError::InvalidOperation(
                "maxpool2d: pool size must be >= 1".to_string(),
            ));
        }

        let n    = self.shape[0];
        let c    = self.shape[1];
        let h_in = self.shape[2];
        let w_in = self.shape[3];

        if h_in < ph || w_in < pw {
            return Err(RuntimeError::InvalidOperation(format!(
                "maxpool2d: input [{}, {}] smaller than pool [{}, {}]",
                h_in, w_in, ph, pw
            )));
        }

        let h_out = h_in / ph;
        let w_out = w_in / pw;

        let inp = self.to_vec();
        let mut result = vec![0.0f64; n * c * h_out * w_out];

        kernel_fns::maxpool2d_raw(&inp, &mut result, n, c, h_in, w_in, ph, pw);

        Tensor::from_vec(result, &[n, c, h_out, w_out])
    }

    /// Applies 2-D average pooling over a `[C, H, W]` tensor.
    ///
    /// # Arguments
    ///
    /// * `kernel_h` / `kernel_w` - Pooling window size
    /// * `stride_h` / `stride_w` - Stride for the pooling window
    ///
    /// # Returns
    ///
    /// Tensor of shape `[C, out_h, out_w]` where `out_h = (H - kernel_h) / stride_h + 1`.
    ///
    /// # Errors
    ///
    /// Returns an error if the tensor is not 3-D or if kernel/stride produce invalid output dimensions.
    pub fn avgpool2d(&self, kernel_h: usize, kernel_w: usize, stride_h: usize, stride_w: usize) -> Result<Tensor, RuntimeError> {
        let shape = self.shape();
        if shape.len() != 3 {
            return Err(RuntimeError::InvalidOperation(format!("avgpool2d requires 3-D [C,H,W], got {:?}", shape)));
        }
        let (c, h, w) = (shape[0], shape[1], shape[2]);
        if kernel_h > h || kernel_w > w {
            return Err(RuntimeError::InvalidOperation("avgpool2d: kernel larger than input".into()));
        }
        let out_h = (h - kernel_h) / stride_h + 1;
        let out_w = (w - kernel_w) / stride_w + 1;
        let data = self.to_vec();
        let mut out = Vec::with_capacity(c * out_h * out_w);
        let pool_size = (kernel_h * kernel_w) as f64;

        for ch in 0..c {
            for oh in 0..out_h {
                for ow in 0..out_w {
                    let mut sum = 0.0f64;
                    for kh in 0..kernel_h {
                        for kw in 0..kernel_w {
                            let ih = oh * stride_h + kh;
                            let iw = ow * stride_w + kw;
                            sum += data[ch * h * w + ih * w + iw];
                        }
                    }
                    out.push(sum / pool_size);
                }
            }
        }
        Tensor::from_vec(out, &[c, out_h, out_w])
    }

    /// Scaled dot-product attention (single head).
    ///
    /// `queries` is `[..., T, d_k]`
    /// `keys`    is `[..., S, d_k]`
    /// `values`  is `[..., S, d_v]`
    ///
    /// Computes: softmax(Q × Kᵀ / √d_k) × V
    /// Returns `[..., T, d_v]`.
    pub fn scaled_dot_product_attention(
        queries: &Tensor,
        keys: &Tensor,
        values: &Tensor,
    ) -> Result<Tensor, RuntimeError> {
        if queries.ndim() < 2 || keys.ndim() < 2 || values.ndim() < 2 {
            return Err(RuntimeError::InvalidOperation(
                "attention: Q, K, V must be at least 2-D".to_string(),
            ));
        }
        let nd = queries.ndim();
        let d_k = queries.shape[nd - 1];
        let scale = 1.0 / (d_k as f64).sqrt();

        // Transpose keys: swap last two dims
        let keys_t = keys.transpose_last_two()?;

        // Q × K^T → [... T, S]
        let scores = queries.bmm(&keys_t)?;

        // Scale
        let scores_scaled = scores.scalar_mul(scale);

        // Softmax along last dim
        let attn_weights = scores_scaled.softmax()?;

        // Attn × V → [... T, d_v]
        attn_weights.bmm(values)
    }

    /// Transpose the last two dimensions of a tensor.
    ///
    /// `[..., A, B]` → `[..., B, A]`
    pub fn transpose_last_two(&self) -> Result<Tensor, RuntimeError> {
        if self.ndim() < 2 {
            return Err(RuntimeError::InvalidOperation(
                "transpose_last_two requires at least 2-D tensor".to_string(),
            ));
        }
        let nd = self.ndim();
        let rows = self.shape[nd - 2];
        let cols = self.shape[nd - 1];
        let data = self.to_vec();
        let batch_size: usize = self.shape[..nd - 2].iter().product::<usize>().max(1);
        let mat_size = rows * cols;
        let mut result = vec![0.0f64; data.len()];

        for b in 0..batch_size {
            let off = b * mat_size;
            for i in 0..rows {
                for j in 0..cols {
                    result[off + j * rows + i] = data[off + i * cols + j];
                }
            }
        }

        let mut out_shape = self.shape.clone();
        out_shape[nd - 2] = cols;
        out_shape[nd - 1] = rows;
        Tensor::from_vec(result, &out_shape)
    }

    // -- Zero-Copy Weight Mapping -------------------------------------------

    /// Create a tensor view from raw bytes — **zero allocation**.
    ///
    /// Interprets `bytes` as a contiguous block of `f64` (8 bytes each) or
    /// `f32` (4 bytes each, promoted to f64) values and maps them into a
    /// `Tensor` with the given shape.
    ///
    /// `dtype` must be `"f64"` or `"f32"`.
    ///
    /// For f64: bytes.len() must equal shape_numel * 8.
    /// For f32: bytes.len() must equal shape_numel * 4.
    ///
    /// The returned tensor **owns** its buffer (copied from the raw bytes)
    /// but performs exactly one allocation for the data vector.
    pub fn from_bytes(bytes: &[u8], shape: &[usize], dtype: &str) -> Result<Tensor, RuntimeError> {
        let numel = Self::shape_numel(shape);
        match dtype {
            "f64" => {
                let expected = numel * 8;
                if bytes.len() != expected {
                    return Err(RuntimeError::ShapeMismatch {
                        expected,
                        got: bytes.len(),
                    });
                }
                let mut data = Vec::with_capacity(numel);
                for i in 0..numel {
                    let off = i * 8;
                    let mut buf = [0u8; 8];
                    buf.copy_from_slice(&bytes[off..off + 8]);
                    data.push(f64::from_le_bytes(buf));
                }
                Ok(Tensor {
                    buffer: Buffer::from_vec(data),
                    shape: shape.to_vec(),
                    strides: Self::compute_strides(shape),
                    offset: 0,
                })
            }
            "f32" => {
                let expected = numel * 4;
                if bytes.len() != expected {
                    return Err(RuntimeError::ShapeMismatch {
                        expected,
                        got: bytes.len(),
                    });
                }
                let mut data = Vec::with_capacity(numel);
                for i in 0..numel {
                    let off = i * 4;
                    let mut buf = [0u8; 4];
                    buf.copy_from_slice(&bytes[off..off + 4]);
                    data.push(f32::from_le_bytes(buf) as f64);
                }
                Ok(Tensor {
                    buffer: Buffer::from_vec(data),
                    shape: shape.to_vec(),
                    strides: Self::compute_strides(shape),
                    offset: 0,
                })
            }
            _ => Err(RuntimeError::InvalidOperation(
                format!("from_bytes: unsupported dtype '{}', expected 'f32' or 'f64'", dtype),
            )),
        }
    }

    // -- Multi-Head Attention Splitting -------------------------------------

    /// Reshape a 3D tensor `[batch, seq, model_dim]` into 4D
    /// `[batch, num_heads, seq, head_dim]` by splitting the last dimension.
    ///
    /// This is a **zero-copy view** — it only changes shape/strides metadata.
    /// `model_dim` must be divisible by `num_heads`.
    pub fn split_heads(&self, num_heads: usize) -> Result<Tensor, RuntimeError> {
        if self.ndim() != 3 {
            return Err(RuntimeError::DimensionMismatch {
                expected: 3,
                got: self.ndim(),
            });
        }
        let batch = self.shape[0];
        let seq = self.shape[1];
        let model_dim = self.shape[2];
        if model_dim % num_heads != 0 {
            return Err(RuntimeError::InvalidOperation(
                format!(
                    "split_heads: model_dim {} not divisible by num_heads {}",
                    model_dim, num_heads
                ),
            ));
        }
        let head_dim = model_dim / num_heads;
        // Need contiguous data for the reshape
        let tensor = if self.is_contiguous() { self.clone() } else { self.to_contiguous() };
        // Reshape [B, S, H*D] -> [B, S, H, D] then transpose to [B, H, S, D]
        let reshaped = Tensor {
            buffer: tensor.buffer.clone(),
            shape: vec![batch, seq, num_heads, head_dim],
            strides: Self::compute_strides(&[batch, seq, num_heads, head_dim]),
            offset: 0,
        };
        // Transpose dims 1 and 2: [B, S, H, D] -> [B, H, S, D]
        // New strides: swap strides[1] and strides[2]
        Ok(Tensor {
            buffer: reshaped.buffer,
            shape: vec![batch, num_heads, seq, head_dim],
            strides: vec![
                reshaped.strides[0], // batch stride unchanged
                reshaped.strides[2], // head stride (was dim 2)
                reshaped.strides[1], // seq stride (was dim 1)
                reshaped.strides[3], // head_dim stride unchanged
            ],
            offset: 0,
        })
    }

    /// Merge heads back: reshape 4D `[batch, num_heads, seq, head_dim]` into
    /// 3D `[batch, seq, model_dim]`. Materializes if non-contiguous.
    pub fn merge_heads(&self) -> Result<Tensor, RuntimeError> {
        if self.ndim() != 4 {
            return Err(RuntimeError::DimensionMismatch {
                expected: 4,
                got: self.ndim(),
            });
        }
        let batch = self.shape[0];
        let num_heads = self.shape[1];
        let seq = self.shape[2];
        let head_dim = self.shape[3];
        // Need [B, H, S, D] -> [B, S, H, D] -> [B, S, H*D]
        // Transpose dims 1 and 2 first
        let transposed = Tensor {
            buffer: self.buffer.clone(),
            shape: vec![batch, seq, num_heads, head_dim],
            strides: vec![
                self.strides[0],
                self.strides[2], // seq stride
                self.strides[1], // head stride
                self.strides[3],
            ],
            offset: self.offset,
        };
        // Materialize contiguous then reshape
        let contig = transposed.to_contiguous();
        let model_dim = num_heads * head_dim;
        Ok(Tensor {
            buffer: contig.buffer,
            shape: vec![batch, seq, model_dim],
            strides: Self::compute_strides(&[batch, seq, model_dim]),
            offset: 0,
        })
    }

    /// View-only reshape: reinterpret shape without copying.
    /// Only works on contiguous tensors. Falls back to copy if non-contiguous.
    pub fn view_reshape(&self, new_shape: &[usize]) -> Result<Tensor, RuntimeError> {
        self.reshape(new_shape)
    }

    // -----------------------------------------------------------------------
    // Phase C4: Sorting & Tensor Indexing
    // -----------------------------------------------------------------------

    /// Returns indices that would sort the flattened tensor in ascending order.
    /// Uses f64::total_cmp for deterministic ordering of NaN.
    pub fn argsort(&self) -> Tensor {
        let data = self.to_vec();
        let mut indices: Vec<usize> = (0..data.len()).collect();
        indices.sort_by(|&a, &b| data[a].total_cmp(&data[b]));
        let result: Vec<f64> = indices.iter().map(|&i| i as f64).collect();
        Tensor::from_vec_unchecked(result, &[data.len()])
    }

    /// Gather elements from the tensor along a dimension using index tensor.
    /// For 1D: result[i] = self[indices[i]]
    /// For 2D dim=0: result[i][j] = self[indices[i][j]][j]
    /// For 2D dim=1: result[i][j] = self[i][indices[i][j]]
    pub fn gather(&self, dim: usize, indices: &Tensor) -> Result<Tensor, RuntimeError> {
        let data = self.to_vec();
        let idx_data = indices.to_vec();
        if self.ndim() == 1 {
            let mut result = Vec::with_capacity(idx_data.len());
            for &idx in &idx_data {
                let i = idx as usize;
                if i >= data.len() {
                    return Err(RuntimeError::InvalidOperation(
                        format!("gather: index {} out of bounds for size {}", i, data.len()),
                    ));
                }
                result.push(data[i]);
            }
            Ok(Tensor::from_vec_unchecked(result, indices.shape()))
        } else if self.ndim() == 2 {
            let rows = self.shape[0];
            let cols = self.shape[1];
            let idx_shape = indices.shape();
            let out_rows = idx_shape[0];
            let out_cols = idx_shape[1];
            let mut result = vec![0.0; out_rows * out_cols];
            for i in 0..out_rows {
                for j in 0..out_cols {
                    let idx = idx_data[i * out_cols + j] as usize;
                    let val = if dim == 0 {
                        if idx >= rows {
                            return Err(RuntimeError::InvalidOperation(
                                format!("gather dim=0: index {} out of bounds for {} rows", idx, rows),
                            ));
                        }
                        data[idx * cols + j]
                    } else {
                        if idx >= cols {
                            return Err(RuntimeError::InvalidOperation(
                                format!("gather dim=1: index {} out of bounds for {} cols", idx, cols),
                            ));
                        }
                        data[i * cols + idx]
                    };
                    result[i * out_cols + j] = val;
                }
            }
            Ok(Tensor::from_vec_unchecked(result, idx_shape))
        } else {
            Err(RuntimeError::InvalidOperation(
                "gather: only 1D and 2D tensors supported".into(),
            ))
        }
    }

    /// Scatter src values into a tensor of given shape at indices along a dimension.
    /// For 1D: result[indices[i]] = src[i]
    /// For 2D dim=0: result[indices[i][j]][j] = src[i][j]
    /// For 2D dim=1: result[i][indices[i][j]] = src[i][j]
    pub fn scatter(&self, dim: usize, indices: &Tensor, src: &Tensor) -> Result<Tensor, RuntimeError> {
        let mut result = self.to_vec();
        let idx_data = indices.to_vec();
        let src_data = src.to_vec();
        if self.ndim() == 1 {
            for (k, &idx) in idx_data.iter().enumerate() {
                let i = idx as usize;
                if i >= result.len() {
                    return Err(RuntimeError::InvalidOperation(
                        format!("scatter: index {} out of bounds for size {}", i, result.len()),
                    ));
                }
                result[i] = src_data[k];
            }
            Ok(Tensor::from_vec_unchecked(result, self.shape()))
        } else if self.ndim() == 2 {
            let cols = self.shape[1];
            let idx_shape = indices.shape();
            let out_cols = idx_shape[1];
            let out_rows = idx_shape[0];
            for i in 0..out_rows {
                for j in 0..out_cols {
                    let idx = idx_data[i * out_cols + j] as usize;
                    let src_val = src_data[i * out_cols + j];
                    if dim == 0 {
                        if idx >= self.shape[0] {
                            return Err(RuntimeError::InvalidOperation(
                                format!("scatter dim=0: index {} out of bounds for {} rows", idx, self.shape[0]),
                            ));
                        }
                        result[idx * cols + j] = src_val;
                    } else {
                        if idx >= cols {
                            return Err(RuntimeError::InvalidOperation(
                                format!("scatter dim=1: index {} out of bounds for {} cols", idx, cols),
                            ));
                        }
                        result[i * cols + idx] = src_val;
                    }
                }
            }
            Ok(Tensor::from_vec_unchecked(result, self.shape()))
        } else {
            Err(RuntimeError::InvalidOperation(
                "scatter: only 1D and 2D tensors supported".into(),
            ))
        }
    }

    /// Select slices along a dimension by index.
    /// For 2D dim=0: selects rows
    /// For 2D dim=1: selects columns
    pub fn index_select(&self, dim: usize, indices: &Tensor) -> Result<Tensor, RuntimeError> {
        let data = self.to_vec();
        let idx_data = indices.to_vec();
        if self.ndim() == 1 {
            let mut result = Vec::with_capacity(idx_data.len());
            for &idx in &idx_data {
                let i = idx as usize;
                if i >= data.len() {
                    return Err(RuntimeError::InvalidOperation(
                        format!("index_select: index {} out of bounds for size {}", i, data.len()),
                    ));
                }
                result.push(data[i]);
            }
            Ok(Tensor::from_vec_unchecked(result, &[idx_data.len()]))
        } else if self.ndim() == 2 {
            let rows = self.shape[0];
            let cols = self.shape[1];
            let n = idx_data.len();
            if dim == 0 {
                let mut result = Vec::with_capacity(n * cols);
                for &idx in &idx_data {
                    let i = idx as usize;
                    if i >= rows {
                        return Err(RuntimeError::InvalidOperation(
                            format!("index_select dim=0: index {} out of bounds for {} rows", i, rows),
                        ));
                    }
                    for j in 0..cols {
                        result.push(data[i * cols + j]);
                    }
                }
                Ok(Tensor::from_vec_unchecked(result, &[n, cols]))
            } else {
                let mut result = Vec::with_capacity(rows * n);
                for i in 0..rows {
                    for &idx in &idx_data {
                        let j = idx as usize;
                        if j >= cols {
                            return Err(RuntimeError::InvalidOperation(
                                format!("index_select dim=1: index {} out of bounds for {} cols", j, cols),
                            ));
                        }
                        result.push(data[i * cols + j]);
                    }
                }
                Ok(Tensor::from_vec_unchecked(result, &[rows, n]))
            }
        } else {
            Err(RuntimeError::InvalidOperation(
                "index_select: only 1D and 2D tensors supported".into(),
            ))
        }
    }

    // -----------------------------------------------------------------------
    // Phase 2: Boolean / Masking Ops
    // -----------------------------------------------------------------------

    /// Element-wise conditional select: `where(condition, other)`.
    /// For each element, returns `self[i]` if `condition[i] != 0.0`, else `other[i]`.
    pub fn tensor_where(&self, condition: &Tensor, other: &Tensor) -> Result<Tensor, RuntimeError> {
        if self.shape() != condition.shape() || self.shape() != other.shape() {
            return Err(RuntimeError::InvalidOperation(
                format!("where: shape mismatch self={:?} cond={:?} other={:?}",
                    self.shape(), condition.shape(), other.shape()),
            ));
        }
        let s = self.to_vec();
        let c = condition.to_vec();
        let o = other.to_vec();
        let result: Vec<f64> = s.iter().zip(c.iter()).zip(o.iter())
            .map(|((&sv, &cv), &ov)| if cv != 0.0 { sv } else { ov })
            .collect();
        Tensor::from_vec(result, self.shape())
    }

    /// Return `true` if any element is non-zero.
    pub fn any(&self) -> bool {
        let data = self.to_vec();
        data.iter().any(|&x| x != 0.0)
    }

    /// Return `true` if all elements are non-zero.
    pub fn all(&self) -> bool {
        let data = self.to_vec();
        data.iter().all(|&x| x != 0.0)
    }

    /// Return a 1-D tensor of flat indices where elements are non-zero.
    ///
    /// If no elements are non-zero, returns an empty tensor of shape `[0]`.
    pub fn nonzero(&self) -> Tensor {
        let data = self.to_vec();
        let indices: Vec<f64> = data.iter().enumerate()
            .filter(|(_, &v)| v != 0.0)
            .map(|(i, _)| i as f64)
            .collect();
        let len = indices.len();
        if len == 0 {
            Tensor::from_vec(vec![], &[0]).unwrap()
        } else {
            Tensor::from_vec(indices, &[len]).unwrap()
        }
    }

    /// Fill elements where `mask` is non-zero with `value`.
    pub fn masked_fill(&self, mask: &Tensor, value: f64) -> Result<Tensor, RuntimeError> {
        if self.shape() != mask.shape() {
            return Err(RuntimeError::InvalidOperation(
                format!("masked_fill: shape mismatch self={:?} mask={:?}",
                    self.shape(), mask.shape()),
            ));
        }
        let data = self.to_vec();
        let m = mask.to_vec();
        let result: Vec<f64> = data.iter().zip(m.iter())
            .map(|(&d, &mv)| if mv != 0.0 { value } else { d })
            .collect();
        Tensor::from_vec(result, self.shape())
    }

    // -----------------------------------------------------------------------
    // Phase 2: Axis Reductions with keepdim
    // -----------------------------------------------------------------------

    /// Generic axis reduction using a caller-provided reduction function.
    ///
    /// Gathers values along the specified axis for each output position
    /// and applies `reduce_fn`. If `keepdim` is `true`, the reduced axis
    /// retains size 1; otherwise it is removed from the output shape.
    fn reduce_axis<F>(&self, axis: usize, keepdim: bool, reduce_fn: F)
        -> Result<Tensor, RuntimeError>
    where
        F: Fn(&[f64]) -> f64,
    {
        let ndim = self.ndim();
        if axis >= ndim {
            return Err(RuntimeError::IndexOutOfBounds {
                index: axis,
                length: ndim,
            });
        }

        let axis_len = self.shape[axis];
        // Build output shape
        let mut out_shape: Vec<usize> = self.shape.clone();
        out_shape[axis] = 1;
        let out_numel = Self::shape_numel(&out_shape);
        let out_strides = Self::compute_strides(&out_shape);

        let data = self.to_vec();
        let mut result = Vec::with_capacity(out_numel);
        let mut indices = vec![0usize; ndim];

        for out_idx in 0..out_numel {
            // Compute N-D index from flat output index
            {
                let mut remaining = out_idx;
                for d in 0..ndim {
                    indices[d] = remaining / out_strides[d];
                    remaining %= out_strides[d];
                }
            }

            // Gather values along the reduction axis
            let mut vals = Vec::with_capacity(axis_len);
            for k in 0..axis_len {
                let mut flat = self.offset;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { indices[d] };
                    flat += idx * self.strides[d];
                }
                vals.push(data[flat]);
            }
            result.push(reduce_fn(&vals));
        }

        let final_shape = if keepdim {
            out_shape
        } else {
            // Remove the axis dimension
            let mut s: Vec<usize> = self.shape.iter().enumerate()
                .filter(|&(i, _)| i != axis)
                .map(|(_, &v)| v)
                .collect();
            if s.is_empty() {
                s.push(1); // scalar result
            }
            s
        };

        Tensor::from_vec(result, &final_shape)
    }

    /// Mean along an axis with optional keepdim.
    ///
    /// Uses [`BinnedAccumulatorF64`](crate::accumulator::BinnedAccumulatorF64)
    /// for deterministic summation before dividing by the axis length.
    pub fn mean_axis(&self, axis: usize, keepdim: bool) -> Result<Tensor, RuntimeError> {
        self.reduce_axis(axis, keepdim, |vals| {
            let mut acc = BinnedAccumulatorF64::new();
            for &v in vals { acc.add(v); }
            acc.finalize() / vals.len() as f64
        })
    }

    /// Max along an axis with optional keepdim. Return `(values, indices)`.
    ///
    /// Ties are broken by choosing the first occurrence (smallest index).
    pub fn max_axis(&self, axis: usize, keepdim: bool) -> Result<(Tensor, Tensor), RuntimeError> {
        let ndim = self.ndim();
        if axis >= ndim {
            return Err(RuntimeError::IndexOutOfBounds { index: axis, length: ndim });
        }
        let axis_len = self.shape[axis];
        let mut out_shape = self.shape.clone();
        out_shape[axis] = 1;
        let out_numel = Self::shape_numel(&out_shape);
        let out_strides = Self::compute_strides(&out_shape);
        let data = self.to_vec();
        let mut values = Vec::with_capacity(out_numel);
        let mut idx_vals = Vec::with_capacity(out_numel);
        let mut indices = vec![0usize; ndim];

        for out_idx in 0..out_numel {
            let mut remaining = out_idx;
            for d in 0..ndim {
                indices[d] = remaining / out_strides[d];
                remaining %= out_strides[d];
            }
            let mut best_val = f64::NEG_INFINITY;
            let mut best_idx = 0usize;
            for k in 0..axis_len {
                let mut flat = self.offset;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { indices[d] };
                    flat += idx * self.strides[d];
                }
                let v = data[flat];
                if v > best_val {
                    best_val = v;
                    best_idx = k;
                }
            }
            values.push(best_val);
            idx_vals.push(best_idx as f64);
        }

        let final_shape = if keepdim {
            out_shape
        } else {
            let mut s: Vec<usize> = self.shape.iter().enumerate()
                .filter(|&(i, _)| i != axis).map(|(_, &v)| v).collect();
            if s.is_empty() { s.push(1); }
            s
        };
        Ok((
            Tensor::from_vec(values, &final_shape)?,
            Tensor::from_vec(idx_vals, &final_shape)?,
        ))
    }

    /// Min along an axis with optional keepdim. Return `(values, indices)`.
    ///
    /// Ties are broken by choosing the first occurrence (smallest index).
    pub fn min_axis(&self, axis: usize, keepdim: bool) -> Result<(Tensor, Tensor), RuntimeError> {
        let ndim = self.ndim();
        if axis >= ndim {
            return Err(RuntimeError::IndexOutOfBounds { index: axis, length: ndim });
        }
        let axis_len = self.shape[axis];
        let mut out_shape = self.shape.clone();
        out_shape[axis] = 1;
        let out_numel = Self::shape_numel(&out_shape);
        let out_strides = Self::compute_strides(&out_shape);
        let data = self.to_vec();
        let mut values = Vec::with_capacity(out_numel);
        let mut idx_vals = Vec::with_capacity(out_numel);
        let mut indices = vec![0usize; ndim];

        for out_idx in 0..out_numel {
            let mut remaining = out_idx;
            for d in 0..ndim {
                indices[d] = remaining / out_strides[d];
                remaining %= out_strides[d];
            }
            let mut best_val = f64::INFINITY;
            let mut best_idx = 0usize;
            for k in 0..axis_len {
                let mut flat = self.offset;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { indices[d] };
                    flat += idx * self.strides[d];
                }
                let v = data[flat];
                if v < best_val {
                    best_val = v;
                    best_idx = k;
                }
            }
            values.push(best_val);
            idx_vals.push(best_idx as f64);
        }

        let final_shape = if keepdim {
            out_shape
        } else {
            let mut s: Vec<usize> = self.shape.iter().enumerate()
                .filter(|&(i, _)| i != axis).map(|(_, &v)| v).collect();
            if s.is_empty() { s.push(1); }
            s
        };
        Ok((
            Tensor::from_vec(values, &final_shape)?,
            Tensor::from_vec(idx_vals, &final_shape)?,
        ))
    }

    /// Variance along an axis with optional keepdim.
    ///
    /// Computes population variance: `Var = sum((x - mean)^2) / N`.
    /// Uses [`BinnedAccumulatorF64`](crate::accumulator::BinnedAccumulatorF64)
    /// for the squared-differences summation.
    pub fn var_axis(&self, axis: usize, keepdim: bool) -> Result<Tensor, RuntimeError> {
        let mean_t = self.mean_axis(axis, true)?;
        let ndim = self.ndim();
        if axis >= ndim {
            return Err(RuntimeError::IndexOutOfBounds { index: axis, length: ndim });
        }
        let axis_len = self.shape[axis];
        let mut out_shape = self.shape.clone();
        out_shape[axis] = 1;
        let out_numel = Self::shape_numel(&out_shape);
        let out_strides = Self::compute_strides(&out_shape);
        let data = self.to_vec();
        let mean_data = mean_t.to_vec();
        let mut result = Vec::with_capacity(out_numel);
        let mut indices = vec![0usize; ndim];

        for out_idx in 0..out_numel {
            let mut remaining = out_idx;
            for d in 0..ndim {
                indices[d] = remaining / out_strides[d];
                remaining %= out_strides[d];
            }
            let mu = mean_data[out_idx];
            let mut acc = BinnedAccumulatorF64::new();
            for k in 0..axis_len {
                let mut flat = self.offset;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { indices[d] };
                    flat += idx * self.strides[d];
                }
                let diff = data[flat] - mu;
                acc.add(diff * diff);
            }
            result.push(acc.finalize() / axis_len as f64);
        }

        let final_shape = if keepdim {
            out_shape
        } else {
            let mut s: Vec<usize> = self.shape.iter().enumerate()
                .filter(|&(i, _)| i != axis).map(|(_, &v)| v).collect();
            if s.is_empty() { s.push(1); }
            s
        };
        Tensor::from_vec(result, &final_shape)
    }

    /// Standard deviation along an axis with optional keepdim.
    ///
    /// Computed as `sqrt(var_axis(axis, keepdim))`.
    pub fn std_axis(&self, axis: usize, keepdim: bool) -> Result<Tensor, RuntimeError> {
        let var = self.var_axis(axis, keepdim)?;
        Ok(var.map(|x| x.sqrt()))
    }

    /// Product along an axis with optional keepdim.
    ///
    /// Computes a simple sequential product (exact for integer-like values).
    pub fn prod_axis(&self, axis: usize, keepdim: bool) -> Result<Tensor, RuntimeError> {
        self.reduce_axis(axis, keepdim, |vals| {
            // Product via exp(sum(ln(abs))) for numerical stability is overkill here;
            // simple product is deterministic and exact for integer-like values.
            let mut product = 1.0f64;
            for &v in vals { product *= v; }
            product
        })
    }

    // -----------------------------------------------------------------------
    // Phase 2: Sort Operations
    // -----------------------------------------------------------------------

    /// Sort along an axis (stable sort). Return the sorted tensor.
    ///
    /// For N-D tensors, independently sorts each 1-D slice along the specified
    /// axis. Uses `f64::partial_cmp` with deterministic tie-breaking by
    /// original index position.
    pub fn sort_axis(&self, axis: usize, descending: bool) -> Result<Tensor, RuntimeError> {
        let ndim = self.ndim();
        if axis >= ndim {
            return Err(RuntimeError::IndexOutOfBounds { index: axis, length: ndim });
        }
        let data = self.to_vec();
        let axis_len = self.shape[axis];
        let out_shape = self.shape.clone();
        let out_numel = Self::shape_numel(&out_shape);

        // Build strides for iterating over all non-axis positions
        let mut iter_shape: Vec<usize> = Vec::new();
        for (i, &s) in self.shape.iter().enumerate() {
            if i != axis { iter_shape.push(s); }
        }
        let n_slices: usize = iter_shape.iter().product::<usize>().max(1);

        let mut result = vec![0.0f64; out_numel];

        // We iterate over all positions with axis index = 0
        let mut pos = vec![0usize; ndim];
        for slice_idx in 0..n_slices {
            // Compute the N-D position (with axis dim = 0)
            let mut remaining = slice_idx;
            let mut dim_idx = 0;
            for d in 0..ndim {
                if d == axis {
                    pos[d] = 0;
                } else {
                    let stride = {
                        let mut s = 1usize;
                        let mut di = 0;
                        for d2 in 0..ndim {
                            if d2 == axis { continue; }
                            if di > dim_idx { s *= self.shape[d2]; }
                            di += 1;
                        }
                        s
                    };
                    pos[d] = remaining / stride;
                    remaining %= stride;
                    dim_idx += 1;
                }
            }

            // Gather values along axis
            let mut vals: Vec<(f64, usize)> = Vec::with_capacity(axis_len);
            for k in 0..axis_len {
                let mut flat = self.offset;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { pos[d] };
                    flat += idx * self.strides[d];
                }
                vals.push((data[flat], k));
            }

            // Stable sort with deterministic tie-breaking by original index
            if descending {
                vals.sort_by(|a, b| b.0.partial_cmp(&a.0).unwrap_or(std::cmp::Ordering::Equal)
                    .then(a.1.cmp(&b.1)));
            } else {
                vals.sort_by(|a, b| a.0.partial_cmp(&b.0).unwrap_or(std::cmp::Ordering::Equal)
                    .then(a.1.cmp(&b.1)));
            }

            // Scatter back
            for (k, &(v, _)) in vals.iter().enumerate() {
                let mut flat = 0;
                let out_strides_local = Self::compute_strides(&out_shape);
                for d in 0..ndim {
                    let idx = if d == axis { k } else { pos[d] };
                    flat += idx * out_strides_local[d];
                }
                result[flat] = v;
            }
        }

        Tensor::from_vec(result, &out_shape)
    }

    /// N-D argsort along an axis. Return a tensor of indices that would sort
    /// each slice along the given axis.
    ///
    /// Deterministic tie-breaking: ties are resolved by original index order.
    pub fn argsort_axis(&self, axis: usize, descending: bool) -> Result<Tensor, RuntimeError> {
        let ndim = self.ndim();
        if axis >= ndim {
            return Err(RuntimeError::IndexOutOfBounds { index: axis, length: ndim });
        }
        let data = self.to_vec();
        let axis_len = self.shape[axis];
        let out_shape = self.shape.clone();
        let out_numel = Self::shape_numel(&out_shape);

        let mut iter_shape: Vec<usize> = Vec::new();
        for (i, &s) in self.shape.iter().enumerate() {
            if i != axis { iter_shape.push(s); }
        }
        let n_slices: usize = iter_shape.iter().product::<usize>().max(1);

        let mut result = vec![0.0f64; out_numel];
        let mut pos = vec![0usize; ndim];

        for slice_idx in 0..n_slices {
            let mut remaining = slice_idx;
            let mut dim_idx = 0;
            for d in 0..ndim {
                if d == axis {
                    pos[d] = 0;
                } else {
                    let stride = {
                        let mut s = 1usize;
                        let mut di = 0;
                        for d2 in 0..ndim {
                            if d2 == axis { continue; }
                            if di > dim_idx { s *= self.shape[d2]; }
                            di += 1;
                        }
                        s
                    };
                    pos[d] = remaining / stride;
                    remaining %= stride;
                    dim_idx += 1;
                }
            }

            let mut vals: Vec<(f64, usize)> = Vec::with_capacity(axis_len);
            for k in 0..axis_len {
                let mut flat = self.offset;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { pos[d] };
                    flat += idx * self.strides[d];
                }
                vals.push((data[flat], k));
            }

            if descending {
                vals.sort_by(|a, b| b.0.partial_cmp(&a.0).unwrap_or(std::cmp::Ordering::Equal)
                    .then(a.1.cmp(&b.1)));
            } else {
                vals.sort_by(|a, b| a.0.partial_cmp(&b.0).unwrap_or(std::cmp::Ordering::Equal)
                    .then(a.1.cmp(&b.1)));
            }

            for (k, &(_, orig_idx)) in vals.iter().enumerate() {
                let out_strides_local = Self::compute_strides(&out_shape);
                let mut flat = 0;
                for d in 0..ndim {
                    let idx = if d == axis { k } else { pos[d] };
                    flat += idx * out_strides_local[d];
                }
                result[flat] = orig_idx as f64;
            }
        }

        Tensor::from_vec(result, &out_shape)
    }

    // -----------------------------------------------------------------------
    // Phase 2: Einsum
    // -----------------------------------------------------------------------

    /// Einstein summation notation.
    /// Supports patterns like "ij,jk->ik" (matmul), "ii->i" (diagonal),
    /// "ij->ji" (transpose), "ijk,ikl->ijl" (batched matmul).
    /// Uses BinnedAccumulator for all reductions.
    pub fn einsum(notation: &str, inputs: &[&Tensor]) -> Result<Tensor, RuntimeError> {
        // Parse notation: "subscripts->output" or just "subscripts" (implicit)
        let parts: Vec<&str> = notation.split("->").collect();
        if parts.len() != 2 {
            return Err(RuntimeError::InvalidOperation(
                format!("einsum: expected 'subscripts->output' notation, got '{}'", notation),
            ));
        }
        let input_specs: Vec<&str> = parts[0].split(',').collect();
        let output_spec = parts[1];

        if input_specs.len() != inputs.len() {
            return Err(RuntimeError::InvalidOperation(
                format!("einsum: {} input specs but {} tensors", input_specs.len(), inputs.len()),
            ));
        }

        // Build label → size mapping
        let mut label_size = std::collections::BTreeMap::new();
        for (i, &spec) in input_specs.iter().enumerate() {
            let chars: Vec<char> = spec.chars().collect();
            if chars.len() != inputs[i].ndim() {
                return Err(RuntimeError::InvalidOperation(
                    format!("einsum: spec '{}' has {} dims but tensor has {}", spec, chars.len(), inputs[i].ndim()),
                ));
            }
            for (d, &c) in chars.iter().enumerate() {
                let sz = inputs[i].shape()[d];
                if let Some(&prev) = label_size.get(&c) {
                    if prev != sz {
                        return Err(RuntimeError::InvalidOperation(
                            format!("einsum: label '{}' has conflicting sizes {} vs {}", c, prev, sz),
                        ));
                    }
                } else {
                    label_size.insert(c, sz);
                }
            }
        }

        // Determine output shape
        let output_chars: Vec<char> = output_spec.chars().collect();
        let output_shape: Vec<usize> = output_chars.iter()
            .map(|c| label_size.get(c).copied().ok_or_else(||
                RuntimeError::InvalidOperation(format!("einsum: unknown label '{}' in output", c))))
            .collect::<Result<_, _>>()?;
        let out_numel = Self::shape_numel(&output_shape);

        // Determine contraction labels (in input but not output)
        let output_set: std::collections::BTreeSet<char> = output_chars.iter().copied().collect();
        let contract_labels: Vec<char> = label_size.keys()
            .filter(|c| !output_set.contains(c))
            .copied()
            .collect();
        let contract_sizes: Vec<usize> = contract_labels.iter()
            .map(|c| label_size[c])
            .collect();
        let contract_numel: usize = contract_sizes.iter().product::<usize>().max(1);

        // Precompute input spec chars
        let input_chars: Vec<Vec<char>> = input_specs.iter().map(|s| s.chars().collect()).collect();

        // For each output position, iterate over contraction indices
        let out_strides = Self::compute_strides(&output_shape);
        let mut result = vec![0.0f64; out_numel];

        // Pre-read input data
        let input_data: Vec<Vec<f64>> = inputs.iter().map(|t| t.to_vec()).collect();
        let input_strides: Vec<Vec<usize>> = inputs.iter().map(|t| t.strides.clone()).collect();
        let input_offsets: Vec<usize> = inputs.iter().map(|t| t.offset).collect();

        for out_idx in 0..out_numel {
            // Compute output label values
            let mut label_vals = std::collections::BTreeMap::new();
            let mut remaining = out_idx;
            for (d, &c) in output_chars.iter().enumerate() {
                let stride = if d < out_strides.len() { out_strides[d] } else { 1 };
                label_vals.insert(c, remaining / stride);
                remaining %= stride;
            }

            let mut acc = BinnedAccumulatorF64::new();
            // Iterate over all contraction index combinations
            for cidx in 0..contract_numel {
                // Compute contraction label values
                let mut cr = cidx;
                for (ci, &cl) in contract_labels.iter().enumerate() {
                    let stride: usize = contract_sizes[ci+1..].iter().product::<usize>().max(1);
                    label_vals.insert(cl, cr / stride);
                    cr %= stride;
                }

                // Compute product of input elements
                let mut product = 1.0f64;
                for (inp_idx, chars) in input_chars.iter().enumerate() {
                    let mut flat = input_offsets[inp_idx];
                    for (d, &c) in chars.iter().enumerate() {
                        flat += label_vals[&c] * input_strides[inp_idx][d];
                    }
                    product *= input_data[inp_idx][flat];
                }
                acc.add(product);
            }
            result[out_idx] = acc.finalize();
        }

        if output_shape.is_empty() {
            Tensor::from_vec(result, &[1])
        } else {
            Tensor::from_vec(result, &output_shape)
        }
    }

    // -----------------------------------------------------------------------
    // Phase 2: Reshape / View Enhancements
    // -----------------------------------------------------------------------

    /// Add a dimension of size 1 at position `dim`.
    ///
    /// For a tensor of shape `[A, B]`, `unsqueeze(0)` yields `[1, A, B]`,
    /// `unsqueeze(1)` yields `[A, 1, B]`, etc.
    pub fn unsqueeze(&self, dim: usize) -> Result<Tensor, RuntimeError> {
        let ndim = self.ndim();
        if dim > ndim {
            return Err(RuntimeError::IndexOutOfBounds { index: dim, length: ndim + 1 });
        }
        let mut new_shape = self.shape.clone();
        new_shape.insert(dim, 1);
        self.reshape(&new_shape)
    }

    /// Remove a dimension of size 1 at position `dim`.
    /// If `dim` is `None`, removes all dimensions of size 1.
    pub fn squeeze(&self, dim: Option<usize>) -> Result<Tensor, RuntimeError> {
        match dim {
            Some(d) => {
                if d >= self.ndim() {
                    return Err(RuntimeError::IndexOutOfBounds { index: d, length: self.ndim() });
                }
                if self.shape[d] != 1 {
                    return Err(RuntimeError::InvalidOperation(
                        format!("squeeze: dimension {} has size {}, not 1", d, self.shape[d]),
                    ));
                }
                let mut new_shape = self.shape.clone();
                new_shape.remove(d);
                if new_shape.is_empty() {
                    new_shape.push(1); // scalar
                }
                self.reshape(&new_shape)
            }
            None => {
                let new_shape: Vec<usize> = self.shape.iter()
                    .filter(|&&s| s != 1)
                    .copied()
                    .collect();
                let new_shape = if new_shape.is_empty() { vec![1] } else { new_shape };
                self.reshape(&new_shape)
            }
        }
    }

    /// Broadcast without copying. Return a view with `stride=0` for broadcasted dims.
    ///
    /// Alias for [`broadcast_to`](Tensor::broadcast_to).
    pub fn expand(&self, target_shape: &[usize]) -> Result<Tensor, RuntimeError> {
        self.broadcast_to(target_shape)
    }

    /// Flatten a range of dimensions [start_dim, end_dim] into a single dimension.
    pub fn flatten(&self, start_dim: usize, end_dim: usize) -> Result<Tensor, RuntimeError> {
        if start_dim > end_dim || end_dim >= self.ndim() {
            return Err(RuntimeError::InvalidOperation(
                format!("flatten: invalid dim range [{}, {}] for {}D tensor", start_dim, end_dim, self.ndim()),
            ));
        }
        let mut new_shape = Vec::new();
        for i in 0..start_dim {
            new_shape.push(self.shape[i]);
        }
        let flat_size: usize = self.shape[start_dim..=end_dim].iter().product();
        new_shape.push(flat_size);
        for i in (end_dim + 1)..self.ndim() {
            new_shape.push(self.shape[i]);
        }
        self.reshape(&new_shape)
    }

    /// Split tensor into `n` roughly equal chunks along dimension `dim`.
    pub fn chunk(&self, n: usize, dim: usize) -> Result<Vec<Tensor>, RuntimeError> {
        if dim >= self.ndim() {
            return Err(RuntimeError::IndexOutOfBounds { index: dim, length: self.ndim() });
        }
        if n == 0 {
            return Err(RuntimeError::InvalidOperation("chunk: n must be > 0".into()));
        }
        let dim_size = self.shape[dim];
        let chunk_size = (dim_size + n - 1) / n;
        let mut sizes = Vec::new();
        let mut remaining = dim_size;
        while remaining > 0 {
            let s = remaining.min(chunk_size);
            sizes.push(s);
            remaining -= s;
        }
        self.split(&sizes, dim)
    }

    /// Split tensor along dimension `dim` according to the given sizes.
    pub fn split(&self, sizes: &[usize], dim: usize) -> Result<Vec<Tensor>, RuntimeError> {
        if dim >= self.ndim() {
            return Err(RuntimeError::IndexOutOfBounds { index: dim, length: self.ndim() });
        }
        let total: usize = sizes.iter().sum();
        if total != self.shape[dim] {
            return Err(RuntimeError::InvalidOperation(
                format!("split: sizes sum {} != dim size {}", total, self.shape[dim]),
            ));
        }

        let mut results = Vec::new();
        let mut offset = 0;

        for &sz in sizes {
            let ranges: Vec<(usize, usize)> = self.shape.iter()
                .enumerate()
                .map(|(i, &s)| {
                    if i == dim { (offset, offset + sz) } else { (0, s) }
                })
                .collect();
            let chunk = self.slice(&ranges)?;
            // Materialize as contiguous
            results.push(chunk.to_contiguous());
            offset += sz;
        }

        Ok(results)
    }

    /// Fused `alpha * self + beta * other` element-wise. Single pass, one allocation.
    ///
    /// Critical for LSTM/GRU gates where `f * c_prev + i * g` would otherwise
    /// create 3 intermediate tensors.
    pub fn scale_add(&self, alpha: f64, other: &Tensor, beta: f64) -> Result<Tensor, RuntimeError> {
        if self.shape != other.shape {
            return Err(RuntimeError::InvalidOperation(
                "scale_add: shape mismatch".to_string(),
            ));
        }
        let a = self.to_vec();
        let b = other.to_vec();
        let result: Vec<f64> = a.iter().zip(b.iter()).map(|(&x, &y)| alpha * x + beta * y).collect();
        Tensor::from_vec(result, &self.shape)
    }
}