axonml_nn/layers/

attention.rs

//! Attention Mechanisms - Multi-Head Attention
//!
//! # File
//! `crates/axonml-nn/src/layers/attention.rs`
//!
//! # Author
//! Andrew Jewell Sr - AutomataNexus
//!
//! # Updated
//! March 8, 2026
//!
//! # Disclaimer
//! Use at own risk. This software is provided "as is", without warranty of any
//! kind, express or implied. The author and AutomataNexus shall not be held
//! liable for any damages arising from the use of this software.

use std::collections::HashMap;

use axonml_autograd::Variable;
#[cfg(feature = "cuda")]
use axonml_autograd::functions::FusedAttentionBackward;
#[cfg(feature = "cuda")]
use axonml_autograd::grad_fn::GradFn;
use axonml_tensor::Tensor;

use crate::layers::Linear;
use crate::module::Module;
use crate::parameter::Parameter;

// =============================================================================
// MultiHeadAttention
// =============================================================================

/// Multi-Head Attention mechanism.
///
/// Allows the model to jointly attend to information from different
/// representation subspaces at different positions.
///
/// # Arguments
/// * `embed_dim` - Total dimension of the model
/// * `num_heads` - Number of parallel attention heads
/// * `dropout` - Dropout probability (default: 0.0)
///
/// # Shape
/// - Query: (L, N, E) or (N, L, E) if batch_first
/// - Key: (S, N, E) or (N, S, E) if batch_first
/// - Value: (S, N, E) or (N, S, E) if batch_first
/// - Output: (L, N, E) or (N, L, E) if batch_first
pub struct MultiHeadAttention {
    /// Query projection.
    q_proj: Linear,
    /// Key projection.
    k_proj: Linear,
    /// Value projection.
    v_proj: Linear,
    /// Output projection.
    out_proj: Linear,
    /// Embedding dimension.
    embed_dim: usize,
    /// Number of attention heads.
    num_heads: usize,
    /// Dimension per head.
    head_dim: usize,
    /// Scaling factor.
    scale: f32,
    /// Whether input is batch first.
    batch_first: bool,
}

impl MultiHeadAttention {
    /// Creates a new MultiHeadAttention module.
    pub fn new(embed_dim: usize, num_heads: usize) -> Self {
        Self::with_options(embed_dim, num_heads, 0.0, true)
    }

    /// Creates MultiHeadAttention with all options.
    pub fn with_options(
        embed_dim: usize,
        num_heads: usize,
        _dropout: f32,
        batch_first: bool,
    ) -> Self {
        assert!(
            embed_dim % num_heads == 0,
            "embed_dim must be divisible by num_heads"
        );

        let head_dim = embed_dim / num_heads;
        let scale = (head_dim as f32).sqrt().recip();

        Self {
            q_proj: Linear::new(embed_dim, embed_dim),
            k_proj: Linear::new(embed_dim, embed_dim),
            v_proj: Linear::new(embed_dim, embed_dim),
            out_proj: Linear::new(embed_dim, embed_dim),
            embed_dim,
            num_heads,
            head_dim,
            scale,
            batch_first,
        }
    }

    /// Expand attention mask to match scores shape via broadcast.
    ///
    /// Handles common mask shapes:
    /// - [T, S] → [B, H, T, S] (same mask for all batches/heads)
    /// - [B, 1, T, S] → [B, H, T, S] (per-batch, shared across heads)
    /// - [B, H, T, S] → no expansion needed
    ///
    /// Works on both CPU and GPU — uses Variable::expand which preserves device.
    #[allow(dead_code)]
    fn expand_mask(
        mask: &Variable,
        batch_size: usize,
        num_heads: usize,
        tgt_len: usize,
        src_len: usize,
    ) -> Variable {
        let mask_shape = mask.shape();
        let target = [batch_size, num_heads, tgt_len, src_len];

        if mask_shape == target {
            return mask.clone();
        }

        // [T, S] → [1, 1, T, S] → expand to [B, H, T, S]
        if mask_shape.len() == 2 {
            let reshaped = mask.reshape(&[1, 1, tgt_len, src_len]);
            return reshaped.expand(&target);
        }

        // [B, 1, T, S] → expand heads dim
        if mask_shape.len() == 4 && mask_shape[1] == 1 {
            return mask.expand(&target);
        }

        // [1, 1, T, S] → expand both
        if mask_shape.len() == 4 && mask_shape[0] == 1 && mask_shape[1] == 1 {
            return mask.expand(&target);
        }

        // Fallback: just clone
        mask.clone()
    }

    /// Computes attention using batched matmul (BLAS-accelerated).
    pub fn attention(
        &self,
        query: &Variable,
        key: &Variable,
        value: &Variable,
        attn_mask: Option<&Variable>,
    ) -> Variable {
        let q_shape = query.shape();
        let (batch_size, tgt_len, _) = if self.batch_first {
            (q_shape[0], q_shape[1], q_shape[2])
        } else {
            (q_shape[1], q_shape[0], q_shape[2])
        };
        let src_len = if self.batch_first {
            key.shape()[1]
        } else {
            key.shape()[0]
        };

        // Project Q, K, V  (all tracked through autograd)
        let q = self.q_proj.forward(query);
        let k = self.k_proj.forward(key);
        let v = self.v_proj.forward(value);

        // Reshape to multi-head: [batch, seq, embed] → [batch, seq, heads, head_dim]
        // Then transpose to:     [batch, heads, seq, head_dim]
        let q = q
            .reshape(&[batch_size, tgt_len, self.num_heads, self.head_dim])
            .transpose(1, 2);
        let k = k
            .reshape(&[batch_size, src_len, self.num_heads, self.head_dim])
            .transpose(1, 2);
        let v = v
            .reshape(&[batch_size, src_len, self.num_heads, self.head_dim])
            .transpose(1, 2);

        // GPU fast path: fused attention kernel avoids materializing the N*N
        // attention matrix. Works for both inference and training when no mask
        // is provided. The kernel computes Q@K^T * scale -> softmax -> @V in
        // one pass per row. In training mode, a FusedAttentionBackward autograd
        // function is attached that uses the CUDA backward kernel (or CPU fallback).
        #[cfg(feature = "cuda")]
        if q.data().device().is_gpu() && attn_mask.is_none() {
            let is_training = axonml_autograd::no_grad::is_grad_enabled();
            let q_tensor = q.data();
            let k_tensor = k.data();
            let v_tensor = v.data();

            if let Some(attn_out) = q_tensor.fused_attention_cuda(
                &k_tensor, &v_tensor, self.scale,
                false, // not causal by default; causal mask would be in attn_mask
            ) {
                let attn_output = if is_training
                    && (q.requires_grad() || k.requires_grad() || v.requires_grad())
                {
                    // Build autograd backward function for training
                    let backward = FusedAttentionBackward::new(
                        q.grad_fn().cloned(),
                        k.grad_fn().cloned(),
                        v.grad_fn().cloned(),
                        q_tensor,
                        k_tensor,
                        v_tensor,
                        attn_out.clone(),
                        self.scale,
                        false,
                    );
                    Variable::from_operation(attn_out, GradFn::new(backward), true)
                } else {
                    Variable::new(attn_out, false)
                };
                let attn_output =
                    attn_output
                        .transpose(1, 2)
                        .reshape(&[batch_size, tgt_len, self.embed_dim]);
                return self.out_proj.forward(&attn_output);
            }
            // Fall through to standard path if fused kernel fails
        }

        // Scaled dot-product attention: scores = Q @ K^T * scale
        // K^T: [batch, heads, head_dim, src_len]
        let k_t = k.transpose(2, 3);
        // scores: [batch, heads, tgt_len, src_len]
        let scores = q.matmul(&k_t).mul_scalar(self.scale);

        // Apply attention mask (0 → -1e9 additive mask)
        // Mask shapes: [tgt_len, src_len] (causal) or [batch, src_len] (padding)
        // Scores shape: [batch, heads, tgt_len, src_len]
        let scores = if let Some(mask) = attn_mask {
            let mask_shape = mask.shape();
            let mask_data = mask.data();
            let scores_shape = scores.shape();
            let total = scores_shape.iter().product::<usize>();

            // GPU fast path: expand mask entirely on GPU via CUDA kernel
            // Avoids GPU→CPU→GPU round-trip (9 mask expansions per forward pass)
            #[cfg(feature = "cuda")]
            if scores.data().device().is_gpu() {
                // Ensure mask is on GPU (it's small, so upload is cheap if needed)
                let mask_gpu = if mask_data.device().is_gpu() {
                    mask_data.clone()
                } else {
                    mask_data.to_device(scores.data().device()).unwrap()
                };

                if let Some(expanded_tensor) = mask_gpu.mask_expand_cuda(
                    &scores_shape,
                    batch_size,
                    self.num_heads,
                    tgt_len,
                    src_len,
                ) {
                    let additive_mask = Variable::new(expanded_tensor, false);
                    return self.finish_attention(
                        scores.add_var(&additive_mask),
                        &v,
                        batch_size,
                        tgt_len,
                    );
                }
                // Fall through to CPU path on unsupported shape
            }

            // CPU fallback: expand mask with nested loops
            let mask_vec = mask_data.to_vec();
            let additive: Vec<f32> = mask_vec
                .iter()
                .map(|&v| if v == 0.0 { -1e9 } else { 0.0 })
                .collect();

            let mut expanded = vec![0.0f32; total];

            if mask_shape.len() == 2 && mask_shape[0] == tgt_len && mask_shape[1] == src_len {
                // Causal mask [tgt_len, src_len] → broadcast over batch & heads
                for b in 0..batch_size {
                    for h in 0..self.num_heads {
                        for i in 0..tgt_len {
                            for j in 0..src_len {
                                let idx = b * self.num_heads * tgt_len * src_len
                                    + h * tgt_len * src_len
                                    + i * src_len
                                    + j;
                                expanded[idx] = additive[i * src_len + j];
                            }
                        }
                    }
                }
            } else if mask_shape.len() == 2
                && mask_shape[0] == batch_size
                && mask_shape[1] == src_len
            {
                // Padding mask [batch, src_len] → broadcast over heads & tgt positions
                for b in 0..batch_size {
                    for h in 0..self.num_heads {
                        for i in 0..tgt_len {
                            for j in 0..src_len {
                                let idx = b * self.num_heads * tgt_len * src_len
                                    + h * tgt_len * src_len
                                    + i * src_len
                                    + j;
                                expanded[idx] = additive[b * src_len + j];
                            }
                        }
                    }
                }
            } else {
                // General: tile mask across scores using modular indexing
                for (i, val) in expanded.iter_mut().enumerate() {
                    *val = additive[i % additive.len()];
                }
            }

            let mut additive_tensor =
                Tensor::from_vec(expanded, &scores_shape).expect("tensor creation failed");
            let scores_device = scores.data().device();
            if scores_device.is_gpu() {
                additive_tensor = additive_tensor
                    .to_device(scores_device)
                    .expect("device transfer failed");
            }
            let additive_mask = Variable::new(additive_tensor, false);
            scores.add_var(&additive_mask)
        } else {
            scores
        };

        self.finish_attention(scores, &v, batch_size, tgt_len)
    }

    /// Softmax → weighted sum → reshape → output projection.
    /// Shared by both GPU and CPU mask expansion paths.
    fn finish_attention(
        &self,
        scores: Variable,
        v: &Variable,
        batch_size: usize,
        tgt_len: usize,
    ) -> Variable {
        let attn_weights = scores.softmax(-1);
        let attn_output = attn_weights.matmul(v);
        let attn_output =
            attn_output
                .transpose(1, 2)
                .reshape(&[batch_size, tgt_len, self.embed_dim]);
        self.out_proj.forward(&attn_output)
    }
}

impl Module for MultiHeadAttention {
    fn forward(&self, input: &Variable) -> Variable {
        // Self-attention: query = key = value = input
        self.attention(input, input, input, None)
    }

    fn parameters(&self) -> Vec<Parameter> {
        let mut params = Vec::new();
        params.extend(self.q_proj.parameters());
        params.extend(self.k_proj.parameters());
        params.extend(self.v_proj.parameters());
        params.extend(self.out_proj.parameters());
        params
    }

    fn named_parameters(&self) -> HashMap<String, Parameter> {
        let mut params = HashMap::new();
        for (name, param) in self.q_proj.named_parameters() {
            params.insert(format!("q_proj.{name}"), param);
        }
        for (name, param) in self.k_proj.named_parameters() {
            params.insert(format!("k_proj.{name}"), param);
        }
        for (name, param) in self.v_proj.named_parameters() {
            params.insert(format!("v_proj.{name}"), param);
        }
        for (name, param) in self.out_proj.named_parameters() {
            params.insert(format!("out_proj.{name}"), param);
        }
        params
    }

    fn name(&self) -> &'static str {
        "MultiHeadAttention"
    }
}

// =============================================================================
// CrossAttention
// =============================================================================

/// Cross-Attention mechanism for encoder-decoder architectures.
///
/// Queries come from the decoder, keys and values come from the encoder.
/// This is the standard cross-attention used in Transformer decoders,
/// seq2seq models, and vision-language models.
///
/// # Shape (batch_first=true)
/// - Query (decoder): (N, T, E)
/// - Memory (encoder): (N, S, E)
/// - Output: (N, T, E)
///
/// where N=batch, T=target seq len, S=source seq len, E=embed_dim.
pub struct CrossAttention {
    /// Underlying multi-head attention.
    mha: MultiHeadAttention,
}

impl CrossAttention {
    /// Creates a new CrossAttention module.
    pub fn new(embed_dim: usize, num_heads: usize) -> Self {
        Self {
            mha: MultiHeadAttention::new(embed_dim, num_heads),
        }
    }

    /// Creates CrossAttention with all options.
    pub fn with_options(
        embed_dim: usize,
        num_heads: usize,
        dropout: f32,
        batch_first: bool,
    ) -> Self {
        Self {
            mha: MultiHeadAttention::with_options(embed_dim, num_heads, dropout, batch_first),
        }
    }

    /// Computes cross-attention.
    ///
    /// # Arguments
    /// * `query` - Decoder hidden states (N, T, E)
    /// * `memory` - Encoder output (N, S, E)
    /// * `attn_mask` - Optional attention mask
    pub fn cross_attention(
        &self,
        query: &Variable,
        memory: &Variable,
        attn_mask: Option<&Variable>,
    ) -> Variable {
        self.mha.attention(query, memory, memory, attn_mask)
    }

    /// Returns the embedding dimension.
    pub fn embed_dim(&self) -> usize {
        self.mha.embed_dim
    }

    /// Returns the number of heads.
    pub fn num_heads(&self) -> usize {
        self.mha.num_heads
    }
}

impl Module for CrossAttention {
    fn forward(&self, input: &Variable) -> Variable {
        // When called as Module (single input), acts as self-attention.
        // Use cross_attention() for encoder-decoder attention.
        self.mha.forward(input)
    }

    fn parameters(&self) -> Vec<Parameter> {
        self.mha.parameters()
    }

    fn named_parameters(&self) -> HashMap<String, Parameter> {
        let mut params = HashMap::new();
        for (name, param) in self.mha.named_parameters() {
            params.insert(format!("mha.{name}"), param);
        }
        params
    }

    fn name(&self) -> &'static str {
        "CrossAttention"
    }
}

// =============================================================================
// Fused Scaled Dot-Product Attention
// =============================================================================

/// Fused scaled dot-product attention on Tensors.
///
/// Computes `softmax(Q @ K^T * scale) @ V` without materializing the full
/// N*N attention matrix. On GPU, this uses a CUDA kernel that processes one
/// query row per thread. On CPU, falls back to standard matmul + softmax.
///
/// # Arguments
/// * `q` - Query tensor `[B, H, Tq, D]`
/// * `k` - Key tensor `[B, H, Tk, D]`
/// * `v` - Value tensor `[B, H, Tk, D]`
/// * `scale` - Scaling factor (typically `1/sqrt(head_dim)`)
/// * `is_causal` - Whether to apply causal masking
///
/// # Returns
/// Output tensor `[B, H, Tq, D]`
///
/// # Note on Flash Attention
/// This is a fused kernel (Option B) — it avoids the N*N memory allocation
/// but each thread still iterates over the full key sequence. True Flash
/// Attention (Option A) would use shared memory tiling with online softmax
/// (the Dao et al. algorithm), processing in blocks of ~64-128 and requiring:
/// - Shared memory for Q/K/V tile loading
/// - Online softmax with running max/sum correction across tiles
/// - Two passes: forward for output + logsumexp, backward with recomputation
/// - ~3x more complex kernel code but O(N) memory and better cache behavior
///
/// For sequences up to ~2048, this fused kernel provides most of the benefit.
/// For longer sequences, use the tiled CPU Flash Attention in `axonml-llm`.
pub fn scaled_dot_product_attention_fused(
    q: &Tensor<f32>,
    k: &Tensor<f32>,
    v: &Tensor<f32>,
    scale: f32,
    is_causal: bool,
) -> Tensor<f32> {
    // Try GPU fused kernel
    #[cfg(feature = "cuda")]
    if q.device().is_gpu() {
        if let Some(result) = q.fused_attention_cuda(k, v, scale, is_causal) {
            return result;
        }
    }

    // CPU fallback: standard matmul-based attention
    let shape = q.shape();
    let batch_size = shape[0];
    let num_heads = shape[1];
    let tgt_len = shape[2];
    let head_dim = shape[3];
    let src_len = k.shape()[2];

    let q_data = q.to_vec();
    let k_data = k.to_vec();
    let v_data = v.to_vec();

    let mut output = vec![0.0f32; batch_size * num_heads * tgt_len * head_dim];

    for b in 0..batch_size {
        for h in 0..num_heads {
            for i in 0..tgt_len {
                // Compute attention scores for row i
                let mut scores = vec![0.0f32; src_len];
                let mut max_score = f32::NEG_INFINITY;

                for j in 0..src_len {
                    if is_causal && j > i {
                        scores[j] = f32::NEG_INFINITY;
                        continue;
                    }
                    let mut score = 0.0f32;
                    for d in 0..head_dim {
                        let q_idx = ((b * num_heads + h) * tgt_len + i) * head_dim + d;
                        let k_idx = ((b * num_heads + h) * src_len + j) * head_dim + d;
                        score += q_data[q_idx] * k_data[k_idx];
                    }
                    score *= scale;
                    scores[j] = score;
                    if score > max_score {
                        max_score = score;
                    }
                }

                // Softmax
                let mut sum_exp = 0.0f32;
                for s in &mut scores {
                    if *s > f32::NEG_INFINITY {
                        *s = (*s - max_score).exp();
                        sum_exp += *s;
                    } else {
                        *s = 0.0;
                    }
                }
                let inv_sum = if sum_exp > 0.0 { 1.0 / sum_exp } else { 0.0 };

                // Weighted sum of V
                for d in 0..head_dim {
                    let mut val = 0.0f32;
                    for j in 0..src_len {
                        let v_idx = ((b * num_heads + h) * src_len + j) * head_dim + d;
                        val += scores[j] * v_data[v_idx];
                    }
                    let out_idx = ((b * num_heads + h) * tgt_len + i) * head_dim + d;
                    output[out_idx] = val * inv_sum;
                }
            }
        }
    }

    Tensor::from_vec(output, &[batch_size, num_heads, tgt_len, head_dim])
        .expect("tensor creation failed")
}

// =============================================================================
// Tests
// =============================================================================

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_multihead_attention_creation() {
        let mha = MultiHeadAttention::new(512, 8);
        assert_eq!(mha.embed_dim, 512);
        assert_eq!(mha.num_heads, 8);
        assert_eq!(mha.head_dim, 64);
    }

    #[test]
    fn test_multihead_attention_forward() {
        let mha = MultiHeadAttention::new(64, 4);
        let input = Variable::new(
            Tensor::from_vec(vec![1.0; 2 * 10 * 64], &[2, 10, 64]).expect("tensor creation failed"),
            false,
        );
        let output = mha.forward(&input);
        assert_eq!(output.shape(), vec![2, 10, 64]);
    }

    #[test]
    fn test_cross_attention() {
        let mha = MultiHeadAttention::new(64, 4);
        let query = Variable::new(
            Tensor::from_vec(vec![1.0; 2 * 5 * 64], &[2, 5, 64]).expect("tensor creation failed"),
            false,
        );
        let key_value = Variable::new(
            Tensor::from_vec(vec![1.0; 2 * 10 * 64], &[2, 10, 64]).expect("tensor creation failed"),
            false,
        );
        let output = mha.attention(&query, &key_value, &key_value, None);
        assert_eq!(output.shape(), vec![2, 5, 64]);
    }

    #[test]
    fn test_multihead_attention_parameters() {
        let mha = MultiHeadAttention::new(64, 4);
        let params = mha.parameters();
        // Q, K, V, Out projections each have weight + bias = 8 total
        assert_eq!(params.len(), 8);
    }

    #[test]
    fn test_cross_attention_creation() {
        let ca = CrossAttention::new(256, 8);
        assert_eq!(ca.embed_dim(), 256);
        assert_eq!(ca.num_heads(), 8);
    }

    #[test]
    fn test_cross_attention_forward() {
        let ca = CrossAttention::new(64, 4);
        // Decoder query: (batch=2, tgt_len=5, embed=64)
        let query = Variable::new(
            Tensor::from_vec(vec![0.1; 2 * 5 * 64], &[2, 5, 64]).expect("tensor creation failed"),
            false,
        );
        // Encoder memory: (batch=2, src_len=10, embed=64)
        let memory = Variable::new(
            Tensor::from_vec(vec![0.2; 2 * 10 * 64], &[2, 10, 64]).expect("tensor creation failed"),
            false,
        );
        let output = ca.cross_attention(&query, &memory, None);
        assert_eq!(output.shape(), vec![2, 5, 64]);
    }

    #[test]
    fn test_cross_attention_self_attention_fallback() {
        let ca = CrossAttention::new(64, 4);
        let input = Variable::new(
            Tensor::from_vec(vec![1.0; 2 * 8 * 64], &[2, 8, 64]).expect("tensor creation failed"),
            false,
        );
        // Module::forward does self-attention
        let output = ca.forward(&input);
        assert_eq!(output.shape(), vec![2, 8, 64]);
    }

    #[test]
    fn test_cross_attention_parameters() {
        let ca = CrossAttention::new(64, 4);
        let params = ca.parameters();
        assert_eq!(params.len(), 8); // Q, K, V, Out × (weight + bias)
        let named = ca.named_parameters();
        assert!(named.contains_key("mha.q_proj.weight"));
        assert!(named.contains_key("mha.out_proj.bias"));
    }

    #[test]
    fn test_fused_attention_cpu() {
        // Test fused attention on CPU (fallback path)
        let batch = 2;
        let heads = 4;
        let seq = 8;
        let dim = 16;
        let scale = 1.0 / (dim as f32).sqrt();

        let q = Tensor::from_vec(
            vec![0.1; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();
        let k = Tensor::from_vec(
            vec![0.1; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();
        let v = Tensor::from_vec(
            vec![0.5; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();

        let out = scaled_dot_product_attention_fused(&q, &k, &v, scale, false);
        assert_eq!(out.shape(), &[batch, heads, seq, dim]);

        // With uniform V=0.5, output should be close to 0.5
        let out_vec = out.to_vec();
        for val in &out_vec {
            assert!((*val - 0.5).abs() < 0.01, "Expected ~0.5, got {}", val);
        }
    }

    #[test]
    fn test_fused_attention_causal() {
        let batch = 1;
        let heads = 1;
        let seq = 4;
        let dim = 4;
        let scale = 1.0 / (dim as f32).sqrt();

        // Q and K are identity-like so attention focuses on matching positions
        let q = Tensor::from_vec(
            vec![0.1; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();
        let k = Tensor::from_vec(
            vec![0.1; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();
        let v = Tensor::from_vec(
            vec![
                1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0,
            ],
            &[batch, heads, seq, dim],
        )
        .unwrap();

        let out = scaled_dot_product_attention_fused(&q, &k, &v, scale, true);
        assert_eq!(out.shape(), &[batch, heads, seq, dim]);

        // First position can only attend to position 0, so output = V[0] = [1,0,0,0]
        let out_vec = out.to_vec();
        assert!(
            (out_vec[0] - 1.0).abs() < 1e-5,
            "row 0, col 0 should be 1.0"
        );
        assert!((out_vec[1]).abs() < 1e-5, "row 0, col 1 should be 0.0");
    }

    #[test]
    fn test_multihead_attention_backward_cpu() {
        // Test that gradients flow through MHA in training mode (CPU path)
        use axonml_autograd::backward;

        let mha = MultiHeadAttention::new(32, 4);
        let input = Variable::new(
            Tensor::from_vec(vec![0.1; 2 * 4 * 32], &[2, 4, 32]).expect("tensor creation failed"),
            true,
        );
        let output = mha.forward(&input);
        assert_eq!(output.shape(), vec![2, 4, 32]);

        // Sum the output and backward
        let loss = output.sum();
        let ones = Tensor::from_vec(vec![1.0f32], &[1]).expect("tensor creation failed");
        backward(&loss, &ones);

        // Input should have gradients
        let grad = input.grad();
        assert!(grad.is_some(), "Input gradient should exist");
        let grad_data = grad.unwrap();
        assert_eq!(grad_data.shape(), &[2, 4, 32]);

        // Gradients should be non-zero
        let grad_vec = grad_data.to_vec();
        let non_zero = grad_vec.iter().any(|&v| v.abs() > 1e-10);
        assert!(non_zero, "Gradients should be non-zero");
    }

    #[test]
    fn test_fused_attention_backward_cpu() {
        // Test the FusedAttentionBackward autograd function directly on CPU
        use axonml_autograd::functions::FusedAttentionBackward;
        use axonml_autograd::grad_fn::GradientFunction;

        let batch = 1;
        let heads = 2;
        let seq = 4;
        let dim = 8;
        let scale = 1.0 / (dim as f32).sqrt();

        // Create random-ish tensors
        let q_data: Vec<f32> = (0..batch * heads * seq * dim)
            .map(|i| ((i as f32) * 0.01).sin())
            .collect();
        let k_data: Vec<f32> = (0..batch * heads * seq * dim)
            .map(|i| ((i as f32) * 0.02).cos())
            .collect();
        let v_data: Vec<f32> = (0..batch * heads * seq * dim)
            .map(|i| ((i as f32) * 0.03).sin() + 0.5)
            .collect();

        let q =
            Tensor::from_vec(q_data, &[batch, heads, seq, dim]).expect("tensor creation failed");
        let k =
            Tensor::from_vec(k_data, &[batch, heads, seq, dim]).expect("tensor creation failed");
        let v =
            Tensor::from_vec(v_data, &[batch, heads, seq, dim]).expect("tensor creation failed");

        // Compute forward output using the fused CPU path
        let output = scaled_dot_product_attention_fused(&q, &k, &v, scale, false);
        assert_eq!(output.shape(), &[batch, heads, seq, dim]);

        // Create backward function
        let backward_fn = FusedAttentionBackward::new(
            None,
            None,
            None,
            q.clone(),
            k.clone(),
            v.clone(),
            output.clone(),
            scale,
            false,
        );

        // Use ones as grad_output
        let grad_output = Tensor::from_vec(
            vec![1.0f32; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();

        let grads = backward_fn.apply(&grad_output);
        assert_eq!(grads.len(), 3);

        let gq = grads[0].as_ref().expect("grad_Q should exist");
        let gk = grads[1].as_ref().expect("grad_K should exist");
        let gv = grads[2].as_ref().expect("grad_V should exist");

        assert_eq!(gq.shape(), &[batch, heads, seq, dim]);
        assert_eq!(gk.shape(), &[batch, heads, seq, dim]);
        assert_eq!(gv.shape(), &[batch, heads, seq, dim]);

        // Gradients should be finite
        for val in gq
            .to_vec()
            .iter()
            .chain(gk.to_vec().iter())
            .chain(gv.to_vec().iter())
        {
            assert!(val.is_finite(), "Gradient should be finite, got {}", val);
        }

        // grad_V should be non-zero (it's P^T @ grad_output)
        let gv_nonzero = gv.to_vec().iter().any(|&v| v.abs() > 1e-10);
        assert!(gv_nonzero, "grad_V should be non-zero");
    }

    #[test]
    fn test_fused_attention_backward_causal_cpu() {
        // Test the backward with causal masking
        use axonml_autograd::functions::FusedAttentionBackward;
        use axonml_autograd::grad_fn::GradientFunction;

        let batch = 1;
        let heads = 1;
        let seq = 4;
        let dim = 4;
        let scale = 1.0 / (dim as f32).sqrt();

        let q = Tensor::from_vec(
            vec![0.1f32; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();
        let k = Tensor::from_vec(
            vec![0.2f32; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();
        let v = Tensor::from_vec(
            vec![0.5f32; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();

        let output = scaled_dot_product_attention_fused(&q, &k, &v, scale, true);

        let backward_fn = FusedAttentionBackward::new(
            None,
            None,
            None,
            q.clone(),
            k.clone(),
            v.clone(),
            output.clone(),
            scale,
            true,
        );

        let grad_output = Tensor::from_vec(
            vec![1.0f32; batch * heads * seq * dim],
            &[batch, heads, seq, dim],
        )
        .unwrap();

        let grads = backward_fn.apply(&grad_output);
        assert_eq!(grads.len(), 3);

        let gq = grads[0].as_ref().unwrap();
        let gk = grads[1].as_ref().unwrap();
        let gv = grads[2].as_ref().unwrap();

        // All grads should be finite
        for val in gq
            .to_vec()
            .iter()
            .chain(gk.to_vec().iter())
            .chain(gv.to_vec().iter())
        {
            assert!(val.is_finite(), "Gradient should be finite, got {}", val);
        }
    }
}