Skip to content

⚡ Bolt: Optimize autograd reverse pass to avoid redundant Tensor clones #127

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
teerthsharma wants to merge 1 commit into
base: master
Choose a base branch
from
Open

⚡ Bolt: Optimize autograd reverse pass to avoid redundant Tensor clones #127

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions .jules/bolt.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,3 +5,7 @@
## 2026-05-01 - Avoid High-Level Tensor Ops in Scalar Reductions
**Learning:** High-level `Tensor` operations like `sub()` and `mul()` trigger intermediate heap allocations for shape and stride metadata. When computing scalar reductions (like MSE, distances, or loss functions), using these operations introduces severe memory overhead inside hot loops. Attempting to use `.min()` length truncation as a safeguard is an anti-pattern as it masks shape mismatch errors.
**Action:** For scalar reductions, assert shape equality (`assert_eq!(a.shape, b.shape)`) and perform a single-pass iteration directly over the underlying borrowed data arrays (`a.data.borrow()`) to eliminate intermediate allocations and safely compute the result.

## 2026-06-05 - Avoid Redundant Tensor Clones in Autograd Backprop
**Learning:** In the `backward` pass of the autograd engine, retrieving gradients from the `grads` vector using `.clone()` causes unnecessary heap allocations for `Tensor` metadata (shape and strides). Because the `grads` vector is a `Vec<Option<Tensor>>`, we can use `.take()` to take ownership of the gradient, perform our calculations, and then place it back in the vector. Furthermore, `accumulate_grad` was passing `Tensor` references and then cloning inside when empty. Taking `Tensor` by value in `accumulate_grad` avoids cloning when placing the new tensor into a `None` slot.
**Action:** When working with vectors of `Option<T>` where `T` is heap-allocated or expensive to clone, use `.take()` to temporarily borrow ownership if the vector does not need to be accessed concurrently. Pass such types by value into accumulator functions to avoid redundant allocations when storing them.
33 changes: 18 additions & 15 deletions crates/aether-core/src/ml/autograd.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,6 @@
// ═══════════════════════════════════════════════════════════════════════════════
//


#[cfg(not(feature = "std"))]
extern crate alloc;

Expand Down Expand Up @@ -201,52 +200,56 @@ impl<'a> Context<'a> {
match op {
Op::Add { out, lhs, rhs } => {
// Solves borrow checker by cloning Option first
let grad_out = grads[out.index].clone();
let grad_out = grads[out.index].take();
if let Some(grad) = grad_out {
// dL/d(lhs) += dL/dout * 1
Self::accumulate_grad(&mut grads, lhs.index, &grad);
Self::accumulate_grad(&mut grads, rhs.index, &grad);
Self::accumulate_grad(&mut grads, lhs.index, grad.clone());
Self::accumulate_grad(&mut grads, rhs.index, grad.clone());
grads[out.index] = Some(grad);
}
}
Op::Mul { out, lhs, rhs } => {
let grad_out = grads[out.index].clone();
let grad_out = grads[out.index].take();
if let Some(grad) = grad_out {
let lhs_val = self.heap.get(*lhs).unwrap();
let rhs_val = self.heap.get(*rhs).unwrap();

// dL/dLhs = grad_out * rhs
let d_lhs: Tensor = grad.mul(rhs_val);
Self::accumulate_grad(&mut grads, lhs.index, &d_lhs);
Self::accumulate_grad(&mut grads, lhs.index, d_lhs);

// dL/dRhs = grad_out * lhs
let d_rhs: Tensor = grad.mul(lhs_val);
Self::accumulate_grad(&mut grads, rhs.index, &d_rhs);
Self::accumulate_grad(&mut grads, rhs.index, d_rhs);
grads[out.index] = Some(grad);
}
}
Op::MatMul { out, lhs, rhs } => {
let grad_out = grads[out.index].clone();
let grad_out = grads[out.index].take();
if let Some(grad) = grad_out {
let lhs_val = self.heap.get(*lhs).unwrap();
let rhs_val = self.heap.get(*rhs).unwrap();

// C = A @ B
// dA = dC @ B^T
let d_lhs: Tensor = grad.matmul(&rhs_val.transpose());
Self::accumulate_grad(&mut grads, lhs.index, &d_lhs);
Self::accumulate_grad(&mut grads, lhs.index, d_lhs);

// dB = A^T @ dC
let d_rhs: Tensor = lhs_val.transpose().matmul(&grad);
Self::accumulate_grad(&mut grads, rhs.index, &d_rhs);
Self::accumulate_grad(&mut grads, rhs.index, d_rhs);
grads[out.index] = Some(grad);
}
}
Op::ReLU { out, input } => {
let grad_out = grads[out.index].clone();
let grad_out = grads[out.index].take();
if let Some(grad) = grad_out {
let input_val = self.heap.get(*input).unwrap();
// dL/dx = grad_out * (1 if x > 0 else 0)
let mask = input_val.map(|x| if x > 0.0 { 1.0 } else { 0.0 });
let d_input: Tensor = grad.mul(&mask);
Self::accumulate_grad(&mut grads, input.index, &d_input);
Self::accumulate_grad(&mut grads, input.index, d_input);
grads[out.index] = Some(grad);
}
}
}
Expand All @@ -255,18 +258,18 @@ impl<'a> Context<'a> {
grads
}

fn accumulate_grad(grads: &mut Vec<Option<Tensor>>, idx: usize, grad: &Tensor) {
fn accumulate_grad(grads: &mut Vec<Option<Tensor>>, idx: usize, grad: Tensor) {
if idx >= grads.len() {
grads.resize(idx + 1 + 256, None);
}

match &mut grads[idx] {
Some(existing) => {
let new = existing.add(grad);
let new = existing.add(&grad);
grads[idx] = Some(new);
}
None => {
grads[idx] = Some(grad.clone());
grads[idx] = Some(grad);
}
}
}
Expand Down