burn-autodiff 0.21.0

use crate::{
    checkpoint::strategy::{CheckpointStrategy, NoCheckpointing},
    grads::Gradients,
    tensor::AutodiffTensor,
};
use alloc::{format, string::String};
use core::marker::PhantomData;

use burn_backend::{
    backend::{AutodiffBackend, Backend, BackendTypes, ExecutionError},
    tensor::{BoolTensor, IntTensor, QuantizedTensor},
};

#[cfg(feature = "distributed")]
use burn_backend::distributed::{DistributedBackend, DistributedParamId, DistributedParams};

/// Enable auto-differentiation on a backend.
///
/// This works as a backend decorator, extending the functionality of any backend with
/// backpropagation.
#[derive(Clone, Copy, Debug, Default)]
pub struct Autodiff<B, C = NoCheckpointing> {
    _b: PhantomData<B>,
    _checkpoint_strategy: PhantomData<C>,
}

impl<B: Backend, C: CheckpointStrategy> BackendTypes for Autodiff<B, C> {
    type Device = B::Device;

    type FloatTensorPrimitive = AutodiffTensor<B>;
    type FloatElem = B::FloatElem;

    type IntTensorPrimitive = B::IntTensorPrimitive;
    type IntElem = B::IntElem;

    type BoolTensorPrimitive = B::BoolTensorPrimitive;
    type BoolElem = B::BoolElem;

    type QuantizedTensorPrimitive = B::QuantizedTensorPrimitive;
}

impl<B: Backend, C: CheckpointStrategy> Backend for Autodiff<B, C> {
    fn ad_enabled(_device: &Self::Device) -> bool {
        true
    }

    fn name(device: &Self::Device) -> String {
        format!("autodiff<{}>", B::name(device))
    }

    fn seed(device: &B::Device, seed: u64) {
        B::seed(device, seed)
    }

    fn sync(device: &B::Device) -> Result<(), ExecutionError> {
        B::sync(device)
    }

    fn memory_persistent_allocations<
        Output: Send,
        Input: Send,
        Func: Fn(Input) -> Output + Send,
    >(
        device: &Self::Device,
        input: Input,
        func: Func,
    ) -> Output {
        B::memory_persistent_allocations(device, input, func)
    }

    fn memory_cleanup(device: &Self::Device) {
        B::memory_cleanup(device)
    }

    fn staging<'a, Iter>(data: Iter, device: &Self::Device)
    where
        Iter: Iterator<Item = &'a mut burn_backend::TensorData>,
    {
        B::staging(data, device);
    }

    fn supports_dtype(device: &Self::Device, dtype: burn_std::DType) -> bool {
        B::supports_dtype(device, dtype)
    }

    fn dtype_usage(device: &Self::Device, dtype: burn_std::DType) -> burn_backend::DTypeUsageSet {
        B::dtype_usage(device, dtype)
    }

    fn device_count(type_id: u16) -> usize {
        B::device_count(type_id)
    }
}

#[cfg(not(feature = "distributed"))]
impl<B: Backend, C: CheckpointStrategy> AutodiffBackend for Autodiff<B, C> {
    type InnerBackend = B;
    type Gradients = Gradients;

    fn backward(tensor: AutodiffTensor<B>) -> Gradients {
        tensor.backward()
    }

    fn grad(tensor: &AutodiffTensor<B>, grads: &Gradients) -> Option<B::FloatTensorPrimitive> {
        tensor.grad(grads)
    }

    fn grad_remove(
        tensor: &AutodiffTensor<B>,
        grads: &mut Gradients,
    ) -> Option<B::FloatTensorPrimitive> {
        tensor.grad_remove(grads)
    }
    fn inner(tensor: AutodiffTensor<B>) -> B::FloatTensorPrimitive {
        tensor.primitive
    }

    fn from_inner(tensor: B::FloatTensorPrimitive) -> AutodiffTensor<B> {
        AutodiffTensor::new(tensor)
    }

    fn grad_replace(
        tensor: &AutodiffTensor<B>,
        grads: &mut Self::Gradients,
        grad: B::FloatTensorPrimitive,
    ) {
        tensor.grad_replace(grads, grad);
    }

    fn int_inner(tensor: IntTensor<Self>) -> IntTensor<Self::InnerBackend> {
        tensor
    }

    fn bool_inner(tensor: BoolTensor<Self>) -> BoolTensor<Self::InnerBackend> {
        tensor
    }

    fn int_from_inner(tensor: IntTensor<Self::InnerBackend>) -> IntTensor<Self> {
        tensor
    }

    fn bool_from_inner(tensor: BoolTensor<Self::InnerBackend>) -> BoolTensor<Self> {
        tensor
    }

    fn q_inner(tensor: QuantizedTensor<Self>) -> QuantizedTensor<Self::InnerBackend> {
        tensor
    }

    fn q_from_inner(tensor: QuantizedTensor<Self::InnerBackend>) -> QuantizedTensor<Self> {
        tensor
    }
}

#[cfg(feature = "distributed")]
impl<B: DistributedBackend, C: CheckpointStrategy> AutodiffBackend for Autodiff<B, C> {
    type InnerBackend = B;
    type Gradients = Gradients;

    fn backward(tensor: AutodiffTensor<B>) -> Gradients {
        tensor.backward()
    }

    fn grad(tensor: &AutodiffTensor<B>, grads: &Gradients) -> Option<B::FloatTensorPrimitive> {
        tensor.grad(grads)
    }

    fn grad_remove(
        tensor: &AutodiffTensor<B>,
        grads: &mut Gradients,
    ) -> Option<B::FloatTensorPrimitive> {
        tensor.grad_remove(grads)
    }
    fn inner(tensor: AutodiffTensor<B>) -> B::FloatTensorPrimitive {
        tensor.primitive
    }

    fn from_inner(tensor: B::FloatTensorPrimitive) -> AutodiffTensor<B> {
        AutodiffTensor::new(tensor)
    }

    fn grad_replace(
        tensor: &AutodiffTensor<B>,
        grads: &mut Self::Gradients,
        grad: B::FloatTensorPrimitive,
    ) {
        tensor.grad_replace(grads, grad);
    }

    fn int_inner(tensor: IntTensor<Self>) -> IntTensor<Self::InnerBackend> {
        tensor
    }

    fn bool_inner(tensor: BoolTensor<Self>) -> BoolTensor<Self::InnerBackend> {
        tensor
    }

    fn int_from_inner(tensor: IntTensor<Self::InnerBackend>) -> IntTensor<Self> {
        tensor
    }

    fn bool_from_inner(tensor: BoolTensor<Self::InnerBackend>) -> BoolTensor<Self> {
        tensor
    }

    fn q_inner(tensor: QuantizedTensor<Self>) -> QuantizedTensor<Self::InnerBackend> {
        tensor
    }

    fn q_from_inner(tensor: QuantizedTensor<Self::InnerBackend>) -> QuantizedTensor<Self> {
        tensor
    }

    fn set_distributed_params(
        tensor: AutodiffTensor<B>,
        param_id: DistributedParamId,
    ) -> AutodiffTensor<B> {
        tensor.grad_distributed(param_id)
    }

    fn distributed_params(tensor: &AutodiffTensor<B>) -> Option<DistributedParams> {
        tensor.node.distributed_params.clone()
    }

    fn is_distributed(tensor: &AutodiffTensor<B>) -> bool {
        tensor.node.distributed_params.is_some()
    }
}