Struct ExactDdnnfProgram 

pub struct ExactDdnnfProgram { /* private fields */ }

Implementations

impl ExactDdnnfProgram

pub fn compile_source(source: &str) -> Result<Self>

pub fn compile_source_with_gpu(source: &str, config: GpuConfig) -> Result<Self>

pub fn compile_from_program( program: &Program, config: GpuConfig, ) -> Result<Self>

pub fn uses_gpu_production_backend(&self) -> bool

pub fn last_compile_profile(&self) -> Option<&CircuitCompileProfile>

Get the latest circuit compilation profile (populated when XLOG_WARMUP_PROFILE=1).

pub fn evaluate(&self) -> Result<ExactResult>

pub fn num_vars(&self) -> usize

pub fn random_var_indices(&self) -> Vec<u32>

Returns the indices of random (probabilistic) variables in order.

Random variables are those with non-trivial weights (not (0.0, 0.0)). These correspond to annotated disjunctions in the source program. The order matches the order variables were assigned during CNF encoding.

pub fn neural_backward_nll_buffers( &self, slots: &GpuWeightSlots, query_idx: usize, probs: &[CudaBuffer], out_grads: &mut [CudaBuffer], cfg: NeuralFastPathConfig, ) -> Result<()>

GPU neural fast-path: compute NLL gradients w.r.t. probability tensors (no host reads).

This implements the design in docs/design/2026-01-22-gpu-native-compilation-design.md §5.3:

• Fill AD conditional-chain log-weights from device-resident p[label].
• Run XGCF forward+backward on GPU.
• Scatter gradients back into probability-space via the correct chain rule (uses both grad_true + grad_false).

The output gradient buffers are updated in-place:

• Base run: out = dlogZ_base/dp
• Query-forced run: out -= dlogZ_query/dp Result: out = dL/dp for L = -log P(query | evidence) (NLL).

pub fn neural_backward_nll_buffers_with_device_loss( &self, slots: &GpuWeightSlots, query_idx: usize, probs: &[CudaBuffer], out_grads: &mut [CudaBuffer], cfg: NeuralFastPathConfig, expected_true: bool, ) -> Result<TrackedCudaSlice<f64>>

Same as Self::neural_backward_nll_buffers, but also returns the device-resident scalar NLL loss: L = -log P(query | evidence).

The returned slice has length 1 and is written on GPU (no device->host reads).

pub fn neural_backward_nll_buffers_batch_with_device_loss( &self, slots: &GpuWeightSlots, query_indices: &[usize], probs_batch: &[Vec<CudaBuffer>], out_grads_batch: &mut [Vec<CudaBuffer>], cfg: NeuralFastPathConfig, expected_true: bool, ) -> Result<TrackedCudaSlice<f64>>

Batched variant of Self::neural_backward_nll_buffers_with_device_loss.

Computes NLL gradients for batch queries that share one compiled circuit template and returns a device-resident vector of batch scalar losses.

On circuits that require free-variable correction, this falls back to the existing per-query path for correctness.

pub fn evaluate_gpu_with_grads(&self) -> Result<ExactResultWithGrads>

Trait Implementations

impl Clone for ExactDdnnfProgram

fn clone(&self) -> ExactDdnnfProgram

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.