Struct GpuCircuitCache 

pub struct GpuCircuitCache { /* private fields */ }

Implementations

impl GpuCircuitCache

pub fn provider(&self) -> &Arc<CudaKernelProvider>

pub fn var_log_weights_mut( &mut self, ) -> (&mut TrackedCudaSlice<f64>, &mut TrackedCudaSlice<f64>)

Returns mutable device-resident log-weight tables.

Device-only callers are responsible for validating numeric inputs and outputs. Positive infinity is valid when it does not enter an undefined normalization; NaN or undefined arithmetic that reaches evaluation is represented by a non-finite device result.

pub fn grad_true(&self) -> &TrackedCudaSlice<f64>

pub fn grad_false(&self) -> &TrackedCudaSlice<f64>

pub fn values(&self) -> &TrackedCudaSlice<f64>

pub fn meta_num_nodes_device(&self) -> &TrackedCudaSlice<u32>

pub fn meta_num_levels_device(&self) -> &TrackedCudaSlice<u32>

pub fn meta_root_device(&self) -> &TrackedCudaSlice<u32>

pub fn meta_max_var_device(&self) -> &TrackedCudaSlice<u32>

pub fn num_slots(&self) -> u32

pub fn new( provider: &Arc<CudaKernelProvider>, config: GpuCircuitCacheConfig, ) -> Result<Self>

pub fn lookup_or_insert(&mut self, key: u64) -> Result<GpuCacheLookup>

pub fn claim_slot(&mut self, key: u64) -> Result<GpuCircuitCacheHandle>

pub fn store_from_xgcf( &mut self, handle: &mut GpuCircuitCacheHandle, xgcf: &GpuXgcf, ) -> Result<()>

pub fn store_weights( &mut self, handle: &GpuCircuitCacheHandle, weights_true: &TrackedCudaSlice<f64>, weights_false: &TrackedCudaSlice<f64>, ) -> Result<()>

pub fn overwrite_weights( &mut self, handle: &GpuCircuitCacheHandle, weights_true: &TrackedCudaSlice<f64>, weights_false: &TrackedCudaSlice<f64>, ) -> Result<()>

pub fn store_free_var_mask( &mut self, handle: &GpuCircuitCacheHandle, mask: &TrackedCudaSlice<u8>, ) -> Result<()>

pub fn eval_log_wmc_device_inplace( &mut self, handle: &GpuCircuitCacheHandle, out_log_z: &mut TrackedCudaSlice<f64>, ) -> Result<()>

Evaluates cached logZ without a host transfer.

Callers must ensure evaluated resident weights are NaN-free. NaN or undefined arithmetic that reaches evaluation is represented by a NaN sentinel in out_log_z; callers must reject that sentinel at an existing readback. Positive infinity remains valid for value-only evaluation.

pub fn eval_log_wmc_device_only( &mut self, handle: &GpuCircuitCacheHandle, out_log_z: &mut TrackedCudaSlice<f64>, ) -> Result<()>

Device-only implementation of cached value evaluation; see Self::eval_log_wmc_device_inplace for its numeric preconditions and sentinel contract.

pub fn eval_grads_inplace( &mut self, handle: &GpuCircuitCacheHandle, ) -> Result<()>

Evaluates cached gradients without a host transfer.

Callers must validate numeric inputs and eventual outputs. Positive infinity is valid when the actual backward path remains finite. This asynchronous API cannot return a host-side numeric error; NaN or undefined normalization that reaches evaluation is observable as a non-finite device result.

pub fn eval_grads_inplace_fused( &mut self, handle: &GpuCircuitCacheHandle, ) -> Result<()>

Like [eval_grads_inplace] but replaces the per-level backward loop with a single launch of xgcf_backward_all_levels_cached, and omits the trailing device().synchronize() so that the caller can batch multiple queries before syncing. The same caller-side numeric validation contract applies.