Struct DequantizePerTensorBackwardPlan 

pub struct DequantizePerTensorBackwardPlan<TIn: Element, TOut: IntElement> { /* private fields */ }

dequantize_per_tensor backward plan.

Straight-through. The dequant FW is exactly linear in the int input (x = scale * (q - zp)), so dq = scale * dy. The int input itself is non-differentiable; this BW propagates the upstream gradient back into FP-typed d_input for autograd continuity.

When to use: backward for DequantizePerTensorPlan.

Dtypes: gradients in {f32, f64, f16, bf16}; TOut carried for SKU parity with the FW Plan.

Shape limits: flat [numel].

Workspace: none.

Precision guarantee: deterministic, bit-stable.

Implementations

impl<TIn: Element, TOut: IntElement> DequantizePerTensorBackwardPlan<TIn, TOut>

pub fn select( _stream: &Stream, desc: &DequantizePerTensorBackwardDescriptor, _pref: PlanPreference, ) -> Result<Self>

Pick a kernel.

pub fn can_implement( &self, args: &DequantizePerTensorBackwardArgs<'_, TIn, TOut>, ) -> Result<()>

Validate args.

pub fn workspace_size(&self) -> usize

Workspace bytes.

pub fn sku(&self) -> KernelSku

Identity.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

pub fn run( &self, stream: &Stream, _workspace: Workspace<'_>, args: DequantizePerTensorBackwardArgs<'_, TIn, TOut>, ) -> Result<()>

Launch.