Struct QuantizePerTensorBackwardPlan 

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

quantize_per_tensor backward plan.

Straight-Through Estimator (STE): dx = (dy / scale) * 1[qmin ≤ round(x/scale)+zp ≤ qmax]. The in-range mask is recomputed in-kernel from the saved input x (no separate mask is saved on FW).

When to use: backward for QuantizePerTensorPlan. Caller must retain the original input x from the FW pass.

Dtypes: gradient dy and dx in input FP {f32, f64, f16, bf16}. TOut is the FW output int dtype, carried for SKU consistency only — BW kernel does not consume an int operand.

Shape limits: flat [numel].

Workspace: none.

Precision guarantee: deterministic, bit-stable. The 1/scale factor is mandatory (omitting it is the most common STE-grad bug).

Implementations

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

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

Pick a kernel.

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

Launch.