Struct QuantizePerTokenBackwardPlan 

pub struct QuantizePerTokenBackwardPlan<TIn: Element> { /* private fields */ }

quantize_per_token backward plan.

STE: dx[n, d] = (dy[n, d] / scale[n]) * 1[qmin ≤ round(x[n,d]/scale[n])+zp[n] ≤ qmax]. Mask recomputed in-kernel.

When to use: backward for QuantizePerTokenPlan. Caller retains FW input, scale, zero_point.

Dtypes: gradients in {f32, f64, f16, bf16}; no int output — hence the single-type-parameter signature.

Shape limits: rank-2 [N, D]; per-row scale and zp of length N.

Workspace: none.

Precision guarantee: deterministic, bit-stable.

Implementations

impl<TIn: Element> QuantizePerTokenBackwardPlan<TIn>

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

Pick a kernel for desc.

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

Validate args.

pub fn workspace_size(&self) -> usize

Workspace bytes — none.

pub fn sku(&self) -> KernelSku

Identity of the kernel.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

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

Launch.