Struct QuantizePerGroupBackwardPlan 

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

quantize_per_group backward plan.

STE: dx = (dy / scale[g]) * 1[in_range(round(x/scale[g])+zp[g])] with g the group index along the quant axis.

When to use: backward for QuantizePerGroupPlan. Caller retains FW input, scale[outer, num_groups], zp[outer, num_groups].

Dtypes: gradients in {f32, f64, f16, bf16}.

Shape limits: rank-2 [outer, axis_size] with axis_size % group_size == 0.

Workspace: none.

Precision guarantee: deterministic, bit-stable.

Implementations

impl<TIn: Element> QuantizePerGroupBackwardPlan<TIn>

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

Pick a kernel for desc.

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

Validate args.

pub fn workspace_size(&self) -> usize

Workspace bytes — none.

pub fn sku(&self) -> KernelSku

Identity.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

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

Launch.