Struct FakeQuantizePlan 

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

fake_quantize forward plan.

y = scale * (clamp(round(x / scale) + zp, q_min, q_max) - zp). The FP-only roundtrip of quantize-then-dequantize, no integer storage (PyTorch torch.fake_quantize_per_tensor_affine).

When to use: QAT (quantization-aware training) — produces a lossy FP output of the same dtype as the input. Pair with FakeQuantizeBackwardPlan for STE autograd.

Dtypes: {f32, f64, f16, bf16} in and out (same dtype).

Shape limits: flat [numel].

Workspace: none.

Precision guarantee: deterministic, bit-stable. Round-ties- even matches FW quantize.

Implementations

impl<TIn: Element> FakeQuantizePlan<TIn>

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

Pick a kernel.

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

Launch.