Struct TensorParallelShim 

pub struct TensorParallelShim<'a> {
    pub base: &'a dyn Shim,
    pub manifest: &'a TpManifest,
}

What a chip allows of a tensor-parallel program: what base allows, less a fused op over more than one chip.

★ A FUSED OP RUNS ON ONE CHIP. Every statement of the program runs where the manifest puts what it binds — a chip’s own tensor on that chip, the replicated spine on every chip — and a section is relocated whole to one chip (relocate). So a fusion whose members bind tensors of different chips, or a chip’s tensor and the spine’s, would move a chip’s work onto another chip; and a fusion holding a collective’s fold would lower that fold as a sum on the spine where the program means an all-reduce over every chip. Both are refused here, by the chip the optimizer asks before it fuses; every other fusion is base’s to answer.

Fields

base: &'a dyn Shim

manifest: &'a TpManifest

Trait Implementations

impl Shim for TensorParallelShim<'_>

fn name(&self) -> String

The chip and the configuration it answers for, e.g. cuda-lane:rtx3060:TwoPassInt8.

fn supports(&self, class: ClassId) -> Support

What the chip’s link does with class. Every chip answers by a match over every class with no catch-all arm, so a class added to the registry fails the build until each chip answers it.

fn fits(&self, op: &Op, g: &TensorGraph) -> Result<(), String>

Whether the chip’s link builds op as it stands — its shapes, its accesses, its destination: Ok, or why not.

fn activation( &self, consumer: &Op, operand: usize, g: &TensorGraph, ) -> Option<QuantFormat>

The quantized form consumer reads its operand operand in, or None for one it reads as its values.

fn rules(&self) -> Vec<(Position, ChipRule<'_>)>

The chip’s own rules, each linked into the table where its position says.

fn numerics(&self) -> Option<Contract>

The numerics contract of the knob point the chip answers for, or None while the chip has not declared it.