Enum TensorOpSpec 

pub enum TensorOpSpec {
    Matmul(MatmulSpec),
    Elementwise(ElementwiseSpec),
    SwiGlu(SwiGluSpec),
    FusedRow(FusedRowSpec),
    Rope(RopeSpec),
    Silu(SiluSpec),
    Sigmoid(SigmoidSpec),
    Gelu(GeluSpec),
    Scale(ScaleSpec),
    Softcap(SoftcapSpec),
    Decay(DecaySpec),
    L2Norm(L2NormSpec),
    Transpose(TransposeSpec),
    HeadSlice(HeadSliceSpec),
    Join(JoinSpec),
    HeadMatmul(HeadMatmulSpec),
    CausalConv(CausalConvSpec),
    GatedDelta(GatedDeltaSpec),
    Attention(AttentionSpec),
    FullAttention(FullAttentionSpec),
    Moe(MoeSpec),
    RowGather(RowGatherSpec),
    RowsGather(RowsGatherSpec),
    ArgMax(ArgMaxSpec),
    SetRow(SetRowSpec),
    IntRequant(IntRequantSpec),
    IntClamp(IntClampSpec),
    Relu(ReluSpec),
    LutAttention(LutAttentionSpec),
}

One recognized tensor op — the typed spec the shared recognizers lift. The variant set is the language’s tensor sugar; the lowering of each to a spatial template lives with the composer.

Variants

Matmul(MatmulSpec)

Elementwise(ElementwiseSpec)

SwiGlu(SwiGluSpec)

FusedRow(FusedRowSpec)

Rope(RopeSpec)

Silu(SiluSpec)

Sigmoid(SigmoidSpec)

Gelu(GeluSpec)

Scale(ScaleSpec)

Softcap(SoftcapSpec)

Decay(DecaySpec)

L2Norm(L2NormSpec)

Transpose(TransposeSpec)

HeadSlice(HeadSliceSpec)

Join(JoinSpec)

HeadMatmul(HeadMatmulSpec)

CausalConv(CausalConvSpec)

GatedDelta(GatedDeltaSpec)

Attention(AttentionSpec)

FullAttention(FullAttentionSpec)

Moe(MoeSpec)

RowGather(RowGatherSpec)

RowsGather(RowsGatherSpec)

ArgMax(ArgMaxSpec)

SetRow(SetRowSpec)

IntRequant(IntRequantSpec)

IntClamp(IntClampSpec)

Relu(ReluSpec)

LutAttention(LutAttentionSpec)

Implementations

impl TensorOpSpec

pub fn kind(&self) -> &'static str

The variant’s name — the kind a ledger or a schedule reads.

pub fn out(&self) -> Symbol

The tensor the op binds.

pub fn ins(&self) -> Vec<Symbol>

The tensors the op reads, in the spec’s operand order: matmul [a, b]; elementwise [a, b]; swiglu [gate, up]; fused row [x, weight?, residual?, bias?] (the present ones); rope [q, rc, rs, position?]; the unary maps and a scale [x]; decay [ar, dt, alog]; l2norm / transpose / head slice [x]; join [a, b]; per-head matmul [x, w]; causal conv [x, wc, cs]; gated delta [q, k, v, z, b, a, st, gn]; attention [q, k, v, position?, sinks?] and full attention [q, k, v]; moe [x, wr, wg, wu, wd, router bias?, gate, up and down biases?]; row gather [table, index?]; rows gather [table, indices]; arg-max [x]; set row [table, source, index?] (a NAMED index is an operand — the data the row comes from; a literal one is part of the op).

pub fn state_inputs(&self) -> Vec<usize>

The positions in Self::ins of the operands the op writes back in place — its carried state: the conv ring, the GDN memory, and the table a row is written into (the KV cache’s append). An attention only READS its cache — masked or not — so it carries nothing; the cache is carried because Set row writes it.

pub fn assembled_rows(&self) -> Option<usize>

The rows of a result the op writes a row at a time into a table it builds, or None: a block gated delta, whose rows each take the step in turn and write their output row. The table is state inside the step — each row’s write updates it in place — which the step assembles whole.

Trait Implementations

impl Clone for TensorOpSpec

fn clone(&self) -> TensorOpSpec

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for TensorOpSpec

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.