Public operation registry (generated)
This page is generated from the operation registries (api_ops! in crates/<family>/src/api/registry.rs; core in tenferro-runtime). Do not edit it by hand: each family’s section is rewritten by UPDATE_API_DOCS=1 cargo test -p <family> api_docs_are_current, and the same test fails when the committed section differs from the registry.
Call grammar
recv.op(operands…, exec)
recv.op_with(operands…, &opts, exec)
recv.op_into(operands…, &opts, exec, dest)
session.op(&x, operands…) // eager: the session is the receiver
traced.op_with(operands…, &opts) // traced: no exec, run through prepare → run
Suffixes
| Suffix | Meaning |
|---|---|
| (none) | Default options; returns a new result. |
_with |
Explicit XxxOptions. |
_into |
Writes into a caller-provided Dest (Overwrite, Accumulate or Scaled), only where the registry declares it. |
_in_place |
Consuming destructive update; exception E01 (in-place FFT). |
_optional |
Eager AD only: None for an inactive input instead of an error; exception E11. |
backward_with(&output, &seed) (exception E11) takes its seed as an operand; it is not the _with options form. Every other spelling of input kind, parameters, layer or output mode is not a suffix: it is a receiver type, an options field or a Dest value.
linalg
| Op | concrete | traced | eager | Suffixes |
|---|---|---|---|---|
svd |
LinalgOps | TracedLinalgOps | EagerLinalgOps | svd, svd_with |
svdvals |
LinalgOps | TracedLinalgOps | EagerLinalgOps | svdvals, svdvals_with |
qr |
LinalgOps | TracedLinalgOps | EagerLinalgOps | qr, qr_with |
householder_qr |
LinalgOps | TracedLinalgOps | EagerLinalgOps | householder_qr |
rank_revealing_qr |
LinalgOps | TracedLinalgOps | EagerLinalgOps | rank_revealing_qr, rank_revealing_qr_with |
lu |
LinalgOps | TracedLinalgOps | EagerLinalgOps | lu |
full_piv_lu |
LinalgOps | TracedLinalgOps | EagerLinalgOps | full_piv_lu |
full_piv_lu_solve |
LinalgOps | TracedLinalgOps | EagerLinalgOps | full_piv_lu_solve |
solve |
LinalgOps | TracedLinalgOps | EagerLinalgOps | solve, solve_into (overwrite) |
lstsq |
LinalgOps | TracedLinalgOps | EagerLinalgOps | lstsq |
triangular_solve |
LinalgOps | TracedLinalgOps | EagerLinalgOps | triangular_solve, triangular_solve_with |
cholesky |
LinalgOps | TracedLinalgOps | EagerLinalgOps | cholesky, cholesky_with |
eig |
LinalgOps | TracedLinalgOps | EagerLinalgOps | eig |
eigvals |
LinalgOps | TracedLinalgOps | EagerLinalgOps | eigvals |
eigh |
LinalgOps | TracedLinalgOps | EagerLinalgOps | eigh, eigh_with |
eigvalsh |
LinalgOps | TracedLinalgOps | EagerLinalgOps | eigvalsh, eigvalsh_with |
slogdet |
LinalgOps | TracedLinalgOps | EagerLinalgOps | slogdet |
det |
LinalgOps | TracedLinalgOps | EagerLinalgOps | det |
inv |
LinalgOps | TracedLinalgOps | EagerLinalgOps | inv |
pinv |
LinalgOps | TracedLinalgOps | EagerLinalgOps | pinv, pinv_with |
norm |
LinalgOps | TracedLinalgOps | EagerLinalgOps | norm, norm_with |
svd
Singular value decomposition a = u diag(s) vh; SvdShape::Full returns the full-matrices factors.
- Options:
SvdOptions - Output:
Svd<V, R> - Dtypes:
a: f32, f64, Complex32 or Complex64.uandvhhave the dtype ofa;sis real. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred. In a lane (
CpuEngine::with_lanes), a negative-stride input is gathered once into the lane’s pool; zero steady-state allocation is promised only for non-negative strides. A batched operand is read in place in a lane. - Errors:
Validation,Unsupported,NumericalFailure(NonConvergence) - Omissions:
SvdShape::Fullcombined with a non-defaultgauge,derivative_epsordriverreturnsUnsupported: the full-matrices route has no gauge or derivative regularization yetSvdShape::Fullhas no derivative: differentiating through it is unsupported- lanes (
CpuEngine::with_lanes):SvdShape::Thin,SvdShape::Fulland the gauge are available
svdvals
Singular values only, in descending order.
- Options:
SvdvalsOptions - Output:
R - Dtypes:
a: f32, f64, Complex32 or Complex64. The result is real. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred. In a lane (
CpuEngine::with_lanes), a negative-stride input is gathered once into the lane’s pool; zero steady-state allocation is promised only for non-negative strides. A batched operand is read in place in a lane. - Errors:
Validation,Unsupported,NumericalFailure(NonConvergence) - Omissions:
driveris a hint for providers that can choose a driver; the current CPU provider ignores it
qr
Thin QR factorization a = q r.
- Options:
QrOptions - Output:
Qr<V> - Dtypes:
a: f32, f64, Complex32 or Complex64.qandrhave the dtype ofa. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred. In a lane (
CpuEngine::with_lanes), a negative-stride input is gathered once into the lane’s pool; zero steady-state allocation is promised only for non-negative strides. A batched operand is read in place in a lane. - Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): only the defaultQrOptionsrun in a lane; a non-default gauge returnsUnsupported
- lanes (
householder_qr
Compact Householder QR: an opaque factorization state from which r and q_columns are extracted.
- Options: none
- Output:
HouseholderQr<V> - Dtypes:
a: f32, f64, Complex32 or Complex64. - Layout: The compact factorization takes a compact operand: an owned tensor is read in place, any other operand (a view or a typed tensor) is first made compact by the backend (same placement and dtype). Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
rank_revealing_qr
Column-pivoted QR with a numerical rank: q, r, column_permutation and rank.
- Options:
RankRevealingQrOptions - Output:
RankRevealingQrResult<V, I> - Dtypes:
a: f32, f64, Complex32 or Complex64.qandrhave the dtype ofa;column_permutationandrankare i64. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
lu
Partial-pivot LU a = p l u; p is a permutation matrix and parity is the real sign of the permutation.
- Options: none
- Output:
Lu<V> - Dtypes:
a: f32, f64, Complex32 or Complex64.p,l,uandparity(+1 or -1) have the dtype ofa. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
full_piv_lu
Complete-pivot LU a = p l u q; p and q are permutation matrices and parity is the real sign of the permutations.
- Options: none
- Output:
FullPivLu<V, R> - Dtypes:
a: f32, f64, Complex32 or Complex64.p,l,uandqhave the dtype ofa;parity(+1 or -1) is real for an ownedTensorand has the dtype ofaon the borrowed route (#2552 Q21). - Layout: A rank-2 matrix with non-negative strides is read in place; a negative stride or a batched input (rank >= 3) is gathered once into a compact buffer of the same placement and dtype. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported - typed Complex32/Complex64 input: returns
Unsupported; the borrowed route returnsparityin the input dtype while the owned route and the typed result type have it real (#2552 Q21); erased inputs keep the dtype their route returns
- lanes (
full_piv_lu_solve
Solves a x = b through the complete-pivot LU of a; a rank-deficient system is rejected.
- Options: none
- Output:
V - Dtypes:
aandb: the same dtype, one of f32, f64, Complex32 or Complex64.xhas that dtype. - Layout:
aandbthat are not owned compact tensors (views, typed tensors) are first made compact (same placement and dtype); owned tensors are read in place. Operands stay on their device; nothing is transferred. - Errors:
Validation,Unsupported,NumericalFailure(Singular) - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
solve
Solves a x = b with a partial-pivot LU of a; solve_into overwrites a destination.
- Options: none
- Output:
V - Dtypes:
aandb: the same dtype, one of f32, f64, Complex32 or Complex64.xhas that dtype. - Layout: A rank-2
awith a rank-1 or rank-2 non-emptybis read in place (bis gathered straight into the right-hand side); any other shape or layout is gathered once into compact buffers of the same placement and dtype.solve_intowrites the solution intodestdirectly whendestis a column-major matrix of the shape ofb; otherwise it computes the solution and copies it in. Operands anddestmust be on the same device; nothing is transferred. - Errors:
Validation,Unsupported,NumericalFailure(Singular) - Omissions:
solve_intoaccepts onlyDest::Overwrite;AccumulateandScaledreturnUnsupportedbeforedestis touched- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
lstsq
Least-squares solution of a x = b for a tall or square, full-column-rank a (thin QR, then a triangular solve).
- Options: none
- Output:
V - Dtypes:
aandb: the same dtype, one of f32, f64, Complex32 or Complex64.xhas that dtype. - Layout:
ais read likeqrreads it (rank-2 with non-negative strides in place, otherwise gathered once into a compact buffer of the same placement and dtype);bis read in place by the contraction. Operands stay on their device; nothing is transferred. - Errors:
Validation,Unsupported - Omissions:
- an underdetermined (wide,
rows < cols) system returns aValidationerror - lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- an underdetermined (wide,
triangular_solve
Solves a triangular system with a (or its transpose) on the chosen side; only the selected triangle of a is read.
- Options:
TriangularSolveOptions - Output:
V - Dtypes:
aandb: the same dtype, one of f32, f64, Complex32 or Complex64.xhas that dtype. - Layout: A rank-2
awith non-negative strides is read in place and a rank-2bis gathered straight into the right-hand side; any other shape or layout is gathered once into compact buffers of the same placement and dtype. Operands stay on their device; nothing is transferred. - Errors:
Validation,Unsupported,NumericalFailure(Singular) - Omissions:
MatOp::Adjointon a complex matrix returnsUnsupported; for a real matrix it equalsMatOp::Transpose- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
cholesky
Lower Cholesky factor l of a Hermitian positive-definite matrix, a = l l^H.
- Options:
CholeskyOptions - Output:
V - Dtypes:
a: f32, f64, Complex32 or Complex64.lhas the dtype ofa. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported,NumericalFailure(NonConvergence) - Omissions:
Uplo::UpperreturnsUnsupported: only the lower factor is computed- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
eig
General eigendecomposition a v = v diag(values); eigenvectors have unit 2-norm.
- Options: none
- Output:
Eig<C> - Dtypes:
a: f32, f64, Complex32 or Complex64.valuesandvectorsare complex (Complex32 for f32 and Complex32 input, Complex64 otherwise). - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported,NumericalFailure(NonConvergence),NumericalFailure(NonFinite) - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
eigvals
Eigenvalues of a general matrix, as complex numbers.
- Options: none
- Output:
C - Dtypes:
a: f32, f64, Complex32 or Complex64. The result is complex. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred. In a lane (
CpuEngine::with_lanes), a negative-stride input is gathered once into the lane’s pool; zero steady-state allocation is promised only for non-negative strides. A batched operand is read in place in a lane. - Errors:
Validation,Unsupported,NumericalFailure(NonConvergence),NumericalFailure(NonFinite) - Omissions: none
eigh
Eigendecomposition of a Hermitian matrix (the lower triangle is read); values are real and ascending.
- Options:
EighOptions - Output:
Eigh<V, R> - Dtypes:
a: f32, f64, Complex32 or Complex64.vectorshas the dtype ofa;valuesis real. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred. In a lane (
CpuEngine::with_lanes), a negative-stride input is gathered once into the lane’s pool; zero steady-state allocation is promised only for non-negative strides. A batched operand is read in place in a lane. - Errors:
Validation,Unsupported,NumericalFailure(NonConvergence) - Omissions:
driveris a hint for providers that can choose a driver; the current CPU provider ignores it
eigvalsh
Eigenvalues of a Hermitian matrix (the lower triangle is read), real and ascending.
- Options:
EigvalshOptions - Output:
R - Dtypes:
a: f32, f64, Complex32 or Complex64. The result is real. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred. In a lane (
CpuEngine::with_lanes), a negative-stride input is gathered once into the lane’s pool; zero steady-state allocation is promised only for non-negative strides. A batched operand is read in place in a lane. - Errors:
Validation,Unsupported,NumericalFailure(NonConvergence) - Omissions:
driveris a hint for providers that can choose a driver; the current CPU provider ignores it
slogdet
Sign and natural log of the absolute determinant of a square matrix.
- Options: none
- Output:
Slogdet<V, R> - Dtypes:
a: f32, f64, Complex32 or Complex64.signhas the dtype ofa;log_abs_detis real. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
det
Determinant of a square matrix, computed as sign * exp(log_abs_det).
- Options: none
- Output:
V - Dtypes:
a: f32, f64, Complex32 or Complex64. The result has the dtype ofa. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the computation. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
inv
Inverse of a square matrix (a solve against the identity).
- Options: none
- Output:
V - Dtypes:
a: f32, f64, Complex32 or Complex64. The result has the dtype ofa. - Layout: A rank-2
ais read in place by the solve (the identity right-hand side is allocated compact); any other shape or layout is gathered once into a compact buffer of the same placement and dtype. Operands stay on their device; nothing is transferred. - Errors:
Validation,Unsupported,NumericalFailure(Singular) - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
pinv
Moore-Penrose pseudoinverse through the SVD; singular values below rtol * max(s) are dropped.
- Options:
PinvOptions - Output:
V - Dtypes:
a: f32, f64, Complex32 or Complex64. The result has the dtype ofa. - Layout: A rank-2 matrix with non-negative strides (compact, transposed or any strided view) is read in place; a negative stride, or a batched input (rank >= 3), is gathered once into a compact buffer of the same placement and dtype before the SVD. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported,NumericalFailure(NonConvergence) - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
norm
Vector or matrix norm over the chosen axes, selected by NormOrd; NormOrd::Zero counts entries != 0 (NaN counts as nonzero).
- Options:
NormOptions - Output:
R - Dtypes:
a: f32, f64, Complex32 or Complex64. The result is real. - Layout:
NormOrd::Frobeniusover all axes of a strided operand is one fused pass that reads it in place; the other orders read the operand through strided reductions, and a matrix norm over axes that are not the leading two first moves them to the front (a layout change of the same placement and dtype). Operands stay on their device; nothing is transferred. - Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): not available; returnsUnsupported
- lanes (
Exceptions
| ID | Status | Symbols | Scope | Category | Owner | Reason | Verification | Exit condition |
|---|---|---|---|---|---|---|---|---|
| E08 | retired | tenferro_linalg::svd_with_exectenferro_linalg::svd_values_with_exectenferro_linalg::qr_with_exectenferro_linalg::eigh_with_exectenferro_linalg::eigh_values_with_exectenferro_linalg::eig_values_with_exec |
free functions that run a factorization on a borrowed execution bundle and buffer pool | legacy | free function | tenferro_linalg::with_exec |
retired: replaced by CpuEngine::with_lanes, which hands each lane an engine that runs the ordinary grammar (view.svd_with(&opts, lane)); the six entries were added for #2462 to split threads into lanes | retired: the P7 prototype verified lifetime, buffer reuse, nesting and concurrency, and no retention after the call; the lane route keeps the steady-state allocation counts of the former entries | met: the six functions, their module and the Borrowed-Exec section of the crate docs were removed (#2552 S3); lanes and the grammar replace them |
core
| Op | concrete | traced | eager | Suffixes |
|---|---|---|---|---|
exp |
CoreOps | described | described | exp |
neg |
CoreOps | described | described | neg |
sign |
CoreOps | described | described | sign |
conj |
CoreOps | described | described | conj |
log |
CoreOps | described | described | log |
expm1 |
CoreOps | described | described | expm1 |
log1p |
CoreOps | described | described | log1p |
erf |
CoreOps | described | described | erf |
sin |
CoreOps | described | described | sin |
cos |
CoreOps | described | described | cos |
tanh |
CoreOps | described | described | tanh |
sqrt |
CoreOps | described | described | sqrt |
rsqrt |
CoreOps | described | described | rsqrt |
abs |
CoreOps | described | described | abs |
sigmoid |
CoreOps | described | described | sigmoid |
silu |
CoreOps | described | described | silu |
softplus |
CoreOps | described | described | softplus |
gelu |
CoreOps | described | described | gelu |
gelu_tanh |
CoreOps | described | described | gelu_tanh |
mul |
CoreOps | described | described | mul |
sub |
CoreOps | described | described | sub |
div |
CoreOps | described | described | div |
rem |
CoreOps | described | described | rem |
pow |
CoreOps | described | described | pow |
maximum |
CoreOps | described | described | maximum |
minimum |
CoreOps | described | described | minimum |
add |
CoreOps | described | described | add, add_into (overwrite) |
matmul |
CoreOps | described | described | matmul |
compare |
CoreOps | described | described | compare |
clamp |
CoreOps | described | described | clamp |
where_select |
CoreOps | described | described | where_select |
dot_general |
CoreOps | described | described | dot_general, dot_general_with |
reduce_sum |
CoreOps | described | described | reduce_sum |
reduce_max |
CoreOps | described | described | reduce_max |
reduce_min |
CoreOps | described | described | reduce_min |
reduce_prod |
CoreOps | described | described | reduce_prod |
reduce_sum_squares |
CoreOps | described | described | reduce_sum_squares |
reduce_mean |
CoreOps | described | described | reduce_mean |
reshape |
CoreOps | described | described | reshape |
broadcast_in_dim |
CoreOps | described | described | broadcast_in_dim |
transpose |
CoreOps | described | described | transpose |
reverse |
CoreOps | described | described | reverse |
tril |
CoreOps | described | described | tril |
triu |
CoreOps | described | described | triu |
extract_diag |
CoreOps | described | described | extract_diag |
embed_diag |
CoreOps | described | described | embed_diag |
slice |
CoreOps | described | described | slice |
pad |
CoreOps | described | described | pad |
gather |
CoreOps | described | described | gather |
scatter |
CoreOps | described | described | scatter |
dynamic_slice |
CoreOps | described | described | dynamic_slice |
dynamic_update_slice |
CoreOps | — | — | dynamic_update_slice |
concatenate |
CoreOps | described | described | concatenate |
convert |
CoreOps | described | described | convert |
cast |
CoreOps | described | described | cast |
scale_real |
CoreOps | described | described | scale_real |
scale_complex |
CoreOps | described | described | scale_complex |
softmax |
CoreOps | described | described | softmax |
log_softmax |
CoreOps | described | described | log_softmax |
masked_softmax |
CoreOps | described | described | masked_softmax |
masked_log_softmax |
CoreOps | described | described | masked_log_softmax |
take_along_axis |
CoreOps | described | described | take_along_axis |
layer_norm |
CoreOps | described | described | layer_norm |
rms_norm |
CoreOps | described | described | rms_norm |
select |
— | described | described | select |
stack |
— | described | described | stack |
index_select |
— | described | described | index_select |
slice_axis |
— | described | described | slice_axis |
take_axis |
— | described | described | take_axis |
take_rows |
— | — | described | take_rows |
take_cols |
— | — | described | take_cols |
take_block |
— | — | described | take_block |
dynamic_truncate |
— | described | — | dynamic_truncate |
pad_to_match |
— | described | — | pad_to_match |
exp
Elementwise exponential.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
neg
Elementwise negation.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
sign
Elementwise sign.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
conj
Elementwise complex conjugate (identity on real dtypes).
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
log
Elementwise natural logarithm.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
expm1
Elementwise exp(x) - 1.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
log1p
Elementwise log(1 + x).
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
erf
Elementwise error function.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
sin
Elementwise sine.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
cos
Elementwise cosine.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
tanh
Elementwise hyperbolic tangent.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
sqrt
Elementwise square root.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
rsqrt
Elementwise reciprocal square root.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
abs
Elementwise absolute value; complex inputs return their modulus.
- Options: none
- Output:
R - Dtypes: Any numeric dtype; complex inputs return the real dtype of the same precision.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
sigmoid
Logistic sigmoid.
- Options: none
- Output:
V - Dtypes: f32 or f64.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
silu
SiLU (x * sigmoid(x)).
- Options: none
- Output:
V - Dtypes: f32 or f64.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
softplus
Softplus log(1 + exp(x)).
- Options: none
- Output:
V - Dtypes: f32 or f64.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
gelu
GELU with the exact error function.
- Options: none
- Output:
V - Dtypes: f32 or f64.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
gelu_tanh
GELU with the tanh approximation.
- Options: none
- Output:
V - Dtypes: f32 or f64.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
mul
Elementwise product with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
sub
Elementwise difference with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
div
Elementwise quotient with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
rem
Elementwise remainder with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
pow
Elementwise power with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
maximum
Elementwise maximum with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
minimum
Elementwise minimum with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
add
Elementwise sum with broadcasting.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel.add_intowrites the result into the destination’s storage. - Errors:
Validation,Unsupported - Omissions:
- traced and eager: only
add;add_intoexists on the concrete layer
- traced and eager: only
matmul
Matrix product of the two leading axes, batched over trailing axes.
- Options: none
- Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
compare
Elementwise comparison with broadcasting; returns a bool tensor.
- Options: none
- Output:
B - Dtypes: Both operands share a dtype; the result is
bool. - Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
clamp
Elementwise clamp to [lower, upper] with broadcasting.
- Options: none
- Output:
V - Dtypes: All operands share a real dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
where_select
Elementwise selection: on_true where the receiver is true, on_false elsewhere, with broadcasting. The receiver is the bool condition; a typed receiver must be TypedTensor<bool>.
- Options: none
- Output:
<Y::Repr as tenferro_tensor::Repr>::Value - Dtypes: Receiver
bool; the branches share a dtype, which is the result dtype. - Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
dot_general
General contraction with batch dimensions (DotGeneralConfig); DotGeneralOptions conjugates an operand.
- Options:
tenferro_tensor::DotGeneralOptions - Output:
V - Dtypes: Both operands share a numeric dtype; the result has that dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions:
- traced: no
dot_general_with; conjugate an operand withconj()first
- traced: no
reduce_sum
Sum over axes (None: every axis); axes are signed, -1 is the last axis.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
reduce_max
Maximum over axes (None: every axis); axes are signed, -1 is the last axis.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
reduce_min
Minimum over axes (None: every axis); axes are signed, -1 is the last axis.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
reduce_prod
Product over axes (None: every axis); axes are signed, -1 is the last axis.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
reduce_sum_squares
Sum of squared magnitudes over axes (None: every axis); axes are signed, -1 is the last axis.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
reduce_mean
Mean over axes (None: every axis); axes are signed, -1 is the last axis.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
reshape
Reshape to shape in column-major element order.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. A view that is not compact in column-major order is gathered by the engine; reshape is a data operation, not a metadata query. - Errors:
Validation,Unsupported - Omissions: none
broadcast_in_dim
Broadcast to shape; dims[i] is the output axis of input axis i (an IR-level dimension map, unsigned like the *Config structs).
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
transpose
Permute axes; perm is signed (-1 is the last axis).
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
reverse
Reverse the order of elements along axes (signed).
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
tril
Lower triangle of the two leading axes relative to diagonal k.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
triu
Upper triangle of the two leading axes relative to diagonal k.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
extract_diag
Extract the diagonal of axes axis_a and axis_b (signed).
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
embed_diag
Embed axis axis_a of the receiver as the diagonal of a new axis pair; axis_b (signed, against the output rank) is the position of the inserted axis.
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
slice
Static strided slice (SliceConfig).
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
pad
Pad with edge and interior padding (PadConfig).
- Options: none
- Output:
V - Dtypes: Any numeric dtype the engine supports for the operation; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
gather
XLA-style gather (GatherConfig); indices is an integer tensor of any representation.
- Options: none
- Output:
V - Dtypes: Operand: any dtype; indices: an integer dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
scatter
XLA-style scatter (ScatterConfig); indices and updates may have any representation.
- Options: none
- Output:
V - Dtypes: Operand and updates share a dtype; indices: an integer dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
dynamic_slice
Slice of sizes starting at the run-time starts (an integer vector).
- Options: none
- Output:
V - Dtypes: Operand: any dtype; starts: an integer dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
dynamic_update_slice
Copy of the receiver with update written at the run-time starts.
- Options: none
- Output:
V - Dtypes: Operand and update share a dtype; starts: an integer dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions:
- traced and eager: not available
concatenate
Concatenate inputs along axis (signed). Called as CoreOps::concatenate(&[&a, &b], axis, exec).
- Options: none
- Output:
V - Dtypes: All inputs share a dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
convert
Checked dtype conversion along tenferro’s promotion lattice. The result is a dtype-erased Tensor because its dtype is a run-time value.
- Options: none
- Output:
tenferro_tensor::Tensor - Dtypes: Any source dtype the conversion supports; the result has dtype
to. - Layout: An owned
Tensorreceiver is converted directly. Any other receiver (a borrowed view or a typed tensor) is first gathered once into a compact tensor of the same placement and dtype, because the engine conversion reads an owned tensor. - Errors:
Validation,Unsupported - Omissions: none
cast
Explicit, possibly lossy dtype projection. The result is a dtype-erased Tensor because its dtype is a run-time value.
- Options: none
- Output:
tenferro_tensor::Tensor - Dtypes: Any source dtype the conversion supports; the result has dtype
to. - Layout: An owned
Tensorreceiver is converted directly. Any other receiver (a borrowed view or a typed tensor) is first gathered once into a compact tensor of the same placement and dtype, because the engine conversion reads an owned tensor. - Errors:
Validation,Unsupported - Omissions: none
scale_real
Multiply by a real scalar.
- Options: none
- Output:
V - Dtypes: f32, f64, Complex32 or Complex64; the result has the input dtype.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
scale_complex
Multiply by a complex scalar.
- Options: none
- Output:
V - Dtypes: Complex32 or Complex64.
- Layout: Every operand is read through
TensorReadby the engine’s*_readentry: owned, borrowed, typed and strided inputs reach the same body. Implicit broadcasting is a zero-stride view, never a copy. Whether a strided or negative-stride view is read in place or gathered once (same placement and dtype) is the engine’s documented choice for that kernel. - Errors:
Validation,Unsupported - Omissions: none
softmax
Softmax along axis (signed).
- Options: none
- Output:
V - Dtypes: f32 or f64.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
log_softmax
Log-softmax along axis (signed).
- Options: none
- Output:
V - Dtypes: f32 or f64.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
masked_softmax
Softmax along axis over the entries where mask is true.
- Options: none
- Output:
V - Dtypes: Input f32 or f64; mask
boolof any representation. - Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
masked_log_softmax
Log-softmax along axis over the entries where mask is true.
- Options: none
- Output:
V - Dtypes: Input f32 or f64; mask
boolof any representation. - Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
take_along_axis
Gather along axis (signed) with per-position integer indices.
- Options: none
- Output:
V - Dtypes: Input: any dtype; indices: an integer dtype.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
layer_norm
Layer normalization over axis (signed) with optional affine weight and bias.
- Options: none
- Output:
V - Dtypes: f32 or f64; weight and bias share the input representation.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
rms_norm
RMS normalization over axis (signed) with optional affine weight and bias.
- Options: none
- Output:
V - Dtypes: f32 or f64; weight and bias share the input representation.
- Layout: Composed from core primitives inside the engine; operands are read through
TensorReadand intermediates are engine allocations. A fused engine kernel is used when the engine offers one. - Errors:
Validation,Unsupported - Omissions: none
select
Alias of where_select with the condition as the first argument.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete: use
where_select
- concrete: use
stack
Stack equally shaped tensors along a new signed axis.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete: not available; use
reshapeandconcatenate
- concrete: not available; use
index_select
Select positions along a signed axis.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete: not available; use
gather
- concrete: not available; use
slice_axis
Slice a range along a signed axis.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete: not available; use
slice
- concrete: not available; use
take_axis
Take the listed positions along a signed axis.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete: not available; use
gather
- concrete: not available; use
take_rows
Take the listed rows of a matrix.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete and traced: not available
take_cols
Take the listed columns of a matrix.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete and traced: not available
take_block
Take a block given row and column positions.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete and traced: not available
dynamic_truncate
Truncate a signed axis to a run-time size.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete and eager: not available
pad_to_match
Zero-pad a signed axis to the size of a reference tensor.
- Options: none
- Output:
V - Dtypes: See the method documentation.
- Layout: Recorded as graph operations (traced) or executed by the eager engine.
- Errors:
Validation,Unsupported - Omissions:
- concrete and eager: not available
Exceptions
| ID | Status | Symbols | Scope | Category | Owner | Reason | Verification | Exit condition |
|---|---|---|---|---|---|---|---|---|
| E03 | retained | tenferro_einsum::ConcreteEinsumPlan::clear_cpu_cachetenferro_einsum::ConcreteEinsumPlan::cpu_cache_stats |
plan-cache owner methods of the concrete einsum plan (clearing and inspecting the cache) | cache owner | inherent method | tenferro_einsum::concrete |
the plan owns its CPU numerical descriptor cache; a cache owner exposes its cache, and clearing or inspecting it is not a tensor operation and has no options, inputs or destination | the einsum registry tests, the cache tests and the doctests of both methods; the plan’s execute/execute_into follow the grammar (inputs, exec, dest) | kept while ConcreteEinsumPlan owns a CPU cache |
| E04 | retained | tenferro_runtime::TraceContexttenferro_runtime::GraphCompiler::compiletenferro_runtime::GraphCompiler::compile_manytenferro_runtime::Runtime::preparetenferro_runtime::Runtime::runtenferro_runtime::TraceContext::finish |
trace, compile, prepare and run lifecycle | lifecycle | builder | tenferro_runtime::{trace, graph, runtime} |
program construction and execution are not tensor operations; their shapes are fixed by the public API snapshot (api/) | scripts/api-snapshot.sh –check and the runtime lifecycle tests | kept |
| E05 | retained | tenferro_tensor::Tensor::shapetenferro_tensor::Tensor::dtypetenferro_tensor::TypedTensor::as_viewtenferro_tensor::TypedTensorView::transpose_viewtenferro_tensor::TypedTensorView::slice_viewtenferro_runtime::TracedTensor::ranktenferro_runtime::TracedTensor::sym_sizetenferro_runtime::TracedTensor::axis_sym_dimtenferro_runtime::TracedTensor::shape_of |
descriptor-only operations and metadata queries | metadata | view | tenferro_tensor, tenferro_runtime::traced |
they never touch element data, so they take no engine and are not operations of the grammar | tenferro-runtime metadata_queries_and_views_do_not_copy (pointer identity) |
kept |
| E06 | retained | tenferro_cpu::with_borrowed_cpu_enginetenferro_tensor::Dest::into_accumulationnum_complex::Complex |
approved external types and owner-scoped doc-hidden bridges | bridge | external type | tenferro_cpu, tenferro_tensor::dest |
the lane bridge lets tenferro-linalg reach the lane engine (#2552 S2b); into_accumulation maps a Dest to the backend accumulation descriptor (#2552 Q3); num_complex is the public complex scalar | listed symbol by symbol in api/allowlist.txt; the bridges are #[doc(hidden)] | kept while the owners need the bridge |
| E09 | retained | tenferro_einsum::Errortenferro_einsum::PlanningErrortenferro_einsum::parse_einsum_subscriptstenferro_einsum::parse_einsum_notationtenferro_einsum::ContractionTree::optimizetenferro_einsum::ContractionTree::optimize_with_optionstenferro_einsum::ContractionTree::from_pairstenferro_linalg::Errortenferro_fft::EagerFftInPlaceError |
domain errors of parsers and planners (einsum parser and contraction planner), and failures that return ownership | error | domain | the owning family crates |
parsers and planners report einsum-domain errors; the grammar’s operations wrap them (concrete: tenferro_tensor::Error::Extension, traced/eager: tenferro_runtime::Error::Extension) as the source() of the layer error |
the error-cause tests of each family; the einsum error-cause tests reach tenferro_einsum::Error::InvalidSubscripts through source() on all three layers | kept |
| E10 | retained | tenferro_tensor::Enginetenferro_tensor::TensorDottenferro_tensor::TensorDot::dot_general_read_into_accumtenferro_tensor::Engine::axpby_read_into_accumtenferro_linalg::LinalgBackendtenferro_fft::FftBackend |
backend and provider traits and their read/accumulate hooks | provider contract | trait | tenferro_tensor::backend and the family backends |
a provider-facing contract, not the user grammar; *_read and *_read_into_accum name backend hooks |
the backend contract tests of each crate | kept |
einsum
| Op | concrete | concrete_tensordot | traced | traced_tensordot | eager | Suffixes |
|---|---|---|---|---|---|---|
einsum |
EinsumOps | — | TracedEinsumOps | — | EagerEinsumOps | einsum, einsum_with, einsum_into (overwrite, accumulate, scaled) |
tensordot |
— | TensordotOps | — | TracedTensordotOps | EagerEinsumOps | tensordot |
einsum
Einstein summation over N inputs: contract, trace, transpose, reduce and broadcast as the spec says.
- Options:
EinsumOptions - Output:
V - Dtypes: Every input has the same dtype (any numeric dtype the backend supports); the result has it too.
- Layout: Strided inputs are read in place by the contraction kernels (any strides, including transposed and negative-stride views); an operand that a kernel cannot read is packed once into a compact buffer of the same placement and dtype. A two-input contraction that maps to one dot_general writes straight into the destination. Inputs stay on their device.
_intochecks the destination’s dtype and shape before any write;AccumulateandScaledread the destination,OverwriteandScaledwithbeta = 0never read it. - Errors:
Validation,Unsupported,RuntimeState - Omissions:
_intois concrete only: traced and eager results are new values- lanes (
CpuEngine::with_lanes): einsum is unsupported in a lane Scaledneedsalphaandbetaof the input dtype
tensordot
Contract a with b over the given axis pairs (or the last/first n axes); free axes of a come first.
- Options: none
- Output:
V - Dtypes:
aandbhave the same dtype; the result has it too. - Layout: Lowered to one dot_general, which reads strided operands in place; an operand that the provider cannot read is packed once into a compact buffer of the same placement and dtype.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): tensordot is unsupported in a lane
- lanes (
Exceptions
| ID | Status | Symbols | Scope | Category | Owner | Reason | Verification | Exit condition |
|---|---|---|---|---|---|---|---|---|
| E07 | retired | tenferro_einsum::TraceContextEinsumExt::einsumtenferro_einsum::TraceContextEinsumExt::einsum_notationtenferro_einsum::TraceContextEinsumExt::einsum_subscriptstenferro_einsum::TraceContextEinsumExt::einsum_withtenferro_einsum::TraceContextEinsumExt::einsum_notation_withtenferro_einsum::TraceContextEinsumExt::einsum_subscripts_with |
einsum entries that took and returned TraceValue on a &mut TraceContext |
legacy | trait | tenferro_einsum::traced |
retired: replaced by TracedEinsumOps for TraceContext (trace.einsum(&[&t1, &t2], spec)), which takes and returns TracedTensor (#2552 decision 4) |
retired: a temporary old-vs-new test recorded the same values before removal (#2552 S4) | met: the trait was removed in #2552 S4 |
fft
| Op | concrete | traced | eager | Suffixes |
|---|---|---|---|---|
fft |
FftOps | TracedFftOps | EagerFftOps | fft, fft_with |
ifft |
FftOps | TracedFftOps | EagerFftOps | ifft, ifft_with |
rfft |
FftOps | TracedFftOps | EagerFftOps | rfft, rfft_with |
irfft |
FftOps | TracedFftOps | EagerFftOps | irfft, irfft_with |
fftn |
FftOps | TracedFftOps | EagerFftOps | fftn, fftn_with |
ifftn |
FftOps | TracedFftOps | EagerFftOps | ifftn, ifftn_with |
rfftn |
FftOps | TracedFftOps | EagerFftOps | rfftn, rfftn_with |
irfftn |
FftOps | TracedFftOps | EagerFftOps | irfftn, irfftn_with |
fft
One-axis discrete Fourier transform along axis (default the last axis).
- Options:
FftOptions - Output:
C - Dtypes:
a: f32, f64, Complex32 or Complex64. The result is complex (Complex32 for f32/Complex32 inputs, Complex64 otherwise); a real input is transformed as a complex one with zero imaginary part. - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
ifft
One-axis inverse discrete Fourier transform along axis.
- Options:
FftOptions - Output:
C - Dtypes:
a: Complex32 or Complex64. The result has the dtype ofa. - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
rfft
One-axis transform of a real signal, returning the one-sided spectrum.
- Options:
FftOptions - Output:
C - Dtypes:
a: f32 or f64. The result is the one-sided complex spectrum (Complex32 or Complex64) of lengthn / 2 + 1alongaxis. - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
irfft
Inverse of rfft: a one-sided spectrum back to a real signal.
- Options:
FftOptions - Output:
R - Dtypes:
a: Complex32 or Complex64 (a one-sided spectrum). The result is real (f32 or f64); its length alongaxisisn, by default2 * (m - 1). - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
fftn
N-dimensional discrete Fourier transform over axes (default every axis).
- Options:
FftnOptions - Output:
C - Dtypes:
a: f32, f64, Complex32 or Complex64. The result is complex, as forfft. - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred. The transform runs one one-axis transform per axis; the intermediate spectra stay on the device and the normalization is applied once to the whole transform.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
ifftn
N-dimensional inverse discrete Fourier transform over axes.
- Options:
FftnOptions - Output:
C - Dtypes:
a: Complex32 or Complex64. The result has the dtype ofa. - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred. The transform runs one one-axis transform per axis; the intermediate spectra stay on the device and the normalization is applied once to the whole transform.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
rfftn
N-dimensional transform of a real signal; the last transformed axis is one-sided.
- Options:
FftnOptions - Output:
C - Dtypes:
a: f32 or f64. The result is complex; the last transformed axis is one-sided (n / 2 + 1). - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred. The transform runs one one-axis transform per axis; the intermediate spectra stay on the device and the normalization is applied once to the whole transform.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
irfftn
Inverse of rfftn.
- Options:
FftnOptions - Output:
R - Dtypes:
a: Complex32 or Complex64. The result is real (f32 or f64); the last transformed axis has length2 * (m - 1)unlessshapesays otherwise. - Layout: On the CPU provider a compact column-major operand is read in place; any other strided view (transposed, sliced, negative strides) is gathered once into a compact buffer of the same placement and dtype before the transform. The CUDA provider canonicalizes on the device. Operands stay on their device; nothing is transferred. The transform runs one one-axis transform per axis; the intermediate spectra stay on the device and the normalization is applied once to the whole transform.
- Errors:
Validation,Unsupported - Omissions:
- lanes (
CpuEngine::with_lanes): an FFT on a lane engine returnsUnsupported
- lanes (
Exceptions
| ID | Status | Symbols | Scope | Category | Owner | Reason | Verification | Exit condition |
|---|---|---|---|---|---|---|---|---|
| E01 | retained | tenferro_fft::EagerFftOps::fft_in_placetenferro_fft::EagerFftOps::ifft_in_placetenferro_fft::EagerFftInPlaceError |
consuming eager FFTs that reuse the input storage | in-place | consuming | tenferro_fft::eager_in_place |
an FFT of a large compact complex tensor can reuse its storage; the grammar’s _into writes to a separate destination and cannot express a consumed input |
src/api/tests/registry.rs: a length change and a real input are Rejected with the input returned unchanged; an accepted call equals the out-of-place fft bit for bit | a consuming Dest form in the grammar |
| E02 | retained | tenferro_fft::FftExecutor::fft_withtenferro_fft::FftExecutor::ifft_withtenferro_fft::FftExecutor::rfft_withtenferro_fft::FftExecutor::irfft_withtenferro_fft::FftExecutor::fftn_withtenferro_fft::FftExecutor::ifftn_withtenferro_fft::FftExecutor::rfftn_withtenferro_fft::FftExecutor::irfftn_withtenferro_fft::FftExecutor::newtenferro_fft::FftExecutor::plan_cachetenferro_fft::FftExecutor::plan_cache_muttenferro_fft::FftExecutor::cache_statstenferro_fft::FftExecutor::clear_cache |
the caller-owned FFT plan cache and its transforms | cache owner | tenferro_fft |
the receiver is the plan cache, not the input; the transforms take the input, the options and the engine in grammar order | src/api/tests/registry.rs: every _with equals the FftOps result bit for bit and fills the caller’s plan cache |
plan caches owned by the engine for every caller |
eager_ad
| Op | eager | Suffixes |
|---|---|---|
grad |
described | grad |
vjp |
described | vjp |
jvp |
described | jvp |
backward |
described | backward |
detach |
described | detach |
duplicate_value |
described | duplicate_value |
constant_from |
described | constant_from |
constant_from_host |
described | constant_from_host |
variable_from |
described | variable_from |
grad
Gradient of a scalar eager output with respect to an eager tensor; returns a composable eager tensor and touches no gradient slot.
- Options: none
- Output:
EagerTensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::grad(&mut self, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor> - Dtypes: See the method documentation.
- Layout: Reads the retained values of the operands; the derivative program is prepared and run on the session’s engine. Every eager operand and seed is checked against the session’s runtime before any value is materialized or any gradient slot is written.
- Errors:
ContextMismatch,NonScalarGrad,Validation,UnsupportedAdRule - Omissions:
- concrete: not available (no AD); traced:
AdContext::grad/vjp/jvpbuild a graph and take no engine
- concrete: not available (no AD); traced:
vjp
Reverse-mode vector-Jacobian product of eager tensors; touches no gradient slot.
- Options: none
- Output:
EagerTensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::vjp(&mut self, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor> - Dtypes: See the method documentation.
- Layout: Reads the retained values of the operands; the derivative program is prepared and run on the session’s engine. Every eager operand and seed is checked against the session’s runtime before any value is materialized or any gradient slot is written.
- Errors:
ContextMismatch,Validation,UnsupportedAdRule - Omissions:
- concrete: not available (no AD); traced:
AdContext::grad/vjp/jvpbuild a graph and take no engine
- concrete: not available (no AD); traced:
jvp
Forward-mode Jacobian-vector product of eager tensors.
- Options: none
- Output:
EagerTensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::jvp(&mut self, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor> - Dtypes: See the method documentation.
- Layout: Reads the retained values of the operands; the derivative program is prepared and run on the session’s engine. Every eager operand and seed is checked against the session’s runtime before any value is materialized or any gradient slot is written.
- Errors:
ContextMismatch,Validation,UnsupportedAdRule - Omissions:
- concrete: not available (no AD); traced:
AdContext::grad/vjp/jvpbuild a graph and take no engine
- concrete: not available (no AD); traced:
backward
Reverse-mode AD from a scalar eager output: returns the cotangent map and accumulates into the grad() slots of the reachable tracked leaves.
- Options: none
- Output:
Gradients - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::backward(&mut self, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<tenferro_ad::Gradients> - Dtypes: See the method documentation.
- Layout: Reads the retained values of the operands; the derivative program is prepared and run on the session’s engine. Every eager operand and seed is checked against the session’s runtime before any value is materialized or any gradient slot is written.
- Errors:
ContextMismatch,NonScalarGrad,Validation,RuntimeState,UnsupportedAdRule - Omissions:
- concrete and traced: not available; traced code differentiates with
AdContext::gradand has no gradient slots
- concrete and traced: not available; traced code differentiates with
detach
Detach an eager tensor from the reverse graph: same value and runtime, no gradient flow.
- Options: none
- Output:
EagerTensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::detach(&mut self, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor> - Dtypes: See the method documentation.
- Layout: Copies the retained value once through the session’s engine (
to_contiguous_read), at the same placement and dtype. - Errors:
ContextMismatch,TensorRuntime,RuntimeState - Omissions:
- concrete and traced: not available; cross-runtime import is
duplicate_valuein the source session plusconstant_fromin the target session
- concrete and traced: not available; cross-runtime import is
duplicate_value
Copy an eager value into an independent owned Tensor (the one owned-copy spelling; EagerTensor::into_value consumes without copying).
- Options: none
- Output:
Tensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::duplicate_value(&mut self, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<tenferro_tensor::types::Tensor> - Dtypes: See the method documentation.
- Layout: Copies the retained value once through the session’s engine (
to_contiguous_read), at the same placement and dtype. - Errors:
ContextMismatch,TensorRuntime,RuntimeState - Omissions:
- concrete and traced: not available
constant_from
Import a tensor the engine can read as an untracked eager leaf.
- Options: none
- Output:
EagerTensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::constant_from(&mut self, tenferro_tensor::types::Tensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor> - Dtypes: See the method documentation.
- Layout: The input is read through the session’s engine, so a device-resident or strided tensor the engine can read becomes a leaf without a host round trip.
- Errors:
TensorRuntime,RuntimeState - Omissions:
- concrete and traced: not available;
EagerRuntime::constant_fromis the engine-free form for owned host tensors (E12)
- concrete and traced: not available;
constant_from_host
Upload a host tensor to the session’s backend and import it as an untracked eager leaf.
- Options: none
- Output:
EagerTensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::constant_from_host(&mut self, tenferro_tensor::types::Tensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor> - Dtypes: See the method documentation.
- Layout: Uploads the host tensor once through the session’s engine (
upload_host_tensor); a host copy on a CPU runtime. - Errors:
TensorRuntime,RuntimeState - Omissions:
- concrete and traced: not available
variable_from
Import a tensor the engine can read as a trainable eager leaf.
- Options: none
- Output:
EagerTensor - Signature (eager):
pub fn tenferro_ad::EagerSession<'_>::variable_from(&mut self, tenferro_tensor::types::Tensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor> - Dtypes: See the method documentation.
- Layout: The input is read through the session’s engine, so a device-resident or strided tensor the engine can read becomes a leaf without a host round trip.
- Errors:
TensorRuntime,RuntimeState - Omissions:
- concrete and traced: not available;
EagerRuntime::variable_fromis the engine-free form for owned host tensors (E12)
- concrete and traced: not available;
Exceptions
| ID | Status | Symbols | Scope | Category | Owner | Reason | Verification | Exit condition |
|---|---|---|---|---|---|---|---|---|
| E11 | retained | pub fn tenferro_ad::EagerSession<'_>::grad_optional(&mut self, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<core::option::Option<tenferro_ad::EagerTensor>>pub fn tenferro_ad::EagerSession<'_>::vjp_optional(&mut self, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<core::option::Option<tenferro_ad::EagerTensor>>pub fn tenferro_ad::EagerSession<'_>::jvp_optional(&mut self, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<core::option::Option<tenferro_ad::EagerTensor>>pub fn tenferro_ad::EagerSession<'_>::backward_with(&mut self, &tenferro_ad::EagerTensor, &tenferro_ad::EagerTensor) -> tenferro_runtime::error::Result<tenferro_ad::Gradients> |
the inactive-input result form of the eager functional transforms, and the seeded form of eager backward | eager AD | result form, seed operand | tenferro_ad::eager |
_optional returns None for an inactive input instead of an error, an output mode the grammar does not otherwise spell as a suffix; backward_with(&output, &cotangent) takes the seed as an operand, not an options value, so it is not the _with options form (#2602) |
scripts/check-api-registry-conformance.py; tenferro-ad eager_session_ad_contract tests (inactive inputs, seeded backward, foreign seeds) |
kept while eager AD distinguishes inactive inputs and seeded backward |
| E12 | retained | pub fn tenferro_ad::EagerRuntime::constant_from(&alloc::sync::Arc<Self>, tenferro_tensor::types::Tensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor>pub fn tenferro_ad::EagerRuntime::variable_from(&alloc::sync::Arc<Self>, tenferro_tensor::types::Tensor) -> tenferro_runtime::error::Result<tenferro_ad::EagerTensor>pub fn tenferro_ad::EagerTensor::grad(&self) -> tenferro_runtime::error::Result<core::option::Option<tenferro_ad::GradientValue>>pub fn tenferro_ad::EagerTensor::into_value(self) -> core::result::Result<tenferro_tensor::types::Tensor, tenferro_ad::IntoValueError<Self>>pub fn tenferro_ad::Gradients::grad(&self, &computegraph::types::ValueKey<tenferro_ops::std_tensor_op::StdTensorOp>) -> core::option::Option<tenferro_tensor::types::TensorView<'_>>pub fn tenferro_ad::GradientValue::to_tensor(&self, &mut dyn tenferro_tensor::backend::Engine) -> tenferro_runtime::error::Result<tenferro_tensor::types::Tensor> |
engine-free eager leaf import, gradient-slot lookup, consuming value extraction, and the stored-gradient copy | eager value | engine-free, concrete record | tenferro_ad::eager |
the runtime constructors import an owned host tensor without an engine and reject device or view inputs (use the session constructors); grad() reads a slot and into_value moves the allocation out, so neither needs an engine; a GradientValue carries no eager-runtime identity and follows the concrete grammar recv.op(exec) (#2602) |
scripts/check-api-registry-conformance.py; tenferro-ad leaf-construction and into_value tests | kept |
| E13 | retained | pub fn tenferro_ad::EagerRuntime::with_eager_session<T, E: core::convert::From<tenferro_runtime::error::Error>>(&alloc::sync::Arc<Self>, &mut dyn tenferro_runtime::runtime::engine::ExecutionEngine, impl core::ops::function::FnOnce(&mut tenferro_ad::EagerSession<'_>) -> core::result::Result<T, E>) -> core::result::Result<T, E>pub fn tenferro_ad::EagerSession<'_>::engine(&mut self) -> &mut dyn tenferro_runtime::runtime::engine::ExecutionEnginepub fn tenferro_ad::extension::apply_eager(&mut dyn tenferro_runtime::runtime::engine::ExecutionEngine, alloc::sync::Arc<dyn tenferro_ops::ext_op::ExtensionOp>, &[&tenferro_ad::EagerTensor]) -> tenferro_runtime::error::Result<alloc::vec::Vec<tenferro_ad::EagerTensor>> |
the eager session boundary, its one engine accessor, and the extension-authoring entry | eager boundary | lifecycle, accessor, bridge | tenferro_ad::eager, tenferro_ad::extension |
with_eager_session is where the caller’s engine enters; engine() lends that engine as &mut dyn ExecutionEngine (which is also an Engine) for AD-free tensor work inside the boundary; apply_eager is the extension-authoring entry that opens its own session for an extension op |
scripts/check-api-registry-conformance.py; tenferro-ad session-boundary and extension tests | apply_eager moves to an EagerSession method when the extension-authoring facade is revised |