tenferro-compute

Load this skill when the task is to use tenferro from downstream Rust code, not when changing tenferro itself.

Fast path

1. Choose the API tier: direct concrete tensors, eager tensors, or traced graphs. The same operation has different receivers and arities at each tier.
2. Add the direct crates for that tier. There is no root tenferro facade; import operation families such as tenferro-einsum and tenferro-linalg directly.
3. Bring the operation's public *Ext trait into scope. An E0599 saying that a method does not exist usually means the right extension trait is missing.
4. If an older example mentions a removed module/free function or a constructor signature that no longer compiles, read the API migration guide.
5. Read only the relevant reference below before writing the program.

Need	Read
Crates, features, CPU providers, scratch crates	crate selection
Tier arities and extension-trait imports	API cheatsheet
Backend/executor reuse and compile-once/run-many	performance idioms
Column-major data, einsum syntax, registration, and setup traps	pitfalls

Non-negotiable defaults

• Column-major storage. Dense buffers are column-major: the leftmost dimension varies fastest. Row-major data passed to from_vec_col_major is silently reinterpreted as column-major — permuted/wrong values, never rejected.
• No facade crate. cargo add tenferro fails by design; depend on the crates you need (tenferro-runtime, tenferro-cpu, and operation crates).
• Explicit execution owner. Concrete operations take a borrowed session inside backend.with_backend_session(...) (BackendSessionHost import). Construct the backend/runtime once and reuse it — per-call construction discards the buffer pool. Eager tensors retain their runtime instead.
• Representation is not reuse. Integer einsum labels still plan. For a repeated compatible equation/input count/dtype/shape, prepare a ConcreteEinsumPlan once, even from a string; see performance idioms. Do not flatten parenthesized contraction order into label arrays.
• Einsum dialect. Equations need the explicit arrow ("ij,jk->ik"). Flat notation supports one right-aligned, broadcastable ... ellipsis per term; EinsumNotation provides the programmatic form.
• Result-returning operators. Traced operators return Result; propagate with ?.
• CPU/GPU transfers are explicit; unsupported GPU operations do not silently fall back to CPU.
• Traced standard extensions need an explicitly installed extension module and a matching registered runtime engine.
• Keep a scratch crate in its own Cargo workspace (use an empty [workspace] table when it lives inside a checkout), and enable exactly one BLAS provider when using cpu-blas.

The executable Rust examples in the references are extracted from docs/tutorial-code/src/bin/tenferro_compute_skill.rs and compiled by the existing tutorial-binary test.