Internals

The source code is the source of truth for internal design. Development assumes AI agentic coding. AGENTS.md is the entry point.

Architecture

Design rationale for each subsystem — what each layer does and why.

• Crate structure
• Computation graph
• Primitive AD traits
• Semantic AD ownership
• End-to-end AD pipeline

Specification

Normative specs — trait signatures, op semantics, backend contracts.

• Primitive catalog
• Backend contract
• AD contract
• Optimizer passes
• Tensor semantics
• Public-boundary overhead inventory

The overhead inventory records source-path responsibilities, not measured CPU performance. CUDA/GPU, XLA, and multi-device execution each remain separately unmeasured follow-ups under #1758. CPU source inspection or future CPU timings do not establish their support, synchronization, transfer costs, or performance. Applicable device-specific evidence is required before the parent program’s final acceptance.

Source Pointers

Topic	Location
Op vocabulary	tenferro-internal-ops/src/std_tensor_op.rs
Backend contract	tenferro-tensor/src/backend.rs
Execution session	tenferro-tensor/src/backend.rs
AD rules	tenferro-internal-ops/src/ad/
Compilation pipeline	tenferro-runtime/src/compiler/
Buffer pool	tenferro-tensor/src/buffer_pool.rs
CPU context	tenferro-tensor/src/cpu/context.rs
GPU design	docs/design/gpu-backend-design.md
Public-boundary inventory	scripts/check-public-boundary-inventory.py, docs/internals/public-boundary-overhead-inventory.json, and benchmark export