Glossary
The words this repository’s documents use, each with the header or file that defines it. Header
paths are relative to include/SushiAI/.
The graph
| Word |
Meaning |
| IR |
SushiAI’s own graph: a single-assignment DAG of tensor operations, built on the host and touching no device. Namespace SushiAI::Graph (graph/ir.hpp) |
| Node |
One operation in the IR: an identity, its operands and its results (Graph::Node) |
| Value |
One single-assignment, tensor-shaped result in the IR (Graph::Value) |
| Op identity |
The name an operation carries, such as sushiai.log_softmax; all of them are declared in core/ops.hpp |
| Lifetime |
The class a value belongs to: PARAMETER, GRADIENT, OPTIMIZER_STATE, INPUT, ACTIVATION or SCRATCH. Only values of one class share an arena |
| Layout |
Where the channel axis sits. NHWC is the native layout; NCHW appears at a boundary only (Graph::TensorLayout) |
| Backward rule |
A function that writes the IR for one operation’s gradient. Autograd is an IR-to-IR transform (autograd/backward.hpp, src/autograd/rules/) |
| gradcheck |
The finite-difference check of a backward rule, run in FLOAT64 (autograd/gradcheck.hpp) |
From graph to device
| Word |
Meaning |
| Lowering |
Translating IR operations into recorded SushiBLAS work (Graph::Lowering, graph/lowering.hpp) |
| Compiled step |
A graph lowered, compiled once and re-armed for every batch (Graph::CompiledStep) |
| Compile once |
The guarantee that compile_count() stays 1 however many batches run; a regression test pins it |
| Replay |
Running the compiled step again for the next batch without tracing or compiling |
| Memory plan |
One placement per value plus the arena sizes, computed from value lifetimes before any allocation (Graph::MemoryPlan, graph/memory_plan.hpp) |
| Arena |
One allocation that holds several values whose live ranges do not overlap (Graph::Arena, Graph::ArenaSet) |
| Tensor pool |
The owner of the device buffers behind graph values across steps, including per-node scratch and shared all-ones constants (Graph::TensorPool) |
| Fusion catalog |
The closed list of fused kernels this build contains (Graph::FusedForm, graph/fusion.hpp) |
| Epilogue |
Work folded into the end of a GEMM: the bias add and, for some entries, the activation |
| Spill |
The second value a fused GELU node writes, holding the pre-activation its gradient reads |
| Rematerialisation |
The pass that recomputes a stored IM2COL result at its later consumers (graph/rematerialise.hpp) |
| Step scalars |
A small device-resident float vector the host rewrites each step, for numbers that change between replays (Optim::StepScalars) |
Training
| Word |
Meaning |
| Mixed precision |
Running products in fp16 while the weights the optimizer updates stay in fp32 (--precision mixed) |
| Master weights |
The fp32 copy of each parameter in a mixed-precision run |
| Precision policy |
The object that picks each operation’s element type (Graph::IPrecisionPolicy, graph/precision.hpp) |
| Cast insertion |
The pass that adds type conversions where a policy changes the element type (graph/cast_insertion.hpp) |
| Loss scale |
The factor the loss is multiplied by before the backward pass, so fp16 gradients do not underflow (--loss-scale) |
| Loss scaler |
The state machine that holds the scale, screens gradients for overflow and backs the scale off (Optim::LossScaler) |
| Objective |
What a model is trained to minimise (Train::IObjective) |
| Objective leaves |
The graph values an objective declared, by role (Train::ObjectiveLeaves) |
| Layer stack |
A sequential list of layers chosen at run time; it enumerates parameters as Sequential does (NN::LayerStack) |
| SACP |
The first four bytes of every checkpoint file, ASCII “SACP”, and the name of the format; version 1 (io/checkpoint_format.hpp) |
| Word |
Meaning |
| SushiStack |
The set of repositories: SushiRuntime, SushiBLAS, SushiEngine, SushiAI and the hub |
| Sibling |
A checkout that sits beside this one, such as ../sushiruntime or ../sushiblas |
| D3 |
The order a dependency is resolved in: an in-tree CMake target, then an installed package through find_package, then a sibling checkout |
sa |
This repository’s CLI, also installed as sushiai (cli/) |
hub |
The SushiStack workspace CLI, which installs module CLIs and provisions several checkouts |
| ST, SB, AI |
The three tracks of the roadmap: stack and CLI wiring, SushiBLAS repair, SushiAI construction. AI-0 to AI-7 are the steps of the third |
| Checker |
A script under tools/ that reads the repository and reports where it breaks a skill |
| Work folder |
docs/agent/<YYYY_MM_DD>_<WORK_NAME>/, holding the SPEC.md, PLAN.md and REPORT.md of one piece of agent work |
