15. Reserved future extensions

Planned extensions and the constraints each must keep. This chapter is informative except where it reserves keywords or explicitly constrains future compatibility.

15.1 Runtime state and KV caches

Structured state (a struct of tensors), state carried by runtime loops, and state passed explicitly through entry signatures are open. Each must keep state flow explicit in the IR.

15.2 Explicit RNG

std.random implements a counter-based PRNG (Threefry-2x32) in Linnet, with Tensor[2; i64] keys that programs split and pass. rng stays reserved for a possible language-level construct.

15.3 Runtime loops and scan

A first-class scan that collects per-iteration outputs, beyond while (§8.4).

15.4 Custom gradients and training

Training works by differentiating exported entries in the target framework. Objectives are entries: module-level ones (§7.6), or a block's entries over packed sequences that return each position's loss or log-probability. std.nn.loss defines the losses as semantic ops (§7.3). A backend MAY compute linear_cross_entropy and linear_token_log_probs a block of rows at a time, with its own backward pass, so the [N, Vocab] logits are never whole. Low-rank adapters, gathering sharded parameters, and recomputing blocks in the backward pass are code-generation options and leave source semantics unchanged.

Possible additions are:

• autodiff as a compiler transform, so StableHLO and ONNX exports can carry gradients;
• @custom_vjp or equivalent annotations;
• gradient-specific semantic ops;
• optimizer-state structures.

Training support must not require model-specific Python classes.

15.5 Quantization

std.quant already stores integer weights beside their scales and dequantizes in source. Quantized dtypes or storage annotations in the type system are future work and must keep dequantization explicit.

15.6 Sparse and structured tensors

Type- or value-level metadata for sparse, diagonal/triangular, low-rank, or block-sparse tensors. It should not bake one backend's storage layout into semantic source.

15.7 User-defined traits

Float, Numeric, and related constraints may generalize to a small, safe trait system. It must not become arbitrary compile-time execution.

15.8 extern op

A trusted escape hatch for operations with externally supplied backend implementations (§13.5). They should be clearly non-portable and must never be required merely to add a model built from expressible tensor primitives.

15.9 Kernel language

A low-level kernel DSL for Triton-like tile programs, targeting the same or a lower IR and kept separate from model-level .linnet semantics.

15.10 Distributed execution

Today a Shards generic sizes each process's weights for tensor parallelism, and std.nn.parallel::all_reduce and all_gather mark where partial results are summed or joined; the program with Shards = 1 is the reference. Fully sharded data parallelism needs nothing in source. A code generator splits a block's parameters across processes or devices by block path and gathers each where the block runs; the program computes what it computes unsplit. Further collectives and placement or sharding annotations in the type system are future work.

15.11 Importers

Importers for formats beyond torch.export, JAX through StableHLO, and ONNX. Importers are frontends and do not redefine Linnet source semantics.