IRPlanner

A tiny execution-plan compiler for AI runtimes.

IRPlanner is not a full TVM or XLA. It compiles a minimal .irplan IR description of LLM hot-path operators into an plan.json execution plan — defining shape inference, memory planning, kernel lowering, and budget propagation.

Power

Execution plan definition authority — whoever defines graph → tensor → shape → memory → kernel → plan controls what the runtime can execute.

IR Syntax (.irplan)

input x: f16[1, S, D]
w = param f16[D]

n = rmsnorm(x, w, eps=1e-5)
r = rope(n, base=10000, dim=64)
s = silu(r)
out = output(s)

Compilation Pipeline

.irplan source
  ↓ Parser → AST
  ↓ pass_shape → shape inference (symbolic dims)
  ↓ pass_budget → token/memory constraint propagation
  ↓ pass_memory → buffer reuse planning
  ↓ lowering → op → kernel mapping
  ↓ plan emit → plan.json

Quick Start

cargo run -- --input examples/hotpath.irplan --output plan.json

plan.json Output

{
  "version": "0.1.0",
  "ops": [
    {"id": "n", "op": "rmsnorm", "kernel": "cuda_rmsnorm_v1"},
    {"id": "r", "op": "rope", "kernel": "cuda_rope_v1"},
    {"id": "s", "op": "silu", "kernel": "cuda_silu_v1"}
  ],
  "memory_plan": {
    "peak_bytes": 33554432,
    "reuse_buffers": true
  }
}

Tech Stack

Layer	Choice
Language	Rust
Parser	Hand-written
Serialization	serde + serde_json
CLI	clap

Project Structure

apeinx-ir/
├── crates/apxir-core/src/     # AST, parser, IR, dtype, shape, passes
├── crates/apxir-cli/src/      # CLI entry point
├── examples/                  # .irplan source files
└── tests/

On MLIR

MLIR is a reusable, extensible compiler infrastructure. But don't touch it for v0.1 — the complexity will bury you. A hand-written parser and pass pipeline is the right start.

Relationship to Other Apeinx Projects

IRPlanner → (plan.json) → LLMSched → (trace.jsonl) → Trace2Train

License

TBD