Changelog

All notable changes to AutoMegaKernel. Dates are when the work landed.

[0.4.0], 2026-06-28

Fused-instruction substrate (InstructionKind.FUSED): a generic megakernel instruction that executes a recipe of primitive ops over on-chip scratch (kernel fusion), so a fused computation never round-trips its intermediates through global buffers. The CPU reference (instructions/reference.ref_fused) interprets the recipe and is bit-identical to the unfused op sequence; a future slice code-gens it to a specialized CUDA device function (the GPU dispatch traps FUSED until then). ABI-synced (vm/abi.h AMK_OP_FUSED = 19). This is the substrate for agent-driven instruction synthesis: a generator fuses op subsequences, the validator gates each kept fusion deadlock/race-free, and the reference oracle then proves the final fused program numerically equivalent to the unfused one.

External proposer seam (proposer.py): an optional, public extension point on the propose -> eval -> keep loop. autoresearch(..., proposer=...) lets an external package supply the next (ScheduleConfig, kernel_knobs) candidate via the Proposer protocol (with a ProposalContext carrying the incumbent, the GpuTarget, the public knob space, the tried-set, and the RNG). A proposal only chooses a point; it still flows through the unchanged lower -> validate -> reference-oracle correctness -> measured keep/revert -> roofline floor gate, so it can never bypass the honesty guarantees. Default proposer=None preserves the built-in epsilon-greedy search byte-identically: no behavior change for existing callers.

Standard megakernel IR (schedule/ir.py): task-DAG, counter sync model, ScheduleConfig edit surface, GpuTarget data records, JSON format, and validate(), a real proof of both deadlock-freedom and race-freedom (shared-counter all-join rule, transitive happens-before provenance, per-SM queue order, on-chip locality, page-alias safety).
Frozen instruction ABI (vm/abi.h) with a drift guard (tests/test_abi_sync.py).
CPU reference VM (vm/reference_vm.py), bit-exact scheduling semantics + real numerics, the GPU-free correctness oracle; vm/verify_vm.py proves it.
CUDA megakernel VM (vm/): persistent cooperative kernel (one block per SM), counter sync, page scratchpad. Runs the full Llama-style decode as one kernel launch, matching eager + the reference to ~1e-7.
Instruction library (instructions/): 7 ABI-conformant CUDA micro-kernels (+ a Triton path), per-op verify, and an AutoKernel-style generation loop.
Scheduler (schedule/): HF/toy graph import, decode lowering, roofline cost model, Loop-2 schedule search.
Eval (eval/): correctness oracle (logit equivalence + token divergence), correctness- gated latency bench, roofline reporting, honest baseline stubs.
Product: compile.py (amk compile) + amk_cli.py + the flywheel log.
Docs: program.md (autonomous brain), docs/IR_SPEC.md (the standard IR spec).

HF importer (schedule/graph.from_hf): imports a real transformers.LlamaForCausalLM (MHA/GQA/tied-embeddings) bit-correct vs HF's own forward; rejects unsupported variants loudly.
bf16 path in the VM (matches the bf16 reference bit-identically).
Self-retargeting: the nvcc gencode is derived from the live device, the same code built and ran correct megakernels on sm_120 (RTX 5090), sm_80 (A100), sm_90 (H100) (measured via Modal; see DATACENTER_RESULTS.md).
Coding-agent harness (harness.py, amk propose/eval/loop, HARNESS.md, schemas/schedule_config.schema.json): AMK is drivable by any coding agent.
Flywheel corpus (flywheel/corpus.jsonl) seeded with real cross-arch measured points.

Two adversarial review passes (a 4-lens panel + Codex) hardened the validator; every finding is pinned by a regression in tests/test_validator_races.py.
Honesty enforced in code: no latency without a correctness PASS; roofline distance always reported; no unmeasured datacenter numbers.

Latency is currently far from the HBM roofline (correct-but-unoptimized). The lever, a bandwidth-efficient tensor-core/vectorized GEMV + coarser sync, is the active production push.