Compiler & On-Device Runtime Research Packet (E1 NPU)

Captured: 2026-05-19 Scope: Research-only inventory and analysis of public compiler stacks, mobile runtimes, quantization toolchains, and tensor-IR dialects that could feed a production compiler backend for the E1 NPU. This packet covers fifteen surfaces: MLIR/CIRCT, StableHLO/OpenXLA, IREE, TVM/Relax, PyTorch 2.x compile + ExecuTorch, TFLite/LiteRT, ONNX Runtime, Triton/GPU JIT, Android AICore/NNAPI successor, quantization toolchains (AIMET/Brevitas/GPTQ/AWQ/SpinQuant/etc.), tensor/graph IRs (LinAlg, Stream, Transform, Halide, BUDA), throughput-oriented runtimes (vLLM/SGLang/MAX/llama.cpp/MLX), compiler papers, attention-lowering practice, and mobile NN abstractions (Core ML, QNN, NeuroPilot, ExecuTorch backends).

Claim boundary: Every item below is a public artifact (arXiv paper, vendor documentation page, open-source repository, conference paper, RFC, or release blog). Nothing in this packet certifies that E1's compiler stack implements any of these techniques. E1 claims require local RTL, simulator, software, PD, and evidence gates per docs/benchmarks/capabilities/README.md. Anchors to E1 artifacts in 03_implementation/e1_compiler_path.md are gap-and-step descriptions, not implementation claims.

Contents

• 01_sources/source_inventory.yaml — fifty-plus public sources across the fifteen target surfaces, mirroring the schema used by research/ai_accelerator_sota/01_sources/source_inventory.yaml.
• 02_analysis/mlir_stack_landscape.md — MLIR/StableHLO/IREE/TVM/PyTorch compile landscape as of mid-2026, including current production users and open NPU-relevant gaps.
• 02_analysis/mobile_runtime_landscape.md — TFLite-to-LiteRT, NNAPI deprecation and AICore/AI Edge migration path, ExecuTorch backends, ONNX Runtime Mobile/Web, MLX, MediaPipe.
• 02_analysis/quantization_toolchains.md — INT8/INT4/INT2/FP8/FP4 software toolchains, their operator assumptions, and which precisions land in E1's current opcode set vs. which are software-only.
• 02_analysis/attention_lowering.md — how attention is lowered in 2026 stacks (FlashAttention-3, paged attention, KV cache layouts, RoPE fusion, GEMV decode kernels) and what E1's attention_qk/attention_av smoke paths would need to absorb.
• 03_implementation/e1_compiler_path.md — ranked High/Med/Low steps to take compiler/runtime/e1_npu_lowering.py from a Python smoke harness toward a real StableHLO/TFLite/ExecuTorch NPU backend, each tied to a docs/spec-db file and a current gap.

Source policy

Primary sources are arXiv, MLSys/ASPLOS/PLDI/OSDI/ATC/ICLR/NeurIPS papers, vendor documentation, and project repositories. Every YAML entry carries url, title, year, and relevance; repositories also carry last_release_or_commit_year. Items prefixed repo_ are GitHub or equivalent source forges. Items prefixed paper_ are venue or arXiv papers. Items prefixed vendor_ are vendor documentation pages. Items prefixed spec_ are RFCs or schema specifications.

How to use this packet

Read 02_analysis/mlir_stack_landscape.md first for the upstream IR picture, then mobile_runtime_landscape.md for the deployment surface E1 will live in. Use attention_lowering.md when extending compiler/runtime/e1_npu_lowering.py's attention smoke paths. Use quantization_toolchains.md when extending the INT4/INT2/FP8 opcode coverage documented in docs/arch/npu.md. Use 03_implementation/e1_compiler_path.md to slot incremental compiler work into the docs/spec-db/npu-2028-target.yaml software targets.