the fusions your compiler can't find

Noeris discovers and optimizes cross-operation kernel fusions that torch.compile and single-op libraries like Liger Kernel can't find. The research engine contains the full operator/search stack; the current public noeris.patch() API is a conservative drop-in surface for RMSNorm and gated MLP activation patches.

Latest layer artifacts: docs/results/gemma4-layer-bench-deeper-fusion-a100-after-geglu-retune.json, docs/results/gemma4-layer-bench-deeper-fusion-h100-after-geglu-retune.json. Canonical results index: docs/results/README.md.

import noeris
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("google/gemma-4-2b")
noeris.patch(model)  # Drop-in RMSNorm + gated MLP activation patches.
# QK-RMSNorm+RoPE kernels are available for custom integrations, but generic
# HuggingFace attention patching is not wired into noeris.patch() yet.

Contributor setup has separate CPU-only and Linux CUDA paths. Start with CONTRIBUTING.md if you are on macOS arm64, working without an NVIDIA GPU, or trying to run local CI parity.

noeris.patch() currently wires two module-level optimizations into supported HuggingFace-style models:

• RMSNorm module replacement, including Gemma's (1+w) affine mode.
• Gated MLP activation fusion for GeGLU/SwiGLU-style gate_proj, up_proj, down_proj blocks.

The project also includes lower-level Triton kernels for QK-RMSNorm+RoPE, cross-entropy, attention, and other operators. Those kernels back the benchmark artifacts below, but QK-RMSNorm+RoPE and cross-entropy are not generic noeris.patch() hooks yet.

python3 -m venv .venv && . .venv/bin/activate
pip install -e .

# Search a specific operator (A100 via Modal)
python -m research_engine.cli triton-iterate \
    --operator rmsnorm --gpu A100 --llm --configs-per-run 8

Run the same core checks used in GitHub CI from repo root on a Linux CUDA-capable development environment:

./scripts/ci_local.sh

If your default python3 does not have test dependencies installed, select one:

PYTHON_BIN=python3.11 ./scripts/ci_local.sh

This runs unit tests, public artifact reference checks, two matmul-speedup benchmark runs, history export, and the history regression gate with --fail-on-missing.

On macOS arm64 or CPU-only machines, use the source-tree setup in CONTRIBUTING.md instead of pip install -e .; Triton does not ship a native macOS arm64 wheel in the checked-in lockfile.

# Kaggle (30 hr/week free T4) or Google Colab
!git clone https://github.com/0sec-labs/noeris && cd noeris
!pip install -e . numpy scikit-learn -q
!python scripts/colab_validate_all.py

Noeris is complementary to Liger Kernel. Liger optimizes individual operations (RMSNorm, RoPE, SwiGLU separately). Noeris fuses operations across boundaries that single-op libraries can't cross -- like combining RMSNorm + RoPE into one kernel pass. Use both for maximum performance.

All 19 shapes pass correctness. Fusion speedup measured on T4 (Kaggle) and A100 (Modal).

Noeris currently registers 22 operator specs. Ten are wired into the generic triton-iterate / ablation CLI, eight run in the default GitHub Actions Triton Iterate matrix, and two are exposed as drop-in noeris.patch() hooks. The full surface split is documented in docs/system/OPERATOR_SURFACE.md.

110+ shape buckets. 606 unit tests. T4/A100/H100 validation status varies by operator and benchmark path; see docs/results/ for the current canonical artifacts.

• Thompson-sampling bandit + gradient-boosted cost model (R^2 = 0.94) + MAP-Elites quality-diversity
• Cross-run shape-indexed config database -- knowledge compounds across sessions
• Cross-hardware transfer -- A100-trained cost model rankings transfer to H100 with ρ=0.967
• Kernel-aware NAS proxy -- ranks architecture candidates against A100/T4/H100 kernel profiles
• ~$0.01 per iteration on Modal

torch.compile splits at the RMSNorm reduction boundary and materializes to HBM between passes. Noeris fuses the entire prologue into a single kernel launch.

• vLLM has an experimental enable_qk_norm_rope_fusion pass (disabled by default due to H100 regression, issue #34391). We make this fusion practical with parameterized Triton + bandit autotuning.
• Mirage (OSDI 2025) demonstrated fused norm+matmul. We implement it independently in Triton with autotuning.
• FlexAttention does block-sparse tile skipping. To our knowledge, we are the first to measure >3x wins on narrow-window shapes.
• Liger Kernel fuses RMSNorm and RoPE backward passes separately. To our knowledge, no existing framework fuses the combined QK-RMSNorm+RoPE backward into a single kernel.
• All novelty claims use "to our knowledge" qualification.

Paper draft: docs/paper/noeris.md. arXiv preprint coming.

@misc{noeris2026,
  title   = {Noeris: Architecture-Agnostic Kernel Fusion and Autotuning},
  author  = {Doruk Tan Ozturk},
  year    = {2026},
  url     = {https://github.com/0sec-labs/noeris}
}

MIT License.