Release Notes

[0.9.0] — 2026-05-01

Added

• New layers: LRN and FakeQuant — Local Response Normalization (LRN(size, alpha, beta, bias)) for AlexNet-style models, and FakeQuant(scale, zero_point, qmin, qmax) for simulated quantization. Both are wired through lexer, parser, IR, shape inference, and codegen.
• Explicit per-side pool padding — MaxPool2D and AvgPool2D now accept a padding: [top, left, bottom, right] parameter for asymmetric padding, propagated through shape inference and codegen.

Changed

• ONNX import: quantized CNN support — nnc import now maps LRN ONNX nodes, fuses Gemm/MatMul → (Quantize/Dequantize) → Add bias chains, lowers QuantizeLinear nodes into FakeQuant layers, and recognizes asymmetric pads attributes for Conv, MaxPool, and AveragePool shape inference.
• ONNX import: tensor data decoding — initializers stored in int32_data / int64_data (instead of raw_data) are now decoded correctly, fixing imports for many torch-exported quantized models.

Fixed

• ONNX Reshape → Flatten lowering — Reshape to a [1, N] target shape is now lowered to Flatten even when the input rank is < 3, matching how PyTorch exporters serialize the post-conv flatten.
• ONNX DequantizeLinear with missing zero-point — empty zero-point initializers are now treated as zeros instead of failing to dequantize.

[0.8.0] — 2026-04-30

Added

• nnc new project scaffolding — generate a starter host-language project around a sample NNL model. Supports --project rust, go, cpp, and python. The scaffold includes a sample model.nnl (configured with io: "none"), host-language boilerplate wired to the generated C ABI, a build script or build file for compiling the model artifact, and a README with run instructions.

Changed

• Improved missing-weight diagnostics (E003) — nnc compile now produces a structured, actionable error when required weights are missing. Errors list every missing tensor with its expected shape, identify whether the source is a directory of .npy files, an .npz archive, or another path, and include a hint: to run nnc inspect <model> to view expected tensors and shapes. All missing weights are reported in a single error instead of stopping at the first one.

[0.7.0] — 2026-04-26

Added

• Compile-time memory check with optional memory_limit config — nnc now computes total static memory (weights + workspace) and emits a W003 warning when it exceeds 256 MB. Add memory_limit: "128MB" to the config block to turn this into a hard compile error (E009). nnc inspect now shows a “Total memory” line. Accepted units: KB, MB, GB.
• io: "none" config option — skips main() generation, producing a pure library artifact. Use with --emit lib, --emit shared, or --emit obj for embedding models in host applications. io: "none" with --emit exe produces a clear compile error.
• Integration examples — new examples/integration/ directory with documented examples showing how to call an NNL-compiled model from C++, Rust, Go, and Python, using static/shared library linking and FFI.

[0.6.0] — 2026-04-23

Added

• New layers: Hardswish, Upsample, Conv1D, MaxPool1D, LayerNorm — five new layer types across all pipeline stages (lexer, parser, IR, shape inference, codegen, ONNX import), completing the Tier 4 roadmap from the ONNX spec.
• Hardswish activation — Hardswish(x) = x * min(max(0, x+3), 6) / 6, unlocks MobileNetV3. ONNX HardSwish op imported automatically.
• Upsample layer — Upsample(scale: N) with nearest-neighbor interpolation for spatial upsampling. ONNX Upsample and Resize ops imported automatically. Unlocks YOLO-Tiny, U-Net, and encoder-decoder models.
• Conv1D layer — 1D convolution with filters, kernel, stride, padding parameters. ONNX Conv ops with 3D weight tensors auto-detected as Conv1D. Enables audio, time-series, and keyword spotting models.
• MaxPool1D layer — 1D max pooling with kernel and optional stride. ONNX MaxPool ops with 1D kernel_shape auto-detected. Enables audio and time-series models.
• LayerNorm layer — Layer normalization with learnable scale and bias over the last dimension, with configurable epsilon. ONNX LayerNormalization op imported with epsilon and weights. Enables transformer-adjacent models.

[0.5.0] — 2026-04-23

Added

• New layers: GlobalAvgPool2D, ReLU6, LeakyReLU, SiLU, Mul — six new layer types across all pipeline stages (lexer, parser, IR, shape inference, codegen, ONNX import), unlocking ResNet-18, MobileNetV1/V2, and EfficientNet model families.
• Grouped / depthwise Conv2D — Conv2D now accepts a groups parameter (default 1) for grouped convolution, including depthwise separable convolution (groups == in_channels). ONNX Conv group attribute is imported automatically.
• ONNX external tensor data support — nnc import can now load weights stored as external data files (ONNX data_location = EXTERNAL) with offset/length support, fixing import failures for models exported with torch.onnx.export(..., use_external_data_format=True).

Fixed

• CHW→HWC weight permutation at Flatten→Dense boundary — nnc import now automatically detects the Flatten→Gemm pattern in ONNX graphs and permutes Dense weight matrix rows from PyTorch’s CHW flatten order to nnc’s HWC order, fixing incorrect inference results for all imported CNNs with Flatten→Dense transitions.
• ONNX import empty tensor error — nnc import now produces a clear error message ("tensor '...' has no data") instead of a cryptic npy shape mismatch when tensor data is missing.

[0.4.0] — 2026-04-23

Added

• --version / -V flag — nnc --version now prints the version from Cargo.toml.
• --emit c flag — nnc compile model.nnl --emit c writes the generated .c and .h files directly without invoking the C compiler, useful for debugging and auditing generated code.

Fixed

• Concat codegen for multi-dimensional tensors — fixed incorrect flat memcpy in Concat codegen that produced wrong results when concatenating 3D (HWC) tensors along the channel axis. Now generates proper strided copies for arbitrary concat axes.
• ONNX import protobuf decode failure — fixed incorrect field tag numbers in AttributeProto that caused all ONNX imports to fail with a protobuf wire type error. Added missing floats field (tag 7).
• Unsupported precision silently accepted — precision: "int8" and precision: "float64" now produce a compile error instead of silently generating incorrect float32 code.
• Website hero demo — the output example now shows the realistic workflow (raw bytes piped through Python) instead of implying the binary outputs formatted text.
• Website copyright year — updated from © 2024 to © 2024–2025.
• README DESIGN.md link — corrected broken link to point to docs/src/DESIGN.md.

[0.3.0] — 2025-04-23

Fixed

• Conv2D rectangular kernel correctness — fixed a bug where non-square kernels (e.g., kernel: [3, 5]) produced incorrect inference results due to a variable shadowing issue in the generated C code. Square kernels were unaffected. The same shadowing fix was applied to MaxPool2D and AvgPool2D codegen for consistency.

[0.2.0] — 2025-04-20

Initial public release.

Added

• NNLang DSL with version 0.2 syntax for defining neural network models
• Layers: Input, Dense, Conv2D, MaxPool2D, AvgPool2D, Flatten, BatchNorm, Dropout, Add, Concat, ReLU, Sigmoid, Softmax
• C code generation backend with static memory allocation (no heap, no runtime dependencies)
• Output formats: exe, obj, lib, shared, header
• Cross-compilation via --target-triple flag
• SIMD target hints: generic, avx2, avx512, arm_neon
• Weight loading from .npy files and .npz archives
• ONNX model import via nnc import
• nnc inspect command for model summary and shape information
• nnc test command for verifying inference correctness against expected outputs
• Explicit graph connections with connections { } block and skip connections
• Liveness-based buffer reuse for minimal activation memory footprint
• mdbook documentation site