Skip to content

Release v2.1.2

Latest

Choose a tag to compare

@github-actions github-actions released this 12 Jun 22:22
680550f

This release provides the standalone Model Compiler package for the SiMa.ai Neat SDK 2.1.2 workflow. It updates the compiler package set, adds ARM64 host support, and includes a post-install smoke-test runner for validating Model Compiler installations.

Use this release for Neat SDK 2.1.2 model compilation workflows. If you are working with the legacy Model SDK 2.0.0 or Neat SDK 2.0.0 workflow, use the Model Compiler v2.0.0 release instead.

Platform Support

This release supports both x86/amd64 and ARM64/aarch64 host installations:

  • tools/model-compiler/amd64 for x86/amd64 Ubuntu host and Neat SDK environments
  • tools/model-compiler/arm64 for ARM64/aarch64 Ubuntu host and Neat SDK environments

For host-side installation, this release is compatible with Ubuntu 24.04. The package set is built around Python 3.12.3 and Ubuntu host dependencies including LLVM 18. The ARM64 package set uses architecture-specific Model Compiler wheels and the ARM64 MLA toolchain bundle.

Installation

Install the Model Compiler package with sima-cli.

For x86/amd64 hosts:

sima-cli install -v 2.1.2 tools/model-compiler/amd64

For ARM64/aarch64 hosts:

sima-cli install -v 2.1.2 tools/model-compiler/arm64

Installing from inside the Neat SDK requires sima-cli 2.1.11 or later. When running sima-cli sdk setup for a compatible Neat SDK environment, sima-cli may prompt you to install the matching Model Compiler package.

Activate the Compiler Environment

After installation, start a new shell session or reload your shell configuration so the activation commands are available:

source ~/.bashrc

Then activate the Model Compiler environment before running compiler tools:

activate-model-compiler

To leave the compiler environment, run:

deactivate-model-compiler

What Changed Since v2.0.0

  • Updated the Model Compiler bundle version to 2.1.2 and moved the default Python runtime to Python 3.12.3.
  • Added ARM64/aarch64 bundle support with architecture-specific Python wheels and MLA toolchain artifacts.
  • Updated core compiler package versions, including sima-frontend, mpk_parser, sima-mppe, sima_lmm[sdk], and neural_compressor package inputs.
  • Added package and dependency handling for direct wheel files and direct wheel URLs when a wheel is not exposed through the configured Python index.
  • Updated dependency overrides for the 2.1.2 package set, including NumPy 1.26.4, Matplotlib 3.10.9, ONNX 1.17.0, ONNX Runtime 1.21.1, PyYAML 6.0.3, Torch 2.3.1, and TorchVision 0.18.1.
  • Added build support for selecting a target architecture with --target-arch x86_64 or --target-arch aarch64.
  • Improved bundle cleanup so generated wheel, archive, manifest, and metadata files are removed from the output directory before a new bundle is built.
  • Sanitized the downloaded MLA toolchain archive during bundle creation so only the expected toolchain payload is installed into the Model Compiler environment.

Post-Install Smoke Tests

This release adds scripts/smoke_test_modelsdk.py, a post-install smoke-test runner for Model Compiler environments.

After installing and activating the compiler environment, run the basic smoke test:

activate-model-compiler
python /path/to/model-compiler/scripts/smoke_test_modelsdk.py --tier basic

The basic tier checks that:

  • the active Python interpreter is the Model Compiler virtual environment
  • the Model Compiler bin/ directory is on PATH
  • core MLA tools such as mla-nm, mla-size, mla-readelf, and mla-isim are available
  • entry points such as afe-replay-compile, onnxsim, and llima-compile can launch
  • required Python modules such as afe, onnx, torch, sima_lmm, gguf, llama_cpp, and safetensors are importable

Longer validation tiers are available for pre-release or scheduled validation, including ResNet50 export, ResNet50 quantize/compile, YOLOv8 quantize/compile, and compiled-artifact checks.

Notes for v2.0.0 Users

Model Compiler v2.1.2 is the next standalone compiler package for Neat SDK 2.1.2 workflows. It keeps the same installation and activation model introduced in v2.0.0: install through sima-cli, reload your shell if needed, activate with activate-model-compiler, and deactivate with deactivate-model-compiler when finished.

Use v2.0.0 when you need functional compatibility with the legacy Model SDK 2.0.0 release. Use v2.1.2 for the updated Neat SDK 2.1.2 package set and ARM64 host support.