Skip to content

Release v2.1.3

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 21 Aug 00:40
v2.1.3
d76708a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Model Compiler 2.1.3 provides the compiler toolchain for Neat 2.1.3 workflows on both AMD64 and ARM64. This release improves installation and offline packaging, strengthens ARM64 and GenAI compilation support, and fixes differences between calibration preprocessing and deployed runtime behavior.

Installation

Install the latest sima-cli, authenticate, and run the command for your environment's architecture. The same commands work inside the Neat SDK and on a supported standalone Ubuntu host.

For ARM64 (aarch64):

sima-cli neat install model-compiler/arm64@v2.1.3

For AMD64 (x86_64):

sima-cli neat install model-compiler/amd64@v2.1.3

After installation, start a new shell or reload your shell configuration, then activate Model Compiler:

source ~/.bashrc
activate-model-compiler

Run deactivate-model-compiler to leave the compiler environment.

What's New

  • Added self-contained Model Compiler bundles for AMD64 and ARM64, with improved support for online and offline installation.
  • Improved ARM64 compilation by configuring the JAX backend for NEON and adding validated INT8 and BF16 ResNet50 compilation coverage.
  • Added a no-JAX activation fallback for environments where JAX cannot be initialized. Use activate-model-compiler --no-jax when needed.
  • Integrated and pinned the LLiMa compiler toolchain used by GenAI workflows, with architecture-aware dependencies and Qwen compilation validation.
  • Added a complete ResNet50 post-training quantization example, including model download, calibration data preparation, compilation, and output validation.
  • Updated the model compatibility guide with clearer constraints, searchable operator information, and support entries for LpNormalization and ReduceL2 in INT8 and BF16.

Fixes and Reliability Improvements

  • Matched calibration-time bilinear resizing to deployed runtime preprocessing, improving consistency between calibration and inference.
  • Corrected the AMD64 LLiMa data-type dependency selection.
  • Preserved unrelated XLA_FLAGS values when activating Model Compiler.
  • Restored the expected AMD64 compiler environment when switching activation modes.
  • Improved package validation and release checks across both supported architectures.

Compatibility

Use Model Compiler 2.1.3 with the corresponding Neat 2.1.3 workflow. Model Compiler 2.0.0 remains available for legacy Model SDK and Neat SDK 2.0.0 environments.

Full changelog: v2.1.2...v2.1.3

Assets 2
Loading