antares

1. Support Win10/11 for CUDA backend;
2. Add example: python3 -m autort.examples.06_diffuser_no_opt without dedicated device type optimization.
3. A bunch of Antares IR enhancements.

Preparing operators to support stable diffusion, turning Pytorch dependencies to 2.2.0

Add llama example for latency tests: python3 -m autort.examples.03_llama_tiny

Add Vulkan Backend for Python 3.12;

Bug fixes for python -m autort.examples.mnist in DirectX platform.

Experimental Release AutoRT for Windows DirectX 12 / Microsoft XBox / Linux CUDA.

A bunch of fixes included.

Installation Command for Linux:

python3 -m pip install --upgrade antares

A bunch of enhancement included.

Installation Command for Linux:

python3 -m pip install --upgrade antares

First Antares version ported to PYPI.

Installation Command for Linux:

python3 -m pip install --upgrade antares

New updates:

1. Detect Windows HLSL ending lines \r\n in c-hlsl_win64/c-hlsl_xbox backend;
2. Refine extra overhead in JIT plugin computation for pytorch;
3. Add IPU/IPU2 evaluator for c-ipu backend;
4. Enhance in evaluating c-sycl_cuda backend;

Thanks for contributions from mzmssg, Michoumichmich.

New updates:

1. HLSL support native erf & pow operators.
2. Extend new backend: c-mcpu_android for aarch64 CPU.
3. TF/Pytotch JIT Plugin: Support extending AVX512 & SYCL kernels.
4. Collective library: MPI Support for TF-Intel; NCCL/RCCL Support for TF-CUDA/TF-ROCM.

New updates:

1. Change of Auto-scheduling: Enhanced CPU(c-scpu/c-mcpu/c-mcpu_avx512) & IPU(renamed from c-gc to c-ipu) auto tuning search space.
2. Change of JIT Tuning: Pytorch/Tensorflow JIT Plugin is changed into local tuning by setting ANTARES_ROOT. (Tuning over rest-server is canceled)
3. Change of Installation: Allow non-root users to install Antares components.
4. Many other fixes.

New features:

1. Supporting More Backends: e.g. SYCL for CPU, ROCm for Windows, OCL for Android, etc.
2. Enhanced Tuning Mechanism OpEvo-2: Faster and Effective Tuner than Legacy Ansor.
3. AB Backend interface for all hardware (e.g. ab::init, ab::launchKernel, ..)
4. Enhanced Antares HLSL library using DXC-6.0, tuning efficiency is much improved.
5. Initial support for inter-op tuning (only for small graph in this version, large graph tuning will be supported in the following releases).

This version is frozen to keep legacy usage of Antares (intra-op optimizations, v0.1 api for directx12).