release 0.9.4

What's Changed

• Enable jit load module and cpu_adagrad op by @jinyouzhi in #1
• remove dependence to builder_names by @delock in #4
• Change interface according to latest deepspeed upstream by @delock in #5
• Support oneAPI 2023.0 basekit by @jinyouzhi in #3
• mapping mpich env variable for multi-node training by @inkcherry in #6
• add empty InfereceBuilder for xpu inference by @dc3671 in #9
• add fsycl flag to support icpx for jit_load by @ys950902 in #7
• fix transformer.py for jit load UT pass by @ys950902 in #10
• update readme to support jit-load by @ys950902 in #11
• Add mapping rules for pmix launcher by @YizhouZ in #12
• update kernel for new compiler by @ys950902 in #14
• builder.py: fix python dict query code by @YizhouZ in #29
• Update the pypi package info by @rogerxfeng8 in #30
• add _vector_add kernel by @baodii in #32
• fix bugs: by @baodii in #33
• Update README.md by @rogerxfeng8 in #34

New Contributors

• @jinyouzhi made their first contribution in #1
• @delock made their first contribution in #4
• @inkcherry made their first contribution in #6
• @dc3671 made their first contribution in #9
• @ys950902 made their first contribution in #7
• @rogerxfeng8 made their first contribution in #30

Full Changelog: rel_2022.4...v0.9.4

rel_2022.4 release

Initial preproduction release of Intel Extension to DeepSpeed that supports GPT-2 3.6 billion parameters Large Language Model training, with below features verified on Intel GPUs:
• DeepSpeed ZeRO stage 2
• Model parallel size is 1