HF Kernel for CPU: AMX, AVX2, AVX512 optimized

[jiqing-feng]

Add HF Kernel for CPU, can get significant speed-up on TTFT compared to torch_fused.

Requires review after kernels-community is ready.

[yao-matrix]

kernel PR merged huggingface/kernels-community#81, pls make it to ready for review once kernel binary propagated, @jiqing-feng

[jiqing-feng]

Hi @Qubitium . This PR is ready to be reviewed.

[Qubitium]

@jiqing-feng Awesome. We now are approaching a threshold where we have more mature CPU kernels than GPU ones thanks to Intel. =)

Please add hf_kernel to ci kernel test: https://github.com/ModelCloud/GPTQModel/blob/main/tests/test_kernel_output_torch_fused.py

And maybe change the test_kernel_output_torch_fused.py test name to test_kernel_output_intel_cpu_xpu.py since it now tests both xpu and cpu kernels for output regressions.

[Qubitium]

@jiqing-feng One thing. Please add kernels hard dependency to both pyproject.toml and requirements.txt. I checked the pkg dependency for https://github.com/huggingface/kernels/blob/main/pyproject.toml and kernel is very small and has very low dependency so i think it's safe for us to add this new hard dependency.

[jiqing-feng]

Hi @Qubitium . I have fixed your comments. Please verify it cause I cannot load the test model /monster/data/model/bloom-560m-gptqmodel-4bit

[Qubitium]

Hi @Qubitium . I have fixed your comments. Please verify it cause I cannot load the test model /monster/data/model/bloom-560m-gptqmodel-4bit

Ok. Thanks. I will run the unit test and merge after it passes.

[Qubitium]

CI tests passed

[jiqing-feng]

Hi @Qubitium . Please let me know what needs to be changed before merging. Thanks.