Check microsoft Llama-2-Onnx repo

[gyulaz-htec]

We got a request to check the https://github.com/microsoft/Llama-2-Onnx repository.
This requires a permission to access Llama 2 model, the details are descibed in the repo's readme.

There are two python examples which we should try with migraphx (more details about these in the repo's readme). In these examples we have to replace the onnx runtime with migraphx api calls.

[attila-dusnoki-htec]

To test it with MIGraphX we can update these two apps:

• https://github.com/microsoft/Llama-2-Onnx/tree/main/MinimumExample
  • Replate ORT with MGX
• https://github.com/microsoft/Llama-2-Onnx/tree/main/ChatApp
  • There is an inference part, which requires the same changes

[attila-dusnoki-htec]

Testing 7B_float32/ONNX/LlamaV2_7B_float32.onnx

Without and with input dims (@x 1 2048 4096 @k_cache 1 32 2048 32 128 @v_cache 1 32 2048 32 128 @pos 1 @attn_mask 1 2048 2048), it will run into
/code/AMDMIGraphX/src/onnx/parse_slice.cpp:155: construct_slice_desc: PARSE_SLICE: steps and variable starts and ends is not supported

[attila-dusnoki-htec]

WIth the latest develop, the above issue is bypassed, because the steps were default (1).

Now it will stop at only_dims(1).

Inputs.size = 3
inputs[0]   = [1, 2048, 1, 64]
inputs[1]   = [1, 1]
inputs[2]   = [1, 1]
axes        = [1]

[attila-dusnoki-htec]

If we would skip that check. It would run into this fail:
/code/AMDMIGraphX/src/common.cpp:83: operator(): COMPUTE_BROADCASTED_DYN_DIMS: dynamic shapes {[ 1, 1, {} ], [ 0, 2048, {} ], [ 1, 1, {} ], [ 64, 64, {} ]} and {[ 1, 1, {} ], [ 2048, 2048, {} ], [ 32, 32, {} ], [ 64, 64, {} ]} mismatch!