feat: add the adapted dataset from JaxBench

[shangkunwang01]

No description provided.

[NinaCai]

1. There are some ops with optimized code e.g. MLA, right? They are not included in this PR?
2. If you are able to get kernel_task.yaml file by using adapt script, would it make more sense to let the user run the script to generate the adapted benchmark? It is hard to keep track of original JaxBench and adapted JaxBench imo.

[shangkunwang01]

1. There are some ops with optimized code e.g. MLA, right? They are not included in this PR?
2. If you are able to get kernel_task.yaml file by using adapt script, would it make more sense to let the user run the script to generate the adapted benchmark? It is hard to keep track of original JaxBench and adapted JaxBench imo.

1. Optimized code is not adapted yet. I will adapt it recently.
2. The uploaded adapted code is verified by me and is consistent with current JaxBench. Adaptation by LLM is risky so I would suggest avoiding adapting the reference code and yaml file as much as possible. This adapted dataset should be the source of truth for our benchmark and any changes in JaxBench should be reflected in the adapted dataset and will be easy to do as we will only need to change single files manually.

[NinaCai]

1. There are some ops with optimized code e.g. MLA, right? They are not included in this PR?
2. If you are able to get kernel_task.yaml file by using adapt script, would it make more sense to let the user run the script to generate the adapted benchmark? It is hard to keep track of original JaxBench and adapted JaxBench imo.

1. Optimized code is not adapted yet. I will adapt it recently.
2. The uploaded adapted code is verified by me and is consistent with current JaxBench. Adaptation by LLM is risky so I would suggest avoiding adapting the reference code and yaml file as much as possible. This adapted dataset should be the source of truth for our benchmark and any changes in JaxBench should be reflected in the adapted dataset and will be easy to do as we will only need to change single files manually.

I think this folder and JaxBench folder are somewhat duplicated and potentially inconsistent. Let's discuss how to handle the adapted JaxBench next week. For now, let's use this adapted folder.

[shangkunwang01]

1. There are some ops with optimized code e.g. MLA, right? They are not included in this PR?
2. If you are able to get kernel_task.yaml file by using adapt script, would it make more sense to let the user run the script to generate the adapted benchmark? It is hard to keep track of original JaxBench and adapted JaxBench imo.

1. Optimized code is not adapted yet. I will adapt it recently.
2. The uploaded adapted code is verified by me and is consistent with current JaxBench. Adaptation by LLM is risky so I would suggest avoiding adapting the reference code and yaml file as much as possible. This adapted dataset should be the source of truth for our benchmark and any changes in JaxBench should be reflected in the adapted dataset and will be easy to do as we will only need to change single files manually.

I think this folder and JaxBench folder are somewhat duplicated and potentially inconsistent. Let's discuss how to handle the adapted JaxBench next week. For now, let's use this adapted folder.

The JaxBench folder may be removed in the future. It will has its own repo according to Steven. The main reason I want to have this folder in github is for us to easily run benchmark.