Hi everyone,
\nIt's the first time I step into the field of ML compilers. At first glance, it's too much information to digest, so I go through the amazing paper provided by the host, which clearly explains how the data is generated and arranged. I decide to record and share some of my understandings about the data here.
\nIn short, the hierarchy is defined as:
\noptimization---\nThat is, collections of data differ in terms of:
\nlayout: The tensor computational graph is the input graph to the layoyt assignment pass.tile: The fused subgraph (kernel) will goes through tile size selection.There are two sources from which the computational graphs are collected:
\nxla: The combination of the XLA regression benchmark.nlp: A variety of BERT for training and inference.The XLA autotuner is used to generate the configurations, and there are two strategies to explore the search space:
\ndefault: Explore the search space using a genetic algorithm starting from the default configuration.random: Pick random candidates in the search space.As usual, the dataset contains train, valid, and test sets.
With the information above, let's take a look at the data folder structure. Some observations are summarized as follows:
\ntile optimization contains graphs only from xla source.tile optimization (at the level of the fused subgraph)./kaggle//predict-ai-model-runtime/npz_all/npz\n├── layout\n│ ├── nlp\n│ │ ├── \n│ │ │ ├── test\n│ │ │ ├── train\n│ │ │ └── \n│ │ └── random\n│ │ ├── test\n│ │ ├── train\n│ │ └── \n│ └── xla\n│ ├── \n│ │ ├── test\n│ │ ├── train\n│ │ └── \n│ └── random\n│ ├── test\n│ ├── train\n│ └── \n└── tile\n └── xla\n ├── test\n ├── train\n └── \nFor detailed EDA, please see Google - Fast or Slow? - Detailed EDA.
\nHope this helps, thanks.
", "messages": [ { "id": "2421774", "postDate": "09/03/2023 14:12:02", "content": "Hi everyone,
\nIt's the first time I step into the field of ML compilers. At first glance, it's too much information to digest, so I go through the amazing paper provided by the host, which clearly explains how the data is generated and arranged. I decide to record and share some of my understandings about the data here.
\nIn short, the hierarchy is defined as:
\noptimization---\nThat is, collections of data differ in terms of:
\nlayout: The tensor computational graph is the input graph to the layoyt assignment pass.tile: The fused subgraph (kernel) will goes through tile size selection.There are two sources from which the computational graphs are collected:
\nxla: The combination of the XLA regression benchmark.nlp: A variety of BERT for training and inference.The XLA autotuner is used to generate the configurations, and there are two strategies to explore the search space:
\ndefault: Explore the search space using a genetic algorithm starting from the default configuration.random: Pick random candidates in the search space.As usual, the dataset contains train, valid, and test sets.
With the information above, let's take a look at the data folder structure. Some observations are summarized as follows:
\ntile optimization contains graphs only from xla source.tile optimization (at the level of the fused subgraph)./kaggle//predict-ai-model-runtime/npz_all/npz\n├── layout\n│ ├── nlp\n│ │ ├── \n│ │ │ ├── test\n│ │ │ ├── train\n│ │ │ └── \n│ │ └── random\n│ │ ├── test\n│ │ ├── train\n│ │ └── \n│ └── xla\n│ ├── \n│ │ ├── test\n│ │ ├── train\n│ │ └── \n│ └── random\n│ ├── test\n│ ├── train\n│ └── \n└── tile\n └── xla\n ├── test\n ├── train\n └── \nFor detailed EDA, please see Google - Fast or Slow? - Detailed EDA.
\nHope this helps, thanks.
", "rawMarkdown": "Hi everyone,\n\nIt's the first time I step into the field of ML compilers. At first glance, it's too much information to digest, so I go through the amazing [paper](https://arxiv.org/abs/2308.13490) provided by the host, which clearly explains how the data is generated and arranged. I decide to record and share some of my understandings about the data here.\n\nIn short, the hierarchy is defined as:\n```\noptimization-source-search-split\n``` \n\nThat is, collections of data differ in terms of:\n### 1. *optimization* - The compiler optimization\n* `layout`: The tensor computational graph is the input graph to the **layoyt assignment pass**.\n* `tile`: The fused subgraph (**kernel**) will goes through **tile size selection**.\n\n###2. *source* - The source of graphs (graph collection)Thank you for sharing your insight and the summary!
", "rawMarkdown": "Thank you for sharing your insight and the summary!", "votes": null }, { "id": "2478972", "postDate": "10/12/2023 10:04:05", "content": "Thanks you so much for well explain the file structure
", "rawMarkdown": "Thanks you so much for well explain the file structure", "votes": null } ], "comments": [ { "id": 2422147, "author_name": "mangpophothilimthana", "author_url": "", "post_date": "09/03/2023 18:24:07", "content": "Thank you for sharing your insight and the summary!
", "votes": null, "replies": [] }, { "id": 2478972, "author_name": "phnghiapro", "author_url": "", "post_date": "10/12/2023 10:04:05", "content": "Thanks you so much for well explain the file structure
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2421774": "Hi everyone,\n\nIt's the first time I step into the field of ML compilers. At first glance, it's too much information to digest, so I go through the amazing [paper](https://arxiv.org/abs/2308.13490) provided by the host, which clearly explains how the data is generated and arranged. I decide to record and share some of my understandings about the data here.\n\nIn short, the hierarchy is defined as:\n```\noptimization-source-search-split\n``` \n\nThat is, collections of data differ in terms of:\n### 1. *optimization* - The compiler optimization\n* `layout`: The tensor computational graph is the input graph to the **layoyt assignment pass**.\n* `tile`: The fused subgraph (**kernel**) will goes through **tile size selection**.\n\n###2. *source* - The source of graphs (graph collection)