Especially Facebook's DLRM, which is said to occupy more than 65% of Facebook's machine learning pipelines. In the papers they listed in their github README, they discussed a lot of open problems: irregular cache access pattern on embedding tables, model co-locating, data/model parallel, etc.
I would be thrilled to see how OneFlow can solve these problems.
Especially Facebook's DLRM, which is said to occupy more than 65% of Facebook's machine learning pipelines. In the papers they listed in their github README, they discussed a lot of open problems: irregular cache access pattern on embedding tables, model co-locating, data/model parallel, etc.
I would be thrilled to see how OneFlow can solve these problems.