[XNNPACK][Partitioner] Migrate completely to new config based partitioner #4798

kirklandsign · 2024-08-20T18:12:03Z

Stack from ghstack (oldest at bottom):

New Config Based Partitioner maintains parity with old partitioners. I believe we can now replace all the old partitioners with the new partitioners

I also added pre configured partitioners like the ones most users use today, so that this would not cause any problems with people now. Despite having "multiple" partitioners It is the same partitioenr with different specified configs. The new xnnpack_partitioner.py file contains a lot of changes since it is essentially getting deleted and replaced with xnnpack_partitioner.py. But if you just look at the new file itself it is pretty straight forward.

https://www.internalfb.com/code/fbsource/[D61250577-V2]/fbcode/executorch/backends/xnnpack/partition/xnnpack_partitioner.py

Differential Revision: D61250577

Pull Request resolved: #4765

In order to maintain parity with the current to_edge and to_backend lowering flow, we need to support source based partitioning. We apply source-based partitioning to the AddMMConfig to partition all the nodes surrounding addmm so that it can be recomposed internally. While this is fine for to_edge and to_backend flow. For more robust flow, we will not have to use this when running to_edge_transform_and_lower. Differential Revision: [D61250576](https://our.internmc.facebook.com/intern/diff/D61250576/) [ghstack-poisoned]

Little bug got past partitioning, because 3d and transposed convolutions were now being partitioned. We expand the scope of the ConvConfig's check_constraint to also fail when the convolutions are either transposed or 3d Differential Revision: [D61368158](https://our.internmc.facebook.com/intern/diff/D61368158/) [ghstack-poisoned]

We add the SDPA Config here for partitioner. Currently there is an issue with SDPA when used from the FairSeq Multihead attention models, so I currently have it disabled for the base partitioner until we resolve that. Otherwise, for our tests, we can use the SDPA correctly from there. We have to track D60553559. Will follow up on this later. Differential Revision: [D60323285](https://our.internmc.facebook.com/intern/diff/D60323285/) [ghstack-poisoned]

…oner New Config Based Partitioner maintains parity with old partitioners. I believe we can now replace all the old partitioners with the new partitioners I also added pre configured partitioners like the ones most users use today, so that this would not cause any problems with people now. Despite having "multiple" partitioners It is the same partitioenr with different specified configs. The new xnnpack_partitioner.py file contains a lot of changes since it is essentially getting deleted and replaced with xnnpack_partitioner.py. But if you just look at the new file itself it is pretty straight forward. https://www.internalfb.com/code/fbsource/[D61250577-V2]/fbcode/executorch/backends/xnnpack/partition/xnnpack_partitioner.py Differential Revision: [D61250577](https://our.internmc.facebook.com/intern/diff/D61250577/) [ghstack-poisoned]

We add the SDPA Config here for partitioner. Currently there is an issue with SDPA when used from the FairSeq Multihead attention models, so I currently have it disabled for the base partitioner until we resolve that. Otherwise, for our tests, we can use the SDPA correctly from there. We have to track D60553559. Will follow up on this later. Differential Revision: [D60323285](https://our.internmc.facebook.com/intern/diff/D60323285/) [ghstack-poisoned]

…to new config based partitioner" New Config Based Partitioner maintains parity with old partitioners. I believe we can now replace all the old partitioners with the new partitioners I also added pre configured partitioners like the ones most users use today, so that this would not cause any problems with people now. Despite having "multiple" partitioners It is the same partitioenr with different specified configs. The new xnnpack_partitioner.py file contains a lot of changes since it is essentially getting deleted and replaced with xnnpack_partitioner.py. But if you just look at the new file itself it is pretty straight forward. https://www.internalfb.com/code/fbsource/[D61250577-V2]/fbcode/executorch/backends/xnnpack/partition/xnnpack_partitioner.py Differential Revision: [D61250577](https://our.internmc.facebook.com/intern/diff/D61250577/) [ghstack-poisoned]

…sed partitioner" New Config Based Partitioner maintains parity with old partitioners. I believe we can now replace all the old partitioners with the new partitioners I also added pre configured partitioners like the ones most users use today, so that this would not cause any problems with people now. Despite having "multiple" partitioners It is the same partitioenr with different specified configs. The new xnnpack_partitioner.py file contains a lot of changes since it is essentially getting deleted and replaced with xnnpack_partitioner.py. But if you just look at the new file itself it is pretty straight forward. https://www.internalfb.com/code/fbsource/[D61250577-V2]/fbcode/executorch/backends/xnnpack/partition/xnnpack_partitioner.py Differential Revision: [D61250577](https://our.internmc.facebook.com/intern/diff/D61250577/) [ghstack-poisoned]