Rick Weyrauch

Remove duplicated permute_pooled_embs from sparsenn.

17

Summary: permute_pooled_embs ops are present in both fbgemm_gpu and sparenn. Remove from sparsenn. Differential Revision: D32113379

Add a warning message for CUB DIR and cuda build vs. non-cuda build in setup.py

2

Summary: Add option to build cpu-only (non-cuda). Reviewed By: jianyuh, rkindi Differential Revision: D31279295

Reduce run-times of fbgemm_gpu unit tests.

2

Summary: Attempt to reduce worst-case run-times of occasionally long running unit tests. Reviewed By: jianyuh Differential Revision: D31998929

Add Github Nightly Build GHA (#891)

24

Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/891 From D33798276: - ~~ https://github.com/pytorch/torchrec/runs/4948455778?check_suite_focus=true. ~~ - https://github.com/pytorch/FBGEMM/actions/runs/1767268123 - How to make this work: - Add a scale_config: https://fburl.com/code/wwke4j6c - Reach out to OSS team...

Refactor at, nbit, fbgemm and fbgemm_gpu namespaces used in fbgemm_gpu.

2

Summary: Move fbgemm_gpu code into fbgemm_gpu namespace. Normalize use of at namespace. First Diff to address task T106696656. Additional Diffs to follow that update the python api. Reviewed By: suphoff...

Port improved build script from torchrec to fbgemm_gpu.

10

Differential Revision: D33956565

Update OSS README build instructions.

2

Reviewed By: suphoff Differential Revision: D33821065

Speed up long running fbgemm_gpu:split_table_batched_embeddings_test.

2

Summary: Split long running tests into smaller tests. Reduced max iterations and data sizes. Reviewed By: jianyuh Differential Revision: D32735791

Simplify the CPU-only build for embedding_ops_cpu.

2

Summary: Use Aten::Context to check for CUDA and GPU availablility rather than rely on compile-time define. Reviewed By: jianyuh Differential Revision: D32708248

Hide LEGEND units from Matched Play roster selections.

2

Legend units are not valid in matched play rosters. Hide them and limit them to 'Open Play' modes only.