torchrec
torchrec copied to clipboard
pytorch
Pytorch domain library for recommendation systems
Summary: To support various types of eviction policy, the `HashZchManagedCollisionModule` needs to be able to calculate a score (e.g., TTL) for each incoming ID and pass it to the kernel....
Summary: Implement a new swapping API which does the following: 1. Takes an exported program and torchrec serializer 2. Constructs torchrec modules based on serialized metadata stored in the exported...
Summary: We parse torchrec logs from individual jobs (e.g. P1542961580). Our analyzer fails when Dense/KJT Storage is not available in the logs. This can be due to different types of...
Summary: Add field optimizer_key in Torchrec EmbeddingFusedOptimizer. During the initialization of embedding module BatchedFusedEmbeddingBag, pass the optimizer_key information from fused parameters when creating the EmbeddingFusedOptimizer. In sparse arch, update the...
Typo on line: [Typo here](https://github.com/pytorch/torchrec/blob/main/docs/source/index.rst?plain=1#L64)
Summary: Research doc: https://docs.google.com/document/d/1nDdQiJDnqJKzjzM3ku__Y5j196uxRVEB00Mj6qAl31k/edit Run ada model: https://www.internalfb.com/vanguard/serving_test_cases/487129480789691 We can see huge cpu time spend on cat, which is unnecessary for ada cases, we only cat one tensor, should be...
We can reproduce this problem using the following command: `torchrun --master_addr=127.0.0.1 --master_port=1234 --nnodes=1 --nproc-per-node=1 --node_rank=0 test_optimizer_state.py --sharding_type $SHARDING_TYPE`, and use the enviroment `torchrec==0.8.0+cu121, torch==2.4.0+cu121, fbgemm-gpu==0.8.0+cu121` when **SHARDING_TYPE=row_wise**, it will print...
Summary: The legacy inference solution had duplicate headers that was causing auto dep issues. All the inference_legacy references now only reference inference_legacy folder. Differential Revision: D62901035
