Ivan Kobzarev
Ivan Kobzarev
Summary: Dynamo has some gaps in support of generators, list comprehension etc. Avoiding them for now with syntactic changes Previous diff was reverted because recat was created on the target...
Summary: Adding dynamo coverage for KJT methods: - permute - split - regroup_as_dict - getitem - todict split and getitem tests need additional checks (similar to pre slice check). Extracted...
Summary: Clients requested to be able to call directly `LazyAwaitable._wait_async(a)` under compiled region. LazyAwaitable.wait_async uses `torch.fx.node.map_aggregate` which is skipped by dynamo => copied inside torchrec, changed immutable_list -> list, immutable_dict...
Summary: Fix missing torch.ops.load_library for the used op Differential Revision: D58026777
Summary: Start testing inductor backend for collective ops instead of just eager Differential Revision: D57558437
Summary: Support non-2d tensors in all_to_one. Sometimes torchrec outputs are empty tensors. This can be solved on python module side guarding empty tensors, but we d better not to have...
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #132755 Making wait_tensor effectful op. - Removing previous solution to mark having side_effect for fx, that prevented from DCE with corresponding tests....
Differential Revision: D59524843
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #977 Hotfix for the autoquant of vision transformer model. What happens in the model: ``` multi_head_attention_forward(.. ao_subclass .. ) { torch.nn.functional.linear(.. ao_subclass...