torchrec

torchrec copied to clipboard

Use permute_1D_input for non-VB inductor compilation

3

Differential Revision: D59524843

IvanKobzarev

add sharding_type argument to pipeline benchmark

4

Summary: # context * add sharding_type argument to the pipeline benchmark * better control of different sharding types Differential Revision: D64676132

TroyGarden

torchrec Build inference library and example server failure

2

I followed the steps in https://github.com/pytorch/torchrec/tree/main/torchrec/inference to test inference. But in 4. Build inference library and example server, the Build server and C++ protobufs failed. In particular, after I input...

Chevolier

GRID_SHARD in planner only if specified in constraints

1

Summary: For a minimally intrusive change that works so users don't unexpectedly get Grid Sharding, it must be specified in parameter constraints for the sharding option to be considered. Otherwise...

iamzainhuda

fix precommit lint

1

Summary: Precommit (https://github.com/pytorch/torchrec/actions/runs/11396841323/job/31711354638) is failing due to formatting issue Differential Revision: D64606855

sarckk

- For planner to use cpu for search random generator

2

Summary: torch.rand() defaults to using the default device. If torch.device has been globally set to 'meta', then this breaks the planner code. Force the device to cpu instead. This ensures...

damianr99

[Question] what is the difference between `ManagedCollisionEmbeddingCollection` and `ITEPEmbeddingBagCollection`

3

What is the difference between `ManagedCollisionEmbeddingCollection` and `ITEPEmbeddingBagCollection`, and when should I use `ManagedCollisionEmbeddingCollection` versus `ITEPEmbeddingBagCollection`?

tiankongdeguiji

Call .wait_tensor() in compiled region for dist.Work created in eager region

1

Summary: In compiled region, instead of calling `dist.Work.wait()`, we will call `torch.ops._c10d_functional.wait_tensor()` on the dist.Work's output tensor. This way, we can capture the `wait_tensor()` op within the torch.compile graph (instead...

Microve

Skip fsdp2 import if running with deploy

1

Summary: title, this breaks deploy models Differential Revision: D64237929

s4ayub

KJT permute - more efficient keys manipulation

1

Summary: Slightly optimizes the way KJT.permute handles keys and lengths - which could come in handy for KJTs with large number of keys (i.e. lots of features bundled into a...

che-sh