Yanming W.

Add regression test for torch.expand

3

This PR adds test to capture the regression due to aten::expand's fallback to aten::as_strided reported in https://github.com/pytorch/pytorch/pull/82010.

[TF pin update] The same RunGraph (Worker) request was received twice

1

## 🐛 Bug This is to give a heads-up about next TF pin update. I found this issue when using a newer version of tensorflow to debug torch_xla GPU DDP...

[DETR support] Lower scipy.optimize.linear_sum_assignment

1

This op is needed to support DETR model. This op is used to compute bipartite matching and does not have pytorch implementation. We'll use this thread to track the progress....

[DETR support] Lower aten::_cdist_forward

2

This op is needed to support DETR model. We'll use this thread to track the progress. - Owner: @codeislife99

This PR fixed a corner case where the `fuse_elementwise` pass may produce two external output tensors with different shapes. This subgraph pattern is seen the huggingface bert model.

## Description ## Per-token text is not supported. This may needs some changes on vllm/lmi_dist side. Per-token cum_logprob is not supported, but should be easy to add.