Yan Wang

[WIP] Add operator to support inplace updating of input tensor

3

Add a prim operator to support nvfuser `fd.add_output(output, alias_input)`, which allows in-place updating of the input tensor. Related context is we want to use it to update the running stats...

Support batch_norm backward and nvfuser (PR2278)

1

Support batch norm: - Add backward support - Add nvFuser support - Type promotion of input/weight/bias are handled in thunder - Running stats are left as-is, and let nvfuserex/torchex to...

Fix `sort_waits` to move `wait` closer to its consumer (#277)

11

Before submitting - [ ] Was this discussed/approved via a Github issue? (no need for typos and docs improvements) - [ ] Did you read the [contributor guideline](https://github.com/Lightning-AI/pytorch-lightning/blob/main/.github/CONTRIBUTING.md), Pull Request...

Add prim operators to query/update CUDA default RNG state

6

Background: Currently, the `mask` calculated for `dropout` in the forward trace is passed to the backward trace, which takes too much memory. We want Thunder to save the seed and...

## 🚀 Feature Add the benchmark for ResNet50 cc @crcrpar

Reducing the weight of NumberProxy used in mincut in rematerialization

3

Before submitting - [ ] Was this discussed/approved via a Github issue? (no need for typos and docs improvements) - [ ] Did you read the [contributor guideline](https://github.com/Lightning-AI/pytorch-lightning/blob/main/.github/CONTRIBUTING.md), Pull Request...

If saved_for_backward returns NumberProxy, the value is taken from compile time, not runtime

11

## 🐛 Bug When I add an operator that returns numbers, the values in the saved_for_backward are the compile time value defined in `Symbol.meta`, not the real value computed at...

uniform_like: The function outputs different values when the input tensor is the same but `requires_grad` is True/False.

2

## 🐛 Bug The same function outputs different values when the input tensor is the same but `requires_grad` is `True/False`. note: if change the last line in `func` to be...

Replacing uniform with uniform_philox , and incrementing PRNG state

2

Before submitting - [ ] Was this discussed/approved via a Github issue? (no need for typos and docs improvements) - [ ] Did you read the [contributor guideline](https://github.com/Lightning-AI/pytorch-lightning/blob/main/.github/CONTRIBUTING.md), Pull Request...

max_pool2d：Shape propagation error for input dimensions used by resnet

2

## 🐛 Bug ```py import torch import torch.nn.functional as F import thunder a = torch.randn(1, 64, 112, 112).cuda().requires_grad_() def func(a): return F.max_pool2d(a, 3, 2, 1, 1, False, False) # t79:...