Animesh Jain

[not for merge - just surface issue][inductor] Fix bug in group batch fusion pass

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #124501 * __->__ #125936 @jackiexu1992 Wondering if you could fix the test case after this change. I think pattern matcher should still work....

[DONT MERGE - JUST RFC] Use hierarchical compilation

1

Just to show how to use hierarchical compilation. On my machine, I see compilation finishing very quickly. Would love for someone from TorchTune to do two things * Measure the...

[DONT MERGE][dynamo] Turn on inlining of inbuilt nn modules

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #126304 * #126723 * #126578 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang

[dynamo][nn-modules] Trace through nn.Module.__init__

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #126304 * __->__ #126578 * #126468 Tracing through `__init__` is important because it initializes (calls STORE_ATTR) on members. By doing that, we kick...

[hg-bigbird-ci-failure] Change fail_accuracy to fail_to_run for export

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #126591 I don't really know what made hf_Bigbird pass earlier, but it should never have. The model uses `numpy` in between the...

[inductor][layout optimization] Extra recompilations because of a guard failure on weight

3

### 🐛 Describe the bug ``` TORCH_LOGS="recompiles" TORCHDYNAMO_INLINE_INBUILT_NN_MODULES=1 python benchmarks/dynamo/torchbench.py --ci --accuracy --timing --backend=inductor --device cuda --inference --bfloat16 --only=moco ``` The recompilation causing problem is ``` [rank0]:V0514 16:27:34.740000 140652138796864 torch/_dynamo/guards.py:2594]...

[dynamo][user-defined/nn_module] Inline __getattr__ and __setattr__

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #126304 * __->__ #126723 * #126578 cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501 @awgu @penguinwu @fegin @XilunWu @wanchaol @fduwjj...

[dynamo] Skip FUNCTION_MATCH on method-wrapper objects

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #118448 * __->__ #129830 Fixes https://github.com/pytorch/pytorch/issues/118563 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang

[dynamo] Validate check_fn

13

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #118448 * #129830 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @aakhundov

[dynamic-shapes][dynamo][recompilation] Bug in automatic dynamic when input is a view in one invocation and not in other

2

### 🐛 Describe the bug Seems like we need automatic dynamic to kick for strides. Or maybe skip strides when size is 1? ``` import torch @torch.compile(backend="eager") def fn(x, y):...