From 73335df7fc4cbe126e8805314cc3fba0be38e5bd Mon Sep 17 00:00:00 2001 From: soulitzer Date: Fri, 19 Sep 2025 17:24:22 -0400 Subject: [PATCH] Autograd release notes --- .../result_autograd_frontend.md | 20 +++++++++++-------- 2.9.0/miscategorized.md | 6 +++--- 2 files changed, 15 insertions(+), 11 deletions(-) rename 2.9.0/{todo => done}/result_autograd_frontend.md (84%) diff --git a/2.9.0/todo/result_autograd_frontend.md b/2.9.0/done/result_autograd_frontend.md similarity index 84% rename from 2.9.0/todo/result_autograd_frontend.md rename to 2.9.0/done/result_autograd_frontend.md index ec43cca..08e0c66 100644 --- a/2.9.0/todo/result_autograd_frontend.md +++ b/2.9.0/done/result_autograd_frontend.md @@ -28,20 +28,24 @@ The categories below are as follows: ### deprecation ### new features ### improvements +- Support deterministic `torch.nn.Upsample` `mode="trilinear"` backward ([#154239](https://github.com/pytorch/pytorch/pull/154239)) ### bug fixes -- Add ownership token when needed on GradientEdge ([#160098](https://github.com/pytorch/pytorch/pull/160098)) +- Fix `torch.autograd.Function` memory leak due to `torch.utils.checkpiont` early stopping ([#161171](https://github.com/pytorch/pytorch/pull/161171)) +- Fix `torch.autograd.graph.GradientEdge` for `torch.autograd.Function` ([#160098](https://github.com/pytorch/pytorch/pull/160098)) +- Match 0-dim gradients device type regardless of subclass-ness ([#160165](https://github.com/pytorch/pytorch/pull/160165)) + ### performance +- Fix SVD forward-mode AD multiplication priority ([#161027](https://github.com/pytorch/pytorch/pull/161027)) + ### docs +- Improve `torch.inference_mode` docs and error message ([#161164](https://github.com/pytorch/pytorch/pull/161164)) + ### devs ### Untopiced -- Add basic torch.hash_tensor op ([#154149](https://github.com/pytorch/pytorch/pull/154149)) -- [autograd] match 0-dim gradients device type regardless of subclassness ([#160165](https://github.com/pytorch/pytorch/pull/160165)) -- Fix SVD forward-mode AD multiplication priority ([#161027](https://github.com/pytorch/pytorch/pull/161027)) -- [BE] Improve torch.inference_mode docs and error message ([#161164](https://github.com/pytorch/pytorch/pull/161164)) -- Clear custom autograd Function ctx.to_save earlier ([#161171](https://github.com/pytorch/pytorch/pull/161171)) -- [while_loop][autograd] add hop while_loop_stack_output ([#160467](https://github.com/pytorch/pytorch/pull/160467)) -- Remove guard_size_oblivious from default contiguity python check, and add aten.sym_is_contiguous. [attempt2] ([#160869](https://github.com/pytorch/pytorch/pull/160869)) ### not user facing - [autograd] torch._C._set_view_replay_enabled state leaking into other tests ([#159840](https://github.com/pytorch/pytorch/pull/159840)) - Add new parameter for gen_pyi.py to make it more configureable. ([#161772](https://github.com/pytorch/pytorch/pull/161772)) +- Remove guard_size_oblivious from default contiguity python check, and add aten.sym_is_contiguous. [attempt2] ([#160869](https://github.com/pytorch/pytorch/pull/160869)) +- [while_loop][autograd] add hop while_loop_stack_output ([#160467](https://github.com/pytorch/pytorch/pull/160467)) + ### security diff --git a/2.9.0/miscategorized.md b/2.9.0/miscategorized.md index 0fe5ac4..d69b1f0 100644 --- a/2.9.0/miscategorized.md +++ b/2.9.0/miscategorized.md @@ -25,9 +25,6 @@ StableABI: - Enable generating generic c_shim that doesn't bypass dispatcher ([#158974](https://github.com/pytorch/pytorch/pull/158974)) - Cut a version of TORCH_ERROR_CODE_CHECK in headeronly from AOTI ([#159604](https://github.com/pytorch/pytorch/pull/159604)) -Autograd: -- Support deterministic upsample trilinear backward ([#154239](https://github.com/pytorch/pytorch/pull/154239)) - CUDA: - [CUDA] Fix missing `__syncthreads` in MultiMarginLoss backward ([#158994](https://github.com/pytorch/pytorch/pull/158994)) @@ -40,6 +37,9 @@ MPS: - Add `avg_pool3d` backward pass for MPS ([#159089](https://github.com/pytorch/pytorch/pull/159089)) - Enable _int_mm on Intel GPU ([#157769](https://github.com/pytorch/pytorch/pull/157769)) +Python Frontend: +- Add basic torch.hash_tensor op ([#154149](https://github.com/pytorch/pytorch/pull/154149)) + ## not user facing - Fix Pandas version mismatch upon reinstalling numpy ([#158584](https://github.com/pytorch/pytorch/pull/158584)) - [CUDA-13] Implement workaround for cudaErrorNotSupported ([#162412](https://github.com/pytorch/pytorch/pull/162412))