Release 2.5.0 validations checklist and cherry-picks

[atalman]

Similar to: #134694

We need to make sure that:

• Python 3.13 wheel validate @atalman
• CUDA
  • pypi binaries with slimmed dependencies are usable in standard AWS containers 2023 regression in 1.13. @kit1980 Verified on amazonlinux:2023:
    dnf update
    dnf -y install python3-pip
    pip3 install torch==2.5.0 --index-url https://download.pytorch.org/whl/test/cu124
  • Check cuda 1.12.1 update issue: torch.linalg.eigh fails on GPU #94772 with small wheels (This is an old issue from way back to 2.1, so I wonder if we need to continue verifying it) @atalman
• torch.compile
  • Basic test works (for example see test mentioned in Search for libdevice relative to shared library triton-lang/triton#1176 ) in PyTorch docker container @atalman
  • torch.compile raises an error if used on Windows. Test: https://github.com/pytorch/builder/actions/runs/11238619993/job/31243853359#step:9:469
  • torch.compile works on 3.11, 3.12 : Test: https://github.com/pytorch/builder/actions/runs/11236853245/job/31237993839#step:12:3709
• MPS
  • Resnet is usable out of the box (i.e. https://github.com/pytorch/builder/blob/main/test/smoke_test/smoke_test.py passes for MPS device). Test: https://github.com/pytorch/builder/actions/runs/11236853245/job/31238016293#step:9:437
• Validate docker release builds @atalman https://github.com/pytorch/builder/actions/runs/10084146628

Cherry-Picks to validate

• PyTorch 2.5.0 exposes statically linked libstdc++ CXX11 ABI symbols. #133437 @atalman
• [Cherry-pick][DSD] Fix distributed state dict full_state_dict option hang during set_state_dict (#135725) and Fix loading uneven full tensor into sharded state dict (#136365) #136903
• [ROCm] Update to AOTriton 0.7b (Cherry-picked) #135869 @jithunnair-amd
• Severe SDPA Performance Regression 2.5.0-RC1 #135778 @jithunnair-amd
• SDPA batching rules need randomness handling #135020
• The order of the parameters of nn.Conv1d(), nn.Conv2d() and nn.Conv3d() should be explained in the actual order of the parameters. #135880 @kit1980
• INTERNAL ASSERT FAILED at "../torch/csrc/autograd/python_torch_functions_manual.cpp":661 when returning a constant tensor in the forward method #132134
• /opt/rocm/lib/libamdhip64.so is hardcoded in Caffe2Targets.cmake in ROCm wheels #131701 @jithunnair-amd
• [ROCm] [BUGFIX] Re-enable rocm-specific tuning parameters v2 (#133852) #136139 @jithunnair-amd
• Add option to configure reduced precision math backend for SDPA #135964
• Clarify that libtorch API is C++17 compatible #136471 @kit1980
• Fix addmm silent correctness on aarch64 #136371
• test_c10d_nccl failure on A100 and H100 #136151 @atalman
• torch.ones(2,4,1,30,1).to('mps').sum(dim=-2) throws "buffer is not large enough" on mps #136132 @atalman
• [ONNX] Fix numpy method to return the correct type #136162
• MPS any() crashes on a Tensor with >4 dims #136134
• [FlexAttention] Fix output layout #135882
• [ONNX] Fix symbolic values and numpy implementation #135786
• [ONNX][high-pri] ONNX symbolic values are unhashable; numpy() implementation is incomplete #135700
• MPS any() crashes on a Tensor with >4 dims #136134
• [FlexAttention] Fix output layout #135882
• [ONNX] Fix symbolic values and numpy implementation #135786
• [ONNX][high-pri] ONNX symbolic values are unhashable; numpy() implementation is incomplete #135700
• [ONNX] Update fake mode usage in onnx docs #135512
• [MPS] Add missing dispatch to rshift.Tensor #135607
• [ONNX] Add assertion nodes to ignoring list #135591
• [ONNX] Verify handling of zero output ops #135419
• [ONNX] Fix scaled_dot_product_attention with float scale #135594
• [ONNX] Drop final None values as inputs for nodes in exporter graph #135520
• torch.fft.fftfreq behaves unexpectedly when run on MPS backend #135223
• [ONNX] Re-raise the exception if the dynamic shapes cannot be refined #135418
• [ONNX] Support FakeTensor in ONNXProgram #135399
• [ONNX] Handle mixed sequence inputs properly #135378
• [ONNX] Refactor exporter errors #135180
• remove _check call on item() for torch.istft #135234 @kit1980
• [ONNX] Enable experimental exporter logic to dynamo_export and support refine dynamic_shapes #134976
• [ONNX] Route dynamo_export with the new logic #135183
• [ONNX] Delete ONNXProgramSerializer #135261
• [ONNX] Remove ONNXProgramSerializer #135182
• RuntimeError: Cannot set version_counter for inference tensor when get_data_attr is called  #134798 @kit1980
• Clean up RemoteCache classes #134032 @kit1980
• [ONNX][DORT] Lazy-import onnxruntime #134662
• [MPS] pytorch nightly lost SDPA support? #134678
• [ONNX] Update exporter logic #134304
• Latest nightly pytorch installs benchmarks/ folder as a global python package #133377 @kit1980
• Always use high precision for SDPA math backend #128922

Versions

2.5.0