# PyTorch 1.9 dev release notes

**URL:** https://dev-discuss.pytorch.org/t/pytorch-1-9-dev-release-notes/248
**Category:** release/packaging
**Created:** [June 17, 2021, 3:49pm UTC](https://dev-discuss.pytorch.org/t/pytorch-1-9-dev-release-notes/248 "2021-06-17T15:49:28Z")
**Posts on this page:** 1
**Page:** 1

<div class="post-metadata">

### Author: ![anjali411](https://yyz2.discourse-cdn.com/flex036/user_avatar/dev-discuss.pytorch.org/anjali411/32/99_2.png) [@anjali411](https://dev-discuss.pytorch.org/u/anjali411)
#### Post date: [June 17, 2021, 3:49pm UTC](https://dev-discuss.pytorch.org/t/pytorch-1-9-dev-release-notes/248/1 "2021-06-17T15:49:28Z")

</div>

PyTorch 1.9 release contains quite a few commits that are not user facing but are interesting to people compiling from source or developing low level extensions for PyTorch. Here is a non-exhaustive list of the most important ones.

### Python API

- Added `cpu_kernel_multiple_outputs` to help developers implement new torch functions that return two or more tensors conveniently ([#51097](https://github.com/pytorch/pytorch/pull/51097))
- Support auto generation of device check ([#56872](https://github.com/pytorch/pytorch/pull/56872))
- Fix bug in self.assertExpectedInline ([#55149](https://github.com/pytorch/pytorch/pull/55149))
- protect destructors of python bindings that can be kept alive by c++ objects ([#57488](https://github.com/pytorch/pytorch/pull/57488))
- **Testing related commits:**
  - See the Developer Wiki article [“Writing tests in PyTorch 1.9”](http://see%20the%20developer%20wiki%20article/) for details on significant testing improvements
    - A prototype torch.testing module (see it documentation [here](https://pytorch.org/docs/master/testing.html?highlight=testing#module-torch.testing)) has been added to facilitate testing libraries built using PyTorch
      - It currently has one function, `torch.testing.assert_close`, which can be useful when comparing PyTorch tensors (add your feedback to its RFC [here](https://github.com/pytorch/pytorch/issues/56544)!)
      - Request more features by filing issues on PyTorch’s Github

  - OpInfo coverage and testing continues to expand!
    - test\_ops.py now verifies the out= argument works correctly for operators with OpInfos ([#53259](https://github.com/pytorch/pytorch/pull/53259))
    - OpInfos now support sample inputs with tensorlist arguments ([#54922](https://github.com/pytorch/pytorch/pull/54922))
    - OpInfos can now be wrapped in a lambda for grad and gradgrad checks ([#54914](https://github.com/pytorch/pytorch/pull/54914))
    - OpInfos can now handle sample inputs where the input is broadcast ([#55771](https://github.com/pytorch/pytorch/pull/55771))

  - [ROCm] Setting TEST\_WITH\_ROCM now skips tests that don’t use GPUs ([#55069](https://github.com/pytorch/pytorch/pull/55069))
  - A new test case method, `assertWarnsOnceRegex`, can be used to test warnings that are usually thrown only once per process ([#52387](https://github.com/pytorch/pytorch/pull/52387))
  - `make_tensor()`, the test suite’s goto mechanism for constructing a random tensor, now supports a `discontiguous` kwarg ([#51985](https://github.com/pytorch/pytorch/pull/51985))

### Distributed

- `torch.distributed.rpc`: Adds a parameter server benchmark for RPC to torch/benchmarks/distributed. ([#57454](https://github.com/pytorch/pytorch/pull/57454))
- `torch.distributed.nn.RemoteModule`: Improve typing for RemoteModule ([#58012](https://github.com/pytorch/pytorch/pull/58012))
- `torch.distributed.rpc`: Assert that GIL is not held in blocking destructors in RPC ([#57030](https://github.com/pytorch/pytorch/pull/57030))
- Add logging when `store_based_barrier` succeeds ([#57711](https://github.com/pytorch/pytorch/pull/57711))
- `torch.distributed.rpc: `Allow to specify a set of device for CUDAFuture ([#56515](https://github.com/pytorch/pytorch/pull/56515))
- `torch.distributed.nn.RemoteModule`: Replace Python Pickler with internal RPC pickler for RemoteModule ([#58019](https://github.com/pytorch/pytorch/pull/58019))
- `torch.distributed.rpc`: Make CUDAFuture handle any kind of device type ([#57051](https://github.com/pytorch/pytorch/pull/57051))
- `torch.distributed`: Remove deprecated use of torch.LongTensor, torch.ByteTensor in distributed APIs ([#55861](https://github.com/pytorch/pytorch/pull/55861))
- `torch.distributed`: Join work clean up thread before aborting communicators ([#55444](https://github.com/pytorch/pytorch/pull/55444))
- `DistributedDataParallel`: Log use of uneven inputs API ([#54919](https://github.com/pytorch/pytorch/pull/54919))
- `DistributedDataParallel`: Deduplicate shared params before constructing Reducer in DDP ([#53279](https://github.com/pytorch/pytorch/pull/53279))
- `torch.distributed`: Log nccl\_async\_error\_handling ([#52965](https://github.com/pytorch/pytorch/pull/52965))
- `torch.distributed.rpc`: Reduce logging verbosity in tensorpipe agent ([#51784](https://github.com/pytorch/pytorch/pull/51784), [#51785](https://github.com/pytorch/pytorch/pull/51785))
- `torch.distributed`: Log nccl debug level in ProcessGroupNCCL ([#52803](https://github.com/pytorch/pytorch/pull/52803))
- `torch.distributed.rpc`: make pickler/unpickler pluggable in RPC ([#53050](https://github.com/pytorch/pytorch/pull/53050))
- `torch.distributed`: make the pickler in distributed\_c10d pluggable ([#53060](https://github.com/pytorch/pytorch/pull/53060))
- `DistributedDataParallel`: log newly added construction and runtime stats at randomly selected iterations ([#51394](https://github.com/pytorch/pytorch/pull/51394))
- `torch.distributed.rpc`: Fix flaky TestTrainingLoop - TestE2ETensorPipe ([#51939](https://github.com/pytorch/pytorch/pull/51939))
- `DistributedDataParallel`: Ensure local\_used\_maps\_tmp is distinct from local\_used\_maps\_[i] ([#54474](https://github.com/pytorch/pytorch/pull/54474))
- `DistributedDataParallel`: Declare NamedTuple at top level to fix typing ([#53273](https://github.com/pytorch/pytorch/pull/53273))
- `torch.distributed` Combine backtrace print in test logging into one string to avoid interleaving ([#56961](https://github.com/pytorch/pytorch/pull/56961)).

### torch.nn

- Reenable `test_nn` tests for Windows ([#52051](https://github.com/pytorch/pytorch/pull/52051))
- Replace `type().backend()` with `device()` ([#52558](https://github.com/pytorch/pytorch/pull/52558))
- Remove annoying warnings from `common_nn.py` ([#55982](https://github.com/pytorch/pytorch/pull/55982))
- Fix ` __torch_function__ ` tests. ([#54492](https://github.com/pytorch/pytorch/pull/54492))
- Fixes new tf32 failures in `test_nn.py` ([#52871](https://github.com/pytorch/pytorch/pull/52871))
- Enable test cases in `test_nn.py` for ROCm ([#52836](https://github.com/pytorch/pytorch/pull/52836))
- Fix compiler warnings from conv.h ([#56181](https://github.com/pytorch/pytorch/pull/56181))
- Update upsample tests in `test_nn.py` to test for memory\_format ([#53665](https://github.com/pytorch/pytorch/pull/53665))
- Lowering NLLLoss/CrossEntropyLoss to ATen code ([#53789](https://github.com/pytorch/pytorch/pull/53789))
- Refactor `multi_head_attention_forward` ([#56674](https://github.com/pytorch/pytorch/pull/56674))
- Convert type annotations in `nn/functional.py` to py3 syntax ([#53656](https://github.com/pytorch/pytorch/pull/53656))
- Migrates some of `test_nn.py` from `assertEqualIgnoreTypes` to `assertEqual` ([#57642](https://github.com/pytorch/pytorch/pull/57642))
- Removes unused `RReLU` code ([#57672](https://github.com/pytorch/pytorch/pull/57672))
- Disable `TestComplexity.test_nn_module_test` in `fbcode` ([#56677](https://github.com/pytorch/pytorch/pull/56677))
- Make `convolution_overrideable` default implementation raise NotImplementedError ([#54707](https://github.com/pytorch/pytorch/pull/54707))
- Remove `ddp_gpu_size` field from SyncBatchNorm ([#55946](https://github.com/pytorch/pytorch/pull/55946))
- Remove `_specify_ddp_gpu_num` method from SyncBatchNorm ([#56425](https://github.com/pytorch/pytorch/pull/56425))
- Check exception messages in embedding\_bag\_proxy unit test (5a1191d050)
- Remove legacy constructor calls from `_torch_` folder. ([#53889](https://github.com/pytorch/pytorch/pull/53889))

### C++ Frontend

- Add NoOpDeviceGuardImpl ([#53142](https://github.com/pytorch/pytorch/pull/53142))
- Lower ReLu6 to aten ([#52723](https://github.com/pytorch/pytorch/pull/52723))
- Prevent VS from emitting ambiguous symbol errors ([#53490](https://github.com/pytorch/pytorch/pull/53490))
- Devirtualize TensorImpl is\_contiguous ([#55333](https://github.com/pytorch/pytorch/pull/55333))
- Update expand\_size API to match expand\_inplace ([#55246](https://github.com/pytorch/pytorch/pull/55246))
- Put llvmMathExtras in c10 namespace ([#55886](https://github.com/pytorch/pytorch/pull/55886))
- Move `flatten_dense_tensors` and `unflatten_dense_tensors` to Native ([#58006](https://github.com/pytorch/pytorch/pull/58006))

### Autograd

- Forward AD: Added systematic testing via gradcheck and OpInfos ([#57633](https://github.com/pytorch/pytorch/pull/57633), [#57701](https://github.com/pytorch/pytorch/pull/57701))
- `torch.autograd.gradcheck`: `fast_mode` is now enabled by default for tests ([#55699](https://github.com/pytorch/pytorch/pull/55699), [#55237](https://github.com/pytorch/pytorch/pull/55237))
- Update autograd kernels and tracing codegen to use redispatch API ([#51363](https://github.com/pytorch/pytorch/pull/51363), [#52009](https://github.com/pytorch/pytorch/pull/52009))
- Move view handling logic to gen\_inplace\_or\_view\_type.py ([#53341](https://github.com/pytorch/pytorch/pull/53341))
- Add getters for attributes on autograd Node([#55225](https://github.com/pytorch/pytorch/pull/55225), [#53205](https://github.com/pytorch/pytorch/pull/53205), [#56499](https://github.com/pytorch/pytorch/pull/56499), [#52451](https://github.com/pytorch/pytorch/pull/52451))
- Eliminate global usage of torch.set\_default\_dtype in test\_autograd ([#56446](https://github.com/pytorch/pytorch/pull/56446))
- Use \_WeakTensorRef over weakref in test\_autograd.py ([#55726](https://github.com/pytorch/pytorch/pull/55726))
- Move view and inplace handling to a separate key ([#53342](https://github.com/pytorch/pytorch/pull/53342))

### Complex Numbers

- Added complex support for `torch.testing.assert_(equal|close)` ([#57162](https://github.com/pytorch/pytorch/pull/57162)).
- Fixed NVCC related build warnings for complex operations in PyTorch ([#55142](https://github.com/pytorch/pytorch/pull/55142)).
- Add eager and jit variant consistency tests for `torch.cfloat` tensor type ([#54854](https://github.com/pytorch/pytorch/pull/54854)).
- Fixed complex mean and reduction tests that weren’t being properly run ([#55640](https://github.com/pytorch/pytorch/pull/55640)).
- [ROCm] Added missing template declarations for complex BLAS ([#52472](https://github.com/pytorch/pytorch/pull/52472)).

### CUDA

- Kernel launch checks for aten/src/ATen ([#52185](https://github.com/pytorch/pytorch/pull/52185))
- Add more kernel launch checks ([#53286](https://github.com/pytorch/pytorch/pull/53286))
- Final kernel launch checks ([#54214](https://github.com/pytorch/pytorch/pull/54214))
- Fix nvcc warnings ([#55367](https://github.com/pytorch/pytorch/pull/55367))
- irange for Indexing.cu ([#57479](https://github.com/pytorch/pytorch/pull/57479))
- reduce number of randperm template instantiations ([#58362](https://github.com/pytorch/pytorch/pull/58362))
- Enforce kernel launch checks ([#58116](https://github.com/pytorch/pytorch/pull/58116))
- fix comments in ATenNVRTC.h ([#57318](https://github.com/pytorch/pytorch/pull/57318))

### AMD

Generalize HIP-specific launch bounds to apply to CUDA ([#56143](https://github.com/pytorch/pytorch/pull/56143))

### Composability

- Dispatcher passes computed dispatch keys to kernels ([#49354](https://github.com/pytorch/pytorch/pull/49354))
- Add `TORCH_CHECK_NOT_IMPLEMENTED`/`c10::NotImplementedError`; make dispatch use it ([#53377](https://github.com/pytorch/pytorch/pull/53377))
- Refactor `tensor_new.cpp` to use `TensorOptions` instead of `DispatchKey` ([#54034](https://github.com/pytorch/pytorch/pull/54034))
- Add `Tensor::is_cpu`, genericize `TensorIterator` ([#54079](https://github.com/pytorch/pytorch/pull/54079))
- Migrate about 100 kernel to C10 full dispatcher ([#54109](https://github.com/pytorch/pytorch/pull/54109))
- Rename `XPLAT_MOBILE_BUILD` to `TEMPLATE_SELECTIVE_BUILD` ([#54217](https://github.com/pytorch/pytorch/pull/54217))
- Migrate kernels with `Tensor?` to C10 full dispatcher ([#54263](https://github.com/pytorch/pytorch/pull/54263))
- Delete all unnecessary singular `Math` entries ([#54436](https://github.com/pytorch/pytorch/pull/54436))
- Rename `Math` to `CompositeImplicitAutograd` ([#54466](https://github.com/pytorch/pytorch/pull/54466))
- Rename `DefaultBackend` to `CompositeExplicitAutograd` ([#54470](https://github.com/pytorch/pytorch/pull/54470))
- Migrate kernels with `TensorOptions` to C10 full dispatcher ([#54539](https://github.com/pytorch/pytorch/pull/54539))
- Expose ops present in dispatcher via `Dispatcher::getAllOpNames()` ([#54791](https://github.com/pytorch/pytorch/pull/54791))
- Make redispatch functions callable from out of tree extensions ([#54966](https://github.com/pytorch/pytorch/pull/54966))
- Remove `use_c10_dispatcher` option ([#54969](https://github.com/pytorch/pytorch/pull/54969))
- Provide a method `ObservedOperators::getUnobservedOperatorList()` so that model tracer can empty it out during tracing ([#55017](https://github.com/pytorch/pytorch/pull/55017))
- Support `needsOutputs` for `RecordFunction` and `ObserverUtil` improvements ([#55012](https://github.com/pytorch/pytorch/pull/55012))
- Strict typecheck all files in `tools/codegen` ([#55227](https://github.com/pytorch/pytorch/pull/55227))
- Add `MaybeOwned::operator*() &&` ([#55244](https://github.com/pytorch/pytorch/pull/55244))
- Allow copy operations on `MaybeOwned` ([#55419](https://github.com/pytorch/pytorch/pull/55419))
- Remove non-const `TensorIterator::tensor()` method ([#55420](https://github.com/pytorch/pytorch/pull/55420))
- Make `as_strided_` use\_const ref for mutable tensors ([#55875](https://github.com/pytorch/pytorch/pull/55875))
- Generate xla codegen in-tree ([#56601](https://github.com/pytorch/pytorch/pull/56601))
- HABANA Device registration key and Autograd key addition ([#57094](https://github.com/pytorch/pytorch/pull/57094))
- Refactor autocast to be extensible for devices ([#57104](https://github.com/pytorch/pytorch/pull/57104))
- Add pybind type caster for `c10::Device` ([#57292](https://github.com/pytorch/pytorch/pull/57292))
- Make `c10::TempFile` non-copyable but movable ([#57308](https://github.com/pytorch/pytorch/pull/57308))
- Fix `string_view::equals_` compilation by CUDA-11.3 ([#57322](https://github.com/pytorch/pytorch/pull/57322))
- Delete move constructor on TensorImpl ([#58048](https://github.com/pytorch/pytorch/pull/58048))
- structured kernels - error check when structured\_delegate is not marked structured ([#52227](https://github.com/pytorch/pytorch/pull/52227))
- fix RegistrationDeclarations.yaml, now that we codegen composite kernels for structured functional/inplace ops ([#56307](https://github.com/pytorch/pytorch/pull/56307))

### TorchScript

- Remove `output_args` from `ReduceOp` ([#52187](https://github.com/pytorch/pytorch/pull/52187))
- Remove `ReduceOp::accumulator` ([#52196](https://github.com/pytorch/pytorch/pull/52196))
- Fix memory dependencies computation to not look at reduction output args ([#52170](https://github.com/pytorch/pytorch/pull/52170))
- Update `rfactor` to not use `ReduceOp->output_args()` ([#52177](https://github.com/pytorch/pytorch/pull/52177))
- Add an initialization expression to `Reduce() `([#53751](https://github.com/pytorch/pytorch/pull/53751))
- Add IRVerifier ([#52901](https://github.com/pytorch/pytorch/pull/52901))
- Add index verifier for `Store` ([#53137](https://github.com/pytorch/pytorch/pull/53137))
- Add new APIs to get loops corresponding to a `Buf` ([#53778](https://github.com/pytorch/pytorch/pull/53778))
- Remove Dropout during frozen optimization ([#51589](https://github.com/pytorch/pytorch/pull/51589))
- Add pure list-producing ops to alias analysis ([#51999](https://github.com/pytorch/pytorch/pull/51999))
- Remove `DepTracker` from LoopNest ([#52405](https://github.com/pytorch/pytorch/pull/52405))
- Use graph executor to run `forward` on a gradient ([#52136](https://github.com/pytorch/pytorch/pull/52136))
- Support `casted_batch_one_hot_lengths` with 4-arg `to` ([#53215](https://github.com/pytorch/pytorch/pull/53215))
- Enable `ClipRangesGatherRangesX2SigridHash` fusion for `SigridHashPrecompute` ([#53324](https://github.com/pytorch/pytorch/pull/53324))
- Convert `to` to `to_copy` ([#53524](https://github.com/pytorch/pytorch/pull/53524))
- Fuse SigridTransforms + ListUnpack ([#53920](https://github.com/pytorch/pytorch/pull/53920))
- Use reshape when possible in broadcasting ([#53326](https://github.com/pytorch/pytorch/pull/53326))
- Lazily initialize `AliasDb` constant prop ([#54640](https://github.com/pytorch/pytorch/pull/54640))
- Fix freezing with MKLDNN tensors ([#54632](https://github.com/pytorch/pytorch/pull/54632))
- Add `EliminateExceptions` pass ([#54730](https://github.com/pytorch/pytorch/pull/54730))
- Lazily initialize `AliasDb` and add `changed` status to CSE ([#54776](https://github.com/pytorch/pytorch/pull/54776))
- Make transformations return whether graph is modified ([#54777](https://github.com/pytorch/pytorch/pull/54777))
- Change `resize_as_` to `resize_` ([#55098](https://github.com/pytorch/pytorch/pull/55098))
- Update to short forms of `splitWithTail` / `splitWithMask` ([#55542](https://github.com/pytorch/pytorch/pull/55542))
- Patch `requires_grad` on `DifferentiableGraph` ([#55701](https://github.com/pytorch/pytorch/pull/55701))
- Replace `AutoNonVariableTypeMode` with `InferenceMode` in static runtime ([#55731](https://github.com/pytorch/pytorch/pull/55731))
- Move tensor implicit conversions to `test_builtins.py` ([#55532](https://github.com/pytorch/pytorch/pull/55532))
- Redesign `Rfactor `loopnest transformation. ([#55324](https://github.com/pytorch/pytorch/pull/55324))
- Remove mask field from `Load` and `Store` classes ([#55825](https://github.com/pytorch/pytorch/pull/55825))
- Switch type of `tensors_` from `Tensor` to `Buf` ([#56318](https://github.com/pytorch/pytorch/pull/56318))
- Merge `ivalue::Future`’s `markCompleted` and `markCompletedWithDataPtrs` ([#56512](https://github.com/pytorch/pytorch/pull/56512))
- Don’t lift tensor constants from fusion groups ([#56756](https://github.com/pytorch/pytorch/pull/56756))
- Use `c10::ScalarType` instead of `tensorexpr::ScalarType` ([#56825](https://github.com/pytorch/pytorch/pull/56825))
- Use JIT Plug-in for coverage to cover JIT’d functions and methods ([#56310](https://github.com/pytorch/pytorch/pull/56310))
- Add all pools, Batchnorm and Tanh (i.e. all ideeped MKLDNN ops) to MKLDNNFuser ([#56541](https://github.com/pytorch/pytorch/pull/56541))
- Inline hooks in `ivalue::Future` ([#57354](https://github.com/pytorch/pytorch/pull/57354))
- Add a pass for annotating a graph with input types derived from sample inputs ([#57076](https://github.com/pytorch/pytorch/pull/57076))
- Add a pass for removing a first (self) argument from a graph if it is unused ([#57169](https://github.com/pytorch/pytorch/pull/57169))
- Remove `dtype_` and add `buf_` fields to `CodeGen::BufferArg`. ([#57382](https://github.com/pytorch/pytorch/pull/57382))
- Add tests for custom state\_dict `save`/`load` methods in TorchScript ([#57886](https://github.com/pytorch/pytorch/pull/57886))
- Add schema check to `aten::repeat` and `fb::fast_gather` ([#58106](https://github.com/pytorch/pytorch/pull/58106))
- Rename `Tensor::call` to `Tensor::load` to be consistent with `Buf` and `Placeholder`. ([#55826](https://github.com/pytorch/pytorch/pull/55826))

### Mobile

- Check in Gradle wrapper for easier pytorch\_android builds. ([#51067](https://github.com/pytorch/pytorch/pull/51067))

### torch.fx

- Hoist custom class .so loading into setUp ([#52883](https://github.com/pytorch/pytorch/pull/52883))
- Test forward reference annotations ([#53713](https://github.com/pytorch/pytorch/pull/53713))
- Add TestConstFold coverage to test\_fx ([#54072](https://github.com/pytorch/pytorch/pull/54072))
- Fix logic in TestFX.test\_get\_torch\_func\_signature\_exhaustive ([#54510](https://github.com/pytorch/pytorch/pull/54510))
- Test tracing into all the standard torch.nn.functional ([#55550](https://github.com/pytorch/pytorch/pull/55550))
- Add more model symbolic tracing tests from torchvision ([#55744](https://github.com/pytorch/pytorch/pull/55744))
- Make stack trace testing less strict ([#58088](https://github.com/pytorch/pytorch/pull/58088))

### Quantization

- Make bundled inputs work with quantized zero inputs ([#47407](https://github.com/pytorch/pytorch/pull/47407))
- Call native resize\_/resize\_as\_ as much as possible ([#53425](https://github.com/pytorch/pytorch/pull/53425))
- Use `expect_contiguous` in `quantized::linear` fbgemm version ([#58221](https://github.com/pytorch/pytorch/pull/58221))
- Add pass in convert to fold quant-dequant sequence ([#54860](https://github.com/pytorch/pytorch/pull/54860))
- Add support for one value being quantized with different qconfigs ([#53586](https://github.com/pytorch/pytorch/pull/53586))
- Store dtype, axis as literals in the graph ([#54624](https://github.com/pytorch/pytorch/pull/54624))
- add \_remove\_qconfig flag to convert\_fx ([#53166](https://github.com/pytorch/pytorch/pull/53166))
- Get first linear use of quantize\_per\_tensor for FQN ([#54859](https://github.com/pytorch/pytorch/pull/54859))
- Factoring out the list of no\_observers ([#50459](https://github.com/pytorch/pytorch/pull/50459))
- Enable test for non quantized input for add/mul ([#52412](https://github.com/pytorch/pytorch/pull/52412))
- Guard the supported quantization type for add/mul ([#52413](https://github.com/pytorch/pytorch/pull/52413))
- Enable test for non quantized input for cat ([#52414](https://github.com/pytorch/pytorch/pull/52414))
- Merge add and mul handler ([#52651](https://github.com/pytorch/pytorch/pull/52651))
- Refactoring binary op tests to split int8 and float16 tests ([#52807](https://github.com/pytorch/pytorch/pull/52807))
- Refactoring binary op tests to split int8 and float16 tests (#52807) ([#53020](https://github.com/pytorch/pytorch/pull/53020))
- Remove reduandent code ([#54073](https://github.com/pytorch/pytorch/pull/54073))
- Change activation\_post\_process\_map to track the observer name instead ([#54643](https://github.com/pytorch/pytorch/pull/54643))
- Separate handling Copy operator to a helper function ([#54644](https://github.com/pytorch/pytorch/pull/54644))
- Factor out insert\_observers\_for\_model to a separate function ([#54733](https://github.com/pytorch/pytorch/pull/54733))
- Factor out insert\_observers\_for\_model to a separate function (#54733) ([#55307](https://github.com/pytorch/pytorch/pull/55307))
- Separate handling Copy operator to a helper function (#54644) ([#55429](https://github.com/pytorch/pytorch/pull/55429))
- Add shape to nontensor op list ([#55529](https://github.com/pytorch/pytorch/pull/55529))
- fx quant:
  - clean up nit in insert\_observer ([#57367](https://github.com/pytorch/pytorch/pull/57367))
  - readability improvements on observer functions ([#57368](https://github.com/pytorch/pytorch/pull/57368))
  - move output obs logic to QuantizeHandler ([#57377](https://github.com/pytorch/pytorch/pull/57377))
  - move input\_output\_observed to qhandler ([#57388](https://github.com/pytorch/pytorch/pull/57388))
  - remove FixedQParamsOpQuantizeHandler from quantize.py ([#57393](https://github.com/pytorch/pytorch/pull/57393))
  - remove unnecessary quants arguments ([#57399](https://github.com/pytorch/pytorch/pull/57399))
  - remove `find_quants` from convert ([#57402](https://github.com/pytorch/pytorch/pull/57402))
  - refactor observer insertion (4f50fdc2a3)
  - remove matching hack for binary qhandler ([#57470](https://github.com/pytorch/pytorch/pull/57470))
  - clean up names of quantize handlers ([#53614](https://github.com/pytorch/pytorch/pull/53614))

- Benchmark for torch.ops.quantized.linear\_prepack\_fp16 operator ([#52229](https://github.com/pytorch/pytorch/pull/52229))
- Add Per Tensor Quantization Support to FXIRImporter ([#55405](https://github.com/pytorch/pytorch/pull/55405))
- Hide warnings for deprecated quantization APIs ([#56291](https://github.com/pytorch/pytorch/pull/56291))
- Remove “Sparsity” from the function names ([#56555](https://github.com/pytorch/pytorch/pull/56555))

### ONNX

- cmake: fix ONNX\_NAMESPACE if USE\_SYSTEM\_ONNX ([#54973](https://github.com/pytorch/pytorch/pull/54973))
- Link onnx\_library when BUILD\_TEST=0 for Windows ([#51937](https://github.com/pytorch/pytorch/pull/51937))
- Fix onnx/constant\_fold.cpp compilation on Windows ([#55770](https://github.com/pytorch/pytorch/pull/55770)) ([#56167](https://github.com/pytorch/pytorch/pull/56167))

### Misc

- Updated PyBind to official v2.6.2 tag ([#52304](https://github.com/pytorch/pytorch/pull/52304))
- Added gdb special command to print tensors ([#54339](https://github.com/pytorch/pytorch/pull/54339))
- Numpy dependency is now only checked when using Numpy features ([#52794](https://github.com/pytorch/pytorch/pull/52794))
- don’t set the same C++ and C standards twice ([#51832](https://github.com/pytorch/pytorch/pull/51832))
- Fix cmake\_minimum\_require in libshm ([#58306](https://github.com/pytorch/pytorch/pull/58306))