# What changed in torch

> 291 changes in torch 2 from its official release notes, upgrade guides and advisories: 36 breaking, 8 deprecations, 7 security, 74 new APIs and options. Each one cited. Sources checked 2026-10-18. HTML: https://unstale.sh/pypi/torch/

To see which of these apply to a repository, run `npx @unstale/cli init` in it.

## torch 2

### Breaking changes

- Replaces the exported is_macos_13_or_newer() helper; no compatibility alias is provided. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The enum members drop the VER and PLUS portions of their names, with no compatibility aliases. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The zero-argument C++ overload was removed; pass includeBool explicitly, using false to preserve its previous behavior. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The zero-argument C++ overload was removed; pass includeBool explicitly, using false to preserve its previous behavior. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Dynamo graph-breaks rather than silently dropping tangents; fullgraph=True now errors. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Individual Python function events are excluded by default; pass include_python_functions=True to retain the previous view. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Custom Python process groups implementing new_group() must accept the backend keyword, which PyTorch now forwards when creating subgroups. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The setting is replaced with memory-headroom controls. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Segments allocated after registering a symmetric-memory pool are no longer automatically upgraded to symmetric windows; collectively deregister and register again after allocations to use symmetric-window algorithms. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- It is now a deprecation shim: install and develop still forward to pip, while build and other commands print replacement commands instead of running setuptools. Build through pip or python -m build. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The deprecated API was removed; use torch.linalg.cholesky(). For an upper-triangular factor, take the conjugate transpose with .mH. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The deprecated API was removed; use torch.linalg.qr() with mode="reduced" or mode="complete" instead of the Boolean some argument. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled and fake/meta execution enforce the same exact output-dtype requirements as eager execution. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The use_cuda argument was removed; use use_device="cuda" instead. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Complex type promotion for bfloat16 now uses bcomplex32 instead of complex64; cast explicitly to complex64 when its precision or operator coverage is required. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The deprecated API was removed; use torch.linalg.cholesky(). For an upper-triangular factor, take the conjugate transpose with .mH. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Boundary subgradients now follow the selected dispatcher schema's input space: scalar-bound equality gives the input gradient 0, while tensor-bound ties split gradients evenly. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- At an ordinary tensor-bound tie, gradients split evenly between the input and bound. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- At a scalar-bound equality the input gradient changes from 1 to 0; at an ordinary tensor-bound tie, gradients split evenly between input and bound. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The tvm backend now requires TVM's relax frontend; older TVM installations raise ImportError. Pass a TVM pipeline through options instead of relay-only scheduler and trials options. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- It is a no-op for fake process groups and warns rather than failing for backends that cannot configure timeouts. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Ranks absent from every requested split now receive GroupMember.NON_GROUP_MEMBER instead of None. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Fractional max pooling raw operations now reject non-positive kernel_size values instead of returning invalid outputs and indices. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Ties now split gradients evenly, and forward-mode AD for the min/max family is aligned with the new rules. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Ties now split gradients evenly, and forward-mode AD for the min/max family is aligned with the new rules. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Functional and module Lp pooling APIs now raise a descriptive ValueError when norm_type=0. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Eligible fused backends are now selected for dense rank-3 inputs on CPU, CUDA/ROCm, and XPU. Selection may change numerics, dropout RNG consumption, view behavior, and higher-order-gradient support; select the math backend when those semantics are required. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The balanced policy is no longer accepted; use compact instead. Passing balanced now raises ValueError. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Invalid dim types are rejected at construction rather than failing later during forward. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Invalid dim types are rejected at construction rather than failing later during forward. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Setting profiler_measure_per_kernel to a non-default value no longer enables CUPTI range profiling, is ignored, and emits a FutureWarning. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Setting profiler_metrics to a non-default value no longer enables CUPTI range profiling, is ignored, and emits a FutureWarning. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The use_cuda argument was removed. Select CUDA through activities instead. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The deprecated API was removed; use torch.linalg.qr() with mode="reduced" or mode="complete" instead of the Boolean some argument. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The relay-only backend entry point no longer exists in torch._dynamo.backends.tvm. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The relay-only backend entry point no longer exists in torch._dynamo.backends.tvm. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)

### Security

- PyTorch is vulnerable to memory corruption through its torch.jit.script function (2.13.0) — [security advisory](https://github.com/advisories/GHSA-RRMF-RVHW-RF47)
- PyTorch is a Python package that provides tensor computation. Prior to version 2.10.0, a vulnerability in PyTorch's `weights_only` unpickler allows an attacker to craft a malicious checkpoint file (`.pth`) that, when loaded with `torch.load(..., weights_only=True)`, can corrupt memory and potentially (2.10.0) — [security advisory](https://github.com/advisories/GHSA-63CW-57P8-FM3P)
- PyTorch is vulnerable to memory corruption through its torch.lstm_cell function (2.10.0) — [security advisory](https://github.com/advisories/GHSA-QFHQ-4F3W-5FPH)
- PyTorch is vulnerable to memory corruption through its unpack_sequence function (2.9.1) — [security advisory](https://github.com/advisories/GHSA-VGRW-7CVW-PWGX)
- An issue in the component torch.linalg.lu of pytorch v2.8.0 allows attackers to cause a Denial of Service (DoS) when performing a slice operation. (2.9.0) — [security advisory](https://osv.dev/vulnerability/PYSEC-2025-203)
- pytorch v2.8.0 was discovered to contain an integer overflow in the component torch.nan_to_num-.long(). (2.9.0) — [security advisory](https://osv.dev/vulnerability/PYSEC-2025-206)
- pytorch v2.8.0 was discovered to display unexpected behavior when the components torch.rot90 and torch.randn_like are used together. (2.9.0) — [security advisory](https://osv.dev/vulnerability/PYSEC-2025-204)

### Deprecations

- The method is now a no-op with a deprecation warning; generator state is registered lazily on first RNG use during capture, so explicit registration calls should be removed. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The method emits a FutureWarning; activate a green context by creating a stream from it and using that stream with torch.cuda.stream(). (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The method emits a FutureWarning; activate a green context by creating a stream from it and using that stream with torch.cuda.stream(). (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The private alias remains available with a deprecation warning; use the public torch.distributed.set_timeout API. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The setting is deprecated and is now a no-op; generators are always traced lazily. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The distributed spelling remains a forwarding alias but now emits a FutureWarning; use torch.compiler.config.compile_on_one_rank. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- TorchScript APIs now emit visible FutureWarnings directing users toward torch.compile or torch.export. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- with_modules is deprecated and emits a FutureWarning; it collected module hierarchy only for TorchScript, while eager users can use with_stack=True to record nn.Module events. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)

### New APIs and options

- When set, AOTInductor prints timing and diagnostic messages for each constant-loading stage without requiring a rebuild. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Supports explicit user-defined streams and emits stream-guard code so compiled kernels run on the selected stream rather than always on the default stream. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Creates an AOTInductor model container from caller-owned weight tensors for zero-copy sharing; constants are not loaded from the package and remain caller-owned. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a protocol for distributed checkpointing tensors that expose global and local shape and offset metadata. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Lets custom autograd.Function implementations declare each output's expected gradient dtype independently of its storage dtype. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds destroy callbacks and object retention. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds replay start and end hooks. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds portable JSON loading of distributed execution logs for comparisons across processes or model configurations. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds portable JSON serialization of distributed execution logs for comparisons across processes or model configurations. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Expanded sharding strategies for matrix, attention, sorting, scanning, softmax, tensor operations, and Conv1d/2d/3d; added redistribution from final-dimension sharding to Partial("sum"). (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a shared keyword accepting ShapesSpec or ParamsSpec for explicit dynamic shapes in export, fake make_fx tracing, and torch.compile. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Profiler events expose Kineto metadata as typed values when expose_kineto_event_metadata is enabled, avoiding JSON reparsing. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds an experimental backend that creates per-peer NCCL point-to-point communicators on demand. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds an explicit legacy NCCL backend name alongside an opt-in for selecting the experimental replacement behind the nccl name. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The experimental backend gains fault-tolerant reconfiguration and one-sided window operations. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The experimental backend adds split_group, complete Work semantics, nonblocking communicators, uneven list collectives, communicator memory suspend/resume, and memory-pool registration and deregistration. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Can consume explicitly pre-split positional inputs, keyword inputs, and targets through arg_mbs, kwarg_mbs, and target_mbs. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Custom Python process groups can use batch_isend_irecv, the coalescing manager, and complete collective coverage, including single-tensor gather/scatter. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds an experimental accessor for custom backend development. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Gloo process groups support fault-tolerant reconfiguration. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Allows readable communicator names in NCCL logs and profiler tools. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Controls the number of cuBLASLt heuristic candidates requested by TunableOp. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds an option to choose whether selective activation checkpointing honors surrounding saved_tensors_hooks; the current default preserves legacy behavior and warns when hooks are active. Requires use_reentrant=False. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Used to provide optional shell-completion generation for torchrun. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Callers can provide an is_impure_node callback so side-effecting nodes are preserved during dead-code elimination. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a prototype higher-order op selecting among N branches by index; autograd is not yet supported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds XPU support for symmetric-memory operations used by asynchronous tensor parallelism, enabling communication and computation overlap on Intel GPUs. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds read-only DLPack export, including copy-on-write-preserving exchange with compatible consumers. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- This is the preferred environment variable for compile_on_one_rank; the older distributed-prefixed variable remains supported for compatibility. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The opt-in selects the experimental replacement backend behind the nccl name. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Enables an autotuning_inputs log artifact reporting Triton autotuning input shapes, dtypes, strides, and scalar values. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds stable-ABI permute. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds stable-ABI Tensor overloads for bitwise operations, index_select, floor_divide, and is_pinned. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds stable-ABI conversion from Python torch.Tensor objects to torch::stable::Tensor. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a stable-ABI query for whether a Tensor has storage. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds stable-ABI conversion from torch::stable::Tensor to Python torch.Tensor objects. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a dtype overload to stable-ABI view. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a length argument so a scan can run a fixed number of steps when xs=None. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Opt-in setting applies locality-based graph reordering to training graphs. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Opt-in CUDA Graph Trees generation cloning preserves live user-visible outputs across generations. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds backend-agnostic accelerator RNG state inspection. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds backend-agnostic inspection of RNG states across accelerators. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds backend-agnostic accelerator RNG seed inspection. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a thread-local context manager whose callback receives fully populated autograd graph nodes created in its scope, including during higher-order differentiation and checkpoint recomputation. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Opts eligible bfloat16 grouped GEMM workloads into the cuBLASLt backend on Hopper and Blackwell with CUDA 13.3 or newer; eligible float16 workloads select it by default. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds support for TVM's Relax frontend and pipeline-based tuning. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Lets code request another CUDA Graph Trees warmup iteration. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a public API for non-strict tracing. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a capture-start hook alongside global CUDA graph capture and replay start/end hooks. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a public CUDA graph annotations module. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The CUDAGraph debug flag is unified, debug_dump moves to Python, and capture hooks are added. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Attaches timestamped metadata events to live CUDA allocations; annotations appear in snapshots when memory history is enabled and only the native CUDA caching allocator supports them. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Memory snapshots can include CPU pinned-memory allocations; pass record_cuda=False to record only pinned host memory. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a control for the number of cuBLASLt heuristic candidates used by TunableOp, which now supports CUDA cuBLASLt. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds an experimental accessor for custom backend development. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a public API for setting distributed timeouts; the private alias remains available with a deprecation warning. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds construction of distributed linearly spaced tensors. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds construction of distributed logarithmically spaced tensors. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Attaches a dynamic-shape specification to a function or nn.Module.forward; compile, export, and make_fx use it automatically, and conflicting call-site specifications error. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Lets Python-authored kernels reserve Philox counter ranges that remain correct across CUDA Graph capture and replay. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Computes the principal square root of symmetric or Hermitian positive-definite matrices, supporting batches, autograd, vmap, and torch.compile. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Computes the polar decomposition A = U @ H for matrices with at least as many rows as columns, using portable SVD and eligible CUDA cuSOLVER QDWH acceleration. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS adds flex_attention auxiliary and dynamic-shape support, polar, large-tensor nonzero, complex Cholesky and nan_to_num, upsampling backward, and geqrf support. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds MPS forward and backward support. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The CUDA memory-efficient backend adds grouped-query attention, implicit multi-query key/value broadcasting, and backward support under vmap. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a maximize parameter. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds the spectral_unclamped scaling option to the adjust_lr_fn parameter. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Can be called without a function to create an eager-mode decorator with checkpoint configuration separated from wrapped-function arguments; direct calls remain the compatible form under torch.compile. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a wrapper for read-only DLPack export and compatible copy-on-write-preserving exchange. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a query for per-process GPU memory usage on XPU. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Device-wide synchronization support is added on XPU. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds a log-line prefix template and hostname variable to identify the host that emitted each worker log line. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds optional shell-completion generation through the shtab package. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- A dedicated error type is added for SYCL compilation failures. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)

### Performance

- Importing this module no longer eagerly compiles TorchScript merely because the module is imported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Conversions are lowered directly on CPU and CUDA instead of relying on fallback conversion code. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Importing this module no longer eagerly compiles TorchScript merely because the module is imported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native float32 and complex64 MPS implementations, retaining CPU fallbacks for small matrices and matrices exceeding threadgroup memory. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native float32 and complex64 MPS implementations, retaining CPU fallbacks for small matrices and matrices exceeding threadgroup memory. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Importing this module no longer eagerly compiles TorchScript merely because the module is imported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)

### Behavior and default changes

- Supports lazy autotuning and compilation of torch.cond and torch.while_loop. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native MPS support. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- A single CUDA graph can now use multiple memory pools. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Supports head dimension 256 on SM90 and SM10.x with cuDNN newer than 9.22 and cuDNN Frontend 1.24 or newer; backward currently requires d_qk and d_v both 256. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Multiword compiler commands such as "zig c++" are accepted when building Inductor-generated C++ code on POSIX systems. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Returns input nodes, output nodes, and parameters in deterministic graph order. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds initial technology-preview AMD support; CK SDPA/GEMM, FP8 grouped GEMM, and int4 matrix multiplication remain unsupported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Enabled on Linux with ROCm 7.14 or newer. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Pair communicators are included in error reporting, suspend/resume operations, and memory statistics, with broader shared backend coverage. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Partitioned-scatter selection is memory- and contention-aware and enabled by default on ROCm; set this setting to False to opt out. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native MPS sampling, eliminating its CPU fallback. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Module functions and constants are accepted as messages. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Explicitly registered logging.Logger methods with supported positional arguments run after the compiled region instead of causing graph breaks. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds XPU FP8 blockwise scaling support for MXFP8, MXFP4, and NVFP4 recipes. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds XPU FP8 blockwise scaling support for MXFP8, MXFP4, and NVFP4 recipes. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA now supports torch.float16 and torch.bfloat16 inputs. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Selecting the CK GEMM backend now works on ROCm gfx90a devices, independently of CK attention support. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compilation support expands for Python constructs and tracing covers additional accelerator, tensor, and distributed operations. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Reuse now supports source-backed objects, symbolic shapes, dataclasses, and namedtuples. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Supports dynamic shapes for dimensions marked with mark_unbacked, allowing one artifact to serve multiple runtime sizes without guarding on that dimension. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA graph memory pools are included in the reported reserved memory. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Can be used inside a compiled region; allocations in the context, including fallback and extern kernels, are routed to the pool. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Device-qualified Gloo backends are accepted when TorchComms is enabled. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds bfloat16 support on CUDA and float16/bfloat16 support on XPU; unsupported CUDA and XPU cases promote to float32, while CPU FFT continues to reject these dtypes. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The eager CPU torch.polar kernel now supports Half inputs. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS dispatch key is preserved through transforms, enabling MPS autocast and autograd under vmap and grad. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Enabled on ROCm 7.14 or newer through hipSOLVER's generic Xgeev API. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native float32 and complex64 MPS implementations, retaining CPU fallbacks for small matrices and matrices exceeding threadgroup memory. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native float32 and complex64 MPS implementations, retaining CPU fallbacks for small matrices and matrices exceeding threadgroup memory. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds MPS support, including complex inputs, on macOS 15 or newer. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Backward support is available on CPU, CUDA, and MPS, and Dynamo can trace the operation. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS gains support for the operation. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native float32 and complex64 MPS implementations, retaining CPU fallbacks for small matrices and matrices exceeding threadgroup memory. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Added to the core ATen decomposition table used by ExportedProgram.run_decompositions(), including correct dim=() handling, and gained sparse COO dispatch. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Out-of-tree backends can define additional out_dtype combinations under fake/meta tracing; CUDA and XPU restrictions remain unchanged. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Can process float32 and float64 inputs larger than 2**24 elements on devices supporting float64 by computing ranks in float64. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The chunked path handles probability targets with mean or sum reduction when target dtype matches input and targets do not require gradients. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Can process float32 and float64 inputs larger than 2**24 elements on devices supporting float64 by computing ranks in float64. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA float16 and bfloat16 support is added, including supported sparse-CSR backward paths. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds bfloat16 support on CUDA and float16/bfloat16 support on XPU; unsupported CUDA and XPU cases promote to float32, while CPU continues to reject these dtypes. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled scans now honor deterministic-algorithm settings. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds scalar searchsorted support, broader copy-view batching, scalar fill and comparison support, and a native in-place masked_fill_ rule. Batched repeat_interleave requires a common output_size. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds CUDA graph support. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Bare backend names are auto-qualified, process-group options pass through custom backend creation, and unknown device-qualified names can register as custom backends. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds native recording mode on non-PVC devices when built with oneAPI 2026.1 or newer. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Profiler overhead is recorded as OVERHEAD activities on a dedicated track in exported traces. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds IPC memory handle sharing support on XPU. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Inductor autotuning subprocesses now honor this XPU environment variable. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)

### Tooling

- Adds the documented memory_format overload so static type checkers accept calls such as module.to(memory_format=torch.channels_last). (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Static typing for indexing now resolves integer keys to Module and slices to Sequential. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Adds mappings for cublasMath_t, its enum values, and CUBLAS_COMPUTE_16F so HIP-ported extensions using these cuBLAS features hipify without project aliases. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)

### Fixes and other changes

- Updates the binaries to CUDA 13.2.2, resolving CUDA issues that could produce incorrect results in NVFP4 matrix multiplications and kernels with nested thread divergence. (2.14.1) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.1)
- Fixes incorrect solutions on MPS for complex batched underdetermined systems. (2.14.1) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.1)
- Fixes failures on MPS with a Metal pipeline-state error for inputs above 8192 elements. (2.14.1) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.1)
- Fixes non-orthogonal U and inaccurate small singular values on MPS for rank-deficient and ill-conditioned inputs. (2.14.1) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.1)
- Fixes failures on MPS with a Metal pipeline-state error for inputs above 8192 elements. (2.14.1) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.1)
- Fixes failures on MPS with a Metal pipeline-state error for inputs above 8192 elements. (2.14.1) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.1)
- Fixes an internal assert on MPS for complex inputs when using out. (2.14.1) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.1)
- Functional backward supports the linear avg reduction instead of rejecting it after a successful forward pass. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The NCCL symmetric-memory backend now implements barrier instead of raising a not-implemented error. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Replacing FileBaton with filelock prevents stale-lock deadlocks. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Its pool is trimmed and allocation is retried once before an out-of-memory error is raised. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Registered custom communication backends are selected instead of incorrectly falling back to NCCL or Gloo when the backend is unspecified. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA IPC-backed dataset storage is released when workers exit, preventing retained producer-side references and allocations. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Post-dispatch make_fx tracing decomposes detach by default to fix silently incorrect second-order gradients; an explicit decomposition table retains previous behavior. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled functions can return DTensor permutation views such as transpose, permute, or movedim. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Quantized byte and 4-bit rowwise lookups no longer fault on ROCm. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Generating guard messages for deeply nested symbolic-shape expressions no longer raises RecursionError. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- summon_full_params with offload_to_cpu no longer accesses freed storage when the flattened parameter is already on CPU. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- HIP C++ extensions on Windows now register hipify-produced .hip sources with MSVC so they are dispatched to hipcc. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Complex inputs are rejected with NotImplementedError instead of an internal MPSGraph error. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Pinned memory correctly appears as a CPU tensor while retaining its shared Metal buffer. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Unsupported dtypes now produce a clear error. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The correct Metal cooperative-tensor ABI is selected on macOS 26 to prevent corrupted output. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Dtype-converting copies no longer overwrite their source and non-dense views copy correctly with matching strides. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Non-contiguous tensors are handled safely, fixing vmap failures and possible out-of-bounds memory accesses. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Non-contiguous tensors are handled safely, and CUDA backward precision errors are fixed by applying missing upcasts. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Ranks are normalized through Python's integer protocol, allowing tensor integer ranks and clearly rejecting non-integral values. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA backward uses 64-bit sample offsets. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Python subclasses can be constructed through the (store, rank, size) constructor, and their virtual overrides are dispatched correctly. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- PyTorch publishes Python 3.15 wheels for Linux x86-64 and aarch64, Windows x86-64, and macOS Apple silicon across applicable builds; torch.compile remains unsupported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- PyTorch publishes free-threaded Python 3.15t wheels across the listed Linux, Windows, and macOS platforms; torch.compile remains unsupported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Dynamo now matches Python behavior for a range of built-ins, conversions, descriptors, errors, operators, subclass types, and attribute probing. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The FBGEMM ARM path now checks overflow, fixing incorrect quantized values on some hardware. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Modules using only this hook no longer emit a warning intended for register_full_backward_hook when forward inputs do not require gradients. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The ROCm meta-kernel now uses the compact log-sum-exp dimension returned by ROCm backends, fixing nested tensor backward and torch.compile. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- A warning is issued when collectives are launched concurrently on multiple streams, which can otherwise deadlock. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS computes integer absolute values exactly instead of rounding through float32. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS handles empty dimensions. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS supports strided output tensors on macOS 14 and 15. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Fractional arguments with an integer output dtype no longer produce the wrong length due to premature truncation. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Validation error messages preserve dynamic type names and argument indices. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS handles empty dimensions and does not propagate NaN or infinity from input when beta=0. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS handles empty dimensions. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Supports second-order gradients, so grad-grad computations no longer fail for lack of a backward derivative. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA clamp preserves signed zero. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Inductor handling of torch.combinations and several other operation and graph-output cases is corrected. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compilation better matches Python built-ins, operators, iterators, container initialization, object behavior, exceptions, and subgenerator closure behavior. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Selected Dynamo, Inductor, and FX tracing state is thread-local to prevent race conditions when compilation is invoked concurrently from multiple threads. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- No-op dropout decompositions preserve eager identity semantics, avoiding replacement of a Parameter with a cloned fake tensor when dropout is disabled. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Graph reuse, eager autograd, graph capture, and transposed captured buffers are fixed. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compilation fixes cover tensor subclasses, torch.func, autocast, metadata guards, fake-mode re-entry, in-place mutation, and several graph-break cases. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Symbolic proxy tracing and repeated lowering edge cases involving contiguous-stride expressions are fixed. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- ROCm graph capture no longer fails or hangs for callables using hipBLASLt; warmup and capture use the same stream. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- A nonzero root argument is respected instead of always broadcasting from the first tensor. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled output and gradients are corrected for overlapping-input operations. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Streams are flushed before fsync so buffered writes are persisted correctly on remote filesystems such as GCS. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Locally synchronized calls consistently return GroupMember.NON_GROUP_MEMBER on nonmember ranks. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Subgroups whose local and global source ranks differ no longer produce a zero source gradient. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Gloo and NCCL no longer crash when the world process group was not the first backend instance created. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS floor division now matches CPU and CUDA nonfinite-value semantics. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Dynamic-shape specifications for functions with **kwargs now accept call-like keys and specs nested under the variadic parameter, and ambiguous name collisions report UserError. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Strict export treats parameters from modules in unregistered Python containers as constants rather than trying to restore them from eager module state. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Non-strict export now handles tensor indexing under vmap when the index is a batched scalar tensor. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Restores parameters, buffers, and constants for non-contiguously numbered repeated module calls. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled MPS operations no longer fail with KeyError for unsigned 16-, 32-, or 64-bit index expressions. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS results are corrected for extreme values. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS handles empty indices. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Length clamping and padding remain symbolic under dynamic shapes, avoiding recompilation and data-dependent guard failures when requested length crosses signal length. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Signed right-shift export rounds negative values toward negative infinity, matching PyTorch. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Gradients for negative integer exponents now return 2.0 ** exponent instead of zero. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Invalid complex norm orders raise a clear ValueError; dynamic compilation also handles p='fro' and p='nuc' with symbolic tensor sizes. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA Graph capture is fixed for this operation. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The CPU gelsy driver no longer returns an incorrect rank because of stale pivot values between batched LAPACK calls. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Backward accuracy improves using a correctly rounded double-precision log(10) constant. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Backward accuracy improves by using named mathematical constants. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Inconsistent LU_pivots and LU_data shapes now raise a clear error instead of segfaulting. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled output strides and channels-last memory format now match eager CPU execution. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS supports strided output tensors on macOS 14 and 15. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Meta output shapes are corrected when dim=() should reduce all dimensions. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS bfloat16 tensors no longer return their input unchanged. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Backward no longer crashes for channels-last MPS tensors. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS output is corrected when a kernel spatial dimension is at least 256. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled execution rejects alpha=0 with the same error as eager execution. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA backward no longer overflows int32 for mode="max". (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- ROCm nearest interpolation no longer fails on large channels-last inputs whose output approaches 2^32 elements. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS linear backward handles inputs with more than four dimensions, and complex high-rank linear operations no longer abort on macOS 27. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS linear forward and backward handle zero in_features without aborting. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Compiled values and gradients are corrected when targets use -1 padding. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Supports second-order gradients, so grad-grad computations no longer fail for lack of a backward derivative. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The attention-mask dtype is declared to cuDNN rather than inherited from graph I/O dtype. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Memory-efficient attention remap extents, causal key bounds, and 32-bit dropout offsets are corrected. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- cuDNN variable-length SDPA is fixed. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CPU flash SDPA now returns correct results when the innermost input dimension is not contiguous. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- MPS attention kernels now propagate NaNs correctly. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- ROCm results are corrected for very large row counts when normalized size is not a multiple of 4. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Decompositions avoid data-dependent guard failures when sequence lengths are symbolic. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Quantized gather export unpacks quantized tensor inputs before lowering. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The constant_pad_nd reference decomposition is fully functional, so dynamo ONNX export no longer fails functionalization for models using pad. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Fused CPU optimizer updates and float16/bfloat16 casts are corrected. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Becomes a no-op for an empty parameter group. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Fused CPU optimizer updates and float16/bfloat16 casts are corrected. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Backward handles Boolean scalar bases and symbolic Python integer exponents without an assertion failure or NYI SymInt equality crash. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Ordinary profiling no longer imports the experimental CUPTI monitor, preventing repeated warnings and tracebacks on systems with incompatible cupti-python versions. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- The qint32 per-channel float_qparams path no longer divide-by-zero crashes, and whole-byte quantized types pack correctly. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- CUDA relu preserves signed zero. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Assigning Python integers greater than INT64_MAX no longer raises an overflow-unpacking error. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Nonempty inputs with a zero-sized shape dimension raise a clear ValueError; empty indices remain supported. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Unsupported third-order derivatives are rejected instead of silently returning an invalid result; second-order derivatives and evaluation mode are unchanged. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)
- Zero-sized spatial outputs now short-circuit to empty output and zero backward gradients, matching cuDNN and CPU behavior. (2.14.0) — [release notes](https://github.com/pytorch/pytorch/releases/tag/v2.14.0)