pytorch

Breaking Changes & Deprecations

• NCCL backend rewrite: nccl-legacy is scheduled for removal in the next version.
• FSDP2 behavior change: When reduce_dtype=None, it now uses grad_dtype; explicit configuration is required to maintain previous behavior (BC-breaking).
• Compatibility restrictions: Python 3.11+ (3.15 supported) is required, CUDA 13.2 wheels are provided (cu126/cu130 removed), ROCm 7.14/10.0 (gfx900/906 removed), and macOS 15+ (MPS) is required.
• Security and stability restrictions: safe_globals is required when loading nested tensors, re-execution of import torch and nested tensors within weights_only=True are blocked, and 2.15 checkpoints are incompatible with versions 2.14 and below.

Major New Features

• AOTI: Support for nested_compile_region, TMA TensorDescriptor, and launch_pdl in the C++ wrapper. Returns AOTI_RUNTIME_FAILURE instead of exceptions, reduces peak memory during compilation, and aoti_load_package directly loads the extracted directory.
• Export & AOTDispatcher: Non-strict export TORCH_TRACE logs now include compile/frame ID. Bypassing the AOTAutograd cache within inference_mode regions has been discontinued.
• Composability: histogramdd meta implementation, preservation of symbolic dimensions, and sparse compressed tensor fake-tensor tracing support.
• Quantization: Support for pruned tables in CUDA quantized embedding_bag.
• C++ Frontend: no-GIL inline C++ extension, TORCH_TARGET_VERSION compatibility check, and Windows ARM64 CUDA build support.
• Release Engineering: CUDA 13.4 (cuBLAS 13.8.1, cuDNN 9.26) Linux/Windows wheels, Triton 3.9 bundle, ROCm gfx1250 (MI450) and ROCm 10.0 CI, XPU wheels updated to Intel Deep Learning Essentials 2026.1.3, and NVSHMEM 3.7.2 upgrade.
• Platform-specific features: CUDA float8 scatter/graph annotation, CPU NEON sum/norm AVX2/SVE256, MPS logspace/complex64/grouped_mm/int_mm, ROCm RDNA4 SDPA/AOTriton 0.14.2b, XPU linalg ops/mixed-precision Adam/SDPA GQA, and ROCm MI300 FP8 2:4 sparsity support.

Performance Improvements

• Apple Silicon: MoE decode improved by 1.32–1.34x.
• AArch64: sin/cos improved by 1.7–2.4x (SLEEF 3.5-ULP), QNNPACK linear improved 2-4x, LLVM BOLT applied by default (up to 30% improvement), and Graviton 4/5 BF16 GEMV improved 5x.
• Windows: RTX 5090 GQA bf16 forward reduced from 11.5ms to 0.49ms.
• Core operations: RMSNorm backward improved 2x, cpp_wrapper overhead reduced by 40-50%, Dynamo compilation time shortened by 4.2s, Inductor GB300 HBM bandwidth increased 2x, gfx950 FlexAttention improved 1.8-2.3x, and AOTDispatcher min-cut partitioner sped up by 5%.
• CUDA/CPU: CUDA upsample improved 5-25x, H100 topk improved 2.16x, and gather/scatter latency fixed by 10-30x.

Key Bug Fixes

• Security: Fixed torch.load NEWOBJ opcode, shm_manager buffer overflow, and builtin exception allowlist vulnerabilities.
• torch.compile/Inductor: Fixed 4x RNG deactivation (2.13 regression), in-place, reshape, reduction, FP8 scaled_mm, SDPA, fusion, compile, log_ndtr, float16 scan float32 accumulation, and boolean index_put. Fixed 1000x TFLOPS error and addmm fusion.
• Triton/Inductor: Fixed float16/bfloat16 scans, XPU, CUDA Graphs, autotuning, FlexAttention, NVGEMM, and CPU backend.
• Other: Fixed fp8 isfinite errors, linspace precision (0.8105→0.8110), RNN/SDPA/grid_sample, Muon, Autograd, c10d, Distributed (port/Etcd/TCPStore/log), RPC (RRef use-after-free), Linear Algebra (2nd order forward AD, matrix_exp, CPU _int_mm), Profiler (device time, Cuspy heap, MemoryViz), FX (symbolic bool, cached graph, pickling, BSR/BSC, replace_pattern, memory leak), and Dynamo (-0.0/NaN guard, autocast, isinstance/Protocol, random stale, gradient, optimizer capturable, out= recompilation, memory leak, crash) issues.
• Platform-specific: Resolved Hopper/B200 NaN and corruption issues by including cuDNN 9.26, fixed AVX2 double-int64 conversion truncation towards 0, Arm CPU uint8 conversion wrap-around and 4-bit matmul compilation errors, and MPS linalg/torch.compile/memory management/multi-threaded crashes and incorrect results. Fixed MPS multi-threaded hang, DLPack, linalg.svd, complex operations, var/std, and int64 truncate. Fixed ROCm attention, SDPA, FP8 GEMM, Inductor BF16, topk, hipify, and RCCL warnings.