Dependencies

Runtime Python dependencies

From requirements.txt (top-level pinning) and setup.py:

Package	Why PyTorch needs it
numpy	Tensor ↔ ndarray interop
typing_extensions	Backports of typing features
sympy	SymInt / dynamic-shape symbolic reasoning in compile stack
networkx	FX graph utilities
jinja2	Codegen templates
fsspec	Async checkpointing (torch.distributed.checkpoint)
filelock	Inductor cache locking
mpmath	Required by sympy
setuptools	Distutils replacement

Optional Python dependencies (specific features)

Package	Feature
triton	Inductor CUDA kernel codegen (CUDA wheels bundle this)
tensorboard	Profiler trace viewer + torch.utils.tensorboard
onnx	torch.onnx.export
pytest	Running test files via pytest (test/run_test.py uses unittest)
expecttest	assertExpectedInline
hypothesis	Property-based tests

Native libraries

Linked at build / runtime:

Library	Backend	Purpose
Intel MKL / oneAPI	CPU	BLAS/LAPACK on x86
OpenBLAS / NVPL	CPU	BLAS/LAPACK on aarch64
oneDNN (MKLDNN)	CPU	Optimized conv / matmul on x86
FBGEMM	CPU	Quantized GEMM, embedding table ops
NNPACK / XNNPACK	CPU	Mobile-oriented conv kernels
OpenMP	CPU	Threaded loops
cuBLAS	CUDA	BLAS
cuDNN	CUDA	Conv / pooling / RNN kernels
cuSPARSE	CUDA	Sparse
cuSolver	CUDA	LAPACK on CUDA
MAGMA	CUDA (opt)	Heterogeneous LAPACK
cuFFT	CUDA	FFT
NCCL	CUDA	Multi-GPU collectives
CUTLASS	CUDA (opt)	GEMM kernel templates
FlashAttention	CUDA	Bundled fused attention
Triton	CUDA	Inductor codegen target
rocBLAS, hipBLAS, MIOpen	ROCm	AMD equivalents
RCCL	ROCm	NCCL equivalent
MPS framework	Apple	macOS GPU
oneAPI / SYCL	XPU	Intel GPU

Submodules

third_party/ contains git submodules. Selected major ones:

Submodule	What
third_party/nccl	NCCL
third_party/cudnn_frontend	cuDNN v8 frontend
third_party/cutlass	CUTLASS
third_party/fbgemm	FBGEMM
third_party/onnx	ONNX schema definitions
third_party/protobuf	Caffe2 + ONNX proto
third_party/pybind11	Python bindings
third_party/eigen	Linear algebra (some legacy paths)
third_party/sleef	CPU vector math (sin, cos, etc.)
third_party/kineto	Kineto profiler
third_party/foxi / onnx-tensorrt	TRT integration
third_party/XNNPACK, NNPACK, pthreadpool, cpuinfo	Mobile kernels
third_party/opentelemetry-cpp	OpenTelemetry
third_party/composable_kernel	AMD kernel library

The .gitmodules file is the source of truth.

Versioning policy

• Submodules are pinned by SHA. Updates require an explicit PR.
• Native dependency versions (cuDNN, NCCL, MKL) are pinned in cmake/External/ and the workflow definitions.
• Python deps are pinned by version range in setup.py; the test/dev requirements have stricter pins in requirements*.txt.

Where to look

File	Purpose
requirements.txt, requirements-build.txt	Build / runtime Python deps
requirements-test.txt, requirements-ci.txt	Test / CI deps
setup.py	Authoritative install_requires
.gitmodules	Native submodules
cmake/External/	External native lib detection
cmake/Dependencies.cmake	Dependency wiring