# Data flow — pytorch/pytorch

> Canonical page: https://codevetter.com/unpack/pytorch/data-flow

# Data flow — pytorch/pytorch

Scanned commit 5711f1a4586cb0a3dbb3d67538df814db48f94f4 (2026-09-22). Generated by codevetter codevetter 1.12.0 on 2026-09-22.

The core data path is: Python call -> torch.* binding (codegen'd _VariableFunctions) or torch.ops namespace -> C++ dispatcher keyed by DispatchKey -> backend kernel (CPU/CUDA/etc.) -> TensorImpl/Storage in c10/ATen, with autograd Nodes attached during forward when grad is required. Compile mode inserts a bytecode interception layer: Dynamo's PEP 523 hook traces Python frames into FX graphs, AOT Autograd splits forward/backward, and Inductor lowers to Triton/C++ kernels. Persistence flows through a custom zip format where storages are extracted via pickle persistent_id and written as raw byte records. Distributed state flows through Store-based rendezvous into backend ProcessGroups.
- torch.ops.<ns>.<op> lazily resolves through _OpNamespace.__getattr__ -> OpOverloadPacket.__getattr__ -> torch._C._get_operation_overload + _get_schema -> cached OpOverload; calling it invokes self._op (the C++ unboxed dispatcher entry). — [torch/_ops.py#L1284-1350](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_ops.py#L1284-L1350) [torch/_ops.py#L1417-1456](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_ops.py#L1417-L1456)
- OpOverload keeps a per-DispatchKey _dispatch_cache and a py_kernels table; decompose() prefers a Python CompositeImplicitAutograd kernel and otherwise redispatches into C++ via _op_dk. — [torch/_ops.py#L84-155](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_ops.py#L84-L155) [torch/_ops.py#L949-966](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_ops.py#L949-L966)
- native_functions.yaml is the schema source of truth: each entry is `func: name(ArgType...) -> Return` plus a `dispatch:` block mapping backend keys (CPU, CUDA, ...) to kernel function names; torchgen parse_native_yaml turns these into NativeFunctions + backend indices that drive all generated bindings. — [aten/src/ATen/native/README.md#L16-27](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/aten/src/ATen/native/README.md#L16-L27) [torchgen/gen.py#L149-259](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torchgen/gen.py#L149-L259)
- Autograd data flow: forward ops record grad_fn Nodes; backward() -> _engine_run_backward stashes the contextvars context in TLS then calls Variable._execution_engine.run_backward; C++ Engine::execute validates outputs, creates a GraphTask and schedules NodeTasks on per-device worker threads. — [torch/autograd/graph.py#L1100-1123](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/autograd/graph.py#L1100-L1123) [torch/csrc/autograd/engine.cpp#L1428-1460](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/csrc/autograd/engine.cpp#L1428-L1460) [torch/csrc/autograd/engine.cpp#L151-240](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/csrc/autograd/engine.cpp#L151-L240)
- torch.save path: save() -> _save() runs a PyTorchPickler whose persistent_id intercepts storages, dedupes them by _cdata into serialized_storages, enforces dtype-consistency per data_ptr, writes data.pkl + .format_version + .storage_alignment + byteorder records, then streams each storage (moved to CPU, optionally via pinned-memory D2H) as data/<key>. — [torch/serialization.py#L1184-1313](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/serialization.py#L1184-L1313)
- torch.load path: stricter _is_zipfile check -> _open_zipfile_reader (torch._C.PyTorchFileReader) -> _load() reads byteorder/.format_version/.storage_alignment, computes restore_location from map_location, and load_tensor() builds UntypedStorages (meta storages when a fake mode is active, recording _checkpoint_offset for layout-preserving reload). — [torch/serialization.py#L429-441](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/serialization.py#L429-L441) [torch/serialization.py#L1999-2128](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/serialization.py#L1999-L2128)
- weights_only=True is the default when pickle_module is unset; _weights_only_unpickler restricts pickle GLOBALs to an allowlist extensible via add_safe_globals()/safe_globals(), and get_unsafe_globals_in_checkpoint scans data.pkl for non-allowlisted globals. — [torch/serialization.py#L77-91](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/serialization.py#L77-L91) [torch/serialization.py#L269-382](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/serialization.py#L269-L382)
- torch.compile data flow: compile() -> torch._dynamo.optimize -> _optimize -> OptimizeContext wrapping the fn; the C++ PEP 523 eval-frame hook intercepts frames, symbolic_convert builds an FX graph in OutputGraph with guards, then the backend (compile_fx) is invoked and generated bytecode calls the compiled artifact with resume functions after graph breaks. — [torch/_dynamo/eval_frame.py#L1786-1925](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_dynamo/eval_frame.py#L1786-L1925) [torch/_dynamo/CLAUDE.md](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_dynamo/CLAUDE.md)
- Inductor's compile_fx orchestrates end-to-end compilation: selects the decomposition table, wakes the AsyncCompile worker pool, wraps inner_compile into fw_compiler/bw_compiler callables, and calls aot_autograd which partitions the graph and invokes those compilers before fx_codegen_and_compile emits runnable code. — [torch/_inductor/compile_fx.py#L3117-3200](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_inductor/compile_fx.py#L3117-L3200) [torch/_inductor/compile_fx.py#L3348-3410](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/_inductor/compile_fx.py#L3348-L3410)
- init_process_group resolves the default backend (Backend(...) validation, device-based default), performs rendezvous through init_method/env:// to get a Store, then _new_process_group_helper constructs the ProcessGroup; all collective payloads traverse the backend with no auth or encryption. — [torch/distributed/distributed_c10d.py#L2370-2626](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/distributed/distributed_c10d.py#L2370-L2626) [SECURITY.md#L85-89](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/SECURITY.md#L85-L89)
- Compile caches are per code object, not per frame: multiple compiled entries accumulate per code object until torch._dynamo.config.recompile_limit (default 8), after which execution falls back to eager; isolate_recompiles gives a compile call its own entry set. — [torch/__init__.py#L3117-3126](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/__init__.py#L3117-L3126) [torch/__init__.py#L3200-3218](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/torch/__init__.py#L3200-L3218)
- Distributed checkpoint (torch.distributed.checkpoint) deliberately bypasses the weights_only protections: checkpoints are assumed to come from the user's own trusted store, so loading may execute arbitrary code. — [SECURITY.md#L89](https://github.com/pytorch/pytorch/blob/5711f1a4586cb0a3dbb3d67538df814db48f94f4/SECURITY.md#L89)

Every claim links to its cited files at the pinned commit. Citations identify source evidence; they do not establish runtime correctness. Independent CodeVetter scan — not reviewed by upstream maintainers.

## Public product links

- [CodeVetter](https://codevetter.com/)
- [Download](https://codevetter.com/download)
- [Documentation](https://codevetter.com/docs/)
- [Source](https://github.com/Codevetter/codevetter)