# pylint: disable=useless-parent-delegation from __future__ import annotations import ctypes import gc import typing import warnings import weakref from collections import OrderedDict from collections.abc import Callable from typing import overload, TYPE_CHECKING, TypeAlias, TypeGuard, Union from typing_extensions import ParamSpec, Self, TypeVar import torch from torch import Tensor from torch.cuda._utils import _check_cuda_bindings try: from cuda.bindings import ( # pyrefly: ignore[missing-import] driver as _cuda_driver, runtime as _cuda_runtime, ) except ImportError: _cuda_driver = None # type: ignore[assignment] _cuda_runtime = None # type: ignore[assignment] if TYPE_CHECKING: # importing _POOL_HANDLE at runtime toplevel causes an import cycle from torch.cuda import _POOL_HANDLE, MemPool from torch.utils._cuda_debug import _CUDAGraphInputLivenessTracker from torch.utils.hooks import RemovableHandle _GraphPool: TypeAlias = _POOL_HANDLE | MemPool else: _GraphPool: TypeAlias = typing.Any from .._utils import _dummy_type __all__ = [ "is_current_stream_capturing", "graph_pool_handle", "CUDAGraph", "graph", "make_graphed_callables", "export_dot", "export_graph_data", "register_graph_capture_start_hook", "register_graph_capture_end_hook", "register_graph_instantiate_hook", "register_graph_replay_start_hook", "register_graph_replay_end_hook", "register_graph_destroy_hook", ] _R = TypeVar("_R") _P = ParamSpec("_P") def _is_mem_pool(pool: object) -> TypeGuard[MemPool]: mempool_cls = getattr(torch.cuda, "MemPool", None) return mempool_cls is not None and isinstance(pool, mempool_cls) def _get_pool_id(pool: _GraphPool | None) -> _POOL_HANDLE | None: if pool is None: return None if _is_mem_pool(pool): return typing.cast("_POOL_HANDLE", pool.id) return typing.cast("_POOL_HANDLE", pool) if not hasattr(torch._C, "_CudaStreamBase"): # Define dummy base classes torch._C.__dict__["_CUDAGraph"] = _dummy_type("_CUDAGraph") torch._C.__dict__["_graph_pool_handle"] = _dummy_type("_graph_pool_handle") torch._C.__dict__["_cuda_isCurrentStreamCapturing"] = _dummy_type( "_cuda_isCurrentStreamCapturing" ) from torch._C import _cuda_isCurrentStreamCapturing, _CUDAGraph, _graph_pool_handle def is_current_stream_capturing() -> bool: r"""Return True if CUDA graph capture is underway on the current CUDA stream, False otherwise. If a CUDA context does not exist on the current device, returns False without initializing the context. """ return _cuda_isCurrentStreamCapturing() # Python shim helps Sphinx process docstrings more reliably. def graph_pool_handle() -> _POOL_HANDLE: r"""Return an opaque token representing the id of a graph memory pool. See :ref:`Graph memory management`. .. warning:: This API is in beta and may change in future releases. """ return torch.cuda._POOL_HANDLE(_graph_pool_handle()) def _require_cuda_bindings() -> None: """Raise a uniform error if the cuda-bindings package is unavailable. ``_cuda_runtime`` and ``_cuda_driver`` are imported together, so checking one suffices. """ if _cuda_runtime is None: raise RuntimeError( "This CUDAGraph API requires the cuda-bindings package; " "install it with `pip install cuda-bindings`." ) class _RetainedCallbacks: r"""Holds destroy callbacks and retained objects for a single capture cycle of a :class:`CUDAGraph`. A :class:`CUDAGraph` owns one of these and arms a :func:`weakref.finalize` bound to :meth:`fire`. The holder deliberately does NOT reference the graph, so the finalizer cannot keep the graph alive; :meth:`fire` can still sync the replay streams because they are mirrored onto the holder, not read off the graph. Callbacks and objects live in ``OrderedDict``s keyed by ``RemovableHandle`` id so registrations can be individually removed. """ def __init__(self) -> None: self.callbacks: OrderedDict[int, Callable[[], None]] = OrderedDict() self.objects: OrderedDict[int, object] = OrderedDict() # Streams the graph was replayed on; drained before firing. A set (not a # single stream) because concurrent replays on multiple streams leave # work in flight on all of them and we cannot know which finishes last. # Stream hashes/compares by value, so same-stream replays dedup to one # entry -- the common case. Only populated when a caller requested sync. self.replay_streams: set[torch.cuda.Stream] = set() self.sync_before_fire: bool = False self._fired: bool = False def fire(self) -> None: # Fired at most once (guards the reset()-then-destroy double trigger). if self._fired: return self._fired = True if self.sync_before_fire: # Drain in-flight replays before releasing anything the graph may # reference. Guarded so a sync failure does not skip the callbacks. for stream in self.replay_streams: try: stream.synchronize() except Exception: pass for cb in list(self.callbacks.values()): try: cb() except Exception: pass # match finalizer semantics: swallow, don't abort the rest self.callbacks.clear() self.objects.clear() # Global CUDA-graph lifecycle hooks (the module-level counterpart of the per-graph # _capture_start_hooks / _capture_end_hooks / _post_instantiate_hooks / _replay_*_hooks). A # consumer (e.g. a profiler observer) registers a hook that fires for every graph, so the graph # code stays free of consumer knowledge and a consumer sees captures of graphs it did not build # (an inductor or NCCL capture); registering a hook is the opt-in. Keyed by RemovableHandle id. # The fan-out (_run_global_hooks) swallows per-hook errors so one consumer cannot break a graph # lifecycle step for another -- unlike the per-graph hooks, whose errors propagate to the owner. _global_capture_start_hooks: OrderedDict[int, Callable[[CUDAGraph], None]] = ( OrderedDict() ) _global_capture_end_hooks: OrderedDict[int, Callable[[CUDAGraph], None]] = OrderedDict() _global_instantiate_hooks: OrderedDict[int, Callable[[CUDAGraph], None]] = OrderedDict() _global_replay_start_hooks: OrderedDict[int, Callable[[CUDAGraph], None]] = ( OrderedDict() ) _global_replay_end_hooks: OrderedDict[int, Callable[[CUDAGraph], None]] = OrderedDict() def _register_global_hook( registry: OrderedDict[int, Callable[[CUDAGraph], None]], fn: Callable[[CUDAGraph], None], ) -> RemovableHandle: from torch.utils.hooks import RemovableHandle handle = RemovableHandle(registry) registry[handle.id] = fn return handle def _run_global_hooks( registry: OrderedDict[int, Callable[[CUDAGraph], None]], torch_cuda_graph: CUDAGraph, ) -> None: for fn in list(registry.values()): try: fn(torch_cuda_graph) except Exception: pass def register_graph_capture_start_hook( fn: Callable[[CUDAGraph], None], ) -> RemovableHandle: """Register a hook run with each CUDA graph as its capture begins. Returns a RemovableHandle; call ``.remove()`` to unregister. .. warning:: The hook runs with capture already live on the current stream, so it must not issue CUDA work: anything it launches is captured into the graph, and under the default ``"global"`` capture error mode an unsafe call (e.g. an allocation) raises. Querying capture state is fine. Do preparation that needs CUDA before the capture instead. """ return _register_global_hook(_global_capture_start_hooks, fn) def register_graph_capture_end_hook( fn: Callable[[CUDAGraph], None], ) -> RemovableHandle: """Register a hook run with each CUDA graph when its capture ends, while the captured ``cudaGraph_t`` is still live (see :meth:`CUDAGraph.register_capture_end_hook`). Returns a RemovableHandle; call ``.remove()`` to unregister.""" return _register_global_hook(_global_capture_end_hooks, fn) def register_graph_instantiate_hook( fn: Callable[[CUDAGraph], None], ) -> RemovableHandle: """Register a hook run with each CUDA graph right after it is instantiated. Returns a RemovableHandle; call ``.remove()`` to unregister.""" return _register_global_hook(_global_instantiate_hooks, fn) def register_graph_replay_start_hook( fn: Callable[[CUDAGraph], None], ) -> RemovableHandle: """Register a hook run with each CUDA graph at the start of every replay, just before it is launched. Returns a RemovableHandle; call ``.remove()`` to unregister. .. note:: Replay is the hot path and this fires for EVERY graph on EVERY replay -- keep it cheap. With nothing registered the cost is a single dict emptiness check. """ return _register_global_hook(_global_replay_start_hooks, fn) def register_graph_replay_end_hook( fn: Callable[[CUDAGraph], None], ) -> RemovableHandle: """Register a hook run with each CUDA graph at the end of every replay, once the replay is *enqueued* (the launch is asynchronous, so the GPU work has not completed). Fires even if the launch raised, so a start hook is always balanced by an end. Returns a RemovableHandle; call ``.remove()`` to unregister. See the hot-path note on :func:`register_graph_replay_start_hook`.""" return _register_global_hook(_global_replay_end_hooks, fn) # Graph-destroy hooks, each handed the destroyed graph's exec ids (tools_id >> 32) so a # consumer can purge its per-graph state. A CUDAGraph arms a single destroy callback (only # while a hook is registered, see _graph_destroy_hooks_active) that fans out here. _global_destroy_hooks: OrderedDict[int, Callable[[set[int]], None]] = OrderedDict() def register_graph_destroy_hook(fn: Callable[[set[int]], None]) -> RemovableHandle: """Register ``fn(exec_ids)`` to run when a CUDA graph is destroyed. Returns a handle whose ``remove()`` unregisters it.""" from torch.utils.hooks import RemovableHandle handle = RemovableHandle(_global_destroy_hooks) _global_destroy_hooks[handle.id] = fn return handle def _graph_destroy_hooks_active() -> bool: """True when any graph-destroy hook is registered -- the gate a CUDAGraph checks before arming its destroy callback.""" return bool(_global_destroy_hooks) def _run_graph_destroy_hooks(exec_graph_ids: set[int]) -> None: """Invoke every registered hook with the destroyed exec graph ids, swallowing per-hook errors so one failure does not abort the rest (matching the destroy-callback fire semantics). The single entry point a graph's destroy callback calls.""" for fn in list(_global_destroy_hooks.values()): try: fn(exec_graph_ids) except Exception: pass # Python shim helps Sphinx process docstrings more reliably. class CUDAGraph(_CUDAGraph): r"""Wrapper around a CUDA graph. Arguments: keep_graph (bool, optional): If ``keep_graph=False``, the cudaGraphExec_t will be instantiated on GPU at the end of ``capture_end`` and the underlying cudaGraph_t will be destroyed. Users who want to query or otherwise modify the underlying cudaGraph_t before instantiation can set ``keep_graph=True`` and access it via ``raw_cuda_graph`` after ``capture_end``. Note that the cudaGraphExec_t will not be instantiated at the end of ``capture_end`` in this case. Instead, it will be instantiated via an explicit call to ``instantiate`` or automatically on the first call to ``replay`` if ``instantiate`` was not already called. Calling ``instantiate`` manually before ``replay`` is recommended to prevent increased latency on the first call to ``replay``. It is allowed to modify the raw cudaGraph_t after first calling ``instantiate``, but the user must call ``instantiate`` again manually to make sure the instantiated graph has these changes. Pytorch has no means of tracking these changes. .. warning:: This API is in beta and may change in future releases. """ _tracker: _CUDAGraphInputLivenessTracker | None # Read-only property exposed from the C++ _CUDAGraph base via pybind; # annotated (not assigned) so the type checker sees it without shadowing it. _has_graph_exec: bool # Stays None unless maybe_stamp_capture_graph_id stamps it during capture_end # (requires annotations enabled and cudaGraphNodeGetToolsId available). _capture_graph_id: int | None # Exec graph id the recorded annotations are currently keyed to, or None # before the first remap. Lets a re-instantiate (which produces a fresh exec # id) rekey annotations from the previous exec id to the new one. _remapped_exec_id: int | None # Exec graph ids a consumer has recorded per-graph state under (one per # instantiate). Handed to the graph-destroy hooks on destruction so consumers # can purge that state and their maps do not grow across the run. _recorded_exec_ids: set[int] _keep_graph: bool # User hooks fired by capture_begin / capture_end / instantiate (see register_*_hook). _capture_start_hooks: dict[int, Callable[[CUDAGraph], None]] _capture_end_hooks: dict[int, Callable[[CUDAGraph], None]] _post_instantiate_hooks: dict[int, Callable[[CUDAGraph], None]] # Transient get_graph_data() cache shared across a single instantiate()'s # post-instantiate hooks, so multiple consumers (dependency recording, an # export_graph_data dump) query the topology once instead of once each. # Only populated while _caching_graph_data is set; dropped right after the # hooks run so the topology is not retained for the graph's lifetime. _caching_graph_data: bool _instantiate_graph_data: dict | None # Replay lifecycle hooks, fired around each replay (hot path -- only iterated # when non-empty). See register_replay_start_hook / register_replay_end_hook. _replay_start_hooks: dict[int, Callable[[CUDAGraph], None]] _replay_end_hooks: dict[int, Callable[[CUDAGraph], None]] # Destroy callbacks / retained objects for the current capture cycle, plus # the weakref.finalize armed on this graph that fires the holder on # collection. reset() fires the current holder and re-arms a fresh pair. _retained: _RetainedCallbacks _retained_finalizer: weakref.finalize | None def __new__(cls, keep_graph: bool = False) -> Self: instance = super().__new__(cls, keep_graph) instance._tracker = None instance._capture_graph_id = None instance._remapped_exec_id = None instance._recorded_exec_ids = set() instance._keep_graph = keep_graph # OrderedDict (not dict): RemovableHandle weak-references the mapping. instance._capture_start_hooks = OrderedDict() instance._capture_end_hooks = OrderedDict() instance._post_instantiate_hooks = OrderedDict() instance._caching_graph_data = False instance._instantiate_graph_data = None instance._replay_start_hooks = OrderedDict() instance._replay_end_hooks = OrderedDict() instance._retained_finalizer = None instance._arm_retained() return instance def _arm_retained(self) -> None: # Install a fresh destroy-callback holder and finalizer. The finalizer is # bound to the holder, NOT to `self`: a finalizer arg referencing the # graph would keep it alive forever and never fire. The holder does not # point back at the graph, so fire() can still sync the last replay # stream off it without pinning the graph. self._retained = _RetainedCallbacks() self._retained_finalizer = weakref.finalize(self, self._retained.fire) # When a consumer (e.g. the CUPTI monitor) has registered graph-destroy # hooks, arm the per-graph state purge for this capture cycle. The callback # captures the exec-id SET OBJECT (empty now, filled as this graph # records/instantiates) and the module fan-out function, never self and never # a hook: a closure reachable to the graph would pin it past collection so it # never fires. reset() re-arms a fresh holder, so this re-registers per cycle; # a graph that records nothing just fires on an empty set (a no-op). if _graph_destroy_hooks_active(): exec_ids = self._recorded_exec_ids self.register_destroy_callback(lambda: _run_graph_destroy_hooks(exec_ids)) def register_capture_start_hook( self, hook: Callable[[CUDAGraph], None] ) -> RemovableHandle: r"""Register ``hook(graph)`` to run when capture begins on this graph, right after capture is under way on the current stream. Hooks fire in registration order. Returns a handle whose ``remove()`` deregisters the hook. .. warning:: The hook runs inside the capture: any CUDA work it issues is captured into the graph, and under the default ``"global"`` capture error mode an unsafe call raises. See :func:`torch.cuda.graphs.register_graph_capture_start_hook`. """ from torch.utils.hooks import RemovableHandle handle = RemovableHandle(self._capture_start_hooks) self._capture_start_hooks[handle.id] = hook return handle def register_capture_end_hook( self, hook: Callable[[CUDAGraph], None] ) -> RemovableHandle: r"""Register ``hook(graph)`` to run when capture ends, after capture completes but before the graph is finalized. The captured ``cudaGraph_t`` is live (via :meth:`raw_cuda_graph`) for both ``keep_graph`` modes. Hooks fire in registration order. Returns a handle whose ``remove()`` deregisters the hook. """ from torch.utils.hooks import RemovableHandle handle = RemovableHandle(self._capture_end_hooks) self._capture_end_hooks[handle.id] = hook return handle def register_post_instantiate_hook( self, hook: Callable[[CUDAGraph], None] ) -> RemovableHandle: r"""Register ``hook(graph)`` to run after each instantiation (including re-instantiation, which produces a fresh exec graph). The instantiated graph is available via :meth:`raw_cuda_graph_exec`. Hooks fire in registration order. Returns a handle whose ``remove()`` deregisters the hook. """ from torch.utils.hooks import RemovableHandle handle = RemovableHandle(self._post_instantiate_hooks) self._post_instantiate_hooks[handle.id] = hook return handle def register_replay_start_hook( self, hook: Callable[[CUDAGraph], None] ) -> RemovableHandle: r"""Register ``hook(graph)`` to run at the start of every :meth:`replay`, just before the graph is launched (after any on-demand instantiation, so :meth:`raw_cuda_graph_exec` is valid). Hooks fire in registration order. Returns a handle whose ``remove()`` deregisters the hook. .. note:: Replay is the hot path and a registered hook runs on every replay -- keep it cheap. With no hook registered the cost is a single dict emptiness check. """ from torch.utils.hooks import RemovableHandle handle = RemovableHandle(self._replay_start_hooks) self._replay_start_hooks[handle.id] = hook return handle def register_replay_end_hook( self, hook: Callable[[CUDAGraph], None] ) -> RemovableHandle: r"""Register ``hook(graph)`` to run at the end of every :meth:`replay`, just after the graph is launched. The launch is asynchronous, so the hook runs once the replay is *enqueued*, not once the GPU work completes. Hooks fire in registration order. Returns a handle whose ``remove()`` deregisters the hook. See the hot-path note on :meth:`register_replay_start_hook`. End hooks fire even if the launch raises -- so a start hook is always balanced by an end -- and the launch error then propagates. (Start hooks that raise abort the replay before launch, and no end hook fires.) """ from torch.utils.hooks import RemovableHandle handle = RemovableHandle(self._replay_end_hooks) self._replay_end_hooks[handle.id] = hook return handle def register_destroy_callback( self, cb: Callable[[], None], *, synchronize_before_release: bool = False, ) -> RemovableHandle: r"""Register ``cb()`` to run when this graph is destroyed (finalized) or explicitly :meth:`reset`, just before its CUDA resources are freed. Callbacks fire once per capture cycle, in registration order; exceptions are swallowed so one failure does not abort the rest. ``cb`` must NOT reference this graph: the finalizer that fires it is held by a global registry, so a callback reachable to the graph keeps the graph alive until interpreter exit (it is never collected, hence never fired). Returns a handle whose ``remove()`` deregisters the callback. Teardown does not synchronize CUDA, and ``cudaGraphExecDestroy`` frees an in-flight graph only asynchronously, so a callback that frees device memory the graph reads/writes is a use-after-free if a replay is still in flight. Pass ``synchronize_before_release=True`` to synchronize every stream this graph was replayed on before firing. Otherwise callbacks must not free anything the graph references. """ from torch.utils.hooks import RemovableHandle if synchronize_before_release: self._retained.sync_before_fire = True handle = RemovableHandle(self._retained.callbacks) self._retained.callbacks[handle.id] = cb return handle def retain_object( self, obj: object, *, synchronize_before_release: bool = False, ) -> RemovableHandle: r"""Keep ``obj`` alive for this graph's current capture cycle and release it when the graph is destroyed (finalized) or explicitly :meth:`reset`. No callback runs; normal refcounting drops ``obj`` when the retained reference is released. Returns a handle whose ``remove()`` drops the retained reference early. As with :meth:`register_destroy_callback`, ``obj`` must NOT reference this graph, or the graph is kept alive until interpreter exit and ``obj`` is never released. ``synchronize_before_release`` has the same meaning and caveats as in :meth:`register_destroy_callback`: set it if releasing ``obj`` frees device memory the graph reads/writes (e.g. ``obj`` is the last reference to a tensor the graph uses) and replays may still be in flight. """ from torch.utils.hooks import RemovableHandle if synchronize_before_release: self._retained.sync_before_fire = True handle = RemovableHandle(self._retained.objects) self._retained.objects[handle.id] = obj return handle def _maybe_remap_annotations(self) -> None: # Remap recorded kernel annotations to the current exec graph id. No-op # unless a capture id was stamped (annotations enabled). Called from # instantiate() -- which replay() routes through for keep_graph=True -- # so every fresh exec graph (each instantiate() produces a new exec id) # rekeys the annotations; remap_to_exec_graph self-skips when the exec id # is unchanged. if self._capture_graph_id is None: return from torch.cuda._graph_annotations import remap_to_exec_graph remap_to_exec_graph(self) # Record the exec id we remapped to so the destroy callback (armed in # capture_end_post) hands it to the graph-destroy hooks for per-graph state # cleanup. A re-instantiate mints a fresh exec id, so the set accumulates each. if self._remapped_exec_id is not None: self._recorded_exec_ids.add(self._remapped_exec_id) def _release_python_resources(self) -> None: # Single source of truth for GC-critical Python resources released by # both reset() and __del__. Destroy callbacks are NOT fired here: on the # destruction path the armed weakref.finalize fires them (after __del__, # before the C++ resources are freed), so user code stays out of # __del__; reset() fires + re-arms the holder itself (see reset()). tracker, self._tracker = self._tracker, None if tracker is not None: tracker.stop() def __del__(self) -> None: try: self._release_python_resources() except Exception: pass # don't raise under GC def capture_begin( self, pool: _GraphPool | None = None, capture_error_mode: str = "global", check_input_liveness: bool = False, ) -> None: r"""Begin capturing CUDA work on the current stream. Typically, you shouldn't call ``capture_begin`` yourself. Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`, which call ``capture_begin`` internally. Arguments: pool (optional): Token (returned by :func:`~torch.cuda.graph_pool_handle` or :meth:`other_Graph_instance.pool()`) or :class:`~torch.cuda.MemPool` that hints this graph may share memory with the indicated pool. See :ref:`Graph memory management`. capture_error_mode (str, optional): specifies the cudaStreamCaptureMode for the graph capture stream. Can be "global", "thread_local" or "relaxed". During cuda graph capture, some actions, such as cudaMalloc, may be unsafe. "global" will error on actions in other threads, "thread_local" will only error for actions in the current thread, and "relaxed" will not error on these actions. Do NOT change this setting unless you're familiar with `cudaStreamCaptureMode `_ check_input_liveness (bool, optional): If ``True``, tracks external tensor inputs during graph capture and raises an error if any are deallocated before replay. This helps debug "use after free" errors where input tensors are garbage collected between capture and replay. Default: ``False``. .. note:: Custom CUDA kernels added outside PyTorch (e.g., via cuLaunchKernel or DLPack) are not tracked by this mechanism. """ if self._tracker is not None: self._tracker.stop() self._tracker = None super().capture_begin( pool=_get_pool_id(pool), capture_error_mode=capture_error_mode ) if check_input_liveness: from torch.utils._cuda_debug import _CUDAGraphInputLivenessTracker self._tracker = _CUDAGraphInputLivenessTracker() self._tracker.start() # Capture is live from here, so a hook must not issue CUDA work (see # register_capture_start_hook). Global hooks run before the per-graph ones, # matching instantiate(). if _global_capture_start_hooks: _run_global_hooks(_global_capture_start_hooks, self) for hook in list(self._capture_start_hooks.values()): hook(self) def capture_end_pre(self) -> None: r"""End capture but do not finalize: leaves the captured ``cudaGraph_t`` live (for both ``keep_graph`` modes) so it can be inspected before :meth:`capture_end_post` instantiates and/or destroys it.""" super().capture_end_pre() def capture_end_post(self) -> None: r"""Finalize a capture started by :meth:`capture_end_pre`: destroy the template when ``keep_graph=False`` (the graph must already be instantiated; :meth:`capture_end` and the context manager do so).""" super().capture_end_post() if self._tracker is not None: self._tracker.stop() def capture_end(self) -> None: r"""End CUDA graph capture on the current stream. After ``capture_end``, ``replay`` may be called on this instance. Typically, you shouldn't call ``capture_end`` yourself. Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`, which call ``capture_end`` internally. """ self.capture_end_pre() # Run the in-window work (stamp, then user capture-end hooks) while the # template is live (both keep_graph modes). Errors here are unexpected # (stamp) or user bugs (hooks) and propagate -- we deliberately don't # wrap them in a finally that calls capture_end_post(), since a failing # finalize would mask the real error. from torch.cuda._graph_annotations import maybe_stamp_capture_graph_id maybe_stamp_capture_graph_id(self) if _global_capture_end_hooks: _run_global_hooks(_global_capture_end_hooks, self) for hook in list(self._capture_end_hooks.values()): hook(self) if not self._keep_graph: # Route keep_graph=False instantiation through Python so the remap # and post-instantiate hooks fire from instantiate(). self.instantiate() self.capture_end_post() def instantiate(self) -> None: r"""Instantiate the CUDA graph. Will be called by ``capture_end`` if ``keep_graph=False``, or by ``replay`` if ``keep_graph=True`` and ``instantiate`` has not already been explicitly called. Does not destroy the cudaGraph_t returned by ``raw_cuda_graph``. """ super().instantiate() self._maybe_remap_annotations() # Fire registered instantiate hooks (e.g. a profiler observer inspecting the freshly # instantiated graph). Registering a hook is the opt-in; no consumer -> a no-op. # get_graph_data() is cached for the duration of the hooks so several consumers share # one query; the cache is dropped afterwards so nothing is retained for the graph. self._caching_graph_data = True try: if _global_instantiate_hooks: _run_global_hooks(_global_instantiate_hooks, self) for hook in list(self._post_instantiate_hooks.values()): hook(self) finally: self._caching_graph_data = False self._instantiate_graph_data = None def replay(self) -> None: r"""Replay the CUDA work captured by this graph.""" if self._tracker is not None: self._tracker.check_alive(self.pools()) if self._retained.sync_before_fire: # Record the replay stream on the holder so fire() can drain it # before releasing graph-referenced resources, without the holder # referencing the graph. Only recorded when a caller opted into sync. self._retained.replay_streams.add(torch.cuda.current_stream()) # With keep_graph=True the exec graph is instantiated on demand here on # the first replay; the C++ replay() requires it to already exist. The # annotation remap rides on instantiate(), so it is handled by that call. if not self._has_graph_exec: self.instantiate() if _global_replay_start_hooks: _run_global_hooks(_global_replay_start_hooks, self) if self._replay_start_hooks: for hook in list(self._replay_start_hooks.values()): hook(self) try: super().replay() finally: if _global_replay_end_hooks: _run_global_hooks(_global_replay_end_hooks, self) if self._replay_end_hooks: for hook in list(self._replay_end_hooks.values()): hook(self) def reset(self) -> None: r"""Delete the graph currently held by this instance.""" self._release_python_resources() # also-fire-on-reset: reset() destroys this capture's CUDA resources and # the graph may be re-captured, so fire the current destroy callbacks and # re-arm a fresh holder + finalizer for the next capture cycle (the old # finalizer is bound to the now-spent holder). detach() cancels it and # drops its reference to the spent holder promptly. self._retained.fire() if self._retained_finalizer is not None: self._retained_finalizer.detach() # Start a fresh id set BEFORE re-arming: _arm_retained captures it into the # next cycle's destroy callback. The just-fired holder dropped the old one. self._recorded_exec_ids = set() self._arm_retained() # Reset-only state: scrubbed here because the object is reused after # reset(); on death these ints die with the object. self._capture_graph_id = None self._remapped_exec_id = None super().reset() def pool(self) -> _POOL_HANDLE: r"""Return an opaque token representing the id of this graph's memory pool. This id can optionally be passed to another graph's ``capture_begin``, which hints the other graph may share the same memory pool. """ return torch.cuda._POOL_HANDLE(super().pool()) def pools(self) -> list[_POOL_HANDLE]: r"""Return opaque tokens for all memory pools retained by this graph.""" return [torch.cuda._POOL_HANDLE(pool) for pool in super().pools()] def _retain_pool(self, pool: _GraphPool) -> None: pool_id = _get_pool_id(pool) if pool_id is None: raise RuntimeError("CUDAGraph._retain_pool expected a memory pool") return super()._retain_pool(pool_id) def enable_debug_mode(self) -> None: r"""Retain the captured graph (equivalent to ``keep_graph=True``) so it can be inspected, e.g. via :meth:`debug_dump`. Kept for backward compatibility.""" super().enable_debug_mode() # Keep the Python-side flag in sync with the C++ keep_graph_ it now sets. self._keep_graph = True def debug_dump(self, debug_path: str, *, verbose: bool = True) -> None: r"""Dump the captured graph to ``debug_path`` in Graphviz DOT format. The graph's template must be live: ``keep_graph=True`` (or :meth:`enable_debug_mode`), or called from a capture-end hook. Requires the ``cuda.bindings`` package. Arguments: debug_path (required): Path to dump the graph to. verbose: If ``True`` (default), use the most verbose DOT output. """ _dump_graph_dot(self, debug_path, verbose=verbose) def raw_cuda_graph(self) -> int: r"""Returns the underlying cudaGraph_t. The template must be live: this requires ``keep_graph=True`` (it persists after ``capture_end``), or access from within a capture-end hook (before the template is destroyed for ``keep_graph=False``). See the following for APIs for how to manipulate this object: `Graph Management `_ and `cuda-python Graph Management bindings `_ """ return super().raw_cuda_graph() def raw_cuda_graph_exec(self) -> int: r"""Returns the underlying cudaGraphExec_t. ``instantiate`` must have been called if ``keep_graph`` is True, or ``capture_end`` must have been called if ``keep_graph`` is False. If you call ``instantiate()`` after ``raw_cuda_graph_exec()``, the previously returned cudaGraphExec_t will be destroyed. It is your responsibility not to use this object after destruction. See the following for APIs for how to manipulate this object: `Graph Execution `_ and `cuda-python Graph Execution bindings `_ """ return super().raw_cuda_graph_exec() def get_graph_data(self) -> dict: r"""Return a dictionary describing the graph's topology and node metadata. ``keep_graph`` must be True. The graph must have been instantiated (via :meth:`instantiate`) before calling this method. Requires the ``cuda.bindings`` package. Returns a dictionary with structure:: { "exec_graph_id": int, "nodes": [ { "index": int, "node_type": str, "tools_id": int, "graph_id": int, "node_id": int, "kernel_name": str or None, "event_ptr": int, "host_fn_addr": int, "host_fn_name": str or None, "dependencies": [int, ...], "dependents": [int, ...], }, ..., ], } ``event_ptr`` is the ``cudaEvent_t`` handle (as an int) an event-record or event-wait node records / waits on -- these nodes produce no timed CUPTI record, so matching a wait to the record that signals it is the only way to reason about the cross-stream sync it encodes. It is ``0`` for other node types. ``host_fn_addr`` / ``host_fn_name`` are populated for host nodes (a CPU callback run as a graph node): the callback address and a best-effort demangled symbol name for it (``None`` when it resolves to no exported symbol). They are ``0`` / ``None`` for other node types. Each node's ``graph_id`` is remapped to the exec graph id so that ``tools_id`` values match those reported by CUPTI-based profilers. ``dependencies`` and ``dependents`` are lists of node indices within the ``nodes`` list. This structure is useful for inspecting a profiler trace and establishing whether a particular dependency observed in the profile is a true dependency (encoded in the graph) or a fake dependency caused by mapping of independent streams to the same hardware channel. Child-graph and conditional nodes are reported as nodes in their own right, but this walk does not descend into their bodies (those are separate ``cudaGraph_t`` objects), so the work inside them is absent from ``nodes`` and a warning is issued. The ids that *are* reported stay valid: the exec graph preserves top-level node ids. """ # Serve the shared cache when instantiate() has it live (see _caching_graph_data). if self._instantiate_graph_data is not None: return self._instantiate_graph_data from torch.cuda._graph_annotations import _is_tools_id_unavailable _require_cuda_bindings() # Narrow for the type checker (cuda bindings are present past the check). if _cuda_runtime is None or _cuda_driver is None: raise AssertionError( "expected _cuda_runtime and _cuda_driver to be not None" ) if _is_tools_id_unavailable(): raise RuntimeError( "get_graph_data requires cudaGraphNodeGetToolsId which needs " "cuda.bindings >= 13.1 and CUDA driver >= 13.1 " "(or cuda-compat >= 13.1 in LD_LIBRARY_PATH)" ) node_type_names = { _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeKernel: "kernel", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeMemcpy: "memcpy", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeMemset: "memset", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeHost: "host", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeGraph: "child_graph", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeEmpty: "empty", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeWaitEvent: "wait_event", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeEventRecord: "event_record", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeMemAlloc: "mem_alloc", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeMemFree: "mem_free", _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeConditional: "conditional", } # Node types whose work lives in a separate cudaGraph_t (see the warning below). nested_graph_types = { _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeGraph, _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeConditional, } nested_node_types: set[str] = set() raw = self.raw_cuda_graph() _, num = _check_cuda_bindings(_cuda_runtime.cudaGraphGetNodes(raw, numNodes=0)) nodes, num = _check_cuda_bindings( _cuda_runtime.cudaGraphGetNodes(raw, numNodes=num) ) handle_to_idx: dict[int, int] = {} node_infos: list[dict] = [] for i in range(num): node = nodes[i] handle_to_idx[int(node)] = i ntype = _check_cuda_bindings(_cuda_runtime.cudaGraphNodeGetType(node)) if ntype in nested_graph_types: nested_node_types.add(node_type_names.get(ntype, ntype)) tools_id = _check_cuda_bindings(_cuda_runtime.cudaGraphNodeGetToolsId(node)) graph_id = tools_id >> 32 node_id = tools_id & 0xFFFFFFFF kernel_name = None if ntype == _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeKernel: cu_node = _cuda_driver.CUgraphNode(init_value=int(node)) err, params = _cuda_driver.cuGraphKernelNodeGetParams(cu_node) if err == _cuda_driver.CUresult.CUDA_SUCCESS and int(params.func): cu_func = _cuda_driver.CUfunction(init_value=int(params.func)) err, name = _cuda_driver.cuFuncGetName(cu_func) if err == _cuda_driver.CUresult.CUDA_SUCCESS: kernel_name = name.decode() if isinstance(name, bytes) else name # Event record/wait nodes carry a cudaEvent_t but emit no timed CUPTI record; # capture the handle so a wait node can be matched to the record that signals it. # Not best-effort like the kernel-name lookup above (a name can legitimately be # unavailable): the node type is already established here, so these calls are # expected to succeed. Swallowing a failure would leave event_ptr 0 and make the # record/wait match quietly wrong rather than loud. event_ptr = 0 if ntype == _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeEventRecord: event_ptr = int( _check_cuda_bindings( _cuda_runtime.cudaGraphEventRecordNodeGetEvent(node) ) ) elif ntype == _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeWaitEvent: event_ptr = int( _check_cuda_bindings( _cuda_runtime.cudaGraphEventWaitNodeGetEvent(node) ) ) # Host nodes carry no name in the CUPTI record; recover the callback address # from the node params and resolve a best-effort symbol name (see # _resolve_host_fn_name). Both are None/0 for other node types. host_fn_addr = 0 host_fn_name = None if ntype == _cuda_runtime.cudaGraphNodeType.cudaGraphNodeTypeHost: err, params = _cuda_runtime.cudaGraphHostNodeGetParams(node) if err == _cuda_runtime.cudaError_t.cudaSuccess: host_fn_addr = int(params.fn) host_fn_name = _resolve_host_fn_name(host_fn_addr) node_infos.append( { "index": i, "node_type": node_type_names.get(ntype, str(ntype)), "tools_id": tools_id, "graph_id": graph_id, "node_id": node_id, "kernel_name": kernel_name, "event_ptr": event_ptr, "host_fn_addr": host_fn_addr, "host_fn_name": host_fn_name, "dependencies": [], "dependents": [], } ) _, _, _, num_edges = _check_cuda_bindings( _cuda_runtime.cudaGraphGetEdges(raw, numEdges=0) ) if num_edges > 0: from_nodes, to_nodes, _edge_data, num_edges = _check_cuda_bindings( _cuda_runtime.cudaGraphGetEdges(raw, numEdges=num_edges) ) for i in range(num_edges): src = handle_to_idx.get(int(from_nodes[i])) dst = handle_to_idx.get(int(to_nodes[i])) if src is not None and dst is not None: node_infos[src]["dependents"].append(dst) node_infos[dst]["dependencies"].append(src) exec_handle = _cuda_runtime.cudaGraphExec_t( init_value=self.raw_cuda_graph_exec() ) exec_graph_id = _check_cuda_bindings( _cuda_runtime.cudaGraphExecGetId(exec_handle) ) for info in node_infos: info["tools_id"] = (exec_graph_id << 32) | info["node_id"] info["graph_id"] = exec_graph_id if nested_node_types: # What is returned stays correct: the exec graph preserves top-level node # ids and numbers the nested ones after them. Those nested nodes do execute # and do show up in a profiler trace (under this exec graph id, with the # appended node ids), they are just missing from the topology below. warnings.warn( f"get_graph_data: this graph contains {sorted(nested_node_types)} " "nodes; the work inside them is in a separate cudaGraph_t that this " "walk does not descend into, so it is absent from the returned nodes", stacklevel=2, ) data = { "exec_graph_id": exec_graph_id, "nodes": node_infos, } if self._caching_graph_data: self._instantiate_graph_data = data return data class _DlInfo(ctypes.Structure): _fields_ = [ ("dli_fname", ctypes.c_char_p), ("dli_fbase", ctypes.c_void_p), ("dli_sname", ctypes.c_char_p), ("dli_saddr", ctypes.c_void_p), ] def _resolve_host_fn_name(addr: int) -> str | None: """Best-effort symbol name for a host-node callback address via dladdr + demangle. Returns None when the address resolves to no exported symbol (e.g. an anonymous trampoline). Host nodes carry no name in the CUPTI record, so this is the only source.""" if not addr: return None info = _DlInfo() if not ctypes.CDLL(None).dladdr(ctypes.c_void_p(addr), ctypes.byref(info)): return None if not info.dli_sname: return None return torch._C._demangle(info.dli_sname.decode()) def _dump_graph_dot(cuda_graph: CUDAGraph, path: str, *, verbose: bool = True) -> None: """Write ``cuda_graph``'s template to ``path`` in Graphviz DOT format.""" _require_cuda_bindings() flags = ( _cuda_runtime.cudaGraphDebugDotFlags.cudaGraphDebugDotFlagsVerbose # pyrefly: ignore[missing-attribute] if verbose else 0 ) _check_cuda_bindings( _cuda_runtime.cudaGraphDebugDotPrint( # pyrefly: ignore[missing-attribute] cuda_graph.raw_cuda_graph(), path.encode(), flags ) ) def export_dot(path: str, *, verbose: bool = True) -> Callable[[CUDAGraph], None]: r"""Return a capture-end hook that dumps the captured graph to ``path`` in Graphviz DOT format. Register it with :meth:`CUDAGraph.register_capture_end_hook`; works for both ``keep_graph`` modes since it runs while the template is still live.""" def _hook(cuda_graph: CUDAGraph) -> None: _dump_graph_dot(cuda_graph, path, verbose=verbose) return _hook def export_graph_data(path: str) -> Callable[[CUDAGraph], None]: r"""Return a post-instantiate hook that pickles :meth:`CUDAGraph.get_graph_data` to ``path``. Register it with :meth:`CUDAGraph.register_post_instantiate_hook`: ``get_graph_data`` needs the graph instantiated (it remaps node ids to the exec graph id), and at post-instantiate time the template is still live, so this works for both ``keep_graph`` modes.""" def _hook(cuda_graph: CUDAGraph) -> None: import pickle with open(path, "wb") as f: pickle.dump(cuda_graph.get_graph_data(), f) return _hook class graph: r"""Context-manager that captures CUDA work into a :class:`torch.cuda.CUDAGraph` object for later replay. See :ref:`CUDA Graphs ` for a general introduction, detailed use, and constraints. Arguments: cuda_graph (torch.cuda.CUDAGraph): Graph object used for capture. pool (optional): Opaque token (returned by a call to :func:`~torch.cuda.graph_pool_handle()` or :meth:`other_Graph_instance.pool()`) or :class:`~torch.cuda.MemPool` hinting this graph's capture may share memory from the specified pool. See :ref:`Graph memory management`. stream (torch.cuda.Stream, optional): If supplied, will be set as the current stream in the context. If not supplied, ``graph`` sets its own internal side stream as the current stream in the context. capture_error_mode (str, optional): specifies the cudaStreamCaptureMode for the graph capture stream. Can be "global", "thread_local" or "relaxed". During cuda graph capture, some actions, such as cudaMalloc, may be unsafe. "global" will error on actions in other threads, "thread_local" will only error for actions in the current thread, and "relaxed" will not error on actions. Do NOT change this setting unless you're familiar with `cudaStreamCaptureMode `_ enable_annotations (bool, optional): If ``True``, enables kernel annotation recording on entry and automatically calls :func:`~torch.cuda._graph_annotations.resolve_pending_annotations` before the capture ends. Annotations are **not** cleared on exit so that multiple graphs in the same workload can accumulate annotations. Requires ``cuda.bindings`` package and cuda-compat >= 13.1 or CUDA driver >= 13.1. check_input_liveness (bool, optional): If ``True``, tracks external tensor inputs during graph capture and raises an error if any are deallocated before replay. This helps debug "use after free" errors where input tensors are garbage collected between capture and replay. Default: ``False``. .. note:: Custom CUDA kernels added outside PyTorch (e.g., via cuLaunchKernel or DLPack) are not tracked by this mechanism. .. note:: For effective memory sharing, if you pass a ``pool`` used by a previous capture and the previous capture used an explicit ``stream`` argument, you should pass the same ``stream`` argument to this capture. .. warning:: This API is in beta and may change in future releases. .. _cudaStreamCaptureMode: https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85 """ default_capture_stream: torch.cuda.Stream | None = None def __init__( self, cuda_graph: CUDAGraph, pool: _GraphPool | None = None, stream: torch.cuda.Stream | None = None, capture_error_mode: str = "global", enable_annotations: bool = False, check_input_liveness: bool = False, ): # Lazy-init of default_capture_stream helps avoid circular-import errors. # Not thread safe, but graphs already have the general (explicitly documented) # restriction that only one capture may be underway at a time in the process. if stream is None and self.__class__.default_capture_stream is None: self.__class__.default_capture_stream = torch.cuda.Stream() pool_id = _get_pool_id(pool) self.pool: tuple[()] | tuple[_POOL_HANDLE] = ( () if pool_id is None else (pool_id,) ) self.capture_stream = ( stream if stream is not None else self.__class__.default_capture_stream ) if self.capture_stream is None: raise AssertionError("capture_stream must not be None") self.stream_ctx = torch.cuda.stream(self.capture_stream) self.cuda_graph = cuda_graph self.capture_error_mode = capture_error_mode self._enable_annotations = enable_annotations self.check_input_liveness = check_input_liveness def __enter__(self) -> None: # Free as much memory as we can for the graph torch.cuda.synchronize() if torch.compiler.config.force_cudagraph_gc: # Originally we unconditionally garbage collected here. On one hand # that's nice because we have a chance to collect more memory, but # on the other hand it is REALLY expensive, especially for doing # multiple cudagraph captures in a row. In theory it will only help # when a dead python cycle is holding onto CUDA memory. gc.collect() torch.cuda.empty_cache() # pyrefly: ignore [missing-attribute] torch._C._host_emptyCache() # Scope annotation recording to this capture: stamp/mark_kernels gate on # this flag, and __exit__ always clears it. from torch.cuda._graph_annotations import _set_annotations_enabled _set_annotations_enabled(self._enable_annotations) # Stackoverflow seems comfortable with this pattern # https://stackoverflow.com/questions/26635684/calling-enter-and-exit-manually#39172487 self.stream_ctx.__enter__() self.cuda_graph.capture_begin( # type: ignore[misc] *self.pool, # pyrefly: ignore [bad-keyword-argument] capture_error_mode=self.capture_error_mode, # pyrefly: ignore [bad-keyword-argument] check_input_liveness=self.check_input_liveness, ) # The capture stream is now capturing into the top-level graph; remember it so # mark_kernels can tell a conditional-node body apart from this graph. from torch.cuda._graph_annotations import maybe_stamp_capture_root maybe_stamp_capture_root(torch.cuda.current_stream()) def __exit__(self, *args: object) -> None: from torch.cuda._graph_annotations import ( _set_annotations_enabled, resolve_pending_annotations, ) try: if self._enable_annotations: resolve_pending_annotations() # capture_end stamps the capture id and, for keep_graph=False, # instantiates (which remaps annotations to the exec id). For # keep_graph=True the remap is owned by the later instantiate()/replay(). self.cuda_graph.capture_end() self.stream_ctx.__exit__(*args) finally: # Annotation recording is capture-scoped; clear it unconditionally. _set_annotations_enabled(False) # returning None should propagate exceptions from either capture_end or stream_ctx.__exit__() _ModuleOrCallable: TypeAlias = Union["torch.nn.Module", Callable[..., object]] @overload def make_graphed_callables( callables: _ModuleOrCallable, sample_args: tuple[Tensor, ...], num_warmup_iters: int = 3, allow_unused_input: bool = False, pool: _GraphPool | None = None, capture_error_mode: str = "global", enable_annotations: bool = False, ) -> _ModuleOrCallable: ... @overload def make_graphed_callables( callables: tuple[_ModuleOrCallable, ...], sample_args: tuple[tuple[Tensor, ...], ...], num_warmup_iters: int = 3, allow_unused_input: bool = False, pool: _GraphPool | None = None, capture_error_mode: str = "global", enable_annotations: bool = False, ) -> tuple[_ModuleOrCallable, ...]: ... def make_graphed_callables( callables: _ModuleOrCallable | tuple[_ModuleOrCallable, ...], sample_args: tuple[Tensor, ...] | tuple[tuple[Tensor, ...], ...], num_warmup_iters: int = 3, allow_unused_input: bool = False, pool: _GraphPool | None = None, capture_error_mode: str = "global", enable_annotations: bool = False, ) -> _ModuleOrCallable | tuple[_ModuleOrCallable, ...]: r"""Accept callables (functions or :class:`nn.Module`\ s) and returns graphed versions. Each graphed callable's forward pass runs its source callable's forward CUDA work as a CUDA graph inside a single autograd node. The graphed callable's forward pass also appends a backward node to the autograd graph. During backward, this node runs the callable's backward work as a CUDA graph. Therefore, each graphed callable should be a drop-in replacement for its source callable in an autograd-enabled training loop. See :ref:`Partial-network capture` for detailed use and constraints. If you pass a tuple of several callables, their captures will use the same memory pool. See :ref:`Graph memory management` for when this is appropriate. Arguments: callables (torch.nn.Module or Python function, or tuple of these): Callable or callables to graph. See :ref:`Graph memory management` for when passing a tuple of callables is appropriate. If you pass a tuple of callables, their order in the tuple must be the same order they'll run in the live workload. sample_args (tuple of Tensors, or tuple of tuples of Tensors): Samples args for each callable. If a single callable was passed, ``sample_args`` must be a single tuple of argument Tensors. If a tuple of callables was passed, ``sample_args`` must be tuple of tuples of argument Tensors. num_warmup_iters (int): The number of warmup iterations. Currently, ``DataDistributedParallel`` needs 11 iterations for warm up. Default: ``3``. allow_unused_input (bool): If False, specifying inputs that were not used when computing outputs (and therefore their grad is always zero) is an error. Defaults to False. pool (optional): Token (returned by :func:`~torch.cuda.graph_pool_handle` or :meth:`other_Graph_instance.pool()`) or :class:`~torch.cuda.MemPool` that hints this graph may share memory with the indicated pool. See :ref:`Graph memory management`. enable_annotations (bool, optional): If ``True``, the forward and backward captures record kernel annotations from :func:`torch.cuda.graph_annotations.mark_kernels` scopes inside the callables (backward kernels are tagged via the scopes' autograd node hooks). See :mod:`torch.cuda.graph_annotations`. Default: ``False``. .. note:: The ``requires_grad`` state of each Tensor in ``sample_args`` must match the state that's expected for the corresponding real input in the training loop. .. warning:: This API is in beta and may change in future releases. .. warning:: ``sample_args`` for each callable must contain only Tensors. Other types are not allowed. .. warning:: Returned callables do not support higher order differentiation (e.g., double backward). .. warning:: In any :class:`~torch.nn.Module` passed to :func:`~make_graphed_callables`, only parameters may be trainable. Buffers must have ``requires_grad=False``. .. warning:: After you pass a :class:`torch.nn.Module` through :func:`~make_graphed_callables`, you may not add or remove any of that Module's parameters or buffers. .. warning:: :class:`torch.nn.Module`\s passed to :func:`~torch.cuda.make_graphed_callables` must not have module hooks registered on them at the time they are passed. However, registering hooks on modules *after* passing them through :func:`~torch.cuda.make_graphed_callables` is allowed. .. warning:: When running a graphed callable, you must pass its arguments in the same order and format they appeared in that callable's ``sample_args``. .. warning:: The automatic mixed precision is supported in :func:`~torch.cuda.make_graphed_callables` only with disabled caching. The context manager `torch.cuda.amp.autocast()` must have `cache_enabled=False`. """ if torch.is_autocast_enabled() and torch.is_autocast_cache_enabled(): raise RuntimeError( "make_graphed_callables does not support the autocast caching. Please set `cache_enabled=False`." ) just_one_callable = False _sample_args: tuple[tuple[Tensor, ...], ...] if not isinstance(callables, tuple): just_one_callable = True callables = (callables,) _sample_args = (typing.cast(tuple[Tensor, ...], sample_args),) else: _sample_args = typing.cast(tuple[tuple[Tensor, ...], ...], sample_args) flatten_sample_args = [] for c, args in zip(callables, _sample_args): if isinstance(c, torch.nn.Module): if not ( len(c._backward_hooks) == 0 and len(c._forward_hooks) == 0 and len(c._forward_pre_hooks) == 0 ): raise AssertionError( "Modules must not have hooks registered at the time they are passed. However, registering hooks " + "on modules after passing them through make_graphed_callables is allowed." ) if not all(b.requires_grad is False for b in c.buffers()): raise AssertionError( "In any :class:`~torch.nn.Module` passed to " + ":func:`~make_graphed_callables`, only parameters may be trainable. All buffers must have " + "``requires_grad=False``." ) flatten_arg = torch.utils._pytree.arg_tree_leaves(*args) flatten_sample_args.append(tuple(flatten_arg)) if not all(isinstance(arg, torch.Tensor) for arg in flatten_arg): raise AssertionError( "In the beta API, sample_args " + "for each callable must contain only Tensors. Other types are not allowed." ) # If a callable is an nn.Module, its graph's full input surface is the args the user explicitly # passes to forward (ie, its sample_args) AND the module's parameter attributes. per_callable_len_user_args = [len(args) for args in flatten_sample_args] per_callable_module_params = [ tuple(c.parameters()) if isinstance(c, torch.nn.Module) else () for c in callables ] per_callable_static_input_surfaces = [ flatten_sample_args[i] + per_callable_module_params[i] for i in range(len(callables)) ] fwd_graphs = [torch.cuda.CUDAGraph() for _ in range(len(callables))] bwd_graphs = [torch.cuda.CUDAGraph() for _ in range(len(callables))] mempool = graph_pool_handle() if pool is None else _get_pool_id(pool) # Warmup # Hopefully prevents cudnn benchmarking and other lazy-initialization cuda work # from ending up in any captures. torch.cuda.synchronize() # hipBLASLt handles are per-(device, stream) on ROCm and lazily created. # Need to use the same stream for warmup and capture to avoid capture errors. stream = torch.cuda.Stream() with torch.cuda.stream(stream): for func, args, static_input_surface in zip( callables, _sample_args, per_callable_static_input_surfaces ): grad_inputs, outputs, outputs_grad = None, None, None for _ in range(num_warmup_iters): outputs = torch.utils._pytree.tree_leaves(func(*args)) outputs_grad = tuple(o for o in outputs if o.requires_grad) if len(outputs_grad) > 0: grad_inputs = torch.autograd.grad( outputs=outputs_grad, inputs=tuple( i for i in static_input_surface if i.requires_grad ), grad_outputs=tuple( torch.empty_like(o) for o in outputs if o.requires_grad ), only_inputs=True, allow_unused=allow_unused_input, ) for v in [outputs, outputs_grad, grad_inputs]: del v torch.cuda.synchronize() # All captures here share a mempool. To avoid replays corrupting each other's memory, # the safest approach is to capture all passes in the same order they'll run: # fwd 1, fwd 2, ... fwd N, then bwd N, bwd N-1, ... bwd 1. # Capture forward graphs per_callable_static_outputs = [] per_callable_output_unflatten_spec = [] for func, args, fwd_graph in zip(callables, _sample_args, fwd_graphs): with torch.cuda.graph( fwd_graph, stream=stream, pool=mempool, capture_error_mode=capture_error_mode, enable_annotations=enable_annotations, ): func_outputs = func(*args) flatten_outputs, spec = torch.utils._pytree.tree_flatten(func_outputs) per_callable_static_outputs.append(tuple(flatten_outputs)) per_callable_output_unflatten_spec.append(spec) # Capture backward graphs in reverse order per_callable_static_grad_outputs = [] per_callable_static_grad_inputs = [] for static_input_surface, static_outputs, bwd_graph in zip( reversed(per_callable_static_input_surfaces), reversed(per_callable_static_outputs), reversed(bwd_graphs), ): # For now, assumes all static_outputs require grad # assert all(o.requires_grad for o in static_outputs), "Outputs of graphed callables must require grad." static_grad_outputs = tuple( torch.empty_like(o) if o.requires_grad else None for o in static_outputs ) outputs_grad = tuple(o for o in static_outputs if o.requires_grad) grad_inputs = None if len(outputs_grad) > 0: with torch.cuda.graph( bwd_graph, stream=stream, pool=mempool, capture_error_mode=capture_error_mode, enable_annotations=enable_annotations, ): grad_inputs = torch.autograd.grad( outputs=outputs_grad, inputs=tuple(i for i in static_input_surface if i.requires_grad), grad_outputs=tuple(o for o in static_grad_outputs if o is not None), only_inputs=True, allow_unused=allow_unused_input, ) # Constructs a tuple suitable for returning from Graphed.backward: # Pads out the actually-needed grads with Nones in gradient slots for inputs that don't require grad. # I couldn't think of a slick one-liner for this pattern. static_grad_inputs = [] grad_idx = 0 for arg in static_input_surface: if arg.requires_grad and grad_inputs is not None: static_grad_inputs.append(grad_inputs[grad_idx]) grad_idx += 1 else: static_grad_inputs.append(None) # type: ignore[arg-type] static_grad_inputs = tuple(static_grad_inputs) # type: ignore[assignment] per_callable_static_grad_outputs.append(static_grad_outputs) per_callable_static_grad_inputs.append(static_grad_inputs) # Reverses the most recent two lists per_callable_static_grad_outputs.reverse() per_callable_static_grad_inputs.reverse() # Now for every per_callable list, per_callable_*[i] holds the stuff for the ith callable. def make_graphed_autograd_function( fwd_graph: CUDAGraph, bwd_graph: CUDAGraph, module_params: tuple[torch.nn.Parameter, ...], len_user_args: int, output_unflatten_spec: torch.utils._pytree.TreeSpec, static_input_surface: tuple[Tensor, ...], static_outputs: tuple[Tensor, ...], static_grad_outputs: tuple[Tensor | None, ...], static_grad_inputs: tuple[Tensor, ...], ) -> Callable[..., object]: class Graphed(torch.autograd.Function): @staticmethod # pyrefly: ignore [bad-override] def forward(ctx: object, *inputs: Tensor) -> tuple[Tensor, ...]: # At this stage, only the user args may (potentially) be new tensors. for i in range(len_user_args): if static_input_surface[i].data_ptr() != inputs[i].data_ptr(): static_input_surface[i].copy_(inputs[i]) fwd_graph.replay() if not isinstance(static_outputs, tuple): raise AssertionError( f"static_outputs must be tuple, got {type(static_outputs)}" ) return tuple(o.detach() for o in static_outputs) @staticmethod @torch.autograd.function.once_differentiable # pyrefly: ignore [bad-override] def backward(ctx: object, *grads: Tensor) -> tuple[Tensor, ...]: if len(grads) != len(static_grad_outputs): raise AssertionError( f"len(grads)={len(grads)} != len(static_grad_outputs)={len(static_grad_outputs)}" ) for g, grad in zip(static_grad_outputs, grads): if g is not None: # don't copy if autograd gods have been kind and the # incoming grad is already in the right place if g.data_ptr() != grad.data_ptr(): g.copy_(grad) bwd_graph.replay() # Input args that didn't require grad expect a None gradient. if not isinstance(static_grad_inputs, tuple): raise AssertionError( f"static_grad_inputs must be tuple, got {type(static_grad_inputs)}" ) return tuple( # pyrefly: ignore [bad-argument-type] b.detach() if b is not None else b for b in static_grad_inputs ) def functionalized(*user_args: object) -> object: # Runs the autograd function with inputs == all inputs to the graph that might require grad # (explicit user args + module parameters) # Assumes module params didn't change since capture. flatten_user_args = torch.utils._pytree.arg_tree_leaves(*user_args) out = Graphed.apply(*(tuple(flatten_user_args) + module_params)) return torch.utils._pytree.tree_unflatten(out, output_unflatten_spec) return functionalized # Put together the final graphed callables ret: list[_ModuleOrCallable] = [] for i, func in enumerate(callables): graphed = make_graphed_autograd_function( fwd_graphs[i], bwd_graphs[i], per_callable_module_params[i], per_callable_len_user_args[i], per_callable_output_unflatten_spec[i], per_callable_static_input_surfaces[i], per_callable_static_outputs[i], per_callable_static_grad_outputs[i], per_callable_static_grad_inputs[i], ) if isinstance(func, torch.nn.Module): def make_graphed_forward( func: torch.nn.Module, graph_training_state: bool, graphed: Callable[_P, _R], orig_fwd: Callable[_P, _R], ) -> Callable[_P, _R]: def new_fwd(*user_args: _P.args, **user_kwargs: _P.kwargs) -> _R: # If the module's training-or-eval state matches what we graphed, # run the graph, otherwise run the original forward method if func.training == graph_training_state: return graphed(*user_args, **user_kwargs) else: return orig_fwd(*user_args, **user_kwargs) return new_fwd func.forward = make_graphed_forward( func, func.training, graphed, func.forward ) ret.append(func) else: ret.append(graphed) if just_one_callable: return ret[0] return tuple(ret)