# mypy: allow-untyped-defs import collections import dataclasses import enum import itertools as it import logging from collections.abc import Iterator from typing import Any, cast, Optional, Union from typing_extensions import Literal import torch from torch._C import FunctionSchema from torch._C._autograd import _ProfilerResult from torch._C._profiler import ( _EventType, _ExtraFields_Allocation, _ExtraFields_TorchOp, _ProfilerEvent, _TensorMetadata, RecordScope, ) from torch._utils import _element_size from torch.profiler import _utils KeyAndID = tuple["Key", int] TensorAndID = tuple["TensorKey", int] log = logging.getLogger(__name__) class Category(enum.Enum): INPUT = enum.auto() TEMPORARY = enum.auto() ACTIVATION = enum.auto() GRADIENT = enum.auto() AUTOGRAD_DETAIL = enum.auto() PARAMETER = enum.auto() OPTIMIZER_STATE = enum.auto() _CATEGORY_TO_COLORS = { Category.PARAMETER: "darkgreen", Category.OPTIMIZER_STATE: "goldenrod", Category.INPUT: "black", Category.TEMPORARY: "mediumpurple", Category.ACTIVATION: "red", Category.GRADIENT: "mediumblue", Category.AUTOGRAD_DETAIL: "royalblue", None: "grey", } _CATEGORY_TO_INDEX = {c: i for i, c in enumerate(_CATEGORY_TO_COLORS)} class Action(enum.Enum): PREEXISTING = enum.auto() CREATE = enum.auto() INCREMENT_VERSION = enum.auto() DESTROY = enum.auto() _ACTION_TO_INDEX = {i: i.value for i in Action} @dataclasses.dataclass(eq=True, unsafe_hash=False, frozen=True) class Key: device: torch.device @dataclasses.dataclass class _Storage: """Bundle storage pointer and id. All profiling logic should use `allocation_id`, however it is useful to print storage pointers for debugging and unit tests sometimes look up values using the storage data pointer of a live Tensor.""" ptr: int allocation_id: int def __repr__(self) -> str: return f"{hex(self.ptr):>18} ({self.allocation_id})" def __eq__(self, other: object) -> bool: return isinstance(other, _Storage) and self.allocation_id == other.allocation_id def __hash__(self) -> int: return hash(self.allocation_id) @dataclasses.dataclass(eq=True, unsafe_hash=True, frozen=True) class TensorKey(Key): """Hashable identifier for a storage which has been asigned an ID. A detailed description of Tensor IDs and why they are needed is given in `torch/csrc/profiler/collection.h` when `TensorID` is declared. To summarize, multiple Storage buffers can map to the same logical Tensor. This dataclass is used to refer to a concrete in-memory StorageImpl of a Tensor. """ id: int storage: _Storage def __repr__(self) -> str: return f"id={self.id}: {repr(self.storage):<24} ({self.device})" def __lt__(self, other: "TensorKey") -> bool: return self._as_sortable < other._as_sortable @staticmethod def _make( tensor_id: Optional[int], storage_ptr: Optional[int], allocation_id: Optional[int], device: torch.device, ) -> Optional["TensorKey"]: if ( tensor_id is not None and storage_ptr is not None and allocation_id is not None ): return TensorKey(device, tensor_id, _Storage(storage_ptr, allocation_id)) return None @classmethod def from_allocation(cls, alloc: _ExtraFields_Allocation) -> Optional["TensorKey"]: return cls._make(alloc.id, alloc.ptr, alloc.allocation_id, alloc.device) @classmethod def from_tensor(cls, t: Optional[_TensorMetadata]) -> Optional["TensorKey"]: if t is not None: return cls._make(t.id, t.storage_data_ptr, t.allocation_id, t.device) return None @property def _as_sortable(self) -> tuple[int, int, str, int]: return self.id, self.storage.allocation_id, self.device.type, self.device.index def _extract_parameters_and_gradients( node: _ProfilerEvent, ) -> Iterator[tuple[Optional[TensorKey], Optional[TensorKey]]]: children = node.children # AccumulateGrad is used in the Autograd engine to handle gradient updates. # There are two possible cases: # 1) This is a newly created gradient Tensor. In that case there is nothing # to accumulate, so autograd simply detaches the Tensor. # # 2) There is a preexisting gradient Tensor and we need to add the newly # computed update. This is done with an in-place add (aten::add_) op. # (The underscore suffix denotes "in-place".) if ( node.typed[0] == _EventType.TorchOp and node.typed[1].scope == RecordScope.BACKWARD_FUNCTION # TODO(robieta): Move away from load bearing names and node.name == "torch::autograd::AccumulateGrad" and children and children[0].typed[0] == _EventType.TorchOp and children[0].name in ("aten::detach", "aten::add_") and children[0].typed[1].inputs and isinstance(children[0].typed[1].inputs[0], _TensorMetadata) ): yield None, TensorKey.from_tensor(children[0].typed[1].inputs[0]) # We directly instrument `torch.nn.Module` and `torch.optim.Optimizer` # NOTE: The values captured by the python tracer are cached; they can be # used to build up labels but do not imply that a Tensor was live at # a particular time. elif node.typed[0] == _EventType.PyCall: typed_fields = node.typed[1] assert typed_fields.module is None or typed_fields.optimizer is None if typed_fields.module is not None: for _, p, p_grad in typed_fields.module.parameters: yield TensorKey.from_tensor(p), TensorKey.from_tensor(p_grad) if typed_fields.optimizer is not None: for p, p_grad, _ in typed_fields.optimizer.parameters: yield TensorKey.from_tensor(p), TensorKey.from_tensor(p_grad) def extract_parameters(node: _ProfilerEvent) -> Iterator[TensorKey]: for p, _p_grad in _extract_parameters_and_gradients(node): if p is not None: yield p def extract_gradients( node: _ProfilerEvent, ) -> Iterator[tuple[Optional[TensorKey], TensorKey]]: for p, p_grad in _extract_parameters_and_gradients(node): if p_grad is not None: yield p, p_grad def get_scopes(event: Optional[_ProfilerEvent]) -> tuple[RecordScope, ...]: scopes = [] while event: if event.typed[0] == _EventType.TorchOp: scopes.append(event.typed[1].scope) event = event.parent return tuple(scopes) class SchemaMatcher: """Lookup operator schema based on profiled name. When profiling we record the operator's name but not the schema. However some analysis requires that information. Fortunately we can look up registered schema from the recorded name. We do not, however, record the overload and so we must compare the profiled arguments with all overloads to determine viable matches. Note: Once https://github.com/pytorch/pytorch/issues/78871 is completed this code will be obsolete. """ @classmethod def inputs_are_mutable(cls, t: _ExtraFields_TorchOp) -> tuple[Optional[bool], ...]: """Determine which inputs may have mutated based on function schema. Note that we don't need to resolve down to a single schema to perform this analysis. An input is mutable if it is mutable in any overload. In practice, however, it is overwhelmingly common to match a single overload. If we cannot find any valid schema then we must be conservative and assume all inputs are mutable. """ mutable: Optional[list[bool]] = None for schema in cls.match_schemas(t): mutable = mutable or [False for _ in schema.arguments] for i, arg in enumerate(schema.arguments): mutable[i] |= getattr(arg.alias_info, "is_write", False) return tuple(mutable or (None for _ in t.inputs)) @classmethod def match_schemas(cls, t: _ExtraFields_TorchOp) -> tuple[FunctionSchema, ...]: signature = tuple( # Tensor TensorKey.from_tensor(i) if isinstance(i, _TensorMetadata) # # TensorList else [TensorKey.from_tensor(j) for j in i] if isinstance(i, list) # # Scalar and uncaptured inputs. else i for i in t.inputs ) def matches(schema) -> bool: return len(schema.arguments) == len(signature) and all( cls._types_match(observed, schema_arg.type) for observed, schema_arg in zip(signature, schema.arguments) ) return tuple(s for s in cls.lookup_schemas(t.name) or () if matches(s)) @classmethod def _types_match(cls, observed, schema_type) -> bool: if isinstance(schema_type, torch._C.OptionalType): schema_type = schema_type.getElementType() return observed is None or cls._types_match(observed, schema_type) if isinstance(schema_type, torch._C.AnyType): return True if schema_type.isSubtypeOf(torch._C.ListType.ofTensors()): return isinstance(observed, list) and all( isinstance(i, TensorKey) for i in observed ) type_map: tuple[tuple[Any, Union[type, tuple[type, ...]]], ...] = ( (torch._C.TensorType, TensorKey), (torch._C.NoneType, type(None)), (torch._C.BoolType, bool), (torch._C.IntType, int), (torch._C.FloatType, float), (torch._C.ComplexType, complex), (torch._C.NumberType, (bool, int, float, complex)), ) for jit_type, py_types in type_map: if isinstance(schema_type, jit_type): return isinstance(observed, py_types) # Profiler only records a subset of possible argument types. If we # reach this point then the schema must call for a type that profiler # does not record. Thus, the schema can only be a match if `observed` # is also None. return observed is None @staticmethod def lookup_schemas(name: str) -> Optional[tuple[FunctionSchema, ...]]: # TODO(robieta): # _jit_get_schemas_for_operator is quite expensive. (~100us / call) # Consider adding `functools.lru_cache` if that becomes an issue. try: # Schema lookup will throw if `name` is malformed. (For example, # schemas must be namespaced and schema lookup will fail if name # does not include "::".) We simply catch the exception and return # `None` to denote that `name` cannot be an operator name. # # Note that record_function annotations also go through this path, # so it is expected that some names will not correspond to PyTorch # operators. if "::" not in name: return None return tuple(torch._C._jit_get_schemas_for_operator(name)) except RuntimeError: return None class OpTree: def __init__(self, result: _ProfilerResult) -> None: self._root_nodes = result.experimental_event_tree() self._sorted_nodes = tuple(sorted(self.dfs(), key=lambda x: x.start_time_ns)) def dfs(self, *args, **kwargs) -> Iterator[_ProfilerEvent]: yield from _utils.traverse_dfs(self._root_nodes, *args, **kwargs) @property def sorted_nodes(self) -> tuple[_ProfilerEvent, ...]: return self._sorted_nodes class SizeMap: def __init__(self, op_tree: OpTree) -> None: self._values: dict[TensorKey, int] = {} for node in op_tree.sorted_nodes: if node.typed[0] == _EventType.TorchOp: for t in self._flat_tensor_inputs(node.typed[1]): self._update_values(t) elif node.typed[0] == _EventType.PyCall: typed_fields = node.typed[1] assert typed_fields.module is None or typed_fields.optimizer is None if typed_fields.module is not None: for _, p, p_grad in typed_fields.module.parameters: self._update_values(p) self._update_values(p_grad) if typed_fields.optimizer is not None: for p, p_grad, state in typed_fields.optimizer.parameters: self._update_values(p) self._update_values(p_grad) for _, t in state: self._update_values(t) allocations: dict[TensorKey, int] = {} for node in op_tree.sorted_nodes: if node.typed[0] == _EventType.Allocation: alloc_fields = node.typed[1] key = TensorKey.from_allocation(alloc_fields) if key: new_size = abs(alloc_fields.alloc_size) prior_size = allocations.setdefault(key, new_size) # It is possible to resize Storage in PyTorch, however we # key on data pointer so most resizes will be treated as a # change in storage. The one corner case that cannot be # handled is `realloc` which successfully resizes the # storage. At time of writing this is not done anywhere in # the core PyTorch codebase. if prior_size != new_size: delta = f"{prior_size} vs. {new_size}" log.warning("Mismatch between allocation and free: %s", delta) self._values.update(allocations) def _update_values(self, t: Optional[_TensorMetadata]) -> None: key = TensorKey.from_tensor(t) if key is not None and t is not None and t.layout == torch.strided: # Scalars are represented as zero dim Tensors n = max(i[0] * i[1] for i in zip(t.sizes or [1], t.strides or [1])) num_bytes = n * _element_size(t.dtype) assert num_bytes >= 0, f"{num_bytes}" self._values[key] = max(self._values.get(key, 0), num_bytes) @staticmethod def _flat_tensor_inputs(op: _ExtraFields_TorchOp) -> Iterator[_TensorMetadata]: for i in op.inputs: if isinstance(i, _TensorMetadata): yield i elif isinstance(i, list): yield from i def __getitem__(self, key: TensorKey): return self._values[key] @dataclasses.dataclass() class DataFlowEdge: input_version: Optional[int] = None mutated: Optional[bool] = False @property def is_allocation(self) -> bool: return self.input_version is None @property def is_deletion(self) -> bool: return self.mutated is None class DataFlowNode: def __init__(self, event: _ProfilerEvent, graph: "DataFlowGraph") -> None: self._event = event self._graph = graph self._edges: dict[TensorKey, DataFlowEdge] = self._determine_edges() for key, edge in self._edges.items(): if edge.mutated and not edge.is_allocation: self._graph.bump(key) # Make sure the version bumping behavior matches what we expect. versions = {k: (v, self._graph.lookup(k)) for k, v in self.outputs.items()} assert all(i == j for i, j in versions.values()), f"{versions}, {self._edges}" def _determine_edges(self) -> dict[TensorKey, DataFlowEdge]: subtree = tuple(_utils.traverse_dfs([self._event])) # Start by populating edges from op inputs and outputs. mutable_by_key: dict[Optional[TensorKey], set[Optional[bool]]] = {} for op in (i.typed[1] for i in subtree if i.typed[0] == _EventType.TorchOp): for op_input, mutable in zip( op.inputs, SchemaMatcher.inputs_are_mutable(op) ): # Tensor if isinstance(op_input, _TensorMetadata): key = TensorKey.from_tensor(op_input) mutable_by_key.setdefault(key, set()).add(mutable) # TensorList elif isinstance(op_input, list): for op_input_i in op_input: key = TensorKey.from_tensor(op_input_i) mutable_by_key.setdefault(key, set()).add(mutable) edges: collections.defaultdict[Optional[TensorKey], DataFlowEdge] edges = collections.defaultdict(DataFlowEdge) for key, mutable_set in mutable_by_key.items(): if key is not None: edges[key].input_version = self._graph.lookup(key) if key else -1 # We consider an op to be mutated if we encounter a schema where it # is a mutable argument OR if it is ambiguous. (We never explicitly # see it in any schema.) mutated = (True in mutable_set) or (tuple(mutable_set) == (None,)) edges[key].mutated = mutated # Then handle deletions. Note that deleting a Tensor implicitly adds # it as an input edge. for i in subtree: if i.typed[0] == _EventType.Allocation and i.typed[1].alloc_size < 0: key = TensorKey.from_allocation(i.typed[1]) edge = edges[key] assert key is None or edge.mutated is not None, f"Double delete: {key}" edge.mutated = None edge.input_version = self._graph.lookup(key) if key else -1 # And finally handle allocations. This step must be last, because the # previous two steps optimistically add input edges. for i in subtree: if i.typed[0] == _EventType.Allocation and i.typed[1].alloc_size > 0: edges[TensorKey.from_allocation(i.typed[1])].input_version = None # We don't need to sort the inputs, but it makes debugging and unit tests nicer. return dict(sorted((k, v) for k, v in edges.items() if k is not None)) @property def inputs(self) -> dict[TensorKey, tuple[bool, int]]: return { # MyPy can't see through `is_allocation` to know that # `v.input_version` is not None. k: (bool(v.mutated), cast(int, v.input_version)) for k, v in self._edges.items() if not v.is_allocation } @property def outputs(self) -> dict[TensorKey, int]: return { k: 0 if v.input_version is None else v.input_version + 1 for k, v in self._edges.items() if (v.is_allocation and not v.is_deletion) or v.mutated } @property def intermediates(self) -> tuple[TensorKey, ...]: return tuple( k for k, v in self._edges.items() if v.is_allocation and v.is_deletion ) @property def start_time(self) -> int: return self._event.start_time_ns class DataFlowGraph: def __init__(self, op_tree: OpTree) -> None: self._op_tree = op_tree self._leaf_events = self._extract_leaf_events(op_tree) self._active_version: dict[TensorKey, Optional[int]] = {} self._flow_nodes = [DataFlowNode(e, self) for e in self.leaf_events] self._flow_nodes.sort(key=lambda x: x.start_time) self.validate() @property def flow_nodes(self) -> tuple[DataFlowNode, ...]: return tuple(self._flow_nodes) def validate(self): # Check that each (Tensor, version) pair has a unique creation node outputs: set[tuple[TensorKey, int]] = set() for node in self.flow_nodes: node_outputs = set(node.outputs.items()) duplicates = outputs & node_outputs assert not duplicates, f"{node._event.name} {node._edges} {duplicates}" outputs |= node_outputs # And check that `self._nodes` forms a valid topologically sorted DAG. tensor_versions: dict[TensorKey, int] = {} for node in self.flow_nodes: for key, (_, version) in node.inputs.items(): expected = tensor_versions.get(key, 0) assert expected == version, (expected, version) for key, version in node.outputs.items(): prior_version = tensor_versions.get(key, version) assert version >= prior_version, (version, prior_version) tensor_versions[key] = version @property def leaf_events(self) -> tuple[_ProfilerEvent, ...]: return self._leaf_events @staticmethod def _extract_leaf_events(op_tree: OpTree) -> tuple[_ProfilerEvent, ...]: """Partially traverse the op tree and extract top level ops. Consider the following code: ``` with record_function("My annotation"): x.zero_() y.zero_() ``` The op tree (assuming no Autograd) will look like: TorchOp: "My annotation" TorchOp: zero_ TorchOp: fill_ TorchOp: zero_ TorchOp: fill_ The recursive structure of operator calls makes data flow unwieldy. In order to simplify analysis we would like to select the highest level ops to represent in the graph. In this case those are the `zero_` ops; the fact that `fill_` is called is an implementation detail. We also do not want to group everything under "My annotation" as this could create overly coarse bundles and lose critical semantics. To address this issue we walk over the graph and select the topmost torch ops ** which match at least one operator schema **. These form the leaves of the first pass through the op tree. (As well as any allocations or frees which do are not part of a kernel.) These events form the logical nodes in our data flow graph. """ leaf_events: list[_ProfilerEvent] = [] def leaf_op(e: _ProfilerEvent) -> bool: return e.typed[0] == _EventType.TorchOp and ( e.typed[1].scope == RecordScope.BACKWARD_FUNCTION or bool(SchemaMatcher.match_schemas(e.typed[1])) ) def children_fn(e: _ProfilerEvent): if leaf_op(e) or e.tag == _EventType.Allocation: leaf_events.append(e) return [] return e.children for _ in op_tree.dfs(children_fn=children_fn): pass return tuple(sorted(leaf_events, key=lambda x: x.start_time_ns)) def lookup(self, key: TensorKey) -> int: version = self._active_version.setdefault(key, 0) assert version is not None return version def bump(self, key: TensorKey) -> None: prior_version = self._active_version.get(key, None) assert prior_version is not None self._active_version[key] = prior_version + 1 def delete(self, key: TensorKey) -> None: assert self._active_version.setdefault(key, 0) is not None self._active_version[key] = None @dataclasses.dataclass class CategoryElement: by_id: Optional[Category] = None by_key: dict[TensorKey, Category] = dataclasses.field(default_factory=dict) by_version: dict[TensorAndID, Category] = dataclasses.field(default_factory=dict) # Used by unit tests to check internals. (And consequently by # MemoryProfile.lookup) This should not be used in any other capacity. _by_id_keyset: set[TensorKey] = dataclasses.field(default_factory=set) @dataclasses.dataclass class CategoryDict: _values: collections.defaultdict[int, CategoryElement] = dataclasses.field( default_factory=lambda: collections.defaultdict(CategoryElement) ) def set_by_id(self, key: TensorKey, category: Category) -> None: self._values[key.id].by_id = category self._values[key.id]._by_id_keyset.add(key) def set_by_key(self, key: TensorKey, category: Category) -> None: self._values[key.id].by_key[key] = category def set_by_version(self, key: TensorKey, version: int, category: Category) -> None: self._values[key.id].by_version[(key, version)] = category def setdefault_by_version( self, key: TensorKey, version: int, category: Category ) -> None: self._values[key.id].by_version.setdefault((key, version), category) def get(self, key: Key, version: int) -> Optional[Category]: if isinstance(key, Key) and not isinstance(key, TensorKey): return None element = self._values[key.id] return ( element.by_id or element.by_key.get(key, None) or element.by_version.get((key, version), None) ) class MemoryProfile: def __init__(self, result: _ProfilerResult) -> None: self._op_tree = OpTree(result) self._data_flow_graph = DataFlowGraph(self._op_tree) self._size_map = SizeMap(self._op_tree) self._categories = CategoryDict() self._set_gradients_and_temporaries() self._set_parameters_using_python_tracer() self._set_inputs() self._set_parameters_using_data_flow() self._set_activations() self._set_optimizer_state() self._set_autograd_detail() @property def timeline(self) -> tuple[tuple[int, Action, KeyAndID, int], ...]: output: list[tuple[int, Action, KeyAndID, int]] = [] allocation_times: dict[tuple[TensorKey, bool], int] = {} live_unknown: dict[tuple[int, torch.device], Literal[True]] = {} for event in self._op_tree.dfs(): if event.typed[0] == _EventType.Allocation: alloc_fields = event.typed[1] alloc_size = alloc_fields.alloc_size is_allocation = alloc_size > 0 t = event.start_time_ns tkey = TensorKey.from_allocation(alloc_fields) if tkey is not None: allocation_times[(tkey, is_allocation)] = t else: key = Key(alloc_fields.device) ptr_and_device = (alloc_fields.ptr, key.device) if is_allocation: if ptr_and_device in live_unknown: output.append( (t, Action.INCREMENT_VERSION, (key, 0), alloc_size) ) else: live_unknown[ptr_and_device] = True output.append((t, Action.CREATE, (key, 0), alloc_size)) else: output.append((t, Action.DESTROY, (key, 0), -alloc_size)) if not live_unknown.pop(ptr_and_device, False): output.append( (-1, Action.PREEXISTING, (key, 0), -alloc_size) ) snapshot = self._category_snapshot() last_version = dict(sorted(snapshot.keys())) events: list[tuple[int, Action, TensorAndID]] = [ (-1, Action.PREEXISTING, (key, version)) for key, version in snapshot.keys() if (key, True) not in allocation_times and version == 0 ] for node in self._data_flow_graph.flow_nodes: for key, edge in node._edges.items(): if edge.is_allocation: t = allocation_times[(key, True)] events.append((t, Action.CREATE, (key, 0))) elif edge.mutated: t = node._event.start_time_ns version = edge.input_version assert version is not None events.append((t, Action.INCREMENT_VERSION, (key, version))) if edge.is_deletion: t = allocation_times[(key, False)] events.append((t, Action.DESTROY, (key, last_version[key]))) output.extend( (time, action, (key, version), self._size_map[key]) for time, action, (key, version) in events ) output.sort(key=lambda x: (x[0], x[1].value)) return tuple(output) def _is_gradient(self, *args, **kwargs) -> bool: return self._categories.get(*args, **kwargs) == Category.GRADIENT def _category_snapshot(self) -> dict[TensorAndID, Optional[Category]]: all_tensor_versions: set[TensorAndID] = set() for node in self._data_flow_graph.flow_nodes: all_tensor_versions.update(((k, v) for k, (_, v) in node.inputs.items())) all_tensor_versions.update((key, 0) for key in node.intermediates) all_tensor_versions.update(node.outputs.items()) for i in self._categories._values.values(): all_tensor_versions.update((key, 0) for key in i._by_id_keyset) return { (key, version): self._categories.get(key, version) for key, version in sorted(all_tensor_versions) } def _any_version_depends_on_gradient(self) -> set[int]: """Extract IDs of Tensors which depend or will depend on a gradient. Note that this weakened definition of "depends" requires us to loop over the data flow graph multiple times because it allows dependency information to flow backward through edges and removes the guarantee that nodes are topologically sorted. (Or indeed, even that a valid topological order exists.) Put another way, we have converted an acyclic data flow graph into a cyclic graph and we are attempting to partition cycles involving a gradient from the rest of the graph. """ depends_on_gradient: set[int] = set() while True: start_size = len(depends_on_gradient) for node in self._data_flow_graph.flow_nodes: ids = tuple( key.id for key, (_, version) in node.inputs.items() if self._categories.get(key, version) in (Category.GRADIENT, Category.PARAMETER) or key.id in depends_on_gradient ) if ids: depends_on_gradient.update(ids) depends_on_gradient.update(key.id for key in node.outputs) # We are guaranteed to exit because there is a finite set of # TensorAndID pairs. In practice we do not expect to loop more than # three times: once to identify the core parameter update loop, # once to fold the first step into that loop, and a third time # where no new elements are added. if len(depends_on_gradient) == start_size: return depends_on_gradient def _set_gradients_and_temporaries(self) -> None: """Mark Tensors which are unambiguous and simple to reason about.""" # Gradients are straightforward to detect. We directly check the # `.grad` property in the Python tracer, and we can detect any new # gradient Tensors from `AccumulateGrad` ops. for event in self._op_tree.dfs(): for _, p_grad in extract_gradients(event): self._categories.set_by_id(p_grad, Category.GRADIENT) # Similarly, temporary Tensors are easy to identify and are useful to # flag since they can make memory use "spikier" than one would # otherwise expect. for node in self._data_flow_graph.flow_nodes: for i in node.intermediates: self._categories.set_by_key(i, Category.TEMPORARY) def _set_parameters_using_python_tracer(self) -> None: for event in self._op_tree.dfs(): for p in extract_parameters(event): if p is not None: self._categories.set_by_id(p, Category.PARAMETER) def _set_inputs(self) -> None: """Mark inputs based on which Tensors are updated using gradients. The process for differentiating between inputs and activations is more involved. Most Tensors in a training loop depend on at least one gradient: parameters depend on them through updates, and activations and optimizer state depend on them transitively through parameters. Critically, we do not need to know which Tensors are parameters to apply this method; we can simply walk the data flow graph to build the set of all values which depend on a gradient and then obtain the set of inputs from the conjugate set. There is, however, one hiccup. The first time we see a parameter is generally on the forward pass of the first step. We know from inspection of the data flow graph that v1 of that Tensor depends on a gradient (provided we profile an optimizer step), but not v0. To address this problem we weaken the definition of "depends on a gradient" to "any version of this Tensor depends on a gradient", which in turn strengthens the criteria for the input set enough to filter the activations in the forward pass of the first step.""" # All of this analysis is predicated on using at least one training # step (or parameters from the python tracer) to partition the graph. # Absent that we cannot determine which Tensors are inputs and which # ones are part of the model. depends_on_gradient = self._any_version_depends_on_gradient() # We only want to annotate Tensors which actually contribute to the # model calculation. produces_gradient: set[TensorAndID] = set() for node in reversed(self._data_flow_graph.flow_nodes): tensors = {(key, version) for key, (_, version) in node.inputs.items()} tensors |= node.outputs.items() if any( self._categories.get(*i) in (Category.GRADIENT, Category.PARAMETER) or i in produces_gradient for i in tensors ): produces_gradient |= tensors # Don't include Tensors created in the backward pass, as these are # generally Autograd implementation details rather than proper inputs. input_candidates = produces_gradient.copy() for node in self._data_flow_graph.flow_nodes: if RecordScope.BACKWARD_FUNCTION in get_scopes(node._event): input_candidates -= set(node.outputs.items()) for key, version in input_candidates: if key.id not in depends_on_gradient: self._categories.setdefault_by_version(key, version, Category.INPUT) def _set_parameters_using_data_flow(self) -> None: """Deduce which Tensors are parameters. Consider the following code for the step of SGD with momentum (nesterov=False), where `d_p` is the gradient of `param` and `buf` is the momentum buffer. ``` buf.mul_(momentum).add_(d_p, alpha=1 - dampening) d_p = buf param.add_(d_p, alpha=-lr) ``` Both `param` and `buf` take a gradient and perform an in-place update. The python tracer will inspect calls to `nn.Module.forward` and `optim.Optimizer.step` to extract parameter and optimizer state respectively (including parameters), so this is generally a non-issue. However as a fallback we can also exploit several properties of parameters to distinguish them from other model state. First, they are directly used in the forward pass. (At this point we haven't established which parts of the graph correspond to the forward pass but we can deduce enough to suffice.) Some mutable state such as batch norm moving averages also contribute to the forward pass, but optimizer state does not. Second, a parameter is by definition used to compute at least one gradient and depends on at least one gradient. """ snapshot = self._category_snapshot() # Determine which Tensors might be parameters based on forward pass # data flow. Note this these are only candidates; we filter nodes that # we know are part of the backward pass but that doesn't guarantee that # they are part of the forward pass. candidate_parameters: set[TensorAndID] = set() candidate_fwd_tensors: set[TensorAndID] = { i for i, category in snapshot.items() if category == Category.INPUT } for node in self._data_flow_graph.flow_nodes: inputs = {(key, value) for key, (_, value) in node.inputs.items()} if ( # Don't check nodes in the backward pass. RecordScope.BACKWARD_FUNCTION not in get_scopes(node._event) and not any(self._is_gradient(*i) for i in inputs) and not any(self._is_gradient(*i) for i in node.outputs.items()) # # and only check nodes which depend on an input. and candidate_fwd_tensors.intersection(inputs) ): candidate_fwd_tensors |= node.outputs.items() candidate_parameters |= inputs.difference(candidate_fwd_tensors) # Require that each parameter eventually contributes to the value of a gradient used_for_gradient: set[TensorAndID] = set() for node in reversed(self._data_flow_graph.flow_nodes): if any( self._is_gradient(*i) or i in used_for_gradient for i in node.outputs.items() ): used_for_gradient.update( (key, version) for key, (_, version) in node.inputs.items() ) candidate_parameters.intersection_update(used_for_gradient) # and depends on a gradient. parameter_keys = {key.id for key, _ in candidate_parameters} parameter_keys &= self._any_version_depends_on_gradient() for key, _ in snapshot.keys(): if key.id in parameter_keys: self._categories.set_by_id(key, Category.PARAMETER) def _set_activations(self) -> None: """Flood the graph to identify activations.""" required = {Category.INPUT, Category.ACTIVATION} also_allowed = {Category.PARAMETER, Category.TEMPORARY} for node in self._data_flow_graph.flow_nodes: inputs = {(key, value) for key, (_, value) in node.inputs.items()} input_categories = {self._categories.get(*i) for i in inputs} if ( (input_categories & required) and not (input_categories - (required | also_allowed)) # # Stop filling when we reach the backward pass. and RecordScope.BACKWARD_FUNCTION not in get_scopes(node._event) ): for i in node.outputs.items(): self._categories.setdefault_by_version(*i, Category.ACTIVATION) def _set_optimizer_state(self) -> None: for event in self._op_tree.dfs(): if event.typed[0] == _EventType.PyCall and event.typed[1].optimizer: parameters = event.typed[1].optimizer.parameters for _, t in it.chain.from_iterable( (state for _, _, state in parameters) ): key = TensorKey.from_tensor(t) if key is not None: self._categories.set_by_id(key, Category.OPTIMIZER_STATE) def _set_autograd_detail(self): prior = {None, Category.AUTOGRAD_DETAIL} for node in self._data_flow_graph.flow_nodes: if RecordScope.BACKWARD_FUNCTION in get_scopes(node._event): for key, version in node.outputs.items(): if version == 0 or self._categories.get(key, version - 1) in prior: self._categories.setdefault_by_version( key, version, Category.AUTOGRAD_DETAIL ) class MemoryProfileTimeline: def __init__(self, memory_profile): """The minimum representation of the memory profile timeline includes the memory timeline and categories. The timeline consists of [timestamp, action, (TensorKey, version), numbytes] elements, to denote any actions (pre-existing, create, destroy, or increment_version) that occurred to a specific Tensor for a chunk of memory. The categories help map each (TensorKey, version) pair into a category.""" self.timeline = memory_profile.timeline self.categories = memory_profile._categories def _coalesce_timeline(self, device_str): """Convert the memory timeline and categories into a memory plot consisting of timestamps and their respective sizes by category for a given device. Input: device Output: [timestamps, sizes by category] """ device = torch.device(device_str) times: list[int] = [] sizes: list[list[int]] = [] def update(key, version, delta): category = ( self.categories.get(key, version) if isinstance(key, TensorKey) else None ) index = _CATEGORY_TO_INDEX[category] + 1 sizes[-1][index] += int(delta) t_min = -1 for t, action, (key, version), numbytes in self.timeline: if key.device != device: continue # Convert timestamps from ns to us, to match trace events. if t != -1: t = int(t / 1000) # Save the smallest timestamp to populate pre-existing allocs. if t_min == -1 or (t < t_min and t > 0): t_min = t # Handle timestep if len(times) == 0: times.append(t) sizes.append([0] + [0 for _ in _CATEGORY_TO_INDEX]) elif t != times[-1]: times.append(t) sizes.append(sizes[-1].copy()) # Handle memory and categories if action in (Action.PREEXISTING, Action.CREATE): update(key, version, numbytes) elif action == Action.INCREMENT_VERSION: update(key, version, -numbytes) update(key, version + 1, numbytes) elif action == Action.DESTROY: update(key, version, -numbytes) else: raise ValueError(f"Unknown action: {action}") times = [t_min if t < 0 else t for t in times] return times, sizes def export_memory_timeline(self, path, device_str) -> None: """Saves the memory timeline as [times, sizes by category] as a JSON formatted file to the given path for the given device.""" times, sizes = self._coalesce_timeline(device_str) # TODO: Write a faster serialize (orjson not available in CI) import json with open(path, "w") as f: json.dump([times, sizes], f) def export_memory_timeline_raw(self, path, device_str) -> None: """Saves the memory timeline as raw memory event tuples in the form of (timestamp, action, numbytes, category) as a JSON formatted file to the given path for the given device.""" device = torch.device(device_str) raw_events: list[tuple[int, int, int, int]] = [] def get_category_index(key, version): category = ( self.categories.get(key, version) if isinstance(key, TensorKey) else None ) return _CATEGORY_TO_INDEX[category] for t, action, (key, version), numbytes in self.timeline: if key.device != device: continue if action in (Action.PREEXISTING, Action.CREATE): raw_events.append( ( t, _ACTION_TO_INDEX[action], numbytes, get_category_index(key, version), ) ) elif action == Action.INCREMENT_VERSION: raw_events.append( ( t, _ACTION_TO_INDEX[action], -numbytes, get_category_index(key, version), ) ) raw_events.append( ( t, _ACTION_TO_INDEX[action], numbytes, get_category_index(key, version + 1), ) ) elif action == Action.DESTROY: raw_events.append( ( t, _ACTION_TO_INDEX[action], -numbytes, get_category_index(key, version), ) ) else: raise ValueError(f"Unknown action: {action}") import json with open(path, "w") as f: json.dump(raw_events, f) def export_memory_timeline_html( self, path, device_str, figsize=(20, 12), title=None ) -> None: """Exports the memory timeline as an HTML file which contains the memory timeline plot embedded as a PNG file.""" # Check if user has matplotlib installed, return gracefully if not. import importlib.util matplotlib_spec = importlib.util.find_spec("matplotlib") if matplotlib_spec is None: print( "export_memory_timeline_html failed because matplotlib was not found." ) return from base64 import b64encode from os import remove from tempfile import NamedTemporaryFile import matplotlib.pyplot as plt import numpy as np mt = self._coalesce_timeline(device_str) times, sizes = np.array(mt[0]), np.array(mt[1]) # For this timeline, start at 0 to match Chrome traces. t_min = min(times) times -= t_min stacked = np.cumsum(sizes, axis=1) / 1024**3 device = torch.device(device_str) max_memory_allocated = torch.cuda.max_memory_allocated(device) max_memory_reserved = torch.cuda.max_memory_reserved(device) # Plot memory timeline as stacked data fig = plt.figure(figsize=figsize, dpi=80) axes = fig.gca() for category, color in _CATEGORY_TO_COLORS.items(): i = _CATEGORY_TO_INDEX[category] axes.fill_between( times / 1e3, stacked[:, i], stacked[:, i + 1], color=color, alpha=0.7 ) fig.legend(["Unknown" if i is None else i.name for i in _CATEGORY_TO_COLORS]) # Usually training steps are in magnitude of ms. axes.set_xlabel("Time (ms)") axes.set_ylabel("Memory (GB)") title = "\n\n".join( ([title] if title else []) + [ f"Max memory allocated: {max_memory_allocated / (1024**3):.2f} GiB \n" f"Max memory reserved: {max_memory_reserved / (1024**3):.2f} GiB" ] ) axes.set_title(title) # Embed the memory timeline image into the HTML file tmpfile = NamedTemporaryFile("wb", suffix=".png", delete=False) tmpfile.close() fig.savefig(tmpfile.name, format="png") with open(tmpfile.name, "rb") as tmp: encoded = b64encode(tmp.read()).decode("utf-8") html = f""" GPU Memory Timeline HTML """ with open(path, "w") as f: f.write(html) remove(tmpfile.name)