# mypy: allow-untyped-defs import contextlib import functools import itertools import logging import os import sys import time import warnings from itertools import count from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union from unittest import mock import torch._inductor.async_compile # noqa: F401 required to warm up AsyncCompile pools import torch.fx import torch.utils._pytree as pytree from functorch.compile import min_cut_rematerialization_partition from torch._dynamo import ( compiled_autograd, config as dynamo_config, logging as dynamo_logging, utils as dynamo_utils, ) from torch._dynamo.utils import ( counters, detect_fake_mode, flatten_graph_inputs, lazy_format_graph_code, ) from torch._functorch import config as functorch_config from torch._functorch.aot_autograd import aot_export_module, make_boxed_func from torch._inductor.codecache import code_hash, CompiledFxGraph, FxGraphCache from torch._inductor.cudagraph_utils import ( BoxedDeviceIndex, get_placeholders, log_cudagraph_skip_and_bump_counter, ) from torch._inductor.debug import save_args_for_compile_fx_inner from torch._inductor.utils import ( BoxedBool, count_tangents, fresh_inductor_cache, should_assume_input_aligned, tensor_is_aligned, ) from torch._logging import trace_structured from torch._ops import OpOverload from torch._subclasses.fake_tensor import FakeTensor from torch._utils_internal import compile_time_strobelight_meta from torch.fx.experimental.symbolic_shapes import free_unbacked_symbols from torch.fx.passes.fake_tensor_prop import FakeTensorProp from .._dynamo.backends.common import aot_autograd from ..fx._lazy_graph_module import _use_lazy_graph_module # type: ignore[attr-defined] from ..fx.graph import _PyTreeCodeGen from . import config, metrics from .debug import DebugContext from .decomposition import select_decomp_table from .fx_passes.joint_graph import joint_graph_passes from .fx_passes.post_grad import post_grad_passes, view_to_reshape from .fx_passes.pre_grad import pre_grad_passes from .graph import GraphLowering from .ir import ExternKernelNode from .utils import ( get_cloned_parameter_buffer_name, has_incompatible_cudagraph_ops, maybe_get_suppress_shape_guards_ctx, output_node, ) from .virtualized import V if config.is_fbcode(): from torch._inductor.fb.utils import log_optimus_to_scuba, time_and_log else: # no-op decorator def time_and_log(attr: str): return dynamo_utils.identity log = logging.getLogger(__name__) perf_hint_log = torch._logging.getArtifactLogger(__name__, "perf_hints") post_grad_graphs_log = torch._logging.getArtifactLogger(__name__, "post_grad_graphs") ALIGNMENT = 16 # copy_ fails when trying to write to tensors with memory overlap, # for expanded dimensions (a dimension which used to have size 1 -> ?) # we can select one element from that dimension and write to it # to achieve writing to all values of that dimension of the input tensor def get_expanded_dims(t): if not isinstance(t, torch.Tensor): return None return [i for i in range(t.ndim) if t.stride(i) == 0 and t.size(i) != 1] def index_expanded_dims(t: torch.Tensor, expanded_dims: List[int]) -> torch.Tensor: for expanded_dim in expanded_dims: t = torch.ops.aten.slice(t, expanded_dim, 0, 1) return t def complex_memory_overlap(t: torch.Tensor) -> bool: # if torch._debug_has_internal_overlap thinks this tensor potentially has # memory overlap internally, let's dig deeper to find out whether it's true. # # Call squeeze() so that dimension with size 1 does not cause false positive. t = index_expanded_dims(t, get_expanded_dims(t)).squeeze() if torch._debug_has_internal_overlap(t) != 0: strides = t.stride() sizes = t.shape indices = list(range(len(strides))) indices = [x for _, x in sorted(zip(strides, indices))] for i in range(len(strides)): prev_stride = 1 if i == 0 else strides[indices[i - 1]] prev_size = 1 if i == 0 else sizes[indices[i - 1]] if strides[indices[i]] < prev_stride * prev_size: return True return False def get_static_input_idxs(num_fixed): # If we are inlining NNModules, we treat all torch.nn.Parameters as static for the purposes # of cudagraphs. Rather than copying these into cudagraph-owned memory # like we do for normal inputs on each run, we will re-record a cudagraph if these # parameter locations change. context = torch._guards.TracingContext.try_get() fixed = list(range(num_fixed)) if not context or not context.fw_metadata: return fixed return fixed + context.fw_metadata.static_parameter_indices @functools.lru_cache(None) def _step_logger(): return dynamo_logging.get_step_logger(log) @functools.lru_cache(None) def _warn_tf32_disabled(): if ( torch.cuda.is_available() and not torch.backends.cuda.matmul.allow_tf32 and torch.cuda.get_device_capability() >= (8, 0) ): warnings.warn( "TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. " "Consider setting `torch.set_float32_matmul_precision('high')` for better performance." ) def _unlift_graph(mod, gm, graph_signature): from torch.export.unflatten import _assign_attr, _AttrKind state_dict = {} for name, param in mod.named_parameters(remove_duplicate=False): state_dict[name] = param _assign_attr( param, gm, name, attr_kind=_AttrKind.PARAMETER, ) for name, buffer in mod.named_buffers(remove_duplicate=False): state_dict[name] = buffer _assign_attr( buffer, gm, name, attr_kind=_AttrKind.BUFFER, ) placeholder_nodes = gm.graph.find_nodes(op="placeholder") lifted_inputs = [] # In AOTI, module parameters and buffers are not lifted as graph inputs. # As a result, mutation to buffers has side effect which makes their initial # values different from Eager. So we clone them here as a copy. # We are not cloning for parameters, although it will be needed if we want to # support training. for node in placeholder_nodes: node_name = node.name if node_name in graph_signature.inputs_to_parameters: parameter_name = graph_signature.inputs_to_parameters[node_name] lifted_inputs.append(parameter_name) elif node_name in graph_signature.inputs_to_buffers: buffer_name = graph_signature.inputs_to_buffers[node_name] lifted_inputs.append(buffer_name) gm.meta[ get_cloned_parameter_buffer_name(buffer_name) ] = clone_preserve_strides(state_dict[buffer_name]) else: assert node_name in graph_signature.user_inputs lifted_inputs.append(None) from torch.export._unlift import _unlift outputs = list(gm.graph.nodes)[-1].args[0] mutated_outputs = [] buffer_mutations = graph_signature.buffers_to_mutate user_input_mutations = graph_signature.user_inputs_to_mutate output_tokens = graph_signature.output_tokens for idx, out in enumerate(outputs): value = None if idx < len(buffer_mutations) + len(user_input_mutations) + len(output_tokens): if out.name in buffer_mutations: value = buffer_mutations[out.name] elif out.name in user_input_mutations: value = user_input_mutations[out.name] mutated_outputs.append(value) unlifted_gm = _unlift( gm, lifted_inputs, mutated_outputs, pytree.LeafSpec(), None, state_dict, {}, ) return unlifted_gm def _get_subgraph_names(gm): for node in sorted( itertools.chain( gm.graph.find_nodes(op="call_function", target=torch.ops.higher_order.cond), gm.graph.find_nodes( op="call_function", target=torch.ops.higher_order.while_loop ), ) ): if node.target == torch.ops.higher_order.cond: true_subgraph_name = node.args[1].name false_subgraph_name = node.args[2].name yield true_subgraph_name yield false_subgraph_name elif node.target == torch.ops.higher_order.while_loop: cond_subgraph_name = node.args[0].name body_subgraph_name = node.args[1].name yield cond_subgraph_name yield body_subgraph_name def _recursive_pre_grad_passes(gm, example_inputs): for subgraph_name in _get_subgraph_names(gm): subgraph = getattr(gm, subgraph_name) # as we don't have recursive example inputs, passing None here new_subgraph = _recursive_pre_grad_passes(subgraph, example_inputs=None) setattr(gm, subgraph_name, new_subgraph) return pre_grad_passes(gm, example_inputs) def _recursive_joint_graph_passes(gm): for subgraph_name in _get_subgraph_names(gm): subgraph = getattr(gm, subgraph_name) _recursive_joint_graph_passes(subgraph) joint_graph_passes(gm) def _recursive_post_grad_passes(gm, is_inference: bool = False): for subgraph_name in _get_subgraph_names(gm): subgraph = getattr(gm, subgraph_name) _recursive_post_grad_passes(subgraph, is_inference) post_grad_passes(gm, is_inference) def split_const_gm( gm: torch.fx.GraphModule, ) -> Tuple[torch.fx.GraphModule, Dict[str, int]]: """ This function takes an GraphModule input "gm". The gm will be split into 2 components, 1) const_gm, which consists the subgraph of gm that can be constant folded. 2) gm (being inplace modified,) which returns the graph after constant folding. const_output_index is a mapping of corresponding node name from gm to the output index of const_gm. Returns (const_gm, const_output_index) """ from torch._inductor.constant_folding import ( CONST_MODULE_TAG, META_TAG, MODULE_TAG, replace_node_with_constant, run_and_get_constant_graph, ) const_gm = run_and_get_constant_graph(gm) const_result = const_gm() const_outputs = { x.name: idx for idx, x in enumerate(tuple(const_gm.graph.nodes)[-1].args[0]) } to_erase_node = [] to_replace_node = [] const_output_index = {} for node in gm.graph.nodes: if node.name in const_outputs: to_replace_node.append(node) elif node.meta[META_TAG] == CONST_MODULE_TAG: to_erase_node.append(node) for node in to_replace_node: new_const_name = "_FOLDED_CONST_" + node.name replace_node_with_constant( gm, node, const_result[const_outputs[node.name]], new_const_name, ) const_output_index[new_const_name] = const_outputs[node.name] for node in to_erase_node[::-1]: if node.users: for n in node.users: assert n.meta[META_TAG] == MODULE_TAG, f"node: {node} user not empty." else: gm.graph.erase_node(node) gm.recompile() return const_gm, const_output_index def is_tf32_warning_applicable(gm: torch.fx.GraphModule): aten = torch.ops.aten tf32_ops = { aten.mm.default, aten.addmm.default, aten.bmm.default, aten.baddbmm.default, } for target in tf32_ops: for node in gm.graph.find_nodes(op="call_function", target=target): if ( isinstance(node.meta.get("val", None), torch.Tensor) and node.meta["val"].dtype == torch.float32 and node.meta["val"].device.type == "cuda" ): return True return False def maybe_disable_comprehensive_padding(example_inputs: List[torch.Tensor]): """ For CPU backend, enable comprehensive padding causes some unit tests fail due to changing number of generated kernels. Skip for now. """ has_cuda = any( t.device.type == "cuda" for t in example_inputs if isinstance(t, torch.Tensor) ) if config.comprehensive_padding and not has_cuda: perf_hint_log.info("Skip comprehensive padding on CPU") return config.patch(comprehensive_padding=False) else: return contextlib.nullcontext() def fake_tensor_prop( gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor], force_allow_non_fake_inputs: bool = False, ): """ If we can not detect fake mode from the context of inputs, create one. The created fake mode will be returned. """ fake_mode = detect_fake_mode(example_inputs) if not fake_mode: fake_mode = torch._subclasses.FakeTensorMode(allow_non_fake_inputs=True) FakeTensorProp(gm, mode=fake_mode).propagate(*example_inputs) else: ctx = ( contextlib.nullcontext() if not force_allow_non_fake_inputs else mock.patch.object(fake_mode, "allow_non_fake_inputs", True) ) with ctx: # type: ignore[attr-defined] FakeTensorProp(gm, mode=fake_mode).propagate_dont_convert_inputs( *example_inputs ) return fake_mode def should_use_remote_fx_graph_cache(): if config.fx_graph_remote_cache: return True if not config.is_fbcode(): return False if torch.version.hip is not None: return False try: from triton.fb.fb_memcache import MEMCACHE_VERSION except ModuleNotFoundError: return False return MEMCACHE_VERSION >= torch._utils_internal.justknobs_getval_int( "pytorch/remote_cache:fx_graph_memcache_version" ) # pass config dict back to user def get_patched_config_dict(config_patches=None) -> Dict[str, Any]: with config.patch(config_patches): return config.get_config_copy() @functools.wraps def with_fresh_cache_if_config(f): if config.force_disable_caches: with fresh_inductor_cache(): return f else: return f @DebugContext.wrap @torch.utils._python_dispatch._disable_current_modes() @time_and_log(attr="compilation time (in seconds)") # Need this decorator for compile_fx_inner even if we already have one for # compile_fx. The reason is the compilation for backward graph may happen after # compile_fx return and we may want to use the _LazyGraphModule for compiling # the backward graph as well. @_use_lazy_graph_module(dynamo_config.use_lazy_graph_module) @with_fresh_cache_if_config @dynamo_utils.dynamo_timed(phase_name="inductor_compile", fwd_only=False) def compile_fx_inner( gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor], cudagraphs: Optional[BoxedBool] = None, static_input_idxs: Optional[List[int]] = None, is_backward: bool = False, graph_id: Optional[int] = None, cpp_wrapper: bool = False, aot_mode: bool = False, is_inference: bool = False, boxed_forward_device_index: Optional[BoxedDeviceIndex] = None, user_visible_outputs: Optional[Dict[str, None]] = None, layout_opt: Optional[bool] = None, extern_node_serializer: Optional[Callable[[List[ExternKernelNode]], Any]] = None, ) -> Union[CompiledFxGraph, str]: """ Inductor API that compiles a single graph. If you change the argument list for this function, make sure you also update the call to save_args_for_compile_fx_inner below accordingly. """ if dynamo_utils.count_calls(gm.graph) == 0 and not aot_mode: # trigger the real recompilation for _LazyGraphModule before returning # the forward method. from torch.fx._lazy_graph_module import _LazyGraphModule _LazyGraphModule.force_recompile(gm) return make_boxed_func(gm.forward) if static_input_idxs is None: static_input_idxs = [] assert isinstance( next(iter(reversed(gm.graph.nodes))).args[0], (tuple, list) ), f"inductor can only compile FX graphs which return a tuple/list, but got {gm.graph}" if config.save_args: save_args_for_compile_fx_inner( gm, example_inputs, cudagraphs=cudagraphs, static_input_idxs=static_input_idxs, is_backward=is_backward, graph_id=graph_id, cpp_wrapper=cpp_wrapper, aot_mode=aot_mode, is_inference=is_inference, boxed_forward_device_index=boxed_forward_device_index, user_visible_outputs=user_visible_outputs, layout_opt=layout_opt, ) if cudagraphs is None: cudagraphs = BoxedBool(config.triton.cudagraphs) # Inputs to fx_codegen_and_compile # Anything that affects codegen should go here, so if the signature # of fx_codegen_and_compile changes, the dict should be updated accordingly graph_kwargs = { "cudagraphs": cudagraphs, "static_input_idxs": static_input_idxs, "is_backward": is_backward, "graph_id": graph_id, "cpp_wrapper": cpp_wrapper, "aot_mode": aot_mode, "is_inference": is_inference, "user_visible_outputs": user_visible_outputs, "layout_opt": layout_opt, "extern_node_serializer": extern_node_serializer, } start = time.time() fx_graph_remote_cache = should_use_remote_fx_graph_cache() inputs_to_check = get_input_idxs_to_check(example_inputs, static_input_idxs) if ( not config.force_disable_caches and (config.fx_graph_cache or fx_graph_remote_cache) and not aot_mode ): for i, input in enumerate(example_inputs): if ( isinstance(input, torch.Tensor) and input.device.type == "cuda" and i in static_input_idxs ): input._is_inductor_static = True # type: ignore[attr-defined] compiled_graph = FxGraphCache.load( fx_codegen_and_compile, gm, example_inputs, graph_kwargs, inputs_to_check, local=config.fx_graph_cache, remote=fx_graph_remote_cache, ) else: compiled_graph = fx_codegen_and_compile( gm, example_inputs, **graph_kwargs # type: ignore[arg-type] ) log.debug("FX codegen and compilation took %.3fs", time.time() - start) # check cudagraph disabling reasons from inductor lowering if cudagraphs and compiled_graph.disabled_cudagraphs_reason: if "cuda" in compiled_graph.device_types: log_cudagraph_skip_and_bump_counter( f"skipping cudagraphs due to {compiled_graph.disabled_cudagraphs_reason}" ) else: counters["inductor"]["cudagraph_skips"] += 1 BoxedBool.disable(cudagraphs) # Return the output strides to the caller via TracingContext context = torch._guards.TracingContext.try_get() if context is not None and context.output_strides is not None: assert len(context.output_strides) == 0 context.output_strides.extend(compiled_graph.output_strides) if aot_mode: return compiled_graph if cudagraphs: # output args are tuple of first argument output = output_node(gm) assert len(output.args) == 1 stack_traces = [ (arg.stack_trace if isinstance(arg, torch.fx.node.Node) else None) for arg in output.args[0] ] complex_memory_overlap_inputs = any( complex_memory_overlap(t) for t in example_inputs if isinstance(t, torch.Tensor) ) if not config.triton.cudagraph_support_input_mutation: # Skip supports for cudagraph-managed tensors from torch._inductor.cudagraph_utils import ( check_for_mutation_ignore_cuda_graph_managed_tensor, ) has_mutation_str = check_for_mutation_ignore_cuda_graph_managed_tensor( gm, compiled_graph, static_input_idxs ) has_mutation = has_mutation_str is not None if has_mutation: compiled_graph.disabled_cudagraphs_reason = has_mutation_str else: # Check mutation later to support cudagraph-managed tensors has_mutation = None cudagraph_tests = [ (not has_mutation, "mutated inputs"), (not has_incompatible_cudagraph_ops(gm), "incompatible ops"), (not complex_memory_overlap_inputs, "complex memory overlap"), ( all( isinstance(t, (torch.Tensor, torch.SymInt)) for t in example_inputs ), "non-Tensor inputs", ), ] cudagraph_fail_reasons = [s for b, s in cudagraph_tests if not b] if not cudagraph_fail_reasons: if not config.triton.cudagraph_trees: # Force specialize all inputs so that CUDA graphs will work for t in example_inputs: if isinstance(t, torch.SymInt): int(t) # guard if ( boxed_forward_device_index is not None and not is_inference and not is_backward ): boxed_forward_device_index.set(next(iter(compiled_graph.device_idxs))) compiled_graph.current_callable = cudagraphify( compiled_graph.current_callable, example_inputs, static_input_idxs=static_input_idxs, device_index=next(iter(compiled_graph.device_idxs)), stack_traces=stack_traces, is_backward=is_backward, is_inference=is_inference, constants=tuple(compiled_graph.constants.values()), placeholders=tuple(get_placeholders(gm.graph)), mutated_input_idxs=tuple(compiled_graph.mutated_input_idxs), ) else: BoxedBool.disable(cudagraphs) # See [Backward Generation Handling] # if cudagraph'd the forward and set the device, we need to let the cudagraph manager # know we are we running the backward even if we will not run it in cudagraphs if is_backward and config.triton.cudagraph_trees: assert boxed_forward_device_index is not None assert boxed_forward_device_index.value is not None compiled_graph_callable = compiled_graph.current_callable manager = torch._inductor.cudagraph_trees.get_manager( boxed_forward_device_index.value, create_if_none_exists=False ) # should already exist from forward assert manager is not None def compiled_artifact(new_inputs): manager.set_to_running_backward() # type: ignore[union-attr] return compiled_graph_callable(new_inputs) compiled_graph.current_callable = compiled_artifact if "cuda" in compiled_graph.device_types: # prefer better disable_cudagraphs_reason bc stack trace # TODO: migrate all disable reasons to stack trace, refactor if compiled_graph.disabled_cudagraphs_reason: log_cudagraph_skip_and_bump_counter( compiled_graph.disabled_cudagraphs_reason ) else: log_cudagraph_skip_and_bump_counter( f"skipping cudagraphs due to {cudagraph_fail_reasons}" ) # cudagraphs does its own aligning of inputs if not cudagraphs: new_callable = align_inputs_from_check_idxs( compiled_graph.current_callable, inputs_to_check ) if new_callable is not compiled_graph.current_callable: compiled_graph.current_callable = new_callable _step_logger()( logging.INFO, "torchinductor done compiling " f"{'BACKWARDS' if is_backward else 'FORWARDS'} " f"graph {graph_id}", ) # aot autograd needs to know to pass in inputs as a list compiled_graph._boxed_call = True return compiled_graph @dynamo_utils.preserve_rng_state() def fx_codegen_and_compile( gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor], cudagraphs: Optional[BoxedBool] = None, static_input_idxs: Optional[List[int]] = None, is_backward: bool = False, graph_id: Optional[int] = None, cpp_wrapper: bool = False, aot_mode: bool = False, is_inference: bool = False, # Use a dict with None value rather than a set for deterministic # iteration order just in case. user_visible_outputs: Optional[Dict[str, None]] = None, layout_opt: Optional[bool] = None, extern_node_serializer: Optional[Callable[[List[ExternKernelNode]], Any]] = None, ) -> Union[CompiledFxGraph, str]: if is_tf32_warning_applicable(gm): _warn_tf32_disabled() # lift the maximum depth of the Python interpreter stack # to adapt large/deep models sys.setrecursionlimit(max(sys.getrecursionlimit(), 2000)) _step_logger()( logging.INFO, "torchinductor compiling " f"{'BACKWARDS' if is_backward else 'FORWARDS'} " f"graph {graph_id}", ) V.debug.fx_graph(gm, example_inputs) # TODO: Should we actually dump this? It should be redundant with the aot # structured logs... # trace_structured("inductor_input_graph", payload_fn=lambda: gm.print_readable(print_output=False)) shape_env = _shape_env_from_inputs(example_inputs) # Convert view to reshape in the graph. This is necessary primarily for # layout optimization. Do it unconditionally for uniformity. # # It's needed because when we do layout optimization, an contiguous tensor # in eager mode may becomes a channels last tensor. A view op previously # can be applied to the contiguous tensor may not be able to be applied # on the channels tensor any more. An error like # RuntimeError: view size is not compatible with input tensor's size and stride # (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead. # will be printed. # # Replace view op to reshape op in this case. # As an example, timm_resnest/botnet26t_256/convnext_base etc. will fail if we don't do this. # # Also this has to be done before FakeTensorProp below to avoid the failed # .view() call. view_to_reshape(gm) # It is safe to run FakeTensorProp under no_grad because by the time # we're in inductor, we assume that AOTAutograd has already "taken care" # of autograd, so there should be no more autograd-related API's in the # graph. with torch.no_grad(): fake_mode = fake_tensor_prop(gm, example_inputs) # pattern matcher passes might not preserve striding information # on node.meta["val"]. if in the future we rely on these being # correct we will need to fix. with V.set_fake_mode(fake_mode): # has some issues with memory in training _recursive_post_grad_passes(gm, is_inference=is_inference) V.debug.fx_graph_transformed(gm, example_inputs) post_grad_graphs_log.debug( "%s", lazy_format_graph_code( "AFTER POST GRAD", gm, include_stride=True, include_device=True ), ) trace_structured( "inductor_post_grad_graph", payload_fn=lambda: gm.print_readable( print_output=False, include_stride=True, include_device=True ), ) if config.is_fbcode(): log_optimus_to_scuba( extra_logging={"pt2_configs": str(get_patched_config_dict())} ) with V.set_fake_mode(fake_mode), maybe_disable_comprehensive_padding( example_inputs ): const_output_index = None const_graph = None const_code = None if aot_mode and config.aot_inductor.use_runtime_constant_folding: const_gm, const_output_index = split_const_gm(gm) const_graph = GraphLowering( const_gm, example_inputs=[], shape_env=shape_env, graph_id=graph_id, cpp_wrapper=cpp_wrapper, aot_mode=aot_mode, user_visible_outputs=user_visible_outputs, extern_node_serializer=extern_node_serializer, is_inference=is_inference, is_const_graph=True, ) with V.set_graph_handler(const_graph): assert cpp_wrapper, "AOT mode only supports C++ wrapper" const_graph.run() const_code, _ = const_graph.codegen_with_cpp_wrapper() graph = GraphLowering( gm, # example_inputs will be used by AOTInductor to dry-run the generated code for Triton kernel tuning. # For the forward pass, we have the real inputs to be used as example_inputs. For the backward pass, # we currently use fake tensors and defake them later. example_inputs=example_inputs, shape_env=shape_env, graph_id=graph_id, cpp_wrapper=cpp_wrapper, aot_mode=aot_mode, user_visible_outputs=user_visible_outputs, extern_node_serializer=extern_node_serializer, is_inference=is_inference, const_output_index=const_output_index, const_code=const_code, const_module=const_graph, ) metrics_helper = metrics.CachedMetricsHelper() with V.set_graph_handler(graph): graph.run(*example_inputs) output_strides: List[Optional[Tuple[int, ...]]] = [] if graph.graph_outputs is not None: # We'll put the output strides in the compiled graph so we # can later return them to the caller via TracingContext for out in graph.graph_outputs: if ( hasattr(out, "layout") and len(free_unbacked_symbols(out.layout.stride)) == 0 ): output_strides.append( tuple( V.graph.sizevars.size_hint(s) for s in out.layout.stride ) ) else: output_strides.append(None) _check_triton_bf16_support(graph) compiled_fn = graph.compile_to_fn() num_bytes, nodes_num_elem, node_runtimes = graph.count_bytes() metrics.num_bytes_accessed += num_bytes metrics.node_runtimes += node_runtimes metrics.nodes_num_elem += nodes_num_elem if ( cudagraphs and config.triton.cudagraph_skip_dynamic_graphs and not V.graph.disable_cudagraphs_reason and torch._inductor.utils.any_is_symbolic(*example_inputs) ): stack_trace = None for node in gm.graph.nodes: meta_val = node.meta.get("val", None) if ( node.op == "placeholder" or not isinstance(meta_val, torch.Tensor) or not torch._inductor.utils.any_is_symbolic(meta_val) ): continue if stack_trace := node.meta.get("stack_trace", None): break disable = "graph with symbolic shapes inputs and config.triton.cudagraph_skip_dynamic_graphs=True." if stack_trace: disable = f"{disable} Found from {stack_trace}\n" else: disable = f"{disable}\n" V.graph.disable_cudagraphs_reason = disable if V.aot_compilation is True: return compiled_fn if cudagraphs and not V.graph.disable_cudagraphs_reason: from torch._inductor.cudagraph_utils import ( check_lowering_disable_cudagraph, ) V.graph.disable_cudagraphs_reason = check_lowering_disable_cudagraph( V.graph.device_node_mapping ) compiled_graph = CompiledFxGraph( compiled_fn, graph, output_strides, V.graph.disable_cudagraphs_reason, metrics_helper.get_deltas(), ) return compiled_graph def clone_preserve_strides(x: torch.Tensor): needed_size = ( sum((shape - 1) * stride for shape, stride in zip(x.size(), x.stride())) + 1 ) buffer = torch.as_strided(x, (needed_size,), (1,)).clone() return torch.as_strided(buffer, x.size(), x.stride()) def copy_misaligned_inputs( new_inputs: List[torch.Tensor], check_inputs_idxs: Sequence[int] ) -> None: for i in check_inputs_idxs: if new_inputs[i].data_ptr() % ALIGNMENT: new_inputs[i] = clone_preserve_strides(new_inputs[i]) def get_input_idxs_to_check( inputs: Union[List[torch.Tensor], Sequence[int]], static_input_idxs: Sequence[int], ) -> Sequence[int]: """ This function runs at compile time, and generates a list of indices for which we might need to do a copy to preserve alignment requirements. """ ids_to_check = [] for i, input in enumerate(inputs): if not isinstance(input, torch.Tensor): # non-tensors don't need alignment continue if input.device.type != "cuda": # right now we only care for cuda tensors continue with maybe_get_suppress_shape_guards_ctx(): # suppress guards so that tensor_is_aligned and should_assume_input_aligned # do not add guards on input's storage offset if i in static_input_idxs and tensor_is_aligned(input): continue if not should_assume_input_aligned(input): continue # if we get here, then # (a) our triton code assumes that the input is aligned # (b) we can't be sure ahead of time that the input will actually be aligned. # therefore, at runtime, we'll need to check that the input is aligned # (and if not, clone it to make it aligned.) ids_to_check.append(i) return ids_to_check def align_inputs_from_check_idxs( model: Callable[[List[torch.Tensor]], Any], inputs_to_check: Sequence[int] ): if len(inputs_to_check) == 0: return model def run(new_inputs): copy_misaligned_inputs(new_inputs, inputs_to_check) return model(new_inputs) return run @dynamo_utils.dynamo_timed def cudagraphify( model: torch.fx.GraphModule, inputs: List[torch.Tensor], static_input_idxs: Sequence[int] = (), *, device_index: int, stack_traces: List[Optional[str]], is_backward: bool, is_inference: bool, constants: Tuple[torch.Tensor, ...] = (), placeholders: Tuple[torch.fx.Node, ...] = (), mutated_input_idxs: Tuple[int, ...] = (), ): from torch._inductor.cudagraph_trees import ( cudagraphify_impl as new_cudagraphify_impl, ) cudagraphify_fn: Callable[..., Any] if config.triton.cudagraph_trees: cudagraphify_fn = functools.partial( new_cudagraphify_impl, device_index=device_index, stack_traces=stack_traces, is_backward=is_backward, is_inference=is_inference, constants=constants, placeholders=placeholders, mutated_input_idxs=mutated_input_idxs, ) else: cudagraphify_fn = cudagraphify_impl # if using fake tensors, defer cudagraphs until we get real inputs at runtime if not any(isinstance(inp, FakeTensor) for inp in inputs): return cudagraphify_fn(model, inputs, static_input_idxs) compiled_fn = None def run(new_inputs): nonlocal compiled_fn if compiled_fn is None: with dynamo_utils.preserve_rng_state(): compiled_fn = cudagraphify_fn(model, new_inputs, static_input_idxs) return compiled_fn(new_inputs) return run def remove_unaligned_input_idxs( inputs: Union[List[torch.Tensor], Sequence[int]], static_input_idxs: Sequence[int], ): """ We require all inputs to be aligned, so introduce a copy for any that aren't. """ aligned_static_input_idxs = [] for idx, input in zip(static_input_idxs, inputs): if isinstance(input, torch.Tensor) and (input.data_ptr() % ALIGNMENT) == 0: aligned_static_input_idxs.append(idx) if len(aligned_static_input_idxs) != len(static_input_idxs): return aligned_static_input_idxs return static_input_idxs def static_input(x: torch.Tensor): """ Copy and input while preserving strides """ # TODO(jansel): figure out why this version doesn't work: # return torch.empty_strided(x.size(), x.stride(), dtype=x.dtype, device=x.device) needed_size = ( sum((shape - 1) * stride for shape, stride in zip(x.size(), x.stride())) + 1 ) buffer = torch.empty(needed_size, dtype=x.dtype, device=x.device) return torch.as_strided(buffer, x.size(), x.stride()) def index_expanded_dims_and_copy_( dst: torch.Tensor, src: torch.Tensor, expanded_dims: List[int], ): "Index into expanded dimensions of both dst and src then copy_" dst = index_expanded_dims(dst, expanded_dims) src = index_expanded_dims(src, expanded_dims) dst.copy_(src) def cudagraphify_impl( model: torch.fx.GraphModule, inputs: List[torch.Tensor], static_input_idxs: Sequence[int] = (), ): """ Assumes inputs[static_input_idxs[i]] are always the same memory address """ check_input_idxs = get_input_idxs_to_check(inputs, static_input_idxs) static_input_idxs = remove_unaligned_input_idxs(inputs, static_input_idxs) copy_misaligned_inputs(inputs, check_input_idxs) assert isinstance(inputs, list) inps_expanded_dims = [ get_expanded_dims(x) if idx not in static_input_idxs else [] for idx, x in enumerate(inputs) ] # allocate static tensor inputs static_inputs = [ x if not isinstance(x, torch.Tensor) else static_input(x) if idx not in static_input_idxs else x.detach() for idx, x in enumerate(inputs) ] # copy over input values for fresh allocations for idx, (x, expanded_dims) in enumerate(zip(inputs, inps_expanded_dims)): if isinstance(x, torch.Tensor) and idx not in static_input_idxs: index_expanded_dims_and_copy_(static_inputs[idx], x, expanded_dims) # warmup torch.cuda.synchronize() stream = torch.cuda.Stream() stream.wait_stream(torch.cuda.current_stream()) # copy static_inputs because it will be cleared in model with torch.cuda.stream(stream): model(list(static_inputs)) stream.synchronize() torch.cuda.current_stream().wait_stream(stream) torch.cuda.synchronize() # record graph = torch.cuda.CUDAGraph() with torch.cuda.graph(graph, stream=stream, capture_error_mode="thread_local"): static_outputs = model(list(static_inputs)) if not isinstance(static_outputs, (list, tuple)): static_outputs = (static_outputs,) if config.size_asserts: def run(new_inputs): assert len(static_inputs) == len(new_inputs) for idx, (dst, src, expanded_dims) in enumerate( zip(static_inputs, new_inputs, inps_expanded_dims) ): if not isinstance(dst, torch.Tensor): pass elif idx in static_input_idxs: assert dst.data_ptr() == src.data_ptr() else: # TODO - could make one single op of multiple slices # and avoid dispatch. # Could also pre-index the `dst` tensors index_expanded_dims_and_copy_(dst, src, expanded_dims) new_inputs.clear() graph.replay() return static_outputs else: copy_indices = [ idx for idx in range(len(static_inputs)) if idx not in static_input_idxs ] def run(new_inputs): for idx in copy_indices: expanded_dims = inps_expanded_dims[idx] index_expanded_dims_and_copy_( static_inputs[idx], new_inputs[idx], expanded_dims ) new_inputs.clear() graph.replay() return static_outputs return align_inputs_from_check_idxs(run, check_input_idxs) def compile_fx_aot( model_: torch.fx.GraphModule, example_inputs_: List[torch.Tensor], inner_compile: Callable[..., Any] = compile_fx_inner, config_patches: Optional[Dict[str, Any]] = None, ): config_patches: Dict[str, Any] = ( {"cpp_wrapper": True} if config_patches is None else {**config_patches, "cpp_wrapper": True} ) if ( "aot_inductor.output_path" not in config_patches and not config.aot_inductor.output_path ): config_patches = { **config_patches, "aot_inductor.output_path": code_hash(model_.code), } extern_node_serializer = config_patches.pop("extern_node_serializer", None) with V.set_aot_compilation(True): compiled_lib_path = compile_fx( model_, example_inputs_, inner_compile=functools.partial( inner_compile, aot_mode=True, extern_node_serializer=extern_node_serializer, ), config_patches=config_patches, ) assert os.path.exists( compiled_lib_path ), f"AOTInductor compiled library does not exist at {compiled_lib_path}" return compiled_lib_path _graph_counter = count(0) def fw_compiler_freezing( aot_autograd_model: torch.fx.GraphModule, aot_example_inputs: List[torch.Tensor], dynamo_model: torch.fx.GraphModule, num_example_inputs: int, inner_compile: Callable[..., Any], cudagraphs: BoxedBool, graph_id: int, forward_device: BoxedDeviceIndex, ): from torch._inductor.freezing import convert_conv_weights_to_channels_last, freeze # partition_fn won't be called _recursive_joint_graph_passes(aot_autograd_model) layout_opt = GraphLowering.decide_layout_opt(aot_autograd_model, is_inference=True) if layout_opt: # make sure meta['val'] is properly setup fake_tensor_prop(aot_autograd_model, aot_example_inputs, True) convert_conv_weights_to_channels_last(aot_autograd_model) opt_model, preserved_arg_indices = freeze( dynamo_model, aot_autograd_model, aot_example_inputs, # type: ignore[arg-type] ) aot_example_inputs = [aot_example_inputs[ind] for ind in preserved_arg_indices] num_fixed = len(preserved_arg_indices) - num_example_inputs fake_mode = detect_fake_mode(aot_example_inputs) # for freezing, all graph outputs should be user visible *_, model_outputs_node = opt_model.graph.nodes model_outputs = model_outputs_node.args[0] user_visible_outputs = dict.fromkeys( n.name for n in model_outputs if isinstance(n, torch.fx.Node) ) static_input_idxs = list(range(num_fixed)) # constant params will be real tensors, not fake tracing_context = torch._guards.TracingContext.try_get() if tracing_context is not None: params_flat = tracing_context.params_flat assert params_flat is not None for i in range(len(params_flat)): if i not in preserved_arg_indices: params_flat[i] = None if tracing_context.fw_metadata: static_input_idxs += tracing_context.fw_metadata.static_parameter_indices with mock.patch.object(fake_mode, "allow_non_fake_inputs", True): optimized_function = inner_compile( opt_model, aot_example_inputs, static_input_idxs=static_input_idxs, cudagraphs=cudagraphs, graph_id=graph_id, is_inference=True, boxed_forward_device_index=forward_device, layout_opt=layout_opt, user_visible_outputs=user_visible_outputs, ) # aot_inductor codegens a call that takes in just the inputs, so we don't return a wrapper # that drops constant-ified params if V.aot_compilation is True: return optimized_function def wrapper(args): args_new = [args[i] for i in preserved_arg_indices] args.clear() return optimized_function(args_new) wrapper._boxed_call = True # type: ignore[attr-defined] return wrapper @_use_lazy_graph_module(dynamo_config.use_lazy_graph_module) def compile_fx( model_: torch.fx.GraphModule, example_inputs_: List[torch.Tensor], inner_compile: Callable[..., Any] = compile_fx_inner, config_patches: Optional[Dict[str, Any]] = None, decompositions: Optional[Dict[OpOverload, Callable[..., Any]]] = None, ): """Main entrypoint to a compile given FX graph""" if config_patches: with config.patch(config_patches): return compile_fx( model_, example_inputs_, # need extra layer of patching as backwards is compiled out of scope inner_compile=config.patch(config_patches)(inner_compile), decompositions=decompositions, ) if config.cpp_wrapper: with config.patch( { "cpp_wrapper": False, "triton.autotune_cublasLt": False, "triton.cudagraphs": False, "triton.store_cubin": True, } ), V.set_real_inputs(example_inputs_): inputs_ = example_inputs_ if isinstance(model_, torch.fx.GraphModule): fake_inputs = [ node.meta.get("val") for node in model_.graph.nodes if node.op == "placeholder" ] if all(v is not None for v in fake_inputs): # Validate devices before switching to fake tensors. for idx, fi, i in zip(count(), fake_inputs, inputs_): if fi.device != i.device: raise ValueError( f"Device mismatch between fake input and example input at position #{idx}: " f"{fi.device} vs {i.device}. If the model was exported via torch.export(), " "make sure torch.export() and torch.aot_compile() run on the same device." ) inputs_ = fake_inputs return compile_fx( model_, inputs_, inner_compile=functools.partial(inner_compile, cpp_wrapper=True), decompositions=decompositions, ) recursive_compile_fx = functools.partial( compile_fx, inner_compile=inner_compile, decompositions=decompositions, ) if not graph_returns_tuple(model_): return make_graph_return_tuple( model_, example_inputs_, recursive_compile_fx, ) if isinstance(model_, torch.fx.GraphModule): if isinstance(model_.graph._codegen, _PyTreeCodeGen): # this graph is the result of dynamo.export() return handle_dynamo_export_graph( model_, example_inputs_, recursive_compile_fx, ) model_ = _recursive_pre_grad_passes(model_, example_inputs_) if any(isinstance(x, (list, tuple, dict)) for x in example_inputs_): return flatten_graph_inputs( model_, example_inputs_, recursive_compile_fx, ) assert not config._raise_error_for_testing num_example_inputs = len(example_inputs_) cudagraphs = BoxedBool(config.triton.cudagraphs) forward_device = BoxedDeviceIndex(None) graph_id = next(_graph_counter) decompositions = ( decompositions if decompositions is not None else select_decomp_table() ) @dynamo_utils.dynamo_timed def fw_compiler_base( model: torch.fx.GraphModule, example_inputs: List[torch.Tensor], is_inference: bool, ): if is_inference: # partition_fn won't be called _recursive_joint_graph_passes(model) fixed = torch._inductor.utils.num_fw_fixed_arguments( num_example_inputs, len(example_inputs) ) user_visible_outputs = {} if config.keep_output_stride: model_outputs_node = output_node(model) model_outputs = pytree.arg_tree_leaves(*model_outputs_node.args) num_model_outputs = len(model_outputs) context = torch._guards.TracingContext.try_get() # See Note [User Outputs in the inductor graph] if context is not None and context.fw_metadata and not is_inference: original_output_start_index = ( context.fw_metadata.num_mutated_inp_runtime_indices ) else: original_output_start_index = 0 if isinstance(model_, torch.fx.GraphModule): *_, orig_model_outputs_node = model_.graph.nodes assert orig_model_outputs_node.op == "output" orig_model_outputs, _ = pytree.tree_flatten( orig_model_outputs_node.args ) num_orig_model_outputs = len(orig_model_outputs) else: num_orig_model_outputs = num_model_outputs assert num_orig_model_outputs <= num_model_outputs # Note [User Outputs in the inductor graph] # We makes the following assumption # For inference # len(orig_model_outputs) == len(model_outputs) # For training # len(orig_model_outputs) <= len(model_outputs) # During training, most of the time the model_outputs starts with # original module's outputs followed by saved activations. # But this can be not true if the model have inplace updated tensors. # AOTAutograd will make those tensors being returned before the original # module's output. # To make things safe, we'll use original_output_start_index field # set by AOTAutograd to decide where the original module outputs start. orig_output_end_idx = original_output_start_index + num_orig_model_outputs # Sanity chec: we are about to splice out the "user" outputs from the full set # of "graph" outputs. Make sure we're within bounds. assert orig_output_end_idx <= num_model_outputs user_visible_outputs = dict.fromkeys( n.name for n in model_outputs[original_output_start_index:orig_output_end_idx] if isinstance(n, torch.fx.Node) ) return inner_compile( model, example_inputs, static_input_idxs=get_static_input_idxs(fixed), cudagraphs=cudagraphs, graph_id=graph_id, is_inference=is_inference, boxed_forward_device_index=forward_device, user_visible_outputs=user_visible_outputs, ) fw_compiler = functools.partial(fw_compiler_base, is_inference=False) if config.freezing and not torch.is_grad_enabled(): inference_compiler = functools.partial( fw_compiler_freezing, dynamo_model=model_, num_example_inputs=num_example_inputs, inner_compile=inner_compile, cudagraphs=cudagraphs, graph_id=graph_id, forward_device=forward_device, ) else: inference_compiler = functools.partial(fw_compiler_base, is_inference=True) def partition_fn(graph, joint_inputs, **kwargs): _recursive_joint_graph_passes(graph) return min_cut_rematerialization_partition( graph, joint_inputs, **kwargs, compiler="inductor" ) @compile_time_strobelight_meta(phase_name="bw_compiler") @dynamo_utils.dynamo_timed def bw_compiler(model: torch.fx.GraphModule, example_inputs: List[torch.Tensor]): user_visible_outputs = {} if config.bw_outputs_user_visible: model_outputs_node = output_node(model) model_outputs = pytree.arg_tree_leaves(*model_outputs_node.args) user_visible_outputs = dict.fromkeys( n.name for n in model_outputs if isinstance(n, torch.fx.Node) ) fixed = count_tangents(model) return inner_compile( model, example_inputs, static_input_idxs=list(range(fixed)), cudagraphs=cudagraphs, is_backward=True, graph_id=graph_id, boxed_forward_device_index=forward_device, user_visible_outputs=user_visible_outputs, ) # TODO: can add logging before/after the call to create_aot_dispatcher_function # in torch._functorch/aot_autograd.py::aot_module_simplified::aot_function_simplified::new_func # once torchdynamo is merged into pytorch fake_mode = detect_fake_mode(example_inputs_) or torch._subclasses.FakeTensorMode( allow_non_fake_inputs=True ) tracing_context = ( torch._guards.TracingContext.try_get() or torch._guards.TracingContext(fake_mode) ) if V.aot_compilation is True: with functorch_config.patch(unlift_effect_tokens=True): gm, graph_signature = aot_export_module( model_, example_inputs_, trace_joint=False, decompositions=decompositions, ) unlifted_gm = _unlift_graph(model_, gm, graph_signature) if "dynamo_flat_name_to_original_fqn" in model_.meta: unlifted_gm.meta["dynamo_flat_name_to_original_fqn"] = model_.meta[ "dynamo_flat_name_to_original_fqn" ] # Disable amp as in aot_dispatch_autograd (https://github.com/pytorch/pytorch/pull/86515) # In inference_compiler (fw_compiler_base), _recursive_joint_graph_passes will call into # _sfdp_init() to register patterns. # When fallback_random is set to True, the sdpa patterns will be traced during runtime. # If amp is turned on, the traced FP32 patterns will have prims.convert_element_type which # will be the same as the generated FP16 patterns. disable_amp = torch._C._is_any_autocast_enabled() context = torch._C._DisableAutocast if disable_amp else contextlib.nullcontext with V.set_fake_mode(fake_mode), compiled_autograd.disable(), context(): return inference_compiler(unlifted_gm, example_inputs_) with V.set_fake_mode(fake_mode), torch._guards.tracing( tracing_context ), compiled_autograd.disable(), functorch_config.patch(unlift_effect_tokens=True): return aot_autograd( fw_compiler=fw_compiler, bw_compiler=bw_compiler, inference_compiler=inference_compiler, decompositions=decompositions, partition_fn=partition_fn, keep_inference_input_mutations=True, )(model_, example_inputs_) def _shape_env_from_inputs(inputs: List[torch.Tensor]): shape_env = None fake_mode = detect_fake_mode(inputs) # TODO(voz): It would be nice to enable this assert, but there are lots of tests that # pass in real inputs for now. # if len(inputs) > 0: # assert fake_mode is not None, breakpoint() if fake_mode is not None: return fake_mode.shape_env # When there are no tensor inputs, get shape_env from the first SymInt. for input in inputs: if isinstance(input, torch.SymInt): return input.node.shape_env # TODO(voz): Should we always have one anyway? return None def graph_returns_tuple(gm: torch.fx.GraphModule): """True if a FX graph returns a tuple""" if not isinstance(gm, torch.fx.GraphModule): return True # can't check this, assume true (rv,) = output_node(gm).args if isinstance(rv, (list, tuple)): return True if ( isinstance(rv, torch.fx.node.Node) and hasattr(rv.target, "_schema") and len(rv.target._schema.returns) > 1 and all(str(ret.type) == "Tensor" for ret in rv.target._schema.returns) ): # for graphs whose result is one node with multiple outputs return True return False def make_graph_return_tuple( gm: torch.fx.GraphModule, inputs: List[torch.Tensor], compile_gm: Callable[..., Any], ): """ Mutate gm so it returns a tuple. This is only needed for graphs not created by torchdynamo that return non-tuples. """ node = output_node(gm) (rv,) = node.args rv, spec = pytree.tree_flatten(rv) with gm.graph.inserting_before(node): gm.graph.output(rv) gm.graph.erase_node(node) assert graph_returns_tuple(gm) compiled_fn = compile_gm(gm, inputs) @functools.wraps(compiled_fn) def wrapper(*args, **kwargs): return pytree.tree_unflatten(compiled_fn(*args, **kwargs), spec) return wrapper def handle_dynamo_export_graph( gm: torch.fx.GraphModule, inputs: List[torch.Tensor], compile_gm: Callable[..., Any], ): """ `torch._dynamo.export` embeds pytrees in the FX graph codegen object, convert that to a normal FX graph so inductor can compile it. """ codegen = gm.graph._codegen gm.graph._codegen = torch.fx.graph.CodeGen() gm.recompile() compiled_fn = compile_gm(gm, codegen.process_inputs(*inputs)) @functools.wraps(compiled_fn) def wrapper(*args): return codegen.process_outputs(compiled_fn(*codegen.process_inputs(*args))) return wrapper def _check_triton_bf16_support(graph: GraphLowering) -> None: def warn_and_skip(device) -> None: from torch._dynamo.exc import SkipFrame device_props = torch.cuda.get_device_properties(device) warnings.warn( f"{device_props.name} does not support bfloat16 compilation natively, skipping" ) raise SkipFrame("BF16 is not supported") for inp in graph.graph_inputs.values(): device = getattr(inp, "get_device", lambda: torch.device("meta"))() if device.type != "cuda" or inp.get_dtype() != torch.bfloat16: continue # Print warning and skip frame if attempting to compile for bfloat16 # on device without hardware support for dtype if torch.cuda.is_bf16_supported(including_emulation=False): return warn_and_skip(device) for out in graph.graph_outputs: device = getattr(out, "get_device", lambda: torch.device("meta"))() if device.type != "cuda" or out.get_dtype() != torch.bfloat16: continue # Print warning and skip frame if attempting to compile for bfloat16 # on device without hardware support for dtype if torch.cuda.is_bf16_supported(including_emulation=False): return warn_and_skip(device)