# mypy: allow-untyped-decorators # mypy: allow-untyped-defs import functools import itertools import logging import operator from collections import Counter, defaultdict from typing import Any, Callable, Optional, TypeVar, Union from typing_extensions import ParamSpec import torch import torch._inductor as inductor import torch.utils._pytree as pytree from torch import fx from torch._decomp import register_decomposition from torch._dynamo.utils import counters from torch._inductor import comms from torch._inductor.virtualized import ops from torch._logging import trace_structured from torch._prims_common import is_boolean_dtype, is_expandable_to, is_integer_dtype from torch.fx.experimental.symbolic_shapes import statically_known_true, sym_eq from torch.utils._ordered_set import OrderedSet from .. import config, ir, pattern_matcher from ..codegen.common import custom_backend_passes from ..comms import remove_fsdp2_unsharded_param_graph_input_usage from ..fx_utils import FakeTensorUpdater, get_fake_args_kwargs, get_node_storage from ..lowering import lowerings as L from ..pattern_matcher import ( _return_true, Arg, CallFunction, CallFunctionVarArgs, filter_nodes, fwd_only, get_arg_value, get_mutation_region_id, Ignored, init_once_fakemode, KeywordArg, ListOf, Match, MultiOutputPattern, MULTIPLE, PatternMatcherPass, register_graph_pattern, register_replacement, stable_topological_sort, ) from ..utils import ( decode_device, get_all_devices, get_gpu_type, is_gpu, is_pointwise_use, OPTIMUS_EXCLUDE_POST_GRAD, ) from ..virtualized import V from .b2b_gemm import B2B_GEMM_PASS from .ddp_fusion import fuse_ddp_communication from .group_batch_fusion import group_batch_fusion_passes, POST_GRAD_FUSIONS from .micro_pipeline_tp import micro_pipeline_tp_pass from .pre_grad import is_same_dict, save_inductor_dict from .reinplace import reinplace_inplaceable_ops from .split_cat import POST_GRAD_PATTERNS _T = TypeVar("_T") _P = ParamSpec("_P") log = logging.getLogger(__name__) aten = torch.ops.aten prims = torch.ops.prims # First pass_patterns[0] are applied, then [1], then [2] pass_patterns = [ PatternMatcherPass(), PatternMatcherPass(), PatternMatcherPass(), ] def post_grad_passes(gm: torch.fx.GraphModule, is_inference: bool): """ Passes that run on after grad. This is called once on the forwards graph and once on the backwards graph. The IR here has been normalized and functionalized. """ GraphTransformObserver = functools.partial( torch.fx.passes.graph_transform_observer.GraphTransformObserver, subsystem="post_grad_passes", ) if not torch._dynamo.config.skip_fsdp_hooks: remove_fsdp2_unsharded_param_graph_input_usage(gm.graph) if config.dce: # has some issues with mutation in inference mode gm.graph.eliminate_dead_code() if is_inference and config.reorder_for_locality: GraphTransformObserver(gm, "reorder_for_locality").apply_graph_pass( reorder_for_locality ) fake_tensor_updater = FakeTensorUpdater(gm.graph) if post_grad_custom_pre_pass := config.post_grad_custom_pre_pass: GraphTransformObserver(gm, "post_grad_custom_pre_pass").apply_graph_pass( post_grad_custom_pre_pass ) if torch._C._has_mkldnn: if ( config.cpp.enable_grouped_gemm_template and config.max_autotune and "CPP" in config.max_autotune_gemm_backends ): from .mkldnn_fusion import grouped_gemm_pass grouped_gemm_pass(gm.graph) if config.cpp.enable_concat_linear: from .quantization import concat_linear_woq_int4 # Concat linear optimization for WOQ int4 concat_linear_woq_int4(gm) if config.pattern_matcher: lazy_init() GraphTransformObserver(gm, "post_grad_custom_pre_pass").apply_graph_pass( functools.partial(group_batch_fusion_passes, pre_grad=False) ) GraphTransformObserver(gm, "remove_noop_ops").apply_graph_pass(remove_noop_ops) GraphTransformObserver(gm, "remove_assert_ops").apply_graph_pass( remove_assert_ops ) for i, patterns in enumerate(pass_patterns): GraphTransformObserver(gm, f"pass_pattern_{i}").apply_graph_pass( patterns.apply ) for pass_name in config.post_grad_fusion_options: # skip all patterns for group batch fusions or quantization patterns if pass_name in POST_GRAD_FUSIONS or pass_name in OPTIMUS_EXCLUDE_POST_GRAD: continue pattern_matcher_pass = POST_GRAD_PATTERNS[pass_name] inductor_before_change = save_inductor_dict( [pattern_matcher_pass.pass_name] ) GraphTransformObserver(gm, pass_name).apply_graph_pass( pattern_matcher_pass.apply ) if not is_same_dict(counters["inductor"], inductor_before_change): trace_structured( "artifact", metadata_fn=lambda: { "name": f"{pattern_matcher_pass.pass_name}_post_grad", "encoding": "string", }, payload_fn=lambda: gm.print_readable( print_output=False, include_stride=True, include_device=True ), ) if config.b2b_gemm_pass: B2B_GEMM_PASS.apply(gm.graph) # type: ignore[arg-type] if config._micro_pipeline_tp: micro_pipeline_tp_pass(gm.graph) if config._fuse_ddp_communication: GraphTransformObserver(gm, "fuse_ddp_communication").apply_graph_pass( lambda graph: fuse_ddp_communication( graph, config._fuse_ddp_communication_passes, config._fuse_ddp_bucket_size, ) ) if post_grad_custom_post_pass := config.post_grad_custom_post_pass: GraphTransformObserver(gm, "post_grad_custom_post_pass").apply_graph_pass( post_grad_custom_post_pass ) GraphTransformObserver(gm, "stable_sort").apply_graph_pass(stable_topological_sort) GraphTransformObserver(gm, "move_constructors_to_cuda").apply_graph_pass( move_constructors_to_gpu ) fake_tensor_updater.incremental_update() for device, custom_backend_pass in custom_backend_passes.items(): if custom_backend_pass is not None: gm_devices = [d.type for d in get_all_devices(gm)] if device in gm_devices: pass_name = "custom_backend_passes_" + device GraphTransformObserver(gm, pass_name).apply_gm_pass(custom_backend_pass) # Keep these last, since they introduces mutation. Look at # ./fx_passes/README.md for a discussion of mutation invariants. GraphTransformObserver(gm, "reinplace_inplaceable_ops").apply_graph_pass( reinplace_inplaceable_ops ) GraphTransformObserver( gm, "decompose_triton_kernel_wrapper_functional" ).apply_graph_pass(decompose_triton_kernel_wrapper_functional) GraphTransformObserver(gm, "decompose_auto_functionalized").apply_graph_pass( decompose_auto_functionalized ) if not torch._dynamo.config.skip_fsdp_hooks: GraphTransformObserver(gm, "reinplace_fsdp_all_gather").apply_graph_pass( comms.reinplace_fsdp_all_gather ) GraphTransformObserver(gm, "decompose_scan_to_while_loop").apply_gm_pass( decompose_scan_to_while_loop ) GraphTransformObserver(gm, "decompose_map_to_while_loop").apply_gm_pass( decompose_map_to_while_loop ) gm.recompile() gm.graph.lint() def prepare_softmax_pattern(x, dim): xmax = x.amax(dim=dim, keepdim=True) xsub = x - xmax xexp = xsub.exp() xsum = xexp.sum(dim=dim, keepdim=True) return xmax, xsum, xsub, xexp def prepare_softmax_replacement(x, dim): """ Return xsub since otherwise log-softmax can not be matched due to a use of this intermediate node. Same reason to return xsub.exp() for softmax. """ from torch._inductor.inductor_prims import prepare_softmax_online xmax, xsum = prepare_softmax_online(x, dim) xsub = x - xmax return xmax, xsum, xsub, xsub.exp() def prepare_softmax_extra_check(match): """ We only have triton online softmax kernels currently. """ return ( config.online_softmax and match.kwargs["x"].meta["val"].device.type == "cuda" and config.cuda_backend == "triton" ) def decompose_map_to_while_loop(gm: torch.fx.GraphModule): """This is similar to decompose_scan_to_while_loop.""" graph_pass = PatternMatcherPass() @register_graph_pattern( CallFunctionVarArgs(torch.ops.higher_order.map_impl), pass_dict=graph_pass, ) def _(match: Match, *args, **kwargs): assert len(kwargs) == 0, ( "kwargs of map are not merged into args before entering decompose_map_to_while_loop_pass" ) subgraph, fx_xs, fx_additional_inputs = args sub_gm: torch.fx.GraphModule = getattr(gm, subgraph.target) cur_node = match.nodes[0] mapped_outputs = cur_node.meta["val"] def lower_to_while_loop(*args, **kwargs): assert len(kwargs) == 0 xs, additional_inputs = pytree.tree_unflatten(args, tree_spec) assert isinstance(xs, (tuple, list)) and isinstance( additional_inputs, (tuple, list) ), (xs, additional_inputs) map_length = xs[0].size(0) loop_idx = torch.zeros([], dtype=torch.int64, device=torch.device("cpu")) # Similar to NOTE [Pre-allocate scan's output buffer] bound_symbols = { arg.node.expr: arg for arg in pytree.tree_leaves((args, map_length)) if isinstance(arg, torch.SymInt) } out_buffers = [ torch.empty_strided( resolve_shape_to_proxy(out.size(), bound_symbols), resolve_shape_to_proxy(out.stride(), bound_symbols), device=out.device, dtype=out.dtype, layout=out.layout, requires_grad=out.requires_grad, ) for out in mapped_outputs ] while_loop_operands = (loop_idx, out_buffers, xs) while_loop_flat_operands, operands_spec = pytree.tree_flatten( while_loop_operands ) while_loop_additional_inputs = additional_inputs _, operands_and_additional_inputs_spec = pytree.tree_flatten( (*while_loop_operands, additional_inputs) ) def cond_fn(*flat_args): loop_idx, _, _, _ = pytree.tree_unflatten( flat_args, operands_and_additional_inputs_spec, ) return loop_idx < map_length def body_fn(*flat_args): loop_idx, out_bufs, xs, additional_inputs = pytree.tree_unflatten( flat_args, operands_and_additional_inputs_spec, ) idx_int = loop_idx.item() torch.ops.aten._assert_scalar.default(idx_int >= 0, "") torch.ops.aten._assert_scalar.default(idx_int < map_length, "") sub_xs = [torch.ops.aten.select.int(x, 0, idx_int) for x in xs] outs = sub_gm(*sub_xs, *additional_inputs) for out, buffer in zip(outs, out_bufs): buffer_slice = torch.ops.aten.select.int(buffer, 0, idx_int) buffer_slice.copy_(out) return loop_idx + 1, *out_bufs, *xs _, final_out, _ = pytree.tree_unflatten( torch.ops.higher_order.while_loop( cond_fn, body_fn, tuple(while_loop_flat_operands), tuple(while_loop_additional_inputs), ), operands_spec, ) return (final_out,) lower_to_while_loop_args, tree_spec = pytree.tree_flatten( (fx_xs, fx_additional_inputs) ) match.replace_by_example( lower_to_while_loop, lower_to_while_loop_args, run_functional_passes=False ) graph_pass.apply(gm) for node in gm.graph.find_nodes( op="call_function", target=torch.ops.higher_order.map_impl ): raise AssertionError("map is not lowered to while_loop") def resolve_shape_to_proxy( shape: list[Union[int, torch.SymInt]], bound_symbols: dict[Any, Any] ): """ Given a list of symints/ints, this function returns a calculated expression of bound_symbols' values. When we trace this function, we'll get a graph with call_function nodes that describes how the shape expr is computed from bound_symbols' values. Suppose shape = (s1*s2, s1+s2) and bound_symbols = {s1: arg0, s2: arg1}, the result will be (arg0 * arg1, arg0 + arg1). """ from torch.utils._sympy.interp import sympy_interp from torch.utils._sympy.reference import PythonReferenceAnalysis ret = [] for s in shape: if isinstance(s, torch.SymInt): ret.append( sympy_interp( PythonReferenceAnalysis, bound_symbols, s.node.expr, ), ) else: assert isinstance(s, int) ret.append(s) return ret def decompose_scan_to_while_loop(gm: torch.fx.GraphModule): """ NOTE [decompose scan to while_loop] This pass decomposes `scan` to `while_loop` by replacing the scan fx_node with a while_loop hop. Suppose we have a function f: def f(): init = torch.zeros([]) xs = torch.arange(4) ys = [] for i in range(xs.size(0)): init = xs[i] + init ys.append(init) # Return the final carry and stack the intermediates return init, torch.stack(ys) We could rewrite it with a scan with the benefits of reducing compilation time/binary size, reducing memory usage, supporting loops over unbacked shapes and cudagraph etc. def g(): def step_fn(init: torch.Tensor, x: torch.Tensor): next_init = x + init return next_init, next_init init = torch.zeros([]) xs = torch.arange(4) final_carry, ys = torch._higher_order.scan(step_fn, init, xs) return final_carry, ys This pass will rewrite scan into: def k(): init = torch.zeros([]) xs = torch.arange(4) # we create a loop_idx and loop through xs.shape[0] loop_idx = torch.zeros([]) ys = torch.empty_strided(_shape_stride_of_ys) def cond_fn(loop_idx, ys, init, xs): return loop_idx < xs.shape[0] # we pre-allocate the output buffer ys and inplace # copy the y of each intermediate into a slice. # NOTE [Pre-allocate scan's output buffer]. def body_fn(loop_idx, ys, init, xs): int_idx = loop_idx.item() next_init, y = step_fn(init, xs[int_idx]) ys[int_idx].copy_(y) return loop_idx + 1, ys, next_init, xs final_carry, _, _, ys = torch._higher_order.while_loop(cond_fn, body_fn, (loop_idx, ys, init, xs)) return final_carry, ys """ graph_pass = PatternMatcherPass() @register_graph_pattern( CallFunctionVarArgs(torch.ops.higher_order.scan), pass_dict=graph_pass, ) def _(match: Match, *args, **kwargs): from torch._higher_order_ops.scan import _extract_carry_and_out assert len(kwargs) == 0, ( "kwargs of scan are not merged into args before entering decompose_scan_to_while_loop_pass" ) combine_subgraph, fx_init, fx_xs, fx_additional_inputs = args assert combine_subgraph.op == "get_attr", "first arg is not combine_subgraph" sub_gm: torch.fx.GraphModule = getattr(gm, combine_subgraph.target) cur_node = match.nodes[0] num_init_leaves = len(fx_init) _, ys_outputs = _extract_carry_and_out(cur_node.meta["val"], num_init_leaves) def lower_to_while_loop(*args, **kwargs): """ The traced graph of this function will be used to replace the original scan fx_node. """ assert len(kwargs) == 0 # Step 1: construct necessary inputs to while_loop based on scan's input. ( init, xs, additional_inputs, ) = pytree.tree_unflatten(args, tree_spec) scan_length = xs[0].size(0) loop_idx = torch.zeros([], dtype=torch.int64, device=torch.device("cpu")) # NOTE [Pre-allocate scan's output buffer] # In order to pre-allocate the output buffer for ys, we rely on the meta of scan's fx_node. # However, the meta consists of concrete symints, we need to bind those symints with # proxies in order to trace the torch.empyt_strided call correctly. # # Also note that basic free symbols of tensor's shapes are guaranteed to be lifted as subgraph inputs # in dynamo so we can always re-construct the sym expression from placeholders. # See Note [Auto lift basic free symbols when create_graph_input] for how this is done. bound_symbols = { arg.node.expr: arg for arg in pytree.tree_leaves((args, scan_length)) if isinstance(arg, torch.SymInt) } ys_outs = [ torch.empty_strided( resolve_shape_to_proxy(ys_out.size(), bound_symbols), resolve_shape_to_proxy(ys_out.stride(), bound_symbols), device=ys_out.device, dtype=ys_out.dtype, layout=ys_out.layout, requires_grad=ys_out.requires_grad, ) for ys_out in ys_outputs ] while_loop_operands = (loop_idx, ys_outs, init, xs) flat_operands, operands_spec = pytree.tree_flatten(while_loop_operands) _, operands_and_additional_inputs_spec = pytree.tree_flatten( (*while_loop_operands, additional_inputs) ) # Step 2: create the cond_fn and body_fn for while_loop def cond_fn(*flat_args): loop_idx, _, _, _, _ = pytree.tree_unflatten( flat_args, operands_and_additional_inputs_spec ) # type: ignore[has-type] return loop_idx < scan_length # type: ignore[has-type] def body_fn(*flat_args): loop_idx, ys_outs, carry, xs, additional_inputs = pytree.tree_unflatten( flat_args, operands_and_additional_inputs_spec, # type: ignore[has-type] ) idx_int = loop_idx.item() torch.ops.aten._assert_scalar.default(idx_int >= 0, "") torch.ops.aten._assert_scalar.default(idx_int < scan_length, "") sub_xs = [torch.ops.aten.select.int(x, 0, idx_int) for x in xs] next_carry, ys = _extract_carry_and_out( sub_gm(*(list(carry) + sub_xs + list(additional_inputs))), num_init_leaves, ) for y, y_out in zip(ys, ys_outs): y_out_slice = torch.ops.aten.select.int(y_out, 0, idx_int) y_out_slice.copy_(y) return loop_idx + 1, *ys_outs, *next_carry, *xs # Step 3: call the while_loop operator _, ys_outs, last_carry, _ = pytree.tree_unflatten( torch.ops.higher_order.while_loop( cond_fn, body_fn, tuple(flat_operands), tuple(additional_inputs), ), operands_spec, ) return list(last_carry) + list(ys_outs) lower_to_while_loop_args, tree_spec = pytree.tree_flatten( ( fx_init, fx_xs, fx_additional_inputs, ) ) match.replace_by_example( lower_to_while_loop, lower_to_while_loop_args, run_functional_passes=False, ) graph_pass.apply(gm) for node in gm.graph.find_nodes( op="call_function", target=torch.ops.higher_order.scan ): raise AssertionError("scan is not lowered to while_loop") @init_once_fakemode def lazy_init(): if torch._C._has_mkldnn: from . import decompose_mem_bound_mm # noqa: F401 from .mkldnn_fusion import _mkldnn_fusion_init _mkldnn_fusion_init() # Put this patterns in post-grad pass rather than joint-graph # pass since otherwise there will be perf/peak-memory regression: # https://github.com/pytorch/pytorch/issues/148141 register_replacement( prepare_softmax_pattern, prepare_softmax_replacement, [torch.empty(4, 8)], scalar_workaround=dict(dim=-1), trace_fn=fwd_only, pass_dicts=pass_patterns[1], extra_check=prepare_softmax_extra_check, ) def reorder_for_locality(graph: torch.fx.Graph): if torch.distributed.is_available(): def check(): # This is a wait node, and `other_node`` is some collective node. # Eager semantics allow waits to be issued in a different order than # the collectives. Reordering this wait node might reorder collectives # which cause hangs. Once we have SPMD mode, we can safely reorder them. # However, increasing the locality between a collective and its wait node # is generally worse for performance. return node.target != torch.ops._c10d_functional.wait_tensor.default else: def check(): return True def visit(other_node): if ( other_node.op == "call_function" and other_node.target != operator.getitem and all((n in seen_nodes) for n in other_node.users) and get_mutation_region_id(graph, node) == get_mutation_region_id(graph, other_node) and check() ): # move node's producers right before it node.prepend(other_node) seen_nodes = OrderedSet[torch.fx.Node]() # only reorder nodes before the first copy_ in the graph. # copy_ will appear at the end of functionalized graphs when there is mutation on inputs, # and this reordering doesn't work well with mutation first_copy = next( iter(graph.find_nodes(op="call_function", target=torch.ops.aten.copy_.default)), None, ) past_mutating_epilogue = True if first_copy is None else False for node in reversed(graph.nodes): seen_nodes.add(node) if not past_mutating_epilogue: past_mutating_epilogue = node is first_copy continue torch.fx.map_arg((node.args, node.kwargs), visit) def register_lowering_pattern( pattern, extra_check=_return_true, pass_number=1 ) -> Callable[[Callable[_P, _T]], Callable[_P, _T]]: """ Register an aten to inductor IR replacement pattern """ return pattern_matcher.register_lowering_pattern( pattern, extra_check, pass_dict=pass_patterns[pass_number] ) ################################################################################ # Actual patterns below this point. # Priority of patterns is: # - later output nodes first # - order patterns are defined in ################################################################################ def is_valid_mm_plus_mm(match: Match): if not (config.max_autotune or config.max_autotune_gemm): return False *_b1, m1, k1 = match.kwargs["mat1"].meta.get("tensor_meta").shape *_b2, k2, n1 = match.kwargs["mat2"].meta.get("tensor_meta").shape if k1 != k2: return False *_b1, m2, k3 = match.kwargs["mat3"].meta.get("tensor_meta").shape *_b2, k4, n2 = match.kwargs["mat4"].meta.get("tensor_meta").shape if k3 != k4: return False if m1 != m2 or n1 != n2: return False return True def scatter_upon_const_tensor_extra_check(m): if not config.optimize_scatter_upon_const_tensor: return False full_shape = m.kwargs["shape"] selector = m.kwargs["selector"] dim = m.kwargs["dim"] if dim < 0: dim += len(full_shape) selector_ft = selector.meta["val"] assert selector_ft.dim() == len(full_shape) for idx, select_sz, full_sz in zip( itertools.count(), selector_ft.shape, full_shape ): if idx == dim: continue # TODO: the pattern can be updated to support the case that index tensor # is shorter. But that will need a more complex condition expression # especially for multi-dimensional tensors. # Skip it for now. if isinstance(full_sz, fx.Node): full_sz = full_sz.meta["val"] if select_sz < full_sz: return False # Actually we can support small size larger than 1. It would be a bit # tedius. E.g., we load all the index values (not many) and compare # them with the position in tensor to decide what value to return. return selector_ft.size(dim) == 1 @register_lowering_pattern( CallFunction( aten.scatter.value, CallFunction( aten.full, KeywordArg("shape"), KeywordArg("background_val"), dtype=KeywordArg("dtype"), ), KeywordArg("dim"), KeywordArg("selector"), KeywordArg("val"), # scalar value ), extra_check=scatter_upon_const_tensor_extra_check, ) def scatter_upon_const_tensor( match: Match, shape, background_val, dtype, dim, selector, val ): """ Match the pattern of full+scatter into a pointwise. TODO: Right now the scatter value must be a scalar. But we could support it when it is a tensor as well. """ from torch._inductor import metrics metrics.num_matches_for_scatter_upon_const_tensor += 1 selector_loader = selector.make_loader() def inner_fn(idx): selector_idx = list(idx) selector_idx[dim] = 0 selector = selector_loader(selector_idx) return ops.where( selector == ops.index_expr(idx[dim], torch.int64), ops.constant(val, dtype), ops.constant(background_val, dtype), ) return ir.Pointwise.create( device=selector.get_device(), dtype=dtype, inner_fn=inner_fn, ranges=shape, ) @register_lowering_pattern( CallFunction( aten.add, CallFunction(aten.mm, KeywordArg("mat1"), KeywordArg("mat2")), CallFunction(aten.mm, KeywordArg("mat3"), KeywordArg("mat4")), ), extra_check=is_valid_mm_plus_mm, ) def mm_plus_mm(match: Match, mat1, mat2, mat3, mat4): return inductor.kernel.mm_plus_mm.tuned_mm_plus_mm(mat1, mat2, mat3, mat4) @register_graph_pattern( CallFunction( aten.cumsum.default, CallFunction( torch.ops.aten.full.default, KeywordArg("shape"), KeywordArg("fill_value"), dtype=KeywordArg("dtype"), layout=Ignored(), device=KeywordArg("device"), pin_memory=False, _users=MULTIPLE, ), KeywordArg("dim"), _users=MULTIPLE, ), pass_dict=pass_patterns[1], ) def pointless_cumsum_replacement(match: Match, shape, fill_value, device, dtype, dim): """Based on a pattern in OPTForCausalLM""" if is_integer_dtype(dtype) or is_boolean_dtype(dtype): # cumsum promotes all integral types to int64 dtype = torch.int64 def repl(*shape): dim_size = shape[dim] idx = torch.arange(1, dim_size + 1, device=device, dtype=dtype) inter_shape = [1] * len(shape) inter_shape[dim] = dim_size return (idx * fill_value).view(inter_shape).expand(shape) # only replace the output node, not all nodes match.nodes = [match.output_node()] match.replace_by_example(repl, list(shape)) _cat_1 = CallFunction(aten.cat, Arg(), 1, _users=2) @register_lowering_pattern( CallFunction( aten.cat, [ _cat_1, CallFunction( aten.slice, _cat_1, 1, 0, KeywordArg("size"), ), ], 1, ) ) def cat_slice_cat(match, cat_input, size, dim=1): """ This is an example of a more complex pattern where cat_1 is used multiple times inside the pattern. We fold 2 calls to cat into one. Matches: cat_1: f32[1024, 4077] = torch.ops.aten.cat.default([add_26, primals_217], 1) slice_1: f32[1024, 4077] = torch.ops.aten.slice.Tensor(cat_1, 0, 0, 9223372036854775807) slice_2: f32[1024, 19] = torch.ops.aten.slice.Tensor(slice_1, 1, 0, 19) cat_2: f32[1024, 4096] = torch.ops.aten.cat.default([cat_1, slice_2], 1) Rewrite to: slice_2 = torch.ops.aten.slice.Tensor(add_26, 1, 0, 19) cat_2 = torch.ops.aten.cat.default([add_26, primals_217, slice2], 1) """ first, *rest = cat_input # Optimization is optional, because we can just not fold the cat # size should be within first.get_size()[dim] such that the optimization is valid. # For negative `end`, we currently fallback to not optimizing. if size >= 0 and V.graph.sizevars.statically_known_leq(size, first.get_size()[dim]): # fold 2 cats into 1 cat return L[aten.cat]( [ first, *rest, L[aten.slice](first, dim, 0, size), ], dim, ) else: # don't expect to hit this case, just fall back tmp = L[aten.cat](cat_input, dim) return L[aten.cat]( [ tmp, L[aten.slice](tmp, dim, 0, size), ], dim, ) def is_valid_splitwithsizes_cat(match): split_nodes = filter_nodes(match.nodes, aten.split_with_sizes) cat_nodes = filter_nodes(match.nodes, aten.cat) get_item_nodes = filter_nodes(match.nodes, operator.getitem) if len(split_nodes) != 1 or len(cat_nodes) != 1: return False split_node, cat_node = split_nodes[0], cat_nodes[0] # The dim of split and cat should match for passthrough if get_arg_value(split_node, 2, "dim") != get_arg_value(cat_node, 1, "dim"): return False get_item_args = OrderedSet( get_arg_value(get_item_node, 1) for get_item_node in get_item_nodes ) assert None not in get_item_args split_sizes = get_arg_value(split_node, 1, "split_sizes") # All parts of split should be included in the cat if get_item_args != OrderedSet(range(len(split_sizes))): return False # The order of get_item_args should same with cat_node used. # For example, if the split_node like split_with_sizes(input, [2, 2, 3], 1), # the cat node should be like cat([get_item(0), get_item(1), get_item(2)], 1). cat_items_args_order = [ get_arg_value(item_node, 1) for item_node in get_arg_value(cat_node, 0) ] if cat_items_args_order != list(range(len(split_sizes))): return False return True def same_meta(node1: torch.fx.Node, node2: torch.fx.Node): """True if two nodes have the same metadata""" val1 = node1.meta.get("val") val2 = node2.meta.get("val") return ( val1 is not None and val2 is not None and statically_known_true(sym_eq(val1.size(), val2.size())) and val1.layout == val2.layout and val1.dtype == val2.dtype and val1.device == val2.device and ( val1.layout != torch.strided or statically_known_true(sym_eq(val1.stride(), val2.stride())) ) ) noop_registry: dict[Any, Any] = {} def register_noop_decomp(targets, nop_arg=0): def register_fun(cond): register_decomposition(targets, registry=noop_registry, unsafe=True)( (cond, nop_arg) # type: ignore[arg-type] ) return cond return register_fun @register_noop_decomp(aten.slice) def slice_noop(self, dim=0, start=None, end=None, step=1): if start is None or end is None: return False slice_dim_size = self.shape[dim] if ( statically_known_true(sym_eq(start, 0)) and ( statically_known_true(end >= 2**63 - 1) or statically_known_true(end >= slice_dim_size) ) and statically_known_true(sym_eq(step, 1)) ): return True return False @register_noop_decomp(aten.slice_scatter, 1) def slice_scatter_noop(self, src, dim=0, start=None, end=None, step=1): if start is None: start = 0 if end is None: end = 2**63 - 1 slice_scatter_dim_size = self.shape[dim] if ( self.shape == src.shape and start == 0 and ( statically_known_true(end >= 2**63 - 1) or statically_known_true(end >= slice_scatter_dim_size) ) and step == 1 ): return True return False @register_noop_decomp(aten.repeat) def repeat_noop(self, repeats): return all(r == 1 for r in repeats) @register_noop_decomp(aten.constant_pad_nd) def constant_pad_nd(x, padding, fill_value=0): return all(p == 0 for p in padding) @register_noop_decomp(torch.ops.prims.convert_element_type) def convert_element_type_noop(x, dtype: torch.dtype): return x.dtype == dtype @register_noop_decomp(torch.ops.prims.device_put) def device_put_noop(x, device, non_blocking=True): return x.device == decode_device(device) @register_noop_decomp([aten.ceil, aten.floor, aten.round, aten.trunc]) def int_noop(x): return is_integer_dtype(x.dtype) @register_noop_decomp([aten.pow]) def pow_noop(a, b): return isinstance(b, int) and b == 1 @register_noop_decomp([aten.cat], lambda args: args[0][0]) def cat_noop(inputs, dim=0): return len(inputs) == 1 @register_noop_decomp(aten.view.default) def view_default_noop(arg, size): return statically_known_true(sym_eq(arg.shape, tuple(size))) @register_noop_decomp(aten.view.dtype) def view_dtype_noop(arg, dtype): return arg.dtype == dtype # Note, we also always have a check for identical metadata, which is why these # are safe @register_noop_decomp([aten.copy], nop_arg=1) @register_noop_decomp([aten.alias, aten.clone]) def true_noop(*args, **kwargs): return True def remove_noop_ops(graph: torch.fx.Graph): """ Removes both operations that are essentially aten.clone and operations that are essentially aten.alias from the graph. """ inputs = OrderedSet[torch.fx.Node]() input_storages = OrderedSet[Union[int, None]]() output_storages = OrderedSet[Union[int, None]]() for node in graph.find_nodes(op="placeholder"): inputs.add(node) input_storages.add(get_node_storage(node)) output_node = next(iter(reversed(graph.nodes))) assert output_node.op == "output" outputs = output_node.args[0] if not isinstance(outputs, (list, tuple)): # nested subgraphs can have singleton outputs outputs = (outputs,) for out in outputs: if isinstance(out, torch.fx.Node): output_storages.add(get_node_storage(out)) for node in graph.nodes: if node.target in noop_registry: cond, src_index = noop_registry[node.target] if isinstance(src_index, int): src = node.args[src_index] else: src = src_index(node.args) if not isinstance(src, torch.fx.Node): continue # Don't introduce new aliasing between inputs and outputs. # See fx_passes/README.md for a discussion of why this is # necessary. node_storage = get_node_storage(node) src_storage = get_node_storage(src) node_is_view = node_storage == src_storage if ( not node_is_view and node_storage in output_storages and (src_storage in input_storages or src_storage in output_storages) ): continue # Even if input and outputs are expected to alias, # don't make "node is src" True if ( node_is_view and node in output_node.args and (src in inputs or src in output_node.args) ): continue is_valid, args, kwargs = get_fake_args_kwargs(node) if not is_valid: continue if same_meta(node, src) and cond(*args, **kwargs): node.replace_all_uses_with(src) graph.erase_node(node) def remove_assert_ops(graph: torch.fx.Graph): """ Removes aten._assert_tensor_metadata.default op because 1) it will be lowered to a no-op in inductor 2) it can block fusion, such as unfuse_bias_add_to_pointwise fusion. This op could come from aten.to functionalization in export. For example, if we have a graph like below %addmm = aten.addmm.default(%linear_bias, %arg3_1, %permute) %_assert_tensor_metadata = aten._assert_tensor_metadata.default(%addmm, None, None, torch.float16) %convert_element_type_3 = prims.convert_element_type.default(%addmm, torch.float32) %pow_1 = aten.pow.Tensor_Scalar(%convert_element_type_3, 2) We still want to fuse add from addmm with pow, instead of fusing add with mm, according to unfuse_bias_add_to_pointwise fusion. However, aten._assert_tensor_metadata.default is not a pointwise op, and would fail the should_prefer_unfused_addmm check. We remove this op so it doesn't block fusion decisions. It's safe because this op is lowered to a no-op with @register_lowering. """ for node in graph.find_nodes( op="call_function", target=torch.ops.aten._assert_tensor_metadata.default ): graph.erase_node(node) def decompose_triton_kernel_wrapper_functional(graph): """Decomposes triton_kernel_wrapper_functional nodes into clones and the underlying mutation node. We assume that the reinplacing pass runs before this; the reinplacing pass tells us (via rewriting the arguments or .meta to those nodes) which Tensors we should clone and which Tensors are safe to reinplace. """ graph_pass = PatternMatcherPass() @register_graph_pattern( CallFunctionVarArgs(torch.ops.higher_order.triton_kernel_wrapper_functional), pass_dict=graph_pass, ) def _(match: Match, *args, **kwargs): from torch._higher_order_ops.triton_kernel_wrap import ( triton_kernel_wrapper_functional_dense, ) flat_args, spec = pytree.tree_flatten((args, kwargs)) # NB: we combine (args, kwargs) into flat args for replacing. # This is replace_by_example uses make_fx which does not support # tracing a function with kwargs. def decomp(*flat_args): args, kwargs = pytree.tree_unflatten(flat_args, spec) return (triton_kernel_wrapper_functional_dense(*args, **kwargs),) match.replace_by_example(decomp, flat_args, run_functional_passes=False) graph_pass.apply(graph) for node in graph.find_nodes( op="call_function", target=torch.ops.higher_order.triton_kernel_wrapper_functional, ): raise AssertionError("triton_kernel_wrapper_functional was not removed") def decompose_auto_functionalized(graph): """Decomposes auto_functionalized nodes into clones and the underlying mutation node. We assume that the reinplacing pass runs before this; the reinplacing pass tells us (via rewriting the arguments or .meta to those nodes) which Tensors we should clone and which Tensors are safe to reinplace. """ graph_pass = PatternMatcherPass() @register_graph_pattern( CallFunctionVarArgs(torch.ops.higher_order.auto_functionalized), pass_dict=graph_pass, ) def _(match: Match, *args, **kwargs): from torch._higher_order_ops.auto_functionalize import auto_functionalized_dense only_clone_these_tensors = tuple( match.nodes[0].meta.get("only_clone_these_tensors", []) ) flat_args, spec = pytree.tree_flatten((args, kwargs)) # NB: we combine (args, kwargs) into flat args for replacing. # This is replace_by_example uses make_fx which does not support # tracing a function with kwargs. def decomp(*flat_args): args, kwargs = pytree.tree_unflatten(flat_args, spec) assert len(args) == 1 mode = args[0] return auto_functionalized_dense(mode, only_clone_these_tensors, **kwargs) match.replace_by_example(decomp, flat_args, run_functional_passes=False) @register_graph_pattern( CallFunctionVarArgs(torch.ops.higher_order.auto_functionalized_v2), pass_dict=graph_pass, ) def _(match: Match, *args, **kwargs): from torch._higher_order_ops.auto_functionalize import ( auto_functionalized_v2_dense, ) only_clone_these_bases = tuple( match.nodes[0].meta.get("only_clone_these_tensors", []) ) flat_args, spec = pytree.tree_flatten((args, kwargs)) def _maybe_resolve_constant_get_attr(node): # Resolve getattr node to its value because they don't always have meta["val"] if ( isinstance(node, torch.fx.Node) and node.op == "get_attr" and "val" not in node.meta ): const_attr = getattr(graph.owning_module, node.target) # type: ignore[arg-type] assert isinstance( const_attr, (torch.fx.GraphModule, pytree.TreeSpec) ), (type(const_attr), const_attr) return const_attr return node flat_args = [_maybe_resolve_constant_get_attr(arg) for arg in flat_args] # NB: we combine (args, kwargs) into flat args for replacing. # This is replace_by_example uses make_fx which does not support # tracing a function with kwargs. def decomp(*flat_args): args, kwargs = pytree.tree_unflatten(flat_args, spec) assert len(args) == 1 mutable_op = args[0] return auto_functionalized_v2_dense( mutable_op, only_clone_these_bases, **kwargs ) match.replace_by_example(decomp, flat_args, run_functional_passes=False) graph_pass.apply(graph) # We need to remove the get_attr registered for _constant_schema and the # auto_functioanlized's graph module (it's replaced with original ) when auto_functionalize a hop. _to_remove = [] for node in graph.nodes: if node.op == "get_attr" and len(node.users) == 0: _to_remove.append(node) if hasattr(graph.owning_module, node.target) and isinstance( getattr(graph.owning_module, node.target), torch.fx.GraphModule ): delattr(graph.owning_module, node.target) for node in _to_remove: graph.erase_node(node) graph.lint() for _ in graph.find_nodes( op="call_function", target=torch.ops.higher_order.auto_functionalized ): raise AssertionError("auto_functionalized was not removed") for _ in graph.find_nodes( op="call_function", target=torch.ops.higher_order.auto_functionalized_v2 ): raise AssertionError("auto_functionalized_v2 was not removed") @register_lowering_pattern( CallFunction( aten.cat, ListOf( CallFunction( operator.getitem, CallFunction( aten.split_with_sizes, KeywordArg("input_"), Ignored(), Ignored(), _users=MULTIPLE, ), Ignored(), ), ), Ignored(), ), pass_number=2, extra_check=is_valid_splitwithsizes_cat, ) def splitwithsizes_cat_replace(match, input_): return input_ def is_valid_cat_splitwithsizes(match): cat_nodes = filter_nodes(match.nodes, aten.cat) split_nodes = filter_nodes(match.nodes, aten.split_with_sizes) if len(split_nodes) != 1 or len(cat_nodes) != 1: return False split_node, cat_node = split_nodes[0], cat_nodes[0] # the cat node has other users: can't eliminate if len(cat_node.users) > 1: return False # the dim of the cat and split should match dim = get_arg_value(split_node, 2, "dim") if dim != get_arg_value(cat_node, 1, "dim"): return False cat_inputs = list(get_arg_value(cat_node, 0)) split_sizes = get_arg_value(split_node, 1, "split_sizes") # the number of input tensors in cat and the # length of the split sizes should match if len(cat_inputs) != len(split_sizes): return False for cat_input, split_size in zip(cat_inputs, split_sizes): # each cat input tensor's size along dim # should match the corresponding split size if "val" not in cat_input.meta: return False cat_input_size = cat_input.meta["val"].size(dim) if cat_input_size != split_size: return False return True @register_lowering_pattern( CallFunction( aten.split_with_sizes, CallFunction( aten.cat, KeywordArg("input_"), Ignored(), _users=MULTIPLE, ), Ignored(), Ignored(), ), pass_number=2, extra_check=is_valid_cat_splitwithsizes, ) def cat_splitwithsizes_replace(match, input_): return input_ def view_to_reshape(gm): """ Replace view ops in the GraphModule to reshape ops. """ subgraph_names: OrderedSet[str] = OrderedSet( x.target for x in gm.graph.find_nodes(op="get_attr") ) for child_name, child_mod in gm.named_children(): if child_name in subgraph_names and isinstance(child_mod, torch.fx.GraphModule): view_to_reshape(child_mod) for nd in gm.graph.find_nodes( op="call_function", target=torch.ops.aten.view.default ): nd.target = torch.ops.aten.reshape.default def should_prefer_unfused_addmm(match): inp = match.kwargs["inp"] if not is_gpu(inp.meta["val"].device.type): return False output = match.output_node() return all(is_pointwise_use(use) for use in output.users) @register_graph_pattern( CallFunction(aten.addmm, KeywordArg("inp"), Arg(), Arg()), pass_dict=pass_patterns[2], extra_check=should_prefer_unfused_addmm, ) def unfuse_bias_add_to_pointwise(match: Match, mat1, mat2, *, inp): def repl(inp, x1, x2): return x1 @ x2 + inp match.replace_by_example(repl, [inp, mat1, mat2]) def is_valid_addmm_fusion(match): mat1, mat2 = match.args inp = match.kwargs["inp"] if not ( isinstance(inp, torch.fx.Node) and isinstance(inp.meta["val"], torch.Tensor) ): return False # Input is a number in_shape = inp.meta["val"].shape mm_shape = mat1.meta["val"].shape[0], mat2.meta["val"].shape[1] matched = is_expandable_to(in_shape, mm_shape) if not matched: return False # Shape mismatch return not should_prefer_unfused_addmm(match) @register_graph_pattern( CallFunction( aten.add, CallFunction(aten.mm, Arg(), Arg()), KeywordArg("inp"), ), pass_dict=pass_patterns[2], extra_check=is_valid_addmm_fusion, ) @register_graph_pattern( CallFunction( aten.add, KeywordArg("inp"), CallFunction(aten.mm, Arg(), Arg()), ), pass_dict=pass_patterns[2], extra_check=is_valid_addmm_fusion, ) def addmm(match, mat1, mat2, *, inp): def repl(inp, mat1, mat2): return aten.addmm(inp, mat1, mat2) match.replace_by_example(repl, [inp, mat1, mat2]) def register_partial_reduction_pattern(): "Reuse partial reductions in complete reductions" # post grad equivalents equiv_red = { aten.amax.default: aten.max.default, aten.amin.default: aten.min.default, } # TODO: to support other reductions like sum, would need to skip # lower precision reductions since partial output would need to be kept at fp32. for red_op in (aten.amax.default, aten.amin.default): inp = KeywordArg("input") partial_reduc = CallFunction( red_op, inp, KeywordArg("reduced_dims"), KeywordArg("keepdim") ) full_reduc = CallFunction([red_op, equiv_red[red_op]], inp) @register_graph_pattern( MultiOutputPattern([partial_reduc, full_reduc]), pass_dict=pass_patterns[2] ) def reuse_partial(match, input, reduced_dims, keepdim): partial_red, full_red = match.output_nodes() # if they're small, reuse not worth it if not statically_known_true(input.meta["val"].numel() >= 4096): return True def replacement(inp: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: partial = partial_red.target(inp, reduced_dims, keepdim) complete = full_red.target(partial) return (partial, complete) counters["inductor"]["partial_reduction_reuse"] += 1 match.replace_by_example(replacement, [input]) register_partial_reduction_pattern() def check_shape_cuda_and_fused_int_mm_mul_enabled(match): return ( config.force_fuse_int_mm_with_mul and len(getattr(match.args[2].meta.get("val"), "shape", [])) == 2 and getattr(match.args[2].meta.get("val"), "is_cuda", False) ) def is_index_put_and_requires_h2d_sync_for_gpu_value(node): from torch.fx.operator_schemas import normalize_function if node.target not in [ torch.ops.aten.index_put.default, torch.ops.aten.index_put_.default, ]: return False # Inductor falls back to aten.index_put_. # index_put_ will will call nonzero() and perform a H2D sync if # any of its indices are bool/byte tensors # However, it will short-circuit this H2D sync and run mask_fill_ # if the value we are putting is a cpu scalar. # Therefore, when inductor sees an index_put_ with byte tensor indices, # it should *not* convert the cpu scalar value into a gpu tensor. args_, _kwargs = normalize_function(node.target, node.args, node.kwargs) # type: ignore[misc] any_byte_bool_indices = False indices = args_[1] for i in indices: if i is not None and i.meta["val"].dtype in [torch.bool, torch.int8]: any_byte_bool_indices = True val = args_[2].meta["val"] val_is_cpu_scalar = val.device.type == "cpu" and val.numel() == 1 # If both these conditions hold, then converting the val # to a gpu tensor will incur a H2D sync when inductor calls aten.index_put_ return any_byte_bool_indices and val_is_cpu_scalar class ConstructorMoverPass: def __init__( self, target: str, allow_outputs: bool = False, allow_inputs: bool = False ) -> None: """ Move constructors from cpu to the target_device. Sweeps through the module, looking for constructor nodes that can be moved to the target_device. A constructor node can be moved to the target_device iff all of its users can also be moved (tested by cannot_be_moved). Otherwise, all dependent constructor nodes won't be moved. - target: target device type - allow_outputs: allow outputs to be moved - allow_inputs: allow inputs to be moved """ self.target = target self.allow_inputs = allow_inputs self.allow_outputs = allow_outputs assert isinstance(target, str), ( "target should be a string representing the device type. " f"Got: {type(target).__name__}" ) def allow_cpu_device(self, node: fx.Node) -> bool: """ Returns whether a node that returns a tensor on the target device may have cpu tensors as input. """ return node.target in ( torch.ops.aten.index.Tensor, torch.ops.aten.index_put.default, torch.ops.aten.index_put_.default, torch.ops.aten.copy.default, torch.ops.aten.copy_.default, torch.ops.aten.slice_scatter.default, ) def is_on_target_device(self, node: fx.Node) -> bool: """ Returns whether a node is on the target device. """ node_device = self.get_node_device(node) return node_device is not None and node_device.type == self.target def is_cpu_scalar_tensor(self, node: fx.Node) -> bool: """ Returns whether a node is a cpu scalar tensor. """ device = self.get_node_device(node) is_cpu = device is not None and device.type == "cpu" ten = node.meta.get("val") is_scalar = isinstance(ten, torch.Tensor) and len(ten.size()) == 0 return is_cpu and is_scalar def all_inputs_are_cpu_scalar_or_on_target_device(self, node: fx.Node) -> bool: """ Returns whether a node's inputs are either cpu scalar tensors or on the target device. """ inputs = ( inp for inp in itertools.chain(node.args, node.kwargs.values()) if isinstance(inp, fx.Node) ) return all( self.is_cpu_scalar_tensor(inp) or self.is_on_target_device(inp) for inp in inputs ) def cannot_be_moved(self, node: fx.Node) -> bool: """ Returns whether a node can be moved to the target device. If this function returns False, it means that this node and all of its users won't be moved into the target device. """ if node.target == "output": return not self.allow_outputs if not ( isinstance(node.target, torch._ops.OpOverload) and node.target.namespace in ("prims", "aten") ): return True if is_index_put_and_requires_h2d_sync_for_gpu_value(node): return True return False def get_node_device(self, node: fx.Node) -> Optional[torch.device]: """ Get the device of a node. """ ten = node.meta.get("val") return None if not isinstance(ten, torch.Tensor) else ten.device def get_cpu_indeg_count(self, graph: fx.Graph) -> dict[fx.Node, int]: """ Get the number of cpu inputs to a node """ cpu_indeg: dict[fx.Node, int] = Counter() for node in graph.nodes: cpu_count = 0 def add_cpu_inp(node): nonlocal cpu_count device = self.get_node_device(node) cpu_count += device is not None and device.type == "cpu" pytree.tree_map_only(fx.Node, add_cpu_inp, (node.args, node.kwargs)) if cpu_count: cpu_indeg[node] = cpu_count return cpu_indeg def __call__(self, graph: fx.Graph) -> None: target_devices = OrderedSet[torch.device]() constructors = [] cpu_placeholders: OrderedSet[fx.Node] = OrderedSet() for node in graph.nodes: device = self.get_node_device(node) if device and device.type == self.target: target_devices.add(device) if ( self.allow_inputs and node.op == "placeholder" and self.is_cpu_scalar_tensor(node) ): cpu_placeholders.add(node) constructors.append(node) continue if not ( isinstance(node.target, torch._ops.OpOverload) and node.target.namespace in ("prims", "aten") ): continue if not torch._subclasses.fake_tensor._is_tensor_constructor(node.target): continue if not node.kwargs.get("device") == torch.device("cpu"): continue constructors.append(node) # not handling multiple target devices initially if not constructors or len(target_devices) != 1: return movable_constructors = self.find_movable_constructors(graph, constructors) target_device = next(iter(target_devices)) for node in movable_constructors: if node in cpu_placeholders: with graph.inserting_after(node): gpu_node = graph.call_function( torch.ops.prims.device_put.default, (node, target_device) ) node.replace_all_uses_with( gpu_node, lambda x: x != gpu_node and x.target != torch.ops.aten.copy_.default, ) # noop elimination if there are other device_put for gpu_node to # target device. Alternatively, we could just move the other device_put # earlier in the graph, but that is not supported in fx graph yet. noop_device_puts = [ user for user in gpu_node.users if user.target == torch.ops.prims.device_put.default and user.args[1] == target_device ] for noop in noop_device_puts: noop.replace_all_uses_with(gpu_node) graph.erase_node(noop) else: kwargs = node.kwargs.copy() kwargs["device"] = target_device node.kwargs = kwargs def find_movable_constructors( self, graph: fx.Graph, constructors: list[fx.Node] ) -> OrderedSet[fx.Node]: """ Starting from the cpu constructors, iterate through the graph and test that all of their downstream uses can safely be moved to cpu. """ cpu_indeg: dict[fx.Node, int] = self.get_cpu_indeg_count(graph) # which constructors cannot be moved to gpu cannot_move_to_gpu = OrderedSet[fx.Node]() # For any node in the graph, which constructors does it have a dependency on constructor_dependencies: dict[fx.Node, OrderedSet[fx.Node]] = defaultdict( OrderedSet ) # if a cpu node has a dependency on two different cpu constructors, # then if either constructor cannot be moved to gpu, the other cannot as well. # In this case any node with a dependency on one will have a dependency on the other equal_constructor_sets: dict[fx.Node, OrderedSet[fx.Node]] = { c: OrderedSet([c]) for c in constructors } def make_dependencies_equivalent( set1: OrderedSet[fx.Node], set2: OrderedSet[fx.Node] ) -> OrderedSet[fx.Node]: # could use union find but not worth complexity here set1.update(set2) for obj in set1: equal_constructor_sets[obj] = set1 return set1 queue: list[fx.Node] = list(constructors) for c in queue: constructor_dependencies[c].add(c) while queue: node = queue.pop() dependencies = constructor_dependencies[node] for user in node.users: if self.cannot_be_moved(user): cannot_move_to_gpu.update(dependencies) break # this node was used on a op which takes in multiple devices and output a gpu # tensor. we can convert its cpu input to gpu without making further changes if self.allow_cpu_device(user) and self.is_on_target_device(user): del cpu_indeg[user] elif ( self.allow_inputs and self.all_inputs_are_cpu_scalar_or_on_target_device(user) ): # this node takes only cpu scalar tensors or gpu tensors as inputs # and outputs a gpu tensor. we can convert its cpu scalar inputs to gpu # without making further changes del cpu_indeg[user] else: # otherwise, we should continue look at its downstream uses cpu_indeg[user] -= 1 if cpu_indeg[user] == 0: del cpu_indeg[user] queue.append(user) unioned_set = make_dependencies_equivalent( dependencies, constructor_dependencies[user] ) constructor_dependencies[user] = unioned_set for node in cpu_indeg: if constructor_dependencies[node]: cannot_move_to_gpu.update(constructor_dependencies[node]) all_cannot_move_to_gpu = cannot_move_to_gpu.copy() for constructor in cannot_move_to_gpu: all_cannot_move_to_gpu.update(equal_constructor_sets[constructor]) return OrderedSet(constructors) - all_cannot_move_to_gpu def move_constructors_to_gpu(graph: fx.Graph) -> None: """ Moves intermediary tensors which are constructed on the cpu to gpu when safe """ # cudagraph does not support cpu tensors. In this pass, we update the graph # by explicitly moving cpu scalar tensors to gpu when profitable, relying on # graph partition to split off this data copy, and cudagraphifying # the remaining gpu ops. allow_inputs_outputs = ( True if ( torch._inductor.config.triton.cudagraphs and torch._inductor.config.graph_partition ) else False ) ConstructorMoverPass( get_gpu_type(), allow_inputs=allow_inputs_outputs, allow_outputs=allow_inputs_outputs, )(graph)