import contextlib import functools import itertools import logging from collections.abc import Callable from typing import Any import sympy import torch import torch._inductor.config as config from torch._dynamo.utils import counters from torch._inductor import ir from torch._inductor.autotune_process import ( SubgraphCPUBenchmarkRequest, SubgraphGPUBenchmarkRequest, TensorMeta, ) from torch._inductor.codegen.common import KernelTemplate from torch._inductor.ir import ( Buffer, FixedLayout, get_free_symbols, get_symbolic_inputs, gm_original_output_strides, ir_node_to_tensor, Layout, ) from torch._inductor.runtime.benchmarking import benchmarker from torch._inductor.utils import do_bench_using_profiling from torch._inductor.virtualized import V from torch.utils._ordered_set import OrderedSet log = logging.getLogger(__name__) def inline_subgraph_to_ir_nodes( gm: torch.fx.GraphModule, inputs: list[Any], name: str ) -> Any: """Inline a subgraph by converting its FX operations to individual IR nodes. This converts a subgraph to multiple ComputedBuffer nodes (fusable), enabling epilogue fusion with subsequent operations. Returns: TensorBox containing the final operation result as individual IR nodes """ from torch._inductor.lowering import process_subgraph_nodes # Temporarily switch V.graph.module to subgraph during processing; restore to prevent IR nodes added to wrong graph original_module = V.graph.module try: V.graph.module = gm return process_subgraph_nodes(gm, inputs) finally: V.graph.module = original_module class SubgraphChoiceCaller(ir.ChoiceCaller): """ Represents a Subgraph Autotuning choice, and the subgraph can be any arbitrary GraphModule. Compiles the Subgraph down to a module for benchmarking. """ def __init__( self, name: str, input_nodes: list[Buffer], layout: Layout, description: str, make_fx_graph: Callable[..., Any], input_gen_fns: dict[int, Callable[[Any], torch.Tensor]] | None = None, ) -> None: super().__init__(name, input_nodes, layout, description) # Create inputs for tracing and benchmarking: # - trace_inputs: Always use symbolic inputs for tracing so the graph works # for any size at runtime. This correctly captures shape-dependent operations # like x * x.shape[0]. # - benchmark_inputs: Use input_gen_fns if provided (for range-specific benchmarks) # or concrete inputs from ir_node_to_tensor with guard_shape=False. trace_inputs = [] self.benchmark_inputs = [] with V.fake_mode: for i, inp in enumerate(self.input_nodes): # Here there will be no unbacked symbols, as SubgraphBuffer does not support them assert len(get_free_symbols(inp.get_size(), unbacked_only=True)) == 0 assert len(get_free_symbols(inp.get_stride(), unbacked_only=True)) == 0 inp.data.freeze_layout() # type: ignore[attr-defined] # Always use symbolic inputs for tracing trace_inputs.append(ir_node_to_tensor(inp)) # Use input_gen_fn for benchmarking if provided, otherwise concrete sizes if input_gen_fns is not None and i in input_gen_fns: self.benchmark_inputs.append(input_gen_fns[i](inp)) else: self.benchmark_inputs.append( ir_node_to_tensor(inp, replace_symbols_with_hints=True) ) # Trace with symbolic inputs for guard detection self.gm = make_fx_graph(*trace_inputs) gm_original_output_strides(self.gm) # Store symbolic inputs for sym_input computation self.example_inputs = trace_inputs self.sym_inputs = get_symbolic_inputs(self.input_nodes) self.sym_input_values = self._compute_sym_input_values() # Cached decomposition info for range-based dispatch (set via cache_decomposition) self.decomposition: Callable[..., Any] | None = None self.decomposition_kwargs: dict[str, Any] = {} # Config patches to apply during kernel codegen (e.g., coordinate_descent_tuning) self.config_patches: dict[str, Any] = {} # Cache compiled module to avoid recompiling on every benchmark call self._compiled_module: Any = None # Cache benchmark request for async autotuning self._bmreq: ( SubgraphGPUBenchmarkRequest | SubgraphCPUBenchmarkRequest | None ) = None # Pre-compile only if using async pipelined autotuning # Must happen in __init__ because compilation requires virtualized context (V.graph, V.debug) if config.pipeline_max_autotune_gemm: with V.fake_mode: self._compiled_module = self._compile_for_benchmarking() self._bmreq = self._create_benchmark_request() def _compute_sym_input_values(self) -> list[int]: """Extract concrete dimension values for sym_inputs from benchmark_inputs. The compiled module expects symbolic dimension values as runtime arguments. This maps each symbolic variable to its concrete value from the benchmark tensors. Used for range based autotuning. """ sym_input_names = OrderedSet( [s.name for s in self.sym_inputs if hasattr(s, "name")] ) # Build mapping: symbolic dimension name -> concrete value sym_name_to_value: dict[str, int] = {} for inp_node, benchmark_inp in zip(self.input_nodes, self.benchmark_inputs): if isinstance(benchmark_inp, torch.Tensor): for sym_dim, actual_dim in zip( inp_node.get_size(), benchmark_inp.shape ): if isinstance(sym_dim, sympy.Symbol): sym_name_to_value[sym_dim.name] = int(actual_dim) elif str(sym_dim) in sym_input_names: sym_name_to_value[str(sym_dim)] = int(actual_dim) result = [] for sym_var in self.sym_inputs: if isinstance(sym_var, sympy.Symbol) and sym_var.name in sym_name_to_value: result.append(sym_name_to_value[sym_var.name]) else: hint = V.graph.sizevars.shape_env.optimization_hint(sym_var, fallback=1) result.append(int(hint)) return result def cache_decomposition( self, decomposition: Callable[..., Any], kwargs: dict[str, Any] ) -> None: """Cache decomposition function and kwargs for range-based dispatch lookup.""" self.decomposition = decomposition self.decomposition_kwargs = kwargs def __str__(self) -> str: return f"SubgraphCaller({self.name})" def _compile_for_benchmarking(self) -> Any: """Compile the subgraph for benchmarking, returns the compiled module.""" from torch._inductor.graph import GraphLowering safe_name = self.name.replace("::", "_").replace(".", "_") # Symbolic inputs produce code with symbolic sizes resolved at runtime # via sym_input_values. Without symbols (static shapes), compile with # benchmark_inputs whose sizes match the actual benchmark tensors. compile_inputs = ( self.example_inputs if self.sym_inputs else self.benchmark_inputs ) log.debug("Benchmark compile %s: sym_inputs=%s", self.name, self.sym_inputs) assert self.gm is not None bm_graph_lowering = GraphLowering( gm=self.gm, example_inputs=compile_inputs, shape_env=V.graph._shape_env, cpp_wrapper=V.graph.cpp_wrapper, aot_mode=V.graph.aot_mode, extern_node_serializer=V.graph.extern_node_serializer, is_inference=V.graph.is_inference, is_backward=V.graph.is_backward, name=f"benchmark_{safe_name}", ) for sym_inp in self.sym_inputs: bm_graph_lowering.graph_inputs[sym_inp.name] = sym_inp bm_graph_lowering.graph_input_names.append(sym_inp.name) with V.set_graph_handler(bm_graph_lowering): # Apply config_patches during benchmarking (e.g., coordinate_descent_tuning) # Also disable max_autotune to avoid nested autotuning benchmark_config: dict[str, Any] = { "max_autotune": False, "max_autotune_gemm": False, "max_autotune_gemm_backends": "ATEN", "benchmark_fusion": False, "pipeline_max_autotune_gemm": False, **self.config_patches, } with config.patch(benchmark_config): bm_graph_lowering.run(*compile_inputs) return bm_graph_lowering.compile_to_module() def _create_benchmark_request( self, ) -> SubgraphGPUBenchmarkRequest | SubgraphCPUBenchmarkRequest: """Create a benchmark request for async autotuning.""" assert self._compiled_module is not None, ( "Module must be compiled before creating benchmark request" ) input_tensor_meta = TensorMeta.from_irnodes(self.input_nodes) output_tensor_meta = TensorMeta.from_irnodes(self.layout) if self.layout.device.type == "cpu": bmreq_cls = SubgraphCPUBenchmarkRequest else: bmreq_cls = SubgraphGPUBenchmarkRequest return bmreq_cls( kernel_name=self.name, input_tensor_meta=input_tensor_meta, output_tensor_meta=output_tensor_meta, extra_args=tuple(), module_path=self._compiled_module.__file__, module_cache_key=self._compiled_module.key, sym_input_values=self.sym_input_values, ) @property def bmreq( self, ) -> SubgraphGPUBenchmarkRequest | SubgraphCPUBenchmarkRequest: """Benchmark request for async autotuning. Pre-compiled when pipeline_max_autotune_gemm is enabled.""" assert self._bmreq is not None, ( "bmreq accessed but pipeline_max_autotune_gemm was not enabled during __init__" ) return self._bmreq def _ensure_compiled(self) -> None: """Ensure the module is compiled. Used for lazy compilation in non-async path.""" if self._compiled_module is None: self._compiled_module = self._compile_for_benchmarking() def benchmark(self, *args: list[Any], out: torch.Tensor) -> float: """Regular benchmarking: compile if needed, then use benchmarker.""" self._ensure_compiled() bm_func = self._compiled_module.call sym_inputs = self.sym_input_values def fn() -> Any: return bm_func([*sym_inputs, *args]) if self._benchmark_with_cudagraphs: return benchmarker.benchmark_gpu_with_cuda_graph(fn) if config.profile_bandwidth_with_do_bench_using_profiling: return do_bench_using_profiling(fn) return benchmarker.benchmark( fn, device=benchmarker.infer_device(*sym_inputs, *args), ) def benchmark_collective(self, *args: list[Any], out: torch.Tensor) -> None: """Run once for collective benchmarking (barrier sync handled by caller).""" self._ensure_compiled() self._compiled_module.call([*self.sym_input_values, *args]) def hash_key(self) -> str: assert self.gm is not None return "-".join( [ self.name.rsplit("_", 1)[0], *[str(inp.get_size()) for inp in self.input_nodes], *[str(inp.get_stride()) for inp in self.input_nodes], str(self.gm.graph), ] ) def output_node(self) -> ir.TensorBox: assert self.gm is not None return ir.TensorBox.create( ir.SubgraphBuffer( layout=self.layout, input_nodes=self.input_nodes, gm=self.gm, example_inputs=self.example_inputs, subgraph_name=self.name, config_patches=self.config_patches if self.config_patches else None, ) ) def info_dict(self) -> dict[str, Any]: """Information returned here is logged to the autotune log file when that is enabled.""" return { "backend": "subgraph", "kernel_name": self.name, } def autoheuristic_id(self) -> str: return f"subgraph_{self.name}" class SubgraphTemplate(KernelTemplate): """ A template for subgraph evaluation to be used in autotuning. This class allows creating customized subgraphs that can be appended as choices during the autotuning process, enabling the selection of optimal implementations for complex operations. """ index_counter = itertools.count() def __init__( self, name: str, ): """ Initialize a subgraph template. Args: name: The name of this template graph: The FX graph """ super().__init__(name=name) def generate( # type: ignore[override] self, name: str, input_nodes: list[Buffer], layout: Layout, make_fx_graph: Callable[..., Any], description: str = "", input_gen_fns: dict[int, Callable[[Any], torch.Tensor]] | None = None, **kwargs: Any, ) -> SubgraphChoiceCaller: """ Generate a SubgraphChoiceCaller instance for autotuning. Args: name: The name for this subgraph choice input_nodes: List of input nodes to the subgraph layout: Memory layout information for the output make_fx_graph: Callable that creates the FX graph for this subgraph description: Optional description of this choice input_gen_fns: Optional dict mapping input indices to tensor generators **kwargs: Additional keyword arguments Returns: SubgraphChoiceCaller: A callable object that can be used for autotuning """ return SubgraphChoiceCaller( name=f"{name}_{next(SubgraphTemplate.index_counter)}", input_nodes=input_nodes, layout=layout, description=description, make_fx_graph=make_fx_graph, input_gen_fns=input_gen_fns, ) def generate_custom_op_choices( self, name: str, decompositions: list[Callable[..., Any]], input_nodes: list[Buffer], non_tensor_args: list[dict[str, Any]], default_impl: Callable[..., Any] | None = None, input_gen_fns: dict[int, Callable[[Any], torch.Tensor]] | None = None, config_patches_list: list[dict[str, Any]] | None = None, ) -> list[SubgraphChoiceCaller]: """ Generate multiple SubgraphChoiceCaller instances for custom op autotuning. This method extends SubgraphTemplate to support custom op decompositions, allowing multiple implementations to compete in autotuning. Args: name: Base name for the choices decompositions: List of decomposition functions to compete in autotuning input_nodes: List of tensor inputs. All tensor arguments must be passed here. non_tensor_args: List of non-tensor kwargs only, one dict per corresponding decomposition. default_impl: Default implementation for layout inference input_gen_fns: Optional dict mapping input indices to tensor generators config_patches_list: Optional list of config patches per decomposition Returns: List of SubgraphChoiceCaller instances for autotuning """ if not decompositions: return [] assert len(decompositions) == len(non_tensor_args), ( f"decompositions and non_tensor_args must have same length, " f"got {len(decompositions)} decompositions and {len(non_tensor_args)} kwargs" ) # Default to empty config_patches if not provided if config_patches_list is None: config_patches_list = [{} for _ in decompositions] # Infer layouts and ensure layout consistency for fair autotuning comparison layouts = [ self._infer_custom_op_layout( input_nodes, decomp, kwargs, default_impl, input_gen_fns ) for decomp, kwargs in zip(decompositions, non_tensor_args) ] # Validate all decompositions produce equivalent layouts for fair comparison self._validate_layout_equivalence(name, decompositions, layouts) layout = layouts[0] # All layouts are now validated to be equivalent choices: list[SubgraphChoiceCaller] = [] for decomp, decomp_kwargs, config_patches in zip( decompositions, non_tensor_args, config_patches_list ): # Create make_fx_graph function for this decomposition # Uses error_on_new_guards to detect impls that add guards from torch.fx.experimental.symbolic_shapes import _ShapeEnvGuardError def make_fx_graph( *args: Any, decomp: Callable[..., Any] = decomp, decomp_kwargs: dict[str, Any] = decomp_kwargs, ) -> Any: # decomp_kwargs contains all merged parameters: CustomOpConfig params + runtime kwargs from torch.fx.experimental.proxy_tensor import make_fx from ..decomposition import select_decomp_table decomposition_table = select_decomp_table() shape_env = V.fake_mode.shape_env # Use error_on_new_guards to detect impls that add guards during tracing guard_ctx = ( shape_env.error_on_new_guards() if shape_env is not None else contextlib.nullcontext() ) with guard_ctx: return make_fx( functools.partial(decomp, **decomp_kwargs), decomposition_table=decomposition_table, tracing_mode="symbolic", )(*args) # Generate descriptive name for this variant variant_name = self._generate_variant_name(decomp, decomp_kwargs) # Try to create choice; skip if it adds guards try: choice = self.generate( name=f"{name}_{variant_name}", input_nodes=input_nodes, layout=layout, make_fx_graph=make_fx_graph, description=f"CustomOp {decomp.__name__}", input_gen_fns=input_gen_fns, ) except _ShapeEnvGuardError: log.info( "Skipping decomposition %s: adds guards during tracing", decomp.__name__, ) counters["inductor"]["custom_op_decomp_guard_skips"] += 1 continue # Cache decomposition info for range-based dispatch choice.cache_decomposition(decomp, decomp_kwargs) # Store config_patches for this choice choice.config_patches = config_patches choices.append(choice) return choices def _generate_variant_name( self, decomp: Callable[..., Any], kwargs: dict[str, Any] ) -> str: """Generate a descriptive name for a decomposition variant with its parameters.""" import re base_name = decomp.__name__ if not kwargs: return base_name def sanitize_value(v: Any) -> str: """Convert a value to a valid Python identifier component.""" s = str(v) # Replace invalid characters with underscores s = re.sub(r"[^a-zA-Z0-9_]", "_", s) # Ensure it doesn't start with a digit if s and s[0].isdigit(): s = "_" + s return s param_suffix = "_".join( f"{k}_{sanitize_value(v)}" for k, v in sorted(kwargs.items()) ) return f"{base_name}_{param_suffix}" def _validate_non_tensor_kwargs(self, kwargs: dict[str, Any]) -> None: """Validate that kwargs contains only non-tensor arguments.""" for key, value in kwargs.items(): assert not isinstance(value, (torch.Tensor, Buffer)), ( f"kwargs['{key}'] contains tensor {type(value)}. " f"Tensor arguments should be in input_nodes, not kwargs. " f"Only scalar/non-tensor parameters should be in kwargs." ) def _validate_layout_equivalence( self, op_name: str, decompositions: list[Callable[..., Any]], layouts: list[Layout], ) -> None: """Ensure all layouts have consistent stride, device, dtype, and sizes for fair autotuning.""" if not layouts: return reference = layouts[0] for i, layout in enumerate(layouts[1:], start=1): if (layout.device, layout.dtype, layout.size, layout.stride) != ( reference.device, reference.dtype, reference.size, reference.stride, ): raise AssertionError( f"Layout mismatch in custom op '{op_name}': " f"decomposition '{decompositions[i].__name__}' produces " f"({layout.device}, {layout.dtype}, {layout.size}, {layout.stride}) " f"but '{decompositions[0].__name__}' produces " f"({reference.device}, {reference.dtype}, {reference.size}, {reference.stride})" ) def _infer_custom_op_layout( self, input_nodes: list[Buffer], function_decomposition: Callable[..., Any], kwargs: dict[str, Any], default_impl: Callable[..., Any] | None = None, input_gen_fns: dict[int, Callable[[Any], torch.Tensor]] | None = None, ) -> Layout: """Infer output layout for custom ops using the default implementation when available. Note that the Subgraph assumes custom ops return exactly one tensor output. TODO: Add support for multiple output custom ops. """ import functools from torch._inductor.virtualized import V # Assert kwargs contain only non-tensor arguments self._validate_non_tensor_kwargs(kwargs) with V.fake_mode: example_inputs = [] for i, inp in enumerate(input_nodes): if input_gen_fns and i in input_gen_fns: fake_tensor = input_gen_fns[i](inp) else: raw_shape = inp.get_size() concrete_shape = V.graph.sizevars.optimization_hints(raw_shape) raw_stride = inp.get_stride() concrete_stride = V.graph.sizevars.optimization_hints(raw_stride) fake_tensor = torch.empty_strided( concrete_shape, concrete_stride, dtype=inp.get_dtype(), device=inp.get_device(), ) example_inputs.append(fake_tensor) fn = functools.partial(function_decomposition, **kwargs) output = fn(*example_inputs) # Assert single output assert isinstance(output, torch.Tensor), ( f"Expected single tensor output, got {type(output)}. " f"Multi-output custom ops not yet supported in autotuning." ) return FixedLayout( device=output.device, dtype=output.dtype, size=output.shape, stride=output.stride(), )