from __future__ import annotations import base64 import copyreg import dataclasses import functools import hashlib import importlib import io import itertools import json import logging import os import pickle import pkgutil import re import shlex import shutil import struct import subprocess import sys import tempfile import textwrap import threading import warnings from bisect import bisect_right from copy import copy from ctypes import c_void_p, CDLL, cdll from datetime import timedelta from functools import partial from pathlib import Path from time import time, time_ns from types import ModuleType from typing import ( Any, Callable, cast, NoReturn, Optional, TYPE_CHECKING, TypeVar, Union, ) from typing_extensions import Self import torch import torch.distributed as dist from torch import SymInt, Tensor from torch._dynamo.utils import CompileEventLogger, counters, dynamo_timed from torch._inductor import config, exc, metrics from torch._inductor.codegen.cuda import cuda_env from torch._inductor.codegen.rocm.compile_command import ( rocm_compile_command, rocm_compiler, ) from torch._inductor.cpp_builder import ( _LINKER_SCRIPT, _set_gpu_runtime_env, _TORCH_PATH, _transform_cuda_paths, CppBuilder, CppOptions, CppTorchDeviceOptions, get_compiler_version_info, get_name_and_dir_from_output_file_path, normalize_path_separator, ) from torch._inductor.cpu_vec_isa import pick_vec_isa from torch._inductor.custom_graph_pass import CustomGraphPass, CustomGraphPassType from torch._inductor.freezing_utils import has_frozen_params, is_frozen_param from torch._inductor.runtime.compile_tasks import ( _reload_python_module, _reload_python_module_in_subproc, ) from torch._inductor.runtime.runtime_utils import cache_dir, default_cache_dir from torch._inductor.utils import ( ALIGN_BYTES, clear_on_fresh_inductor_cache, is_linux, is_windows, ) from torch._logging import trace_structured from torch._subclasses.fake_tensor import ( extract_tensor_metadata, FakeTensor, TensorMetadata, ) from torch._utils_internal import log_cache_bypass from torch.compiler import config as cconfig from torch.compiler._cache import CacheArtifactManager, CacheArtifactType from torch.fx.experimental.symbolic_shapes import has_hint, hint_int, ShapeEnv from torch.utils._ordered_set import OrderedSet from .package.pt2_archive_constants import CUSTOM_OBJ_FILENAME_PREFIX from .remote_cache import create_cache from .runtime import autotune_cache from .runtime.autotune_cache import AutotuneCacheBundler from .triton_bundler import TritonBundler if config.is_fbcode(): from triton.fb import build_paths from torch._inductor.fb.utils import ( log_global_cache_errors, log_global_cache_stats, log_global_cache_vals, use_global_cache, ) else: def log_global_cache_errors(*args: Any, **kwargs: Any) -> None: # type: ignore[misc] pass def log_global_cache_stats(*args: Any, **kwargs: Any) -> None: # type: ignore[misc] pass def log_global_cache_vals(*args: Any, **kwargs: Any) -> None: # type: ignore[misc] pass def use_global_cache() -> bool: # type: ignore[misc] return False if TYPE_CHECKING: from collections.abc import Generator, KeysView, Sequence from concurrent.futures import Future from .compile_fx import _CompileFxKwargs, CompiledFxGraph from .graph import GraphLowering from .ir import ChoiceCaller from .output_code import CompiledFxGraphConstants, OutputCode from .remote_cache import JsonDataTy, RemoteCache from .runtime.hints import HalideInputSpec, HalideMeta from .runtime.triton_heuristics import CachingAutotuner from .utils import InputType T = TypeVar("T") _IS_WINDOWS = sys.platform == "win32" LOCK_TIMEOUT = 600 output_code_log = torch._logging.getArtifactLogger(__name__, "output_code") log = logging.getLogger(__name__) def get_cpp_wrapper_cubin_path_name() -> str: return "cubin_path" if torch.version.hip is None else "hsaco_path" @functools.lru_cache(None) def get_global_cache_path_impl(global_cache_dir: str) -> Optional[Path]: return ( Path(os.path.join(global_cache_dir, CacheBase.get_system()["hash"])) if global_cache_dir is not None else None ) class CacheBase: @staticmethod @functools.lru_cache(None) def get_system() -> dict[str, Any]: try: from triton.compiler.compiler import triton_key # Use triton_key instead of triton.__version__ as the version # is not updated with each code change triton_version = triton_key() except ModuleNotFoundError: triton_version = None try: system: dict[str, Any] = { "device": {"name": None}, "version": { "triton": triton_version, }, } device_properties = torch.cuda.get_device_properties( torch.cuda.current_device() ) if torch.version.cuda is not None: system["device"]["name"] = device_properties.name system["version"]["cuda"] = torch.version.cuda else: system["device"]["name"] = device_properties.gcnArchName system["version"]["hip"] = torch.version.hip except (AssertionError, RuntimeError): # If cuda is not installed, none of the above config is relevant. system = {} system["hash"] = hashlib.sha256( json.dumps(system, sort_keys=True).encode("utf-8") ).hexdigest() return system @staticmethod @clear_on_fresh_inductor_cache @functools.lru_cache(None) def get_local_cache_path() -> Path: return Path(os.path.join(cache_dir(), "cache", CacheBase.get_system()["hash"])) @staticmethod def get_global_cache_path() -> Optional[Path]: return get_global_cache_path_impl(config.global_cache_dir) def __init__(self) -> None: self.system = CacheBase.get_system() def get_local_cache(self) -> dict[str, Any]: local_cache_path = self.get_local_cache_path() if not local_cache_path.is_file(): return {} with open(local_cache_path) as local_cache_fp: local_cache = json.load(local_cache_fp) return local_cache["cache"] def update_local_cache(self, local_cache: dict[str, Any]) -> None: local_cache_path = self.get_local_cache_path() write_atomic( str(local_cache_path), json.dumps({"system": self.system, "cache": local_cache}, indent=4), make_dirs=True, ) class LocalCache(CacheBase): def lookup(self, *keys: str) -> Optional[dict[str, Any]]: cache = self.get_local_cache() sub_cache = cache for key in keys: if key in cache: sub_cache = cache[key] else: return None return sub_cache def set_value(self, *keys: str, value: Any) -> None: cache = self.get_local_cache() sub_cache = cache for key in keys[0:-1]: sub_cache.setdefault(key, {}) sub_cache = sub_cache[key] sub_cache[keys[-1]] = value self.update_local_cache(cache) class PersistentCache(CacheBase): @functools.lru_cache(None) # noqa: B019 def get_global_cache(self) -> dict[str, Any]: global_cache_path = self.get_global_cache_path() if global_cache_path is None or not global_cache_path.is_file(): return {} with open(global_cache_path) as global_cache_fp: global_cache = json.load(global_cache_fp) return global_cache["cache"] def lookup( self, choices: list[ChoiceCaller], op: str, inputs: str, benchmark: Optional[Callable[[Any], dict[ChoiceCaller, float]]], ) -> dict[ChoiceCaller, float]: """ Check to see if we have benchmarked the given choice callers. For each choice caller: 1. Check global_cache[op][inputs][choice][precision], return benchmark if cached. 2. Check local_cache[op][inputs][choice][precision], return benchmark if cached. 3. If benchmark is not None: a. `max_autotune_gemm=True`: benchmark the choice, update local_cache[op][inputs][choice], and return the benchmark. b. `max_autotune_gemm=False`: don't benchmark the choice, return nothing. """ precision = torch.get_float32_matmul_precision() log_stats = partial(log_global_cache_stats, self.system, op, inputs, precision) log_vals = partial(log_global_cache_vals, self.system, op, inputs, precision) log_errors = partial( log_global_cache_errors, self.system, op, inputs, precision ) timings = {} def check_cache(cache: dict[str, Any], callback: Any = None) -> bool: """Check if `cache` contains data for all the choices""" hit = True for choice in choices: choice_hash = choice.hash_key() if choice_hash in cache.get(op, {}).get(inputs, {}).get(precision, {}): # cache hit timings[choice] = cache[op][inputs][precision][choice_hash] else: # cache miss hit = False break if callback: callback(cached=hit) return hit if config.max_autotune or config.max_autotune_gemm: local_cache = self.get_local_cache() if config.autotune_local_cache else {} # check local cache first since it is data specific to the current machine if ( not check_cache(local_cache) and not ( use_global_cache() and check_cache(self.get_global_cache(), callback=log_stats) ) and benchmark is not None ): try: # re-benchmark everything to try to get consistent numbers from the same machine timings = benchmark(choices) assert all(choice in timings for choice in choices) local_cache.setdefault(op, {}) local_cache[op].setdefault(inputs, {}).setdefault(precision, {}) for choice, timing in timings.items(): local_cache[op][inputs][precision][choice.hash_key()] = timing except RuntimeError as e: # catch and log autotuning failures log_errors(e) raise e self.update_local_cache(local_cache) timings_to_log = { choice.hash_key(): timings[choice] for choice in choices } log_vals(timings_to_log) elif use_global_cache(): # only check global cache, not local one check_cache(self.get_global_cache(), callback=log_stats) # may have a partial cache hit, where not everything is benchmarked return timings def get_lock_dir() -> str: lock_dir = os.path.join(cache_dir(), "locks") if not os.path.exists(lock_dir): os.makedirs(lock_dir, exist_ok=True) return lock_dir def sha256_hash(data: bytes) -> str: # [:51] to strip off the "Q====" suffix common to every hash value. return base64.b32encode(hashlib.sha256(data).digest())[:51].decode("utf-8").lower() def code_hash(code: Union[str, bytes], extra: Union[str, bytes] = "") -> str: hashing_str = code if isinstance(code, bytes) else code.encode("utf-8") if extra: extra_b = extra if isinstance(extra, bytes) else extra.encode("utf-8") hashing_str = hashing_str + b"||" + extra_b return "c" + sha256_hash(hashing_str) def get_path( basename: str, extension: str, specified_dir: str = "" ) -> tuple[str, str, str]: if specified_dir: if os.path.isabs(specified_dir): subdir = specified_dir else: subdir = os.path.join(cache_dir(), specified_dir) else: subdir = os.path.join(cache_dir(), basename[1:3]) path = os.path.join(subdir, f"{basename}.{extension}") return basename, subdir, path def get_hash( content: Union[str, bytes], extra: str = "", hash_type: str = "code" ) -> str: if hash_type == "code": return code_hash(content, extra) if hash_type in ["cubin", "hsaco", "spv"]: return code_hash(repr(content)) raise AssertionError(f"Unknown hash type {hash_type}") def write( content: Union[str, bytes], extension: str, extra: str = "", hash_type: str = "code", specified_dir: str = "", ) -> tuple[str, str]: # use striped content to compute hash so we don't end up with different # hashes just because the content begins/ends with different number of # spaces. key: str = get_hash(content.strip(), extra, hash_type) basename, _subdir, path = get_path(key, extension, specified_dir) if not os.path.exists(path): write_atomic(path, content, make_dirs=True) return basename, path def write_text(text: str) -> str: """ Write the `text` to a file and return the path computed based on the hash. """ return write(text, "txt")[1] def write_atomic( path_: str, content: Union[str, bytes], make_dirs: bool = False, encode_utf_8: bool = False, ) -> None: # Write into temporary file first to avoid conflicts between threads # Avoid using a named temporary file, as those have restricted permissions assert isinstance(content, (str, bytes)), ( "Only strings and byte arrays can be saved in the cache" ) path = Path(path_) if make_dirs: path.parent.mkdir(parents=True, exist_ok=True) tmp_path = path.parent / f".{os.getpid()}.{threading.get_ident()}.tmp" write_mode = "w" if isinstance(content, str) else "wb" with tmp_path.open(write_mode, encoding="utf-8" if encode_utf_8 else None) as f: f.write(content) try: tmp_path.rename(target=path) except FileExistsError: if not _IS_WINDOWS: raise # On Windows file exist is expected: https://docs.python.org/3/library/pathlib.html#pathlib.Path.rename # Below two lines code is equal to `tmp_path.rename(path)` on non-Windows OS. # 1. Copy tmp_file to Target(Dst) file. shutil.copy2(src=tmp_path, dst=path) # 2. Delete tmp_file. os.remove(tmp_path) @dataclasses.dataclass class TensorMetadataAndValues: """ TensorMetadata plus the elements as a list of raw values. Used for hashing inlined constants. """ tensor_metadata: TensorMetadata values: list[Any] def _ident(x: T) -> T: return x def extract_tensor_metadata_for_cache_key(t: Tensor) -> TensorMetadata: """ Extracts the tensor metadata and removes fields of the TensorMetadata that are not needed for caching """ meta = extract_tensor_metadata(t) if not hasattr(t, "_is_inductor_static"): meta = dataclasses.replace(meta, storage_offset=0, storage_bytes=None) return meta class FxGraphCachePickler(pickle.Pickler): """ Custom pickler to customize the pickling of some objects (Tensors), only for the purpose of computing a hash for keying into the FxGraphCache. Tensors contain objects that don't pickle and/or vary between runs, and we want to capture the data that allow us to compute a stable, but safe hash. """ def __init__( self, gm: torch.fx.GraphModule, has_user_defined_triton_kernels: bool = False, ) -> None: """ Create an FX graph pickler. If include_non_inlined=True, then pickling will include the _values_ for all Tensors. (Note that any tensors are constants attached as attributes to the GraphModule). Otherwise, pickling will include only the metadata for these tensors. """ self._stream = io.BytesIO() super().__init__(self._stream) self.dispatch_table = copyreg.dispatch_table.copy() self.dispatch_table.update( { FakeTensor: functools.partial(self._reduce_fake_tensor), torch.Tensor: functools.partial(self._reduce_tensor), torch.nn.parameter.Parameter: functools.partial(self._reduce_tensor), torch.SymInt: functools.partial(self._reduce_symint), torch.fx.experimental._backward_state.BackwardState: functools.partial( self._reduce_unsupported ), } ) if has_user_defined_triton_kernels: # Need to use runtime type as GraphModule generates a singleton in __new__ function self.dispatch_table[gm.__class__] = functools.partial( self._reduce_graph_module ) # Run with pickler.fast so it doesn't intern strings, making the hash result more predictable # TODO: pickler.fast is technically deprecated. Will this work on new python versions? self.fast = True def _reduce_fake_tensor( self, t: Tensor ) -> tuple[Callable[[T], T], tuple[TensorMetadata]]: """ Custom reducer to pickle FakeTensors. """ metadata = extract_tensor_metadata_for_cache_key(t) return (_ident, (metadata,)) def _reduce_tensor( self, t: Tensor ) -> tuple[Callable[[T], T], tuple[Union[TensorMetadata, TensorMetadataAndValues]]]: """ Custom reducer to pickle Tensors. If we see tensors, we know they're constants stored as attributes on the GraphModule. """ from .graph import GraphLowering if t.is_mkldnn: # TODO: These tensors don't currently pickle, so we can't cache a compiled # graph containing them. Just fail now. If mkldnn tensors get pickling # support, we can remove this. raise BypassFxGraphCache("mkldnn tensors unpickleable") metadata = extract_tensor_metadata_for_cache_key(t) # If this is a non-inlined frozen parameter, we consider the metadata only. if is_frozen_param(t) and not GraphLowering.can_inline_constant(t): return (_ident, (metadata,)) # Very large tensors will be expensive to copy to cpu and hash. Let's at least # report any slowness. start = time() values = t.tolist() elapsed = time() - start if elapsed > 1.0: warnings.warn( f"FX graph cache copying of a large constant took {elapsed:.1}s. " "Please file an issue." ) return (_ident, (TensorMetadataAndValues(metadata, values),)) def _reduce_symint(self, s: SymInt) -> tuple[Callable[[T], T], tuple[str]]: """ Custom reducer to pickle SymInts. """ # For hashing purposes, we only care about the name of the symbol and not the # backed value. We evaluate guards stored with a cached graph to ensure a cached # entity with SymInt args is safe to reuse. return (_ident, (str(s),)) def _reduce_unsupported(self, s: Any) -> NoReturn: """ Custom reducer to handle any objects that we don't support and therefore raise to bypass caching. """ raise BypassFxGraphCache("Reduce unsupported") def _reduce_graph_module( self, gm: torch.fx.GraphModule ) -> tuple[Any, tuple[dict[str, Any], str]]: """ Custom reducer for graph module to handle irrelevant data for user defined triton kernels Essentially what we are doing here is a huge hack where user defined triton kernel contain a dynamo time side table and the arguments to the call_function are indicies into this side table. These arguments are not for hashing purposes since we included the source code into the cache key and the numbers are prone to give false negatives due to ordering. """ fn, (data, imports) = gm.__reduce__() code = data["_code"] code = re.sub(r"kernel_idx = \d+", "", code) code = re.sub(r"constant_args_idx = \d+", "", code) data["_code"] = code return fn, (data, imports) def dumps(self, obj: Any) -> bytes: """ Pickle an object and return a byte string. """ try: self.dump(obj) return self._stream.getvalue() except (TypeError, AttributeError) as e: # Some configs options may not pickle. log.warning("Failed to pickle cache key", exc_info=True) raise BypassFxGraphCache("Failed to pickle cache key") from e finally: # Reset our stream for the next dump. self._stream.seek(0) self._stream.truncate(0) def get_hash(self, obj: Any) -> str: """ Serialize an object and return a hash of the bytes. """ serialized_data = self.dumps(obj) return sha256_hash(serialized_data) def debug_lines(self, inp: FxGraphHashDetails) -> list[str]: """ Get a printable string describing in more detail all the attributes comprising an object. Useful for debugging when one graph hashes to a different value than another. """ def get_str(obj: Any) -> str: if isinstance(obj, torch.Tensor): return str(extract_tensor_metadata_for_cache_key(obj)) elif isinstance(obj, bytes): return "" elif type(obj) in self.dispatch_table: # Run the reducer on the object return str(self.dispatch_table[type(obj)](obj)[1]) else: return str(obj) lines = [] for attr, obj in vars(inp).items(): if isinstance(obj, list): for ii in range(len(obj)): h = self.get_hash(obj[ii]) lines.append(f"[{h}] {attr}[{ii}]: {get_str(obj[ii])}") elif isinstance(obj, dict): for k, v in obj.items(): h = self.get_hash(v) lines.append(f"[{h}] {attr}[{k}]: {get_str(v)}") else: h = self.get_hash(obj) lines.append(f"[{h}] {attr}: {get_str(obj)}") return lines def build_code_hash( roots: list[str] | None, prefix: str, hasher: hashlib._Hash ) -> None: for lib in sorted(pkgutil.iter_modules(roots, prefix), key=lambda x: x.name): spec = lib.module_finder.find_spec(lib.name, None) assert spec is not None module = spec.origin assert module is not None with open(module, "rb") as f: hasher.update(spec.name.encode("utf-8")) hasher.update(f.read()) if lib.ispkg: # need to also hash submodules build_code_hash(spec.submodule_search_locations, f"{spec.name}.", hasher) @functools.lru_cache(None) def torch_key() -> bytes: """ Compute a key that contains relevant information about torch source files """ with dynamo_timed("inductor_codecache_torch_key", log_pt2_compile_event=True): if not config.is_fbcode(): def get_code_hash(root: str) -> bytes: # This function isn't meant to be used outside of torch_key, just a # helper for clarity. Instead, use torch_key() directly when you need # a hash representing the state of the source code. extra_files = ( "codegen/aoti_runtime/interface.cpp", "codegen/cpp_prefix.h", "script.ld", ) inductor_root = os.path.dirname(__file__) extra_files = [os.path.join(inductor_root, x) for x in extra_files] hasher = hashlib.sha256() hasher.update(torch.__version__.encode("utf-8")) build_code_hash([root], "", hasher) for path in extra_files: if os.path.exists(path): with open(path, "rb") as f: hasher.update(f.read()) return hasher.digest() return get_code_hash(_TORCH_PATH) from libfb.py import parutil return parutil.get_file_contents("torch/src_hash.txt").rstrip().encode("ascii") def get_inductor_root() -> str: return os.path.dirname(__file__) @dataclasses.dataclass class OrderedSetHolder: """ See FxGraphHashDetails. Holds a sorted list to support stable hashing of set kwargs. """ items: list[Any] class BypassFxGraphCache(Exception): """ Exception to indicate that the FxGraphCache should be bypassed. """ class FxGraphHashDetails: """ Object to capture all the details for a compiled FX graph relevant to computing a safe and stable cache key. """ # Excluded kwargs param that are not stable between runs EXCLUDED_KWARGS = ["graph_id"] def __init__( self, gm: torch.fx.GraphModule, example_inputs: Sequence[InputType], fx_kwargs: _CompileFxKwargs, inputs_to_check: Sequence[int], ) -> None: self.gm = gm self.example_inputs = example_inputs self.cache_key_tag = cconfig.cache_key_tag # Order kwargs so hashing is stable to changes in kwarg order. Although # it's technically a _CompileFxKwargs we don't actually need it typed as # such since we're just using it to generate a hash. self.fx_kwargs: dict[str, object] = {} for k, v in sorted(fx_kwargs.items()): if k not in self.EXCLUDED_KWARGS: if type(v) in (set, OrderedSet): # noqa: set_linter # Special case to handle set params. Python sets can't be # ordered, so sort the elements and store them in a proxy. self.fx_kwargs[k] = OrderedSetHolder(sorted(v)) # type: ignore[call-overload] else: self.fx_kwargs[k] = v from torch._higher_order_ops.triton_kernel_wrap import ( kernel_side_table, triton_kernel_wrapper_functional, triton_kernel_wrapper_mutation, ) from torch._inductor.codegen.wrapper import ( user_defined_triton_kernel_transitive_closure_source_code, ) # Node meta will not be part of gm's reduce function, so lets remember # the kernel source code separately self.user_defined_triton_source: list[Any] = [] if gm is not None: for module in gm.modules(): if not isinstance(module, torch.fx.GraphModule): continue for node in itertools.chain( module.graph.find_nodes( op="call_function", target=triton_kernel_wrapper_functional ), module.graph.find_nodes( op="call_function", target=triton_kernel_wrapper_mutation ), ): from triton.runtime.autotuner import Autotuner kernel = kernel_side_table.get_kernel(node.kwargs["kernel_idx"]) configs = None if isinstance(kernel, Autotuner): if kernel.configs: configs = str( sorted( sorted(str(kv) for kv in c.all_kwargs().items()) for c in kernel.configs ) ) kernel = kernel.fn kernel_source = ( user_defined_triton_kernel_transitive_closure_source_code( kernel ) ) constant_args = kernel_side_table.get_constant_args( node.kwargs["constant_args_idx"] ) self.user_defined_triton_source.append( (kernel_source, constant_args, configs) ) # Alignment checks self.inputs_to_check = inputs_to_check no_tensor_inputs = not any(isinstance(x, torch.Tensor) for x in example_inputs) # This device index is usually already encoded by the device of the inputs # but fx graphs don't necessarily have tensor inputs. If there aren't any, # we need to guard on the device index in case we allocate cuda tensors if no_tensor_inputs and torch.accelerator.is_available(): self.default_cuda_device_index = torch.accelerator.current_device_index() # 'Deterministic algorithms' can affect codegen via lowering to cuda kernels. self.deterministic_algorithms_settings = ( torch.are_deterministic_algorithms_enabled(), torch.is_deterministic_algorithms_warn_only_enabled(), torch.utils.deterministic.fill_uninitialized_memory, # type: ignore[attr-defined] ) # Global settings affecting matmul codegen. self.cuda_matmul_settings = ( torch.backends.cuda.matmul.allow_tf32, torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction, torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction, ) # Also hash on various system info (including the triton compiler version). self.torch_version = torch_key() self.system_info = CacheBase.get_system() self.inductor_config = config.save_config_portable() # Custom post grad passes should provide an ID to hash. self.post_grad_custom_pre_pass = self._get_custom_pass_detail( config.post_grad_custom_pre_pass ) self.post_grad_custom_post_pass = self._get_custom_pass_detail( config.post_grad_custom_post_pass ) def _get_custom_pass_detail( self, custom_pass: CustomGraphPassType ) -> Optional[Any]: if not custom_pass: return None assert isinstance(custom_pass, CustomGraphPass) return custom_pass.uuid() def compiled_fx_graph_hash( gm: torch.fx.GraphModule, example_inputs: Sequence[InputType], fx_kwargs: _CompileFxKwargs, inputs_to_check: Sequence[int], ) -> tuple[str, list[str]]: """ Generate a unique hash of the FX graph for caching. """ details = FxGraphHashDetails(gm, example_inputs, fx_kwargs, inputs_to_check) has_user_defined_triton_kernels = len(details.user_defined_triton_source) != 0 pickler = FxGraphCachePickler(gm, has_user_defined_triton_kernels) # The prefix distinguishes among the other kinds of objects we # cache in this module. key = "f" + pickler.get_hash(details) debug_lines = pickler.debug_lines(details) debug_str = "\n".join(debug_lines) log.debug(f"FX graph cache hash details for key {key}:\n{debug_str}") # noqa: G004 return key, debug_lines def add_ephemeral_timeout_increase_for_distributed(time_saved_ns: int) -> int: """ Ephemerally increases the NCCL timeout when compiling for a distributed job Returns amount of seconds increased """ if not torch.distributed.is_available() or not torch.distributed.is_initialized(): return 0 increased_timeout_sec = int(time_saved_ns // 1e9) # convert to seconds if config.is_fbcode(): fudge_factor = torch._utils_internal.justknobs_getval_int( "pytorch/remote_cache:ephemeral_timeout_fudge_factor_percentage" ) log.info( "Ephemeral NCCL timeout increase fudge factor %d and original increase value %d", fudge_factor, increased_timeout_sec, ) increased_timeout_sec += int(increased_timeout_sec * fudge_factor / 100) log.info("Increasing NCCL timeout by %d", increased_timeout_sec) dist.distributed_c10d._add_ephemeral_timeout_for_all_pgs( timedelta(seconds=increased_timeout_sec) ) return increased_timeout_sec class FxGraphCache: """ Supports caching and reusing compiled Fx graphs. The overall strategy is as follows: - This cache stores entries on disk. When saving an entry, we can't serialize callables (that could be C++, Triton, etc.), so we serialize their own disk cache location. We then recreate the compiled artifact after fetching from disk. - For indexing the cache, we gather the fields relevant to identifying an FxGraph (the graph module, graph inputs, system settings etc.) into an FxGraphCacheDetails object, pickle it, and compute a hash for the key. See FxGraphCachePickler. - Among the metadata we store, we also include a guards expression that's appropriate for validating any symbols for Tensor arguments that have symbolic bounds. On cache lookup then, we evaluate those guards in the current context to validate that a cached entry can be served. - A given graph could have multiple compiled versions, corresponding to different sets of guards. Therefore, we store cache entries in the form: // - On lookup, we compute the key from the graph details, iterate over all leaf files in the corresponding subdirectory, deserialize the entry, and evaluate its guards expression. If the evaluation succeeds, we have a cache hit. If it fails, we compile the graph and store a new entry. - Finally, on a cache hit, we need to make sure any guards that would have been created during compilation are added to the current context. """ # TODO(masnesral): Investigate whether it's beneficial to store compiled graphs # in an in-memory cache after loading from disk. @staticmethod def _get_tmp_dir() -> str: """ Get the toplevel temporary directory for storing compiled graphs. """ return os.path.join(cache_dir(), "fxgraph") @staticmethod def _get_tmp_dir_for_key(key: str) -> str: """ Return the disk location for a given cache key. """ return os.path.join(FxGraphCache._get_tmp_dir(), key[1:3], key) @staticmethod def _filter_backed_symints(inputs: Sequence[InputType]) -> list[torch.SymInt]: """ Get the backed SymInt objects from the input list. Note that we can never have guards that depend on unbacked symint. """ return [s for s in inputs if isinstance(s, torch.SymInt) and has_hint(s)] @staticmethod def _get_shape_env() -> Optional[ShapeEnv]: """ Helper to get the shape env from the tracing context. """ ctx = torch._guards.TracingContext.try_get() if not ctx: return None return ctx.fake_mode.shape_env @staticmethod def _lookup_graph( key: str, example_inputs: Sequence[InputType], local: bool, remote_cache: Optional[RemoteCache[JsonDataTy]], constants: CompiledFxGraphConstants, ) -> tuple[Optional[CompiledFxGraph], dict[str, Any]]: """ Lookup a compiled graph in the cache by key. On a hit, return the deserialized CompiledFxGraph object. On a miss, return None. """ shape_env = FxGraphCache._get_shape_env() assert shape_env is not None symints = FxGraphCache._filter_backed_symints(example_inputs) hints = [hint_int(s) for s in symints] def iterate_over_candidates() -> Generator[ tuple[CompiledFxGraph, bytes], None, None ]: if local: subdir = FxGraphCache._get_tmp_dir_for_key(key) if os.path.exists(subdir): for path in sorted(os.listdir(subdir)): try: with open(os.path.join(subdir, path), "rb") as f: content = f.read() yield pickle.loads(content), content except Exception: log.warning( "fx graph cache unable to load compiled graph", exc_info=True, ) if remote_cache: try: if (cache_data := remote_cache.get(key)) is not None: assert isinstance(cache_data, dict) data = cache_data["data"] assert isinstance(data, (str, bytes)) content = base64.b64decode(data) yield pickle.loads(content), content except Exception: log.warning( "fx graph cache unable to load compiled graph", exc_info=True ) # Iterate over any entries in the subdir for this key and evaluate # their guards to determine whether there's a hit. graph = None pickled_content = None cache_info: dict[str, Any] = dict() for candidate, pickled_content in iterate_over_candidates(): if not candidate.guards_expr: # No guards to evaluate, so this is a hit. graph = candidate break # Evaluate the guard expression in the current context. # If there's not a cache hit, we don't want the evaluation to # affect the current env, e.g., cause the creation of new guards, # so we evaluate with the hints instead of the symbols. hit = bool( shape_env.evaluate_guards_expression(candidate.guards_expr, hints) ) log.debug( "fx graph cache key %s evaluating guards [%s] with values %s => hit=%s", key, candidate.guards_expr, hints, hit, ) if hit: graph = candidate break if graph is None: return None, cache_info if pickled_content is not None: CacheArtifactManager.record_artifact( CacheArtifactType.INDUCTOR, key, pickled_content ) if bundle := graph._triton_bundle: triton_bundler_meta = TritonBundler.read_and_emit(bundle) if (meta := triton_bundler_meta) is not None: cache_info["triton_bundler_meta"] = str(meta) # TODO: Clean up autograd cache integration CompileEventLogger.try_add_pt2_compile( "inductor_compile", cached_kernel_names=meta.cached_kernel_names ) if len(meta.cached_kernel_names) > 0: CompileEventLogger.increment_toplevel("num_triton_bundles") try: artifact_path = graph.after_deserialization(constants) from .graph import GraphLowering # This is used by tests to check the output for specific details. if GraphLowering.save_output_code is not None: GraphLowering.save_output_code(graph.source_code) except OSError: # Not expected, but in case the PyCodeCache entry is removed from # underneath us, treat it as a cache miss and recompile. return None, cache_info inductor_meta = autotune_cache.inductor_meta_from_config() code = graph.source_code AutotuneCacheBundler.begin_compile(inductor_meta, code=code) # Now re-evaluate with the symints to add any guards to the current env. if graph.guards_expr: check = bool( shape_env.evaluate_guards_expression(graph.guards_expr, symints) ) assert check is True log.debug( "fx graph cache key %s post-load guards: %s", key, shape_env.guards ) # Increment the cached metrics/counters by the amounts recorded when the FX # graph was compiled for this cache entry. Pretending these counters # were incremented normally is useful for testing with the cache enabled. metrics.CachedMetricsHelper.apply_deltas(graph.metrics_deltas) counters["inductor"] += graph.counter_deltas output_code_log.debug("Output code: \n%s", code) output_code_log.debug("Output code written to: %s", artifact_path) # On cache hit, use artifact path as filename trace_structured( "inductor_output_code", lambda: {"filename": artifact_path}, payload_fn=lambda: code, ) return graph, cache_info @staticmethod def _write_to_local_cache(key: str, content: bytes) -> None: subdir = FxGraphCache._get_tmp_dir_for_key(key) if not os.path.exists(subdir): os.makedirs(subdir, exist_ok=True) # Use a hash of the serialized CompiledFxGraph to get a unique file # name. The specific name doesn't matter since a lookup involves # iterating over all entries in the parent subdir. path = os.path.join(subdir, sha256_hash(content)) write_atomic(path, content, make_dirs=True) @staticmethod def _save_graph( key: str, compiled_graph: OutputCode, example_inputs: Sequence[InputType], local: bool, remote_cache: Optional[RemoteCache[JsonDataTy]], ) -> None: """ Store a serialized CompiledFxGraph on disk. """ from .compile_fx import CompiledFxGraph assert isinstance(compiled_graph, CompiledFxGraph), ( f"serialization for {type(compiled_graph)} NYI" ) disk_compiled_graph = copy(compiled_graph) disk_compiled_graph.prepare_for_serialization() # Before serializing, compute the guard expression that will be used to # ensure that a CompiledFxGraph is valid when loaded from the cache. It's # sufficient to consider only the SymInt args to the fx graph since the # Tensor shapes are already captured in the hash for the cache key. Any # Tensor arg with a symbolic shape will have a SymInt arg for the graph. shape_env = FxGraphCache._get_shape_env() assert shape_env is not None symints = FxGraphCache._filter_backed_symints(example_inputs) guards = shape_env.get_pruned_guards(symints) disk_compiled_graph.guards_expr = shape_env.produce_guards_expression( placeholders=symints, guards=guards ) try: content = pickle.dumps(disk_compiled_graph) except Exception: log.warning( "fx graph cache unable to serialize compiled graph", exc_info=True ) counters["inductor"]["fxgraph_cache_pickle_error"] += 1 return try: CacheArtifactManager.record_artifact( CacheArtifactType.INDUCTOR, key, content ) if local: FxGraphCache._write_to_local_cache(key, content) if remote_cache: time_taken_ms = int((disk_compiled_graph._time_taken_ns or 0) // 1e6) cache_data: JsonDataTy = { "data": base64.b64encode(content).decode("ascii"), "time_taken_ms": time_taken_ms, } remote_cache.put(key, cache_data) except Exception: log.warning("fx graph unable to write to cache", exc_info=True) counters["inductor"]["fxgraph_cache_write_error"] += 1 @staticmethod def _check_for_hop(gm: torch.fx.GraphModule) -> None: for module in gm.modules(): if not isinstance(module, torch.fx.GraphModule): continue for node in module.graph.nodes: if ( isinstance(node.target, torch._ops.HigherOrderOperator) and not node.target.cacheable() ): raise BypassFxGraphCache( f"Can't cache HigherOrderOperator: {node.target.name()}" ) if node.op == "getattr" and isinstance( getattr(gm, node.target), torch._C.ScriptObject ): raise BypassFxGraphCache("Can't cache torchbind objects") @staticmethod def _check_can_cache(gm: torch.fx.GraphModule) -> None: """ Check some conditions that would preclude caching and raise BypassFxGraphCache to bypass in case caching is not possible. """ # Post grad custom passes must implement the CustomGraphPass or we don't # know how to include them in the cache key calculation. for p in (config.post_grad_custom_pre_pass, config.post_grad_custom_post_pass): if p and (not isinstance(p, CustomGraphPass) or not p.uuid()): raise BypassFxGraphCache("Unsupported post grad custom pass") # Freezing can embed constants that wouldn't be static across runs. if has_frozen_params(gm) and not torch._utils_internal.justknobs_check( "pytorch/inductor:allow_freezing_with_caching" ): raise BypassFxGraphCache("Skipping graph with frozen constants") if config.aot_inductor.use_runtime_constant_folding: raise BypassFxGraphCache( "Runtime constant folding can introduce constants that aren't " "static across runs" ) from torch._inductor.compiler_bisector import CompilerBisector if CompilerBisector.bisection_enabled: log.debug("dont cache graph when bisect enabled") raise BypassFxGraphCache # The treatment of guards in the caching implementation requires that # we have a shape env. if FxGraphCache._get_shape_env() is None: log.debug("fx graph cache no shape env") raise BypassFxGraphCache("No shape env") # We skip caching if there are any HOPs or torchbind objects. FxGraphCache._check_for_hop(gm) @staticmethod def prepare_key( gm: torch.fx.GraphModule, example_inputs: Sequence[InputType], fx_kwargs: _CompileFxKwargs, inputs_to_check: Sequence[int], remote: bool, ) -> tuple[Optional[tuple[str, list[str]]], dict[str, Any]]: """ Checks that the inductor input is cacheable, then computes and returns the cache key for the input. Returns (key_info, cache_info) where: - key_info is (hash_key, debug_lines), and - cache_info will contain debug info in the event of BypassFxGraphCache. NB: It is possible to have this function return a union instead. But I personally believe it is more annoying/difficult to read in that format. """ try: FxGraphCache._check_can_cache(gm) key, debug_lines = compiled_fx_graph_hash( gm, example_inputs, fx_kwargs, inputs_to_check ) except BypassFxGraphCache as e: counters["inductor"]["fxgraph_cache_bypass"] += 1 log.info("Bypassing FX Graph Cache because '%s'", e) if remote: log_cache_bypass("bypass_fx_graph", str(e)) cache_info = { "cache_state": "bypass", "cache_bypass_reason": str(e), "cache_event_time": time_ns(), } return None, cache_info # If key exists, then cache_info will come from load_with_key return (key, debug_lines), {} @staticmethod def get_remote_cache() -> Optional[RemoteCache[JsonDataTy]]: """ Attempts to load the remote cache, returns None on error. """ cache_id = "fx-graph-v1" return create_cache( cache_id, config.is_fbcode(), "FbRemoteFxGraphCache", "RemoteFxGraphCache", ) @staticmethod def load_with_key( key: str, debug_lines: list[str], example_inputs: Sequence[InputType], local: bool, remote_cache: Optional[RemoteCache[JsonDataTy]], is_backward: bool, constants: CompiledFxGraphConstants, ) -> tuple[Optional[CompiledFxGraph], dict[str, Any]]: """ Lookup the graph with the given key, and return results and metadata. Doesn't do any logging on its own, because AOTAutograd handles a cache miss differently from FXGraphCache. """ compiled_graph, cache_info = FxGraphCache._lookup_graph( key, example_inputs, local, remote_cache, constants ) cache_info = { **cache_info, "key": key, "components": debug_lines, "cache_event_time": time_ns(), } if compiled_graph is not None: log.info("fx graph cache hit for key %s", key) counters["inductor"]["fxgraph_cache_hit"] += 1 cache_info["cache_state"] = "hit" if remote_cache: # Count remote cache hit stats CompileEventLogger.increment_toplevel( "inductor_fx_remote_cache_hit_count" ) CompileEventLogger.add_to_set_toplevel( "inductor_fx_remote_cache_hit_keys", key ) if (time_saved_ns := compiled_graph._time_taken_ns) is not None: cache_info["time_saved_ns"] = time_saved_ns CompileEventLogger.increment_toplevel( "distributed_ephemeral_timeout_us", time_saved_ns // 1000 ) if ( ephemeral_increase := add_ephemeral_timeout_increase_for_distributed(time_saved_ns) ) != 0: cache_info["ephemeral_timeout_increase"] = ephemeral_increase else: if remote_cache: # Count remote cache miss stats CompileEventLogger.increment_toplevel( "inductor_fx_remote_cache_miss_count" ) CompileEventLogger.add_to_set_toplevel( "inductor_fx_remote_cache_miss_keys", key ) log.info("fx graph cache miss for key %s", key) counters["inductor"]["fxgraph_cache_miss"] += 1 cache_info["cache_state"] = "miss" return compiled_graph, cache_info @staticmethod def clear() -> None: """ Clear out the on-disk cache. """ try: shutil.rmtree(FxGraphCache._get_tmp_dir()) except FileNotFoundError: pass @functools.lru_cache(None) def split_aot_inductor_output_path(path: str) -> tuple[str, str]: """Returns the path where the AOT Inductor compiled kernels are stored.""" if path.endswith(".so"): return os.path.split(path) elif path.endswith(".pt2"): return os.path.split(path) else: return path, "" @clear_on_fresh_inductor_cache class CudaKernelParamCache: cache: dict[str, dict[str, Any]] = {} cache_clear = staticmethod(cache.clear) @classmethod def set(cls, key: str, params: dict[str, str], cubin: str, bin_type: str) -> None: _, path = write( cubin, bin_type, hash_type=bin_type, specified_dir=split_aot_inductor_output_path( config.aot_inductor.output_path )[0], ) params[get_cpp_wrapper_cubin_path_name()] = path cls.cache[key] = params @classmethod def get(cls, key: str) -> Optional[dict[str, Any]]: return cls.cache.get(key, None) @classmethod def get_keys(cls) -> KeysView[str]: return cls.cache.keys() class AotCodeCompiler: @classmethod def compile( cls, graph: GraphLowering, wrapper_code: str, kernel_code: str, serialized_extern_kernel_nodes: Optional[str], *, device_type: str, additional_files: list[str], ) -> Union[list[str], str]: """ Returns the .so path, or returns a list of files that were generated if config.aot_inductor.package=True. """ generated_files = additional_files if sys.platform == "win32": raise RuntimeError("AotCodeCompiler not yet supported for inductor") _set_gpu_runtime_env() # cpp_extension consults the env picked_vec_isa = pick_vec_isa() vec_isa_cmd_gen = CppBuilder( name="o", sources="i", BuildOption=CppTorchDeviceOptions( vec_isa=picked_vec_isa, device_type=device_type, aot_mode=graph.aot_mode, ), ) # write function will calc source_code hash, the same source code with different # ISA level should be generate different hash. # So we need get a command_line which contains isa related parameter as a part of hash key. # And then pass the command_line to below write function as extra parameter to # guarantee the source code hash contains ISA difference. cpp_command = repr(vec_isa_cmd_gen.get_command_line()) # Meta internal AOTInductor CPU use_relative_path = ( config.is_fbcode() and device_type == "cpu" and graph.aot_mode ) ( specified_output_path, specified_artifact_name, ) = split_aot_inductor_output_path(config.aot_inductor.output_path) # TODO (benjaminglass1): the CMake packaging path doesn't support linking files # built with different flags. Until that's implemented, append the kernel code # to the wrapper and build everything at max optimization. if config.aot_inductor.package_cpp_only: wrapper_code = "\n".join((wrapper_code, kernel_code)) kernel_code = "" wrapper_key, wrapper_path = write( wrapper_code, "wrapper.cpp", extra=cpp_command, specified_dir=specified_output_path, ) _, kernel_path = write( kernel_code, "kernel.cpp", extra=cpp_command, specified_dir=specified_output_path, ) if config.aot_inductor.package: generated_files.append(wrapper_path) if not config.aot_inductor.package_cpp_only: generated_files.append(kernel_path) output_code_log.info("Wrapper code written to: %s", wrapper_path) output_code_log.info("Kernel code written to: %s", kernel_path) trace_structured( "graph_dump", lambda: { "name": "inductor_aot_wrapper_code", "type": "cpp", "filename": wrapper_path, }, payload_fn=lambda: wrapper_code, ) trace_structured( "graph_dump", lambda: { "name": "inductor_aot_kernel_code", "type": "cpp", "filename": kernel_path, }, payload_fn=lambda: kernel_code, ) # We use a file lock below to protect FS operations. The lock file # is scoped to the 'key', so make sure the consts_s is protected # by the same lock: wrapper_path_operator = Path(wrapper_path) kernel_path_operator = Path(kernel_path) specified_sub_dir = wrapper_path_operator.parent / wrapper_key if not specified_sub_dir.exists(): specified_sub_dir.mkdir(exist_ok=True) cmake_path = str(Path(specified_sub_dir) / "CMakeLists.txt") def _compile_consts(consts: bytes, platform: str) -> str: if platform == "linux": if graph.mutated_buffers & OrderedSet(graph.constants.keys()): # .data section is between .text and .bss. When the size of .data is large, # during the linking, the relocation of .text against .bss may overflow. # Rename it to .ldata so that it won't be in between the .text and .bss section if len(consts) > 2_000_000_000: raise ValueError( "Models with buffer mutation included doesn't support constants greater than 2GB!" ) section_attr = '.ldata, "aw"' else: section_attr = '.lrodata, "a"' symbol_prefix = "" elif platform == "darwin": section_attr = "__DATA,__data" symbol_prefix = "_" else: raise RuntimeError(f"Unsupported platform: {platform}") is_large_consts = len(consts) > 1024 consts_asm = f"\t.section\t{section_attr}\n" consts_asm += f"\t.balign {ALIGN_BYTES}\n" consts_asm += f"\t.globl\t{symbol_prefix}_binary_constants_bin_start\n" consts_asm += f"{symbol_prefix}_binary_constants_bin_start:\n" if not is_large_consts: for c in consts: consts_asm += f"\t.byte {c}\n" # Add one element even if constants are empty # Otherwise assembler will not put them in data section if not consts: consts_asm += "\t.space 1\n" else: consts_asm += "\t.quad 0x1234567899abcdef\n" consts_asm += f"\t.space {len(consts) - 8}\n" consts_asm += f".globl\t{symbol_prefix}_binary_constants_bin_end\n" consts_asm += f"{symbol_prefix}_binary_constants_bin_end:\n" _, consts_s = write( consts_asm, "S", specified_dir=str(specified_sub_dir), ) consts_s = Path(consts_s) object_build_options = CppTorchDeviceOptions( # Intel compiler failed to compile this manully constructed assembly file. # it is ok to use gcc to compile the .S to a .o and linked with Intel comiler . device_type=device_type if device_type != "xpu" else "cpu", aot_mode=graph.aot_mode, compile_only=True, use_relative_path=use_relative_path, ) object_builder = CppBuilder( name=str(consts_s.stem), sources=str(consts_s), output_dir=str(consts_s.parent), BuildOption=object_build_options, ) consts_o = object_builder.get_target_file_path() object_builder.build() if is_large_consts: with open(consts_o, "r+b") as f: f.seek(0) hdr = f.read(1024) # Search for magic number and write the actual data over it start_idx = hdr.find(b"\xef\xcd\xab\x99\x78\x56\x34\x12") assert start_idx != -1 f.seek(start_idx) pos = 0 while pos < len(consts): rc = f.write(consts[pos:]) pos += rc # Remove the .S file to save space os.remove(consts_s) return consts_o from torch.utils._filelock import FileLock lock_dir = get_lock_dir() lock = FileLock( os.path.join(lock_dir, wrapper_key + ".lock"), timeout=LOCK_TIMEOUT ) with lock: if serialized_extern_kernel_nodes: extern_kernel_nodes_json = str( wrapper_path_operator.with_suffix(".json") ) with open(extern_kernel_nodes_json, "w") as f: f.write(serialized_extern_kernel_nodes) if config.aot_inductor.package: generated_files.append(extern_kernel_nodes_json) metadata = config.aot_inductor.metadata metadata["AOTI_DEVICE_KEY"] = device_type # Save user provided metadata meta_json = str( wrapper_path_operator.with_name( f"{wrapper_path_operator.stem}_metadata.json" ) ) for k, v in config.aot_inductor.metadata.items(): assert isinstance(k, str) and isinstance(v, (str)), ( "Metadata must only contain strings" ) with open(meta_json, "w") as f: f.write(json.dumps(config.aot_inductor.metadata)) kernel_meta_json = str( kernel_path_operator.with_name( f"{kernel_path_operator.stem}_metadata.json" ) ) shutil.copy(meta_json, kernel_meta_json) if config.aot_inductor.package: generated_files.append(meta_json) if not config.aot_inductor.package_cpp_only: generated_files.append(kernel_meta_json) output_so = ( config.aot_inductor.output_path if specified_artifact_name else str(wrapper_path_operator.with_suffix(".so")) ) all_cuda = all( graph.get_original_value_of_constant(name).is_cuda for name in graph.constants.keys() if name not in graph.folded_constants ) def _to_bytes(t: torch.Tensor, all_cuda: bool) -> bytes: def _pad_to_alignment(raw_bytes: bytes) -> bytes: padded_bytes = raw_bytes.ljust( (len(raw_bytes) + ALIGN_BYTES - 1) // ALIGN_BYTES * ALIGN_BYTES, b"\x00", ) return padded_bytes # This serializes the tensor's untyped_storage to bytes by accessing # the raw data of the underlying structure. import ctypes if t.numel() == 0: return b"" if t.is_mkldnn: data_ptr = torch.ops.mkldnn.data_ptr(t) nbytes = torch.ops.mkldnn._nbytes(t) else: t_cpu = t.untyped_storage().cpu() data_ptr = t_cpu.data_ptr() nbytes = t_cpu.nbytes() raw_array = ctypes.cast( data_ptr, ctypes.POINTER(ctypes.c_ubyte * nbytes), ) raw_bytes = bytes(raw_array.contents) return raw_bytes if all_cuda else _pad_to_alignment(raw_bytes) if config.aot_inductor.package_constants_in_so: serialized_weights = b"".join( _to_bytes(graph.get_original_value_of_constant(name), all_cuda) for name in graph.constants.keys() if name not in graph.folded_constants ) else: serialized_weights = b"" consts_size = len(serialized_weights) # TODO: Fix mmap weights with cuda use_mmap_weights = not config.is_fbcode() and consts_size > 2_000_000_000 if config.aot_inductor.force_mmap_weights: use_mmap_weights = True compile_command: dict[str, Any] = { "aot_mode": graph.aot_mode, "device_type": device_type, "use_mmap_weights": use_mmap_weights, "use_relative_path": config.is_fbcode(), "vec_isa": picked_vec_isa, } # If we're packaging via CMake, we build the whole code at max optimization. wrapper_build_options = CppTorchDeviceOptions( compile_only=True, min_optimize=not config.aot_inductor.package_cpp_only, **compile_command, ) kernel_build_options = CppTorchDeviceOptions( compile_only=True, **compile_command, ) wrapper_builder = CppBuilder( name=str(wrapper_path_operator.stem), sources=wrapper_path, output_dir=str(wrapper_path_operator.parent), BuildOption=wrapper_build_options, ) wrapper_compile_cmd = wrapper_builder.get_command_line() wrapper_o = wrapper_builder.get_target_file_path() kernel_builder = CppBuilder( name=str(kernel_path_operator.stem), sources=kernel_path, output_dir=str(wrapper_path_operator.parent), BuildOption=kernel_build_options, ) kernel_compile_cmd = kernel_builder.get_command_line() kernel_o = kernel_builder.get_target_file_path() log.debug("aot wrapper compilation command: %s", wrapper_compile_cmd) log.debug("aot kernel compilation command: %s", kernel_compile_cmd) if config.aot_inductor.package_cpp_only: # Not doing the actual compilation here compile_flags = str( wrapper_path_operator.with_name( f"{wrapper_path_operator.stem}_compile_flags.json" ) ) wrapper_build_options.save_flags_to_json(compile_flags) generated_files.append(compile_flags) wrapper_builder.save_compile_cmd_to_cmake(cmake_path) wrapper_builder.save_src_to_cmake(cmake_path, wrapper_path) generated_files.append(cmake_path) else: wrapper_builder.build() kernel_builder.build() if not use_mmap_weights: aot_constants = serialized_weights magic_number = 0 else: magic_number = cast( int, torch.randint(0, torch.iinfo(torch.int64).max, (1,)).item() ) aot_constants = struct.pack("qq", consts_size + 8, magic_number) consts_o = _compile_consts(aot_constants, sys.platform) custom_obj_idx = 0 # Note that custom_objs_config.json file is different from the model_constants_config.json file produced # in package_sigmoid(). The keys in custom_objs_config.json directly correspond to the arg name in extern # nodes json. The key in model_constants_config.json produced by package_sigmoid is the attribute name in the # user model code. qual_name_to_id = {} # Map from constant name to its name in constants folder for custom_obj_idx, (name, constant) in enumerate( graph.torchbind_constants.items() ): assert isinstance(constant, torch._C.ScriptObject) custom_obj_name = f"{CUSTOM_OBJ_FILENAME_PREFIX}{custom_obj_idx}" log.debug("saving script object %s as %s", name, custom_obj_name) qual_name_to_id[name] = custom_obj_name custom_obj_bytes = torch._C._pickle_save(constant) custom_obj_path = os.path.join( wrapper_path_operator.parent, custom_obj_name ) write_atomic(custom_obj_path, custom_obj_bytes, True) generated_files.append(custom_obj_path) constants_config_json = os.path.join( wrapper_path_operator.parent, "custom_objs_config.json" ) with open(constants_config_json, "w") as f: f.write(json.dumps(qual_name_to_id)) generated_files.append(constants_config_json) gpu_codecache: Union[ROCmCodeCache, CUDACodeCache] = ( ROCmCodeCache() if torch.version.hip else CUDACodeCache() ) gpu_kernels_o = [ entry.output_path for entry in gpu_codecache.cache.values() if entry.output_path.endswith(".o") ] gpu_kernels_o = " ".join(gpu_kernels_o) output_name, output_dir = get_name_and_dir_from_output_file_path(output_so) so_build_options = CppTorchDeviceOptions( vec_isa=picked_vec_isa, device_type=device_type, aot_mode=graph.aot_mode, use_relative_path=use_relative_path, ) so_builder = CppBuilder( name=output_name, sources=[wrapper_o, kernel_o, consts_o, gpu_kernels_o] if gpu_kernels_o else [wrapper_o, kernel_o, consts_o], output_dir=output_dir, BuildOption=so_build_options, ) link_cmd = so_builder.get_command_line() output_so = so_builder.get_target_file_path() log.debug("aot linkage command: %s", link_cmd) # Append cmds to the end of codegen-ed wrapper file with open(wrapper_path, "a") as f: f.write("\n") f.write(f"// Compile cmd\n// {wrapper_compile_cmd}\n") f.write(f"// Link cmd\n// {link_cmd}\n") with open(kernel_path, "a") as f: f.write("\n") f.write(f"// Compile cmd\n// {kernel_compile_cmd}\n") f.write(f"// Link cmd\n// {link_cmd}\n") if config.aot_inductor.package_cpp_only: linker_flags = str( wrapper_path_operator.with_name( f"{wrapper_path_operator.stem}_linker_flags.json" ) ) so_build_options.save_flags_to_json(linker_flags) generated_files.append(linker_flags) generated_files.append(_LINKER_SCRIPT) # If we only want to package the cpp, then we need to save the # weights separately into a bin, and we also need to prevent compiling the so if use_mmap_weights: weight_file = str( wrapper_path_operator.with_name( f"{wrapper_path_operator.stem}_serialized_weights.bin" ) ) with open(weight_file, "wb") as f_weights: f_weights.write(serialized_weights) f_weights.write(struct.pack("q", magic_number)) generated_files.append(weight_file) generated_files.append(consts_o) generated_files.append(gpu_kernels_o) so_builder.save_src_to_cmake(cmake_path, consts_o) for gpu_o in gpu_kernels_o.split(): so_builder.save_src_to_cmake(cmake_path, gpu_o) so_builder.save_link_cmd_to_cmake(cmake_path) else: so_builder.build() for o_file in [wrapper_o, kernel_o, consts_o]: # Remove these as they are not needed anymore os.remove(o_file) if use_mmap_weights: import resource page_size_ = resource.getpagesize() page_size = max(16384, page_size_) with open(output_so, "a+b") as f_so: so_size = f_so.tell() # Page align the weights f_so.write(b" " * (page_size - so_size % page_size)) f_so.write(serialized_weights) f_so.write(struct.pack("q", magic_number)) if config.aot_inductor.package: generated_files.append(output_so) if config.aot_inductor.package: # We want to return the directory that contains all the AOTI # generated files, not just the so # return os.path.split(output_so)[0] return generated_files return output_so # Putting this fn in cpp.py (unfortunately) causes a deadlock, which is why it's in codecache.py. # Why? importing from cpp.py invokes codecache.pick_vec_isa(), which takes out a lock. # Cycle goes: # - CppCodeCache.load() # - pick_vec_isa() # - valid_vec_isa_list() # - VecISA.__bool__() <-- takes out a lock # - compile_file() <-- imports cpp_prefix_path from cpp, which causes us to try to take out the same lock. @clear_on_fresh_inductor_cache @functools.lru_cache def cpp_prefix_path() -> str: path = Path(__file__).parent / "codegen/cpp_prefix.h" with path.open() as f: content = f.read() _, filename = write( content, "h", ) return normalize_path_separator(filename) def cpp_prefix() -> str: filename = cpp_prefix_path() if config.is_fbcode(): # We need relative paths, since we bundle up # everything that we compile into a folder for remote compilation. return f'#include "{os.path.basename(filename)}"' else: return f'#include "{filename}"' _libgomp: Optional[CDLL] = None def custom_op_wrapper(op: str, *args: Any) -> Union[list[c_void_p], c_void_p]: # This function will be called from generated cpp wrapper code in the JIT mode. # Because tensors will be passed in as AtenTensorHandle, we need to explicitly convert them. def convert_arg(arg: Any) -> Any: if str(type(arg)) == "": # No easy way to do isinstance check on PyCapsule return torch._C._aoti.alloc_tensor_by_stealing_from_void_ptr(arg) elif isinstance(arg, (list, tuple)): return type(arg)(convert_arg(a) for a in arg) else: return arg converted_args = [convert_arg(arg) for arg in args] assert op.startswith("torch.ops."), ( op + " can not be called through custom_op_wrapper" ) func = None for i, s in enumerate(op.split(".")): if i == 0: func = importlib.import_module(s) func = getattr(func, s) assert callable(func), op + " can not be loaded through custom_op_wrapper" # convert any kwarg-only arguments to kwargs kwargs = dict() for func_arg, conv_arg in zip(func._schema.arguments, converted_args): if func_arg.kwarg_only: kwargs[func_arg.name] = conv_arg if kwargs: del converted_args[-len(kwargs) :] result = func(*converted_args, **kwargs) if isinstance(result, (list, tuple)): # unsafe_alloc_void_ptrs_from_tensors expects result contains tensor only result = [torch.tensor([]) if r is None else r for r in result] for i, r in enumerate(result): assert isinstance(r, torch.Tensor), op + " returns a list of non-tensors" return torch._C._aoti.unsafe_alloc_void_ptrs_from_tensors(result) # type: ignore[arg-type] else: assert isinstance(result, torch.Tensor), op + " returns a non-tensor" return torch._C._aoti.unsafe_alloc_void_ptr_from_tensor(result) @clear_on_fresh_inductor_cache class CppCodeCache: cache: dict[str, Callable[[], Union[CDLL, ModuleType]]] = {} cache_clear = staticmethod(cache.clear) cpp_compile_command_flags: dict[str, Any] = {} @staticmethod def _load_library_inner(path: str, key: str) -> Union[CDLL, ModuleType]: return cdll.LoadLibrary(path) @classmethod def _load_library(cls, path: str, key: str) -> Union[CDLL, ModuleType]: try: result = cls._load_library_inner(path, key) result.key = key # type: ignore[union-attr] return result except (ImportError, OSError) as e: if "gomp" in str(e) and os.path.exists("/usr/lib64/libgomp.so.1"): # hacky workaround for fbcode/buck global _libgomp _libgomp = cdll.LoadLibrary("/usr/lib64/libgomp.so.1") result = cls._load_library_inner(path, key) result.key = key # type: ignore[union-attr] return result if "failed to map segment from shared object" in str(e): raise OSError( f"{e}. The most common reason this may occur is if the {tempfile.gettempdir()} folder " "is mounted with noexec (e.g., by default Docker mounts tmp file systems " f"as noexec). Please remount {tempfile.gettempdir()} with exec enabled, or set another " "temporary directory with TORCHINDUCTOR_CACHE_DIR environment variable." ) from e raise @classmethod def load_async( cls, source_code: str, device_type: str = "cpu", submit_fn: Any = None, extra_flags: Sequence[str] = (), ) -> Any: compile_command = { **cls.cpp_compile_command_flags, "device_type": device_type, "vec_isa": pick_vec_isa(), "extra_flags": extra_flags, } _set_gpu_runtime_env() # cpp_extension consults the env command_gen = CppBuilder( name="o", sources="i", BuildOption=CppTorchDeviceOptions(**compile_command) ) # write function will calc source_code hash, the same source code with different # ISA level should be generate different hash. # So we need get a command_line which contains isa related parameter as a part of hash key. # And then pass the command_line to below write function as extra parameter to # guarantee the source code hash contains ISA difference. vec_isa_cmd = repr(command_gen.get_command_line()) key, input_path = write(source_code, "cpp", extra=vec_isa_cmd) if key not in cls.cache: from torch.utils._filelock import FileLock lock_path = os.path.join(get_lock_dir(), key + ".lock") output_name, output_dir = get_name_and_dir_from_output_file_path(input_path) future: Optional[Future[Any]] = None lib = None cpp_build_option = CppTorchDeviceOptions( **compile_command, use_relative_path=(config.is_fbcode() and device_type == "cpu"), ) cpp_builder = CppBuilder( name=output_name, sources=input_path, output_dir=output_dir, BuildOption=cpp_build_option, ) worker_fn = functools.partial( _worker_compile_cpp, lock_path, cpp_builder, ) binary_path = normalize_path_separator(cpp_builder.get_target_file_path()) def load_fn() -> Any: nonlocal lib if lib is None: if future is not None: future.result() result = worker_fn() assert result is None lib = cls._load_library(binary_path, key) assert lib is not None return lib if submit_fn is not None: with FileLock(lock_path, timeout=LOCK_TIMEOUT): if not os.path.exists(binary_path): future = submit_fn(worker_fn) cls.cache[key] = load_fn return cls.cache[key] @classmethod def load(cls, source_code: str, device_type: str = "cpu") -> Any: return cls.load_async(source_code, device_type)() def _worker_compile_cpp( lock_path: str, cpp_builder: CppBuilder, ) -> None: from torch.utils._filelock import FileLock with FileLock(lock_path, timeout=LOCK_TIMEOUT): if not os.path.exists(cpp_builder.get_target_file_path()): cpp_builder.build() # Customized Python binding for cpp kernels @clear_on_fresh_inductor_cache class CppPythonBindingsCodeCache(CppCodeCache): cache: dict[str, Callable[[], Union[CDLL, ModuleType]]] = {} cache_clear = staticmethod(cache.clear) cpp_compile_command_flags = { # kernels have no dependency on libtorch "include_pytorch": False, "shared": True, } entry_function = "kernel" call_entry_function = "kernel({}); Py_RETURN_NONE;" extra_parse_arg = "" suffix_template = textwrap.dedent( """ // Python bindings to call {entry_func}(): #define PY_SSIZE_T_CLEAN #include #include #include #ifndef _MSC_VER #if __cplusplus < 202002L // C++20 (earlier) code // https://en.cppreference.com/w/cpp/language/attributes/likely #define likely(x) __builtin_expect(!!(x), 1) #define unlikely(x) __builtin_expect(!!(x), 0) #endif #else #define likely(x) (x) #define unlikely(x) (x) #endif // This is defined in guards.cpp so we don't need to import PyTorch headers that are slooow. // We manually link it below to workaround issues with fbcode build. static void* (*_torchinductor_pyobject_tensor_data_ptr)(PyObject* obj); template static inline T parse_arg(PyObject* args, size_t n) {{ static_assert(std::is_pointer_v, "arg type must be pointer or long"); return static_cast(_torchinductor_pyobject_tensor_data_ptr(PyTuple_GET_ITEM(args, n))); }} template <> inline int64_t parse_arg(PyObject* args, size_t n) {{ auto result = PyLong_AsSsize_t(PyTuple_GET_ITEM(args, n)); if(unlikely(result == -1 && PyErr_Occurred())) throw std::runtime_error("expected int arg"); return result; }} template <> inline uintptr_t parse_arg(PyObject* args, size_t n) {{ auto result = PyLong_AsVoidPtr(PyTuple_GET_ITEM(args, n)); if(unlikely(result == reinterpret_cast(-1) && PyErr_Occurred())) throw std::runtime_error("expected int arg"); return reinterpret_cast(result); }} {extra_parse_arg} static PyObject* {entry_func}_py(PyObject* self, PyObject* args) {{ try {{ if(unlikely(!PyTuple_CheckExact(args))) throw std::runtime_error("tuple args required"); if(unlikely(PyTuple_GET_SIZE(args) != {arg_len})) throw std::runtime_error("requires {arg_len} args"); {call_entry_func} }} catch(std::exception const& e) {{ PyErr_SetString(PyExc_RuntimeError, e.what()); return nullptr; }} catch(...) {{ PyErr_SetString(PyExc_RuntimeError, "unhandled error"); return nullptr; }} }} static PyMethodDef py_methods[] = {{ {{"{entry_func}", {entry_func}_py, METH_VARARGS, ""}}, {{NULL, NULL, 0, NULL}}}}; static struct PyModuleDef py_module = {{PyModuleDef_HEAD_INIT, "{entry_func}", NULL, -1, py_methods}}; PyMODINIT_FUNC PyInit_{entry_func}(void) {{ const char* str_addr = std::getenv("_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR"); if(!str_addr) {{ PyErr_SetString(PyExc_RuntimeError, "_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR must be set"); return nullptr; }} std::istringstream iss(str_addr); uintptr_t addr = 0; iss >> addr; _torchinductor_pyobject_tensor_data_ptr = reinterpret_cast(addr); PyObject* module = PyModule_Create(&py_module); if (module == NULL) {{ return NULL; }} #ifdef Py_GIL_DISABLED PyUnstable_Module_SetGIL(mod, Py_MOD_GIL_NOT_USED); #endif return module; }} """ ) @classmethod def _load_library_inner(cls, path: str, key: str) -> ModuleType: os.environ["_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR"] = str( torch._C._dynamo.guards._torchinductor_pyobject_tensor_data_ptr # type: ignore[attr-defined] ) module_name = f"{key}.{cls.entry_function}" try: return sys.modules[module_name] except KeyError: pass spec = importlib.util.spec_from_file_location(module_name, path) assert spec is not None module = importlib.util.module_from_spec(spec) sys.modules[module_name] = module spec.loader.exec_module(module) # type: ignore[union-attr] return module @classmethod def load_pybinding_async( cls, argtypes: list[str], source_code: str, device_type: str = "cpu", num_outputs: int = -1, submit_fn: Any = None, extra_flags: Sequence[str] = (), ) -> Any: """ Wrap a C++ function in fast Python bindings. Args: argtypes: The types of args to ENTRY_FUNCTION(), e.g. ["float*", "long"] source_code: C++ source code containing a ENTRY_FUNCTION() function Returns: A python version of ENTRY_FUNCTION() """ parseargs = ", ".join( f"parse_arg<{argtype.replace('const ', '')}>(args, {n})" for n, argtype in enumerate(argtypes) ) suffix = cls.suffix_template.format( arg_len=len(argtypes), call_entry_func=cls.call_entry_function.format(parseargs), entry_func=cls.entry_function, extra_parse_arg=cls.extra_parse_arg.format(array_len=num_outputs), ) get_result = cls.load_async( source_code + suffix, device_type, submit_fn=submit_fn, extra_flags=extra_flags, ) result = None def future() -> Any: nonlocal result if result is None: result = get_result() assert isinstance(result, ModuleType) return getattr(result, cls.entry_function) return future @classmethod def load_pybinding(cls, *args: Any, **kwargs: Any) -> Any: return cls.load_pybinding_async(*args, **kwargs)() @clear_on_fresh_inductor_cache class CppWrapperCodeCache(CppPythonBindingsCodeCache): cache: dict[str, Callable[[], Union[CDLL, ModuleType]]] = {} cache_clear = staticmethod(cache.clear) cpp_compile_command_flags = { "include_pytorch": True, "shared": True, } entry_function = "inductor_entry_cpp" call_entry_function = "return inductor_entry_cpp({});" extra_parse_arg = textwrap.dedent( """ #include static inline std::vector unpack_tensor_handle_list(PyObject* pyvec) {{ std::vector result; size_t result_len = PyList_GET_SIZE(pyvec); result.reserve(result_len); for (size_t i = 0; i < result_len; i++) {{ // AtenTensorHandle is essentially a pointer void* elem = PyCapsule_GetPointer(PyList_GET_ITEM(pyvec, i), NULL); result.push_back(reinterpret_cast(elem)); }} return result; }} static inline PyObject* pack_tensor_handle_list(const std::array& arr) {{ PyObject* result = PyList_New({array_len}); for (size_t i = 0; i < {array_len}; i++) {{ PyObject *elem = arr[i] == nullptr ? Py_None // Store AtenTensorHandle as PyCapsulate : PyCapsule_New(reinterpret_cast(arr[i]), NULL, NULL); PyList_SET_ITEM(result, i, elem); }} return result; }} template <> inline std::vector parse_arg>(PyObject* args, size_t n) {{ return unpack_tensor_handle_list(PyTuple_GET_ITEM(args, n)); }} PyObject* inductor_entry_cpp(std::vector&& input_handles) {{ // For outputs, we only allocate an array to hold returned tensor handles, // not the actual output tensor storage. std::array output_handles{{}}; try {{ inductor_entry_impl(input_handles.data(), output_handles.data()); if (PyErr_Occurred()) {{ return nullptr; }} return pack_tensor_handle_list(output_handles); }} catch(std::exception const& e) {{ PyErr_SetString(PyExc_RuntimeError, e.what()); return nullptr; }} catch(...) {{ PyErr_SetString(PyExc_RuntimeError, "unhandled error"); return nullptr; }} }} """ ) @clear_on_fresh_inductor_cache class HalideCodeCache(CppPythonBindingsCodeCache): cache: dict[str, Callable[[], Union[ModuleType, CDLL]]] = {} cache_clear = staticmethod(cache.clear) _standalone_runtime_path: Optional[str] = None prefix = textwrap.dedent( """ #include "{halideruntime_h}" #include "{headerfile}" #include #include namespace c10 {{ inline long div_floor_integer(long a, long b) {{ if ((a<0) != (b<0)) {{ const auto quot = a / b; const auto rem = a % b; return rem ? quot - 1 : quot; }} return a / b; }} }} """ ) glue_template_cpp = prefix + textwrap.dedent( """ void kernel({argdefs}) {{ {buffers} int err = halide_kernel({buffer_names}); if(err != 0) throw std::runtime_error("halide_kernel failed"); }} """ ) glue_template_cuda = prefix + textwrap.dedent( """ #include static const halide_device_interface_t* cuda_interface = halide_cuda_device_interface(); void kernel({argdefs}, uintptr_t stream) {{ {buffers} int err = halide_kernel(reinterpret_cast(stream), {buffer_names}); if(err != 0) throw std::runtime_error("halide_kernel failed"); }} """ ) standalone_runtime_cuda_init = textwrap.dedent( """ #include "{}" #include static int acquire_context(void* user_context, void** cuda_context_out, bool create) {{ return cuCtxGetCurrent(reinterpret_cast(cuda_context_out)); }} static int release_context(void* user_context) {{ return 0; }} static int get_stream(void* user_context, void* cuda_context, void** stream_out) {{ *stream_out = user_context; return 0; }} static int register_halide_hooks() {{ halide_set_cuda_acquire_context(&acquire_context); halide_set_cuda_release_context(&release_context); halide_set_cuda_get_stream(&get_stream); return 0; }} int inductor_register_halide_hooks_result = register_halide_hooks(); """ ) @classmethod def _codegen_buffer(cls, name: str, arg: HalideInputSpec, cuda: bool) -> list[str]: assert arg.shape is not None assert arg.stride is not None and len(arg.shape) == len(arg.stride) assert arg.offset is not None data_ptr = f"{arg.alias_of or arg.name} + {arg.offset}" if cuda: device = f"reinterpret_cast({data_ptr})" device_interface = "cuda_interface" host = "nullptr" flags = "halide_buffer_flag_device_dirty" else: device = "0" device_interface = "nullptr" host = f"reinterpret_cast({data_ptr})" flags = "halide_buffer_flag_host_dirty" dims = [] for size, stride in zip(arg.shape, arg.stride): dims.append(f"halide_dimension_t(0, {size}, {stride})") return [ f"halide_buffer_t {name};", f"halide_dimension_t {name}_dims[] = {{{', '.join(dims)}}};", f"{name}.device = {device};", f"{name}.device_interface = {device_interface};", f"{name}.host = {host};", f"{name}.flags = {flags};", f"{name}.type = {arg.halide_type()};", f"{name}.dimensions = {len(dims)};", f"{name}.dim = {name}_dims;", f"{name}.padding = nullptr;", ] @classmethod def _codegen_glue(cls, meta: HalideMeta, headerfile: object) -> str: is_cuda = meta.is_cuda() assert is_cuda is ("user_context" in meta.target) assert "no_runtime" in meta.target buffers = [] buffer_names = [] for i, arg in enumerate(meta.argtypes): if arg.is_buffer(): buffer_names.append(f"&hl_buf_{i}") buffers.extend(cls._codegen_buffer(f"hl_buf_{i}", arg, is_cuda)) else: assert "*" not in arg.ctype buffer_names.append(arg.name) buffers = "\n".join([f" {line}" for line in buffers]).lstrip() glue_template = cls.glue_template_cuda if is_cuda else cls.glue_template_cpp glue_code = glue_template.format( halideruntime_h=cls.find_header( "HalideRuntimeCuda.h" if is_cuda else "HalideRuntime.h" ), headerfile=headerfile, argdefs=", ".join( f"{a.bindings_type()} {a.name}" for a in meta.argtypes if a.alias_of is None ), buffers=buffers, buffer_names=", ".join(buffer_names), ) return glue_code @classmethod @functools.lru_cache(None) def config_hash(cls) -> str: command_gen = CppBuilder( name="O", sources="I", BuildOption=CppOptions(), ) command_line = command_gen.get_command_line() return sha256_hash( "\n".join( [ cls.glue_template_cpp, cls.glue_template_cuda, cls.standalone_runtime_cuda_init, command_line, ] ).encode("utf-8") ) @staticmethod def _search_for_file(suffix: str, errmsg: str) -> str: spec = importlib.machinery.PathFinder.find_spec("halide") if spec is None or not spec.submodule_search_locations: raise RuntimeError("halide python bindings not installed") try: search = spec.submodule_search_locations[0] for file in os.listdir(search): if file.endswith(".so"): try: out = subprocess.check_output( ["ldd", os.path.join(search, file)] ) except subprocess.SubprocessError: continue m = re.search(r"(/.*)/libHalide.so", out.decode("utf-8")) if m: path = os.path.join(os.path.abspath(m.group(1)), suffix) if os.path.exists(path): return os.path.abspath(path) except Exception as e: raise RuntimeError(errmsg) from e raise RuntimeError(errmsg) @staticmethod @functools.lru_cache(None) def find_libautoschedule(name: str) -> str: sofile = f"libautoschedule_{name.lower()}.so" if "HALIDE_LIB" in os.environ: path = os.path.join(os.environ["HALIDE_LIB"], sofile) if os.path.exists(path): return path errmsg = ( f"Can't find {sofile}, set env HALIDE_LIB to the directory containing it" ) return HalideCodeCache._search_for_file(sofile, errmsg) @staticmethod @functools.lru_cache(None) def find_header(name: str) -> str: if "HALIDE_INCLUDE" in os.environ: path = os.path.join(os.environ["HALIDE_INCLUDE"], name) if os.path.exists(path): return path if "HALIDE_LIB" in os.environ: path = os.path.abspath( os.path.join(os.environ["HALIDE_LIB"], f"../include/{name}") ) if os.path.exists(path): return path errmsg = ( f"Can't find {name}, set env HALIDE_INCLUDE to the directory containing it" ) return HalideCodeCache._search_for_file(f"../include/{name}", errmsg) @classmethod def generate_halide_async( cls, meta: HalideMeta, source_code: str, submit_fn: Any = None ) -> Callable[[], Any]: dirpath = Path( get_path( code_hash( source_code, extra=repr((cls.config_hash(), meta)), ), "halide", )[2] ) os.makedirs(dirpath, exist_ok=True) wait_for_compile = None genfile = str(dirpath / "generate_kernel.py") libfile = str(dirpath / "halide_kernel.a") headerfile = str(dirpath / "halide_kernel.h") donefile = str(dirpath / "done") lockfile = str(dirpath / "lock") need_compile = not os.path.exists(donefile) jobs: list[Any] = [] if need_compile: write_atomic(genfile, source_code) cmd = [ sys.executable, genfile, "-g", "kernel", "-o", f"{dirpath}", "-f", "halide_kernel", "-e", "static_library,h,schedule", ] if meta.scheduler: cmd.extend(["-p", cls.find_libautoschedule(meta.scheduler)]) cmd.extend(meta.args()) jobs.append(functools.partial(subprocess.check_call, cmd)) binding_types = [ arg.bindings_type() for arg in meta.argtypes if arg.alias_of is None ] if meta.is_cuda(): binding_types.append("uintptr_t") # stream bindings_future = cls.load_pybinding_async( binding_types, cls._codegen_glue(meta, headerfile), extra_flags=(libfile, cls.build_standalone_runtime()), submit_fn=jobs.append if need_compile else None, device_type="cuda" if meta.is_cuda() else "cpu", ) if need_compile: jobs.append(functools.partial(touch, donefile)) task = functools.partial(_worker_task_halide, lockfile, jobs) if submit_fn: wait_for_compile = submit_fn(task).result else: task() def load() -> Callable[[], Any]: if wait_for_compile: wait_for_compile() return bindings_future() return load @classmethod def generate_halide(cls, *args: Any, **kwargs: Any) -> Callable[[], Any]: return cls.generate_halide_async(*args, **kwargs)() @classmethod def build_standalone_runtime(cls) -> str: if cls._standalone_runtime_path and os.path.exists( cls._standalone_runtime_path ): return cls._standalone_runtime_path device_type = "cuda" if torch.cuda.is_available() else "cpu" libname = "libStandaloneHalideRuntime.so" target = "host-cuda" if device_type == "cuda" else "host" if cls._standalone_runtime_path: assert not os.path.exists(cls._standalone_runtime_path) # We hit this case in unittests when we run with fresh_inductor_cache() # Generating a fresh runtime over and over causes errors because we initialize # cuda hundreds of times in the same process and run out of file descriptors. # Workaround by jail breaking the current fresh_inductor_cache(). base = default_cache_dir() else: base = cache_dir() dirpath = Path(base) / f"halide-runtime-{target}-{cls.config_hash()}" os.makedirs(dirpath, exist_ok=True) donefile = str(dirpath / "done") lockfile = str(dirpath / "lock") hookfile = str(dirpath / "hooks.cpp") afile = str(dirpath / "standalone_halide_runtime.a") sofile = str(dirpath / libname) if not os.path.exists(donefile): import halide as hl # type: ignore[import-untyped,import-not-found] from torch.utils._filelock import FileLock with FileLock(lockfile, LOCK_TIMEOUT): if not os.path.exists(donefile): with open(hookfile, "w") as f: if device_type == "cuda": f.write( cls.standalone_runtime_cuda_init.format( cls.find_header("HalideRuntimeCuda.h") ) ) hl.compile_standalone_runtime(afile, hl.Target(target)) name, output_dir = get_name_and_dir_from_output_file_path(sofile) halide_cmd_gen = CppBuilder( name=name, sources=[hookfile, afile], output_dir=output_dir, BuildOption=CppTorchDeviceOptions( device_type=device_type, ), ) subprocess.check_call( shlex.split(halide_cmd_gen.get_command_line()) ) touch(donefile) assert os.path.exists(sofile) cls._standalone_runtime_path = sofile return sofile def _worker_task_halide(lockfile: str, jobs: list[partial[Any]]) -> None: from torch.utils._filelock import FileLock try: with FileLock(lockfile, LOCK_TIMEOUT): for job in jobs: job() except subprocess.SubprocessError as e: if os.environ.get("HALIDE_REPRO") == "1": python, script, *cmd = getattr(e, "cmd", ("", "", "")) if os.path.basename(python).startswith("python"): code = open(script).read() main = " hl.main()" assert code.count(main) == 1 class Out: def __repr__(self) -> str: return "out" cmd[cmd.index("-o") + 1] = Out() # type: ignore[call-overload] repl = textwrap.indent( textwrap.dedent( f"""\ import sys, tempfile with tempfile.TemporaryDirectory() as out: sys.argv = {["repro.py", *cmd]!r} hl.main() """ ), " ", ) code = code.replace(main, repl) with open("repro.py", "w") as fd: fd.write(code.lstrip()) raise RuntimeError(f"wrote repro.py: {e}") from e raise def touch(filename: str) -> None: open(filename, "a").close() @clear_on_fresh_inductor_cache class PyCodeCache: # Track the loaded modules so we can remove the on-disk artifacts when # clearing the cache. Note also that we may load the same path more # than once, but attach different attributes, i.e., due to different # constant values. modules: list[ModuleType] = [] linemaps: dict[str, list[tuple[Any, ...]]] = {} @classmethod def write(cls, source_code: str, extra: str = "") -> tuple[str, str]: return write(source_code, "py", extra=extra) @classmethod def load( cls, source_code: str, extra: str = "", linemap: Optional[list[tuple[int, str]]] = None, attrs: Optional[dict[str, Any]] = None, ) -> ModuleType: key, path = write(source_code, "py", extra=extra) return cls.load_by_key_path(key, path, linemap, attrs) @classmethod def load_by_key_path( cls, key: str, path: str, linemap: Optional[list[tuple[int, str]]] = None, attrs: Optional[dict[str, Any]] = None, ) -> ModuleType: if linemap is None: linemap = [] mod = _reload_python_module(key, path) # unzip into separate lines/nodes lists cls.linemaps[path] = list(zip(*linemap)) if attrs is not None: for k, v in attrs.items(): setattr(mod, k, v) if not (linemap or attrs): mod._reload_in_subproc = functools.partial( # type: ignore[attr-defined] _reload_python_module_in_subproc, key, path ) cls.modules.append(mod) return mod @classmethod def cache_clear(cls, purge: bool = False) -> None: """ Clear the in-memory module cache. If purge=True, also delete all the corresponding on-disk source files. """ if purge: for mod in cls.modules: try: assert mod.__file__ os.remove(mod.__file__) except FileNotFoundError: pass cls.modules.clear() @classmethod @functools.lru_cache(None) def stack_frames_for_code( cls, path: str, lineno: int ) -> Optional[list[dict[str, Any]]]: if path not in cls.linemaps: return None # [(starting_line, ), ...] lines, nodes = cls.linemaps[path] p = bisect_right(lines, lineno) if p == 0: return None entry = nodes[p - 1] if not entry: return None def parse_stack_trace(stack_trace: str) -> list[dict[str, Any]]: # ideally fx stores stack traces as data rather than a string # but this is not along a performance critical path regex = r'File "(.+)", line (\d+), in (.+)\n' matches = re.findall(regex, stack_trace) return [ {"filename": f, "line": int(l), "name": n} for f, l, n in reversed(matches) ] return parse_stack_trace(entry) def _load_triton_kernel_from_source( kernel_name: str, source_code: str ) -> CachingAutotuner: return getattr(PyCodeCache.load(source_code), kernel_name) def _cuda_compiler() -> Optional[str]: if cuda_env.nvcc_exist(config.cuda.cuda_cxx): return config.cuda.cuda_cxx if config.is_fbcode(): return os.path.join(build_paths.sdk_home, "bin", "nvcc") if cuda_env.nvcc_exist(os.getenv("CUDACXX")): return os.getenv("CUDACXX", "") if cuda_env.nvcc_exist(os.getenv("CUDA_HOME")): return os.path.realpath(os.path.join(os.getenv("CUDA_HOME", ""), "bin/nvcc")) return "nvcc" def _cutlass_include_paths() -> list[str]: if config.is_fbcode(): from libfb.py import parutil cutlass_path = parutil.get_dir_path("cutlass-3-headers") else: cutlass_path = config.cuda.cutlass_dir return [ # Use realpath to get canonical absolute paths, in order not to mess up cache keys os.path.realpath(os.path.join(cutlass_path, "include")), os.path.realpath(os.path.join(cutlass_path, "tools/library/include")), os.path.realpath(os.path.join(cutlass_path, "tools/library/src")), os.path.realpath(os.path.join(cutlass_path, "tools/util/include")), ] def _cuda_lib_options() -> list[str]: _set_gpu_runtime_env() # cpp_extension consults the env from torch.utils import cpp_extension lpaths = cpp_extension.library_paths(device_type="cuda") extra_ldflags: list[str] = [] if is_linux(): _transform_cuda_paths(lpaths) for path in lpaths: # -rpath ensures the DLL can find its dependencies when loaded, even # if the library path is non-standard. extra_ldflags.extend([f"-L{path}", "-Xlinker", f"-rpath={path}"]) extra_ldflags.append("-lcuda") extra_ldflags.append("-lcudart") else: raise NotImplementedError( "Unsupported env, failed to find cuda libs! Currently only Linux is supported." ) return extra_ldflags def _nvcc_host_compiler_options() -> list[str]: return [ "-fPIC", "-fno-strict-aliasing", "-fvisibility=hidden", "-Wconversion", ] def _nvcc_compiler_options() -> list[str]: arch = cuda_env.get_cuda_arch() if arch == "90": # Required by cutlass compilation. arch = "90a" code = [f"sm_{arch}", f"compute_{arch}"] if config.cuda.enable_cuda_lto: code += [f"lto_{arch}"] options = [ "-t=0", "-DCUTLASS_ENABLE_TENSOR_CORE_MMA=1", "-DCUTLASS_ENABLE_SM90_EXTENDED_MMA_SHAPES=1", "-DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED", "-w", f"-gencode=arch=compute_{arch},code=[{','.join(code)}]", config.cuda.compile_opt_level, "-std=c++17", "--expt-relaxed-constexpr", "-DNDEBUG", ] if config.is_fbcode(): options.extend(["-ccbin", os.path.dirname(build_paths.gcc)]) if config.cuda.enable_debug_info: options.extend(["-lineinfo", "-g", "-DCUTLASS_DEBUG_TRACE_LEVEL=1"]) if config.cuda.enable_ptxas_info: options.extend( [ "--keep", # Keep the intermediate files for debugging (including ptx, sass, cubin etc.) "--ptxas-options=--warn-on-local-memory-usage", # warn us if local memory is used in CUDA Kernels "--ptxas-options=--warn-on-spills", # warn us if register spilling happens in CUDA Kernels "--resource-usage", # Report on CUDA resource usage (shared mem, registers etc.) "--source-in-ptx", ] ) # Annotate the ptx file with source information if config.cuda.use_fast_math: options.extend( [ "--use_fast_math", "-DCUTLASS_USE_TANH_FOR_SIGMOID=1", ] ) return options def cuda_compile_command( src_files: list[str], dst_file: str, dst_file_ext: str, extra_args: Optional[list[str]] = None, ) -> str: if extra_args is None: extra_args = [] include_paths = _cutlass_include_paths() cuda_lib_options = _cuda_lib_options() nvcc_host_compiler_options = _nvcc_host_compiler_options() nvcc_compiler_options = _nvcc_compiler_options() options = ( nvcc_compiler_options + extra_args + [ f"-Xcompiler {opt}" if "=" in opt else f"-Xcompiler={opt}" for opt in nvcc_host_compiler_options ] + ["-I" + path for path in include_paths] + cuda_lib_options ) src_file = " ".join(src_files) res = "" if dst_file_ext == "o": res = f"{_cuda_compiler()} {' '.join(options)} -c -o {dst_file} {src_file}" elif dst_file_ext == "so": options.append("-shared") res = f"{_cuda_compiler()} {' '.join(options)} -o {dst_file} {src_file}" elif dst_file_ext == "exe": res = f"{_cuda_compiler()} {' '.join(options)} -o {dst_file} {src_file}" else: raise NotImplementedError(f"Unsupported output file suffix {dst_file_ext}!") log.debug("CUDA command: %s", res) return res class DLLWrapper: """A wrapper for a dynamic library.""" def __init__( self, lib_path: str, ) -> None: self.lib_path = lib_path self.is_open = False self.DLL = cdll.LoadLibrary(lib_path) self.is_open = True def close(self) -> None: if self.is_open: self._dlclose() self.is_open = False def _dlclose(self) -> None: f_dlclose = None if is_linux(): syms = CDLL(None) if not hasattr(syms, "dlclose"): # Apline Linux syms = CDLL("libc.so") if hasattr(syms, "dlclose"): f_dlclose = syms.dlclose elif is_windows(): import ctypes kernel32 = ctypes.CDLL("kernel32", use_last_error=True) f_dlclose = kernel32.FreeLibrary else: raise NotImplementedError("Unsupported env, failed to do dlclose!") if f_dlclose is not None: if is_linux(): f_dlclose.argtypes = [c_void_p] f_dlclose(self.DLL._handle) elif is_windows(): import ctypes from ctypes import wintypes f_dlclose.argtypes = [wintypes.HMODULE] f_dlclose(self.DLL._handle) else: log.warning( "dll unloading function was not found, library may not be unloaded properly!" ) def __getattr__(self, name: str) -> Callable[..., None]: if not self.is_open: raise RuntimeError(f"Cannot use closed DLL library: {self.lib_path}") method = getattr(self.DLL, name) def _wrapped_func(*args: Any) -> None: err = method(*args) if err: raise RuntimeError(f"Error in function: {method.__name__}") return _wrapped_func def __enter__(self) -> Self: return self def __exit__(self, *args: Any) -> None: self.close() def __del__(self) -> None: self.close() @clear_on_fresh_inductor_cache class CUDACodeCache: @dataclasses.dataclass class CacheEntry: input_path: str output_path: str cache: dict[str, CacheEntry] = {} cache_clear = staticmethod(cache.clear) _SOURCE_CODE_SUFFIX = "cu" @classmethod def write(cls, source_code: str, dst_file_ext: str) -> tuple[str, str]: """ Writes source code into a file with dst_file_ext as the file extension. Returns the hash key of source code, and the path to the file. """ cuda_command = repr( cuda_compile_command(["dummy_input"], "dummy_output", dst_file_ext) ) key, input_path = write( source_code, cls._SOURCE_CODE_SUFFIX, extra=cuda_command ) return key, input_path @classmethod def compile( cls, source_code: str, dst_file_ext: str, extra_args: Optional[list[str]] = None ) -> tuple[str, str, str]: """ Compiles CUDA source_code into a file with dst_file_ext extension. Returns a tuple of dst_file_path, hash_key, source_code_path """ key, input_path = cls.write(source_code, dst_file_ext) if key not in cls.cache: from torch.utils._filelock import FileLock lock_dir = get_lock_dir() lock = FileLock(os.path.join(lock_dir, key + ".lock"), timeout=LOCK_TIMEOUT) with lock: output_path = input_path[: -len(cls._SOURCE_CODE_SUFFIX)] + dst_file_ext if not os.path.exists(output_path): cmd = cuda_compile_command( [input_path], output_path, dst_file_ext, extra_args ) with open(input_path, "a") as f: f.write("\n") f.write(f"// CUDA Compile cmd\n// {cmd}\n") start_time = time() log.debug("CUDA Compilation: %s", cmd) cmd_parts = cmd.split(" ") try: subprocess.check_output( cmd_parts, stderr=subprocess.STDOUT, env=os.environ ) except subprocess.CalledProcessError as error: raise exc.CUDACompileError(cmd_parts, error.output) from error end_time = time() log_duration_msg = f"CUDA Compilation took {end_time - start_time} seconds. Compile command: {cmd}" log.info(log_duration_msg) else: log.debug( "CUDA Compilation skipped: %s since output already exists", input_path, ) cls.cache[key] = CUDACodeCache.CacheEntry(input_path, output_path) return (cls.cache[key].output_path, key, input_path) @classmethod def load(cls, source_code: str, dst_file_ext: str) -> tuple[DLLWrapper, str, str]: """ Compiles source code and loads the generated .so file. Returns a tuple of DLLWrapper, hash_key, source_code_path """ if dst_file_ext != "so": raise RuntimeError( f"Only support loading a .so file for now. " f"Requested file extension: {dst_file_ext}. Source code: {source_code}" ) dst_file_path, hash_key, source_code_path = cls.compile( source_code, dst_file_ext ) return (DLLWrapper(dst_file_path), hash_key, source_code_path) @clear_on_fresh_inductor_cache class ROCmCodeCache: @dataclasses.dataclass class CacheEntry: input_path: str output_path: str cache: dict[str, CacheEntry] = {} cache_clear = staticmethod(cache.clear) _SOURCE_CODE_SUFFIX = "cpp" _logged_compiler_version = False @classmethod def write(cls, source_code: str, dst_file_ext: str) -> tuple[str, str]: """ Writes source code into a file with dst_file_ext as the file extension. Returns the hash key of source code, and the path to the file. """ cuda_command = repr( rocm_compile_command(["dummy_input"], "dummy_output", dst_file_ext) ) key, input_path = write( source_code, cls._SOURCE_CODE_SUFFIX, extra=cuda_command ) return key, input_path @classmethod def compile( cls, source_code: str, dst_file_ext: str, extra_args: Optional[list[str]] = None ) -> tuple[str, str, str]: """ Compiles source_code into a file with dst_file_ext extension, using the compile command specific for the ROCm platform. Returns a tuple of dst_file_path, hash_key, source_code_path """ if not cls._logged_compiler_version: cls._logged_compiler_version = True log.debug(get_compiler_version_info(str(rocm_compiler()))) key, input_path = cls.write(source_code, dst_file_ext) if key not in cls.cache: from torch.utils._filelock import FileLock lock_dir = get_lock_dir() lock = FileLock(os.path.join(lock_dir, key + ".lock"), timeout=LOCK_TIMEOUT) with lock: output_path = input_path[: -len(cls._SOURCE_CODE_SUFFIX)] + dst_file_ext if not os.path.exists(output_path): cmd = rocm_compile_command( [input_path], output_path, dst_file_ext, extra_args ) start_time = time() cmd_parts = cmd.split(" ") try: output = subprocess.check_output( cmd_parts, stderr=subprocess.STDOUT, text=True, env=os.environ, ) log.debug("Compilation output: %s", output) except subprocess.CalledProcessError as error: raise exc.CUDACompileError(cmd_parts, error.output) from error end_time = time() log_duration_msg = f"Compilation took {end_time - start_time} seconds. Compile command: {cmd}" log.info(log_duration_msg) else: log.debug( "Skip compiling %s: output %s already exists", input_path, output_path, ) cls.cache[key] = ROCmCodeCache.CacheEntry(input_path, output_path) return (cls.cache[key].output_path, key, input_path) @classmethod def load(cls, source_code: str, dst_file_ext: str) -> tuple[DLLWrapper, str, str]: """ Compiles source code and loads the generated .so file. Returns a tuple of DLLWrapper, hash_key, source_code_path """ if dst_file_ext != "so": raise RuntimeError( f"Only support loading a .so file for now. " f"Requested file extension: {dst_file_ext}. Source code: {source_code}" ) dst_file_path, hash_key, source_code_path = cls.compile( source_code, dst_file_ext ) return (DLLWrapper(dst_file_path), hash_key, source_code_path) class CodeCacheFuture: def result(self) -> Callable[..., Any]: raise NotImplementedError class LambdaFuture(CodeCacheFuture): def __init__( self, result_fn: Callable[..., Any], future: Optional[Future[Any]] = None ) -> None: self.result_fn = result_fn self.future = future def result(self) -> Callable[..., Any]: # type: ignore[override] return self.result_fn()