# Copyright 2023 The JAX Authors. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """JAX APIs for exporting JAX functions for interoperation. """ from __future__ import annotations from collections.abc import Callable, Sequence import copy import dataclasses import functools import itertools import re from typing import Any, Union, cast import warnings from absl import logging import numpy as np import jax from jax import sharding from jax._src import ad_util from jax._src import config from jax._src import core from jax._src import dispatch from jax._src import dtypes from jax._src import effects from jax._src.interpreters import mlir from jax._src.interpreters import pxla from jax._src.lib import xla_client from jax._src.lib.mlir import ir from jax._src.lib.mlir.dialects import hlo from jax._src.lib.mlir.dialects import func as func_dialect from jax._src import pjit from jax._src import sharding_impls from jax._src import source_info_util from jax._src import stages from jax._src import tree_util from jax._src import util from jax._src import xla_bridge as xb from jax._src.export import shape_poly map = util.safe_map zip = util.safe_zip DType = Any Shape = jax._src.core.Shape # The values of input and output sharding from the lowering. LoweringSharding = Union[sharding.Sharding, pxla.UnspecifiedValue] HloSharding = xla_client.HloSharding # The minimum and maximum supported calling convention version. # See https://jax.readthedocs.io/en/latest/export/export.html#export-calling-convention-version minimum_supported_calling_convention_version = 9 maximum_supported_calling_convention_version = 9 class DisabledSafetyCheck: """A safety check that should be skipped on (de)serialization. Most of these checks are performed on serialization, but some are deferred to deserialization. The list of disabled checks is attached to the serialization, e.g., as a sequence of string attributes to `jax.export.Exported` or of `tf.XlaCallModuleOp`. When using jax2tf, you can disable more deserialization safety checks by passing `TF_XLA_FLAGS=--tf_xla_call_module_disabled_checks=platform`. """ _impl: str @classmethod def platform(cls) -> DisabledSafetyCheck: """Allows the compilation platform to differ from the export platform. Has effect only on deserialization. """ return DisabledSafetyCheck("platform") @classmethod def custom_call(cls, target_name: str) -> DisabledSafetyCheck: """Allows the serialization of a call target not known to be stable. Has effect only on serialization. Args: target_name: the name of the custom call target to allow. """ return DisabledSafetyCheck(f"custom_call:{target_name}") @classmethod def shape_assertions(cls) -> DisabledSafetyCheck: """DEPRECATED: A noop. Was used previously to allow invocations with shapes that do not meet the constraints. Has no effect anymore, shape assertions cannot be disabled. """ # TODO(necula): remove this after compatibility period. Was deprecated in # May 2024. warnings.warn( "DisabledSafetyCheck.shape_assertions is deprecated, has no effect anymore", DeprecationWarning, stacklevel=2) return DisabledSafetyCheck("shape_assertions") def is_custom_call(self) -> str | None: """Returns the custom call target allowed by this directive.""" m = re.match(r'custom_call:(.+)$', self._impl) return m.group(1) if m else None def __init__(self, _impl:str): # Do not use directly, use builders `platform`, `custom_call`. self._impl = _impl def __str__(self): return self._impl __repr__ = __str__ def __eq__(self, other) -> bool: return isinstance(other, DisabledSafetyCheck) and self._impl == other._impl def __hash__(self) -> int: return hash(self._impl) @dataclasses.dataclass(frozen=True) class Exported: """A JAX function lowered to StableHLO. Attributes: fun_name: the name of the exported function, for error messages. in_tree: a PyTreeDef describing the tuple (args, kwargs) of the lowered JAX function. The actual lowering does not depend on the `in_tree`, but this can be used to invoke the exported function using the same argument structure. in_avals: the flat tuple of input abstract values. May contain dimension expressions in the shapes. out_tree: a PyTreeDef describing the result of the lowered JAX function. out_avals: the flat tuple of output abstract values. May contain dimension expressions in the shapes, with dimension variables among those in `in_avals`. in_shardings_hlo: the flattened input shardings, a sequence as long as `in_avals`. `None` means unspecified sharding. Note that these do not include the mesh or the actual devices used in the mesh. See `in_shardings_jax` for a way to turn these into sharding specification that can be used with JAX APIs. out_shardings_hlo: the flattened output shardings, a sequence as long as `out_avals`. `None` means unspecified sharding. Note that these do not include the mesh or the actual devices used in the mesh. See `out_shardings_jax` for a way to turn these into sharding specification that can be used with JAX APIs. nr_devices: the number of devices that the module has been lowered for. platforms: a tuple containing the platforms for which the function should be exported. The set of platforms in JAX is open-ended; users can add platforms. JAX built-in platforms are: 'tpu', 'cpu', 'cuda', 'rocm'. See https://jax.readthedocs.io/en/latest/export/export.html#cross-platform-and-multi-platform-export. ordered_effects: the ordered effects present in the serialized module. This is present from serialization version 9. See https://jax.readthedocs.io/en/latest/export/export.html#module-calling-convention for the calling convention in presence of ordered effects. unordered_effects: the unordered effects present in the serialized module. This is present from serialization version 9. mlir_module_serialized: the serialized lowered VHLO module. calling_convention_version: a version number for the calling convention of the exported module. See more versioning details at https://jax.readthedocs.io/en/latest/export/export.html#calling-convention-versions. module_kept_var_idx: the sorted indices of the arguments among `in_avals` that must be passed to the module. The other arguments have been dropped because they are not used. uses_global_constants: whether the `mlir_module_serialized` uses shape polymorphism or multi-platform export. This may be because `in_avals` contains dimension variables, or due to inner calls of Exported modules that have dimension variables or platform index arguments. Such modules need shape refinement before XLA compilation. disabled_safety_checks: a list of descriptors of safety checks that have been disabled at export time. See docstring for `DisabledSafetyCheck`. _get_vjp: an optional function that takes the current exported function and returns the exported VJP function. The VJP function takes a flat list of arguments, starting with the primal arguments and followed by a cotangent argument for each primal output. It returns a tuple with the cotangents corresponding to the flattened primal inputs. See a [description of the calling convention for the `mlir_module`](https://jax.readthedocs.io/en/latest/export/export.html#module-calling-convention). """ fun_name: str in_tree: tree_util.PyTreeDef in_avals: tuple[core.ShapedArray, ...] out_tree: tree_util.PyTreeDef out_avals: tuple[core.ShapedArray, ...] in_shardings_hlo: tuple[HloSharding | None, ...] out_shardings_hlo: tuple[HloSharding | None, ...] nr_devices: int platforms: tuple[str, ...] ordered_effects: tuple[effects.Effect, ...] unordered_effects: tuple[effects.Effect, ...] disabled_safety_checks: Sequence[DisabledSafetyCheck] mlir_module_serialized: bytes calling_convention_version: int module_kept_var_idx: tuple[int, ...] uses_global_constants: bool _get_vjp: Callable[[Exported], Exported] | None def mlir_module(self) -> str: return xla_client._xla.mlir.deserialize_portable_artifact(self.mlir_module_serialized) def __str__(self): # This is called to make a MLIR source location when we call an Exported, and we # do not want the entire serialized module to end up in locations. return f"Exported(fun_name={self.fun_name}, ...)" # For backwards compatibility # TODO(necula): remove after September 2024. @property def in_shardings(self): return self.in_shardings_hlo @property def out_shardings(self): return self.out_shardings_hlo def in_shardings_jax( self, mesh: sharding.Mesh) -> Sequence[sharding.Sharding | None]: """Creates Shardings corresponding to self.in_shardings_hlo. The Exported object stores `in_shardings_hlo` as HloShardings, which are independent of a mesh or set of devices. This method constructs Sharding that can be used in JAX APIs such as `jax.jit` or `jax.device_put`. Example usage: >>> from jax import export >>> exp_mesh = sharding.Mesh(jax.devices(), ("a",)) >>> exp = export.export(jax.jit(lambda x: jax.numpy.add(x, x), ... in_shardings=sharding.NamedSharding(exp_mesh, sharding.PartitionSpec("a"))) ... )(np.arange(jax.device_count())) >>> exp.in_shardings_hlo ({devices=[8]<=[8]},) # Create a mesh for running the exported object >>> run_mesh = sharding.Mesh(jax.devices()[::-1], ("b",)) >>> # Put the args and kwargs on the appropriate devices >>> run_arg = jax.device_put(np.arange(jax.device_count()), ... exp.in_shardings_jax(run_mesh)[0]) >>> res = exp.call(run_arg) >>> res.addressable_shards [Shard(device=CpuDevice(id=7), index=(slice(0, 1, None),), replica_id=0, data=[0]), Shard(device=CpuDevice(id=6), index=(slice(1, 2, None),), replica_id=0, data=[2]), Shard(device=CpuDevice(id=5), index=(slice(2, 3, None),), replica_id=0, data=[4]), Shard(device=CpuDevice(id=4), index=(slice(3, 4, None),), replica_id=0, data=[6]), Shard(device=CpuDevice(id=3), index=(slice(4, 5, None),), replica_id=0, data=[8]), Shard(device=CpuDevice(id=2), index=(slice(5, 6, None),), replica_id=0, data=[10]), Shard(device=CpuDevice(id=1), index=(slice(6, 7, None),), replica_id=0, data=[12]), Shard(device=CpuDevice(id=0), index=(slice(7, 8, None),), replica_id=0, data=[14])] """ return tuple(_hlo_sharding_to_xla_compatible_sharding(s, mesh) for s in self.in_shardings_hlo) def out_shardings_jax( self, mesh: sharding.Mesh) -> Sequence[sharding.Sharding | None]: """Creates Shardings corresponding to self.out_shardings_hlo. See documentation for in_shardings_jax. """ return tuple(_hlo_sharding_to_xla_compatible_sharding(s, mesh) for s in self.out_shardings_hlo) # For backwards compatibility # TODO(necula): remove after September 2024. @property def lowering_platforms(self): """DEPRECATED.""" warnings.warn("lowering_platform is deprecated. Use .platforms instead.", DeprecationWarning, stacklevel=2) return self.platforms # For backwards compatibility # TODO(necula): remove after September 2024. @property def mlir_module_serialization_version(self): """DEPRECATED.""" warnings.warn("mlir_module_serialization_version is deprecated. Use .calling_convention_version instead.", DeprecationWarning, stacklevel=2) return self.calling_convention_version # For backwards compatibility # TODO(necula): remove after September 2024. @property def uses_shape_polymorphism(self): """DEPRECATED.""" warnings.warn("uses_shape_polymorphism is deprecated. Use .uses_global_constants instead.", DeprecationWarning, stacklevel=2) return self.uses_global_constants def has_vjp(self) -> bool: """Returns if this Exported supports VJP.""" return self._get_vjp is not None def vjp(self) -> Exported: """Gets the exported VJP. Returns None if not available, which can happen if the Exported has been loaded from an external format without a VJP. """ if self._get_vjp is None: raise ValueError("No VJP is available") return self._get_vjp(self) def serialize(self, vjp_order: int = 0) -> bytearray: """Serializes an Exported. Args: vjp_order: The maximum vjp order to include. E.g., the value 2 means that we serialize the primal functions and two orders of the `vjp` function. This should allow 2nd order reverse mode differentiation of the deserialized function. i.e., `jax.grad(jax.grad(f)).` """ # Lazy load the serialization module, since flatbuffers is an optional # dependency. from jax._src.export.serialization import serialize return serialize(self, vjp_order=vjp_order) def call(self, *args, **kwargs): return call_exported(self)(*args, **kwargs) def deserialize(blob: bytearray) -> Exported: """Deserializes an Exported. Args: blob: a bytearray obtained from `Exported.serialize`. """ # Lazy load the serialization module, since flatbuffers is an optional # dependency. from jax._src.export.serialization import deserialize return deserialize(blob) def default_export_platform() -> str: """Retrieves the default export platform. One of: `tpu`, `cpu`, `cuda`, `rocm`. """ # Canonicalize to turn 'gpu' into 'cuda' or 'rocm' return xb.canonicalize_platform(jax.default_backend()) default_lowering_platform = default_export_platform def shape_and_dtype_jax_array(a) -> tuple[Sequence[int | None], DType]: """Returns the shape and dtype of a jax.Array or a j""" if isinstance(a, jax.ShapeDtypeStruct): return a.shape, a.dtype aval = core.raise_to_shaped(core.get_aval(a)) return aval.shape, aval.dtype def args_specs( args, # pytree of arguments polymorphic_shapes, # prefix pytree of strings get_shape_and_dtype=shape_and_dtype_jax_array, ): # TODO: deprecated in January 2024, to be removed. warnings.warn( "export.args_specs is deprecated in favor of export.symbolic_args_specs", DeprecationWarning, stacklevel=2) if get_shape_and_dtype is not shape_and_dtype_jax_array: # This was needed in some older jax2tf implementations args = tree_util.tree_map(lambda a: jax.ShapeDtypeStruct(* get_shape_and_dtype(a)), args) return shape_poly.symbolic_args_specs(args, polymorphic_shapes) # TODO(necula): remove this once we remove jax.experimental.export. def export_back_compat( fun_jax: Callable, *, lowering_platforms: Sequence[str] | None = None, disabled_checks: Sequence[DisabledSafetyCheck] = (), _device_assignment_for_internal_jax2tf_use_only = None, ) -> Callable[..., Exported]: """Exports native serialization for a JAX function. Note: this function exists only for internal usage by jax2tf and for backwards compatibility with jax.experimental.export. Use `jax.export` instead. See https://jax.readthedocs.io/en/latest/export/export.html Args: fun_jax: the function to lower and serialize. lowering_platforms: Optional sequence containing a subset of 'tpu', 'cpu', 'cuda', 'rocm'. If more than one platform is specified, then the lowered code takes an argument specifying the platform. If None, then use the default JAX backend. The calling convention for multiple platforms is explained at https://jax.readthedocs.io/en/latest/export/export.html#module-calling-convention. disabled_checks: the safety checks to disable. See docstring of `DisabledSafetyCheck`. Returns: a function that takes args and kwargs pytrees of jax.ShapeDtypeStruct, or values with `.shape` and `.dtype` attributes, and returns an `Exported`. Usage: def f_jax(*args, **kwargs): ... exported = jax_export.export(f_jax)(*args, **kwargs) """ def do_export(*args_specs, **kwargs_specs) -> Exported: if hasattr(fun_jax, "trace"): # If we have a pjit or pmap already we do not wrap with another, and we # allow shardings. wrapped_fun_jax = fun_jax else: # We support convert(pjit(f_jax)) and convert(jit(f_jax)) but also # convert(f_jax), in which case a "jit" is implied. In that case we raise # an error if the lowered function contains non-replicated sharding annotations. wrapped_fun_jax = jax.jit(fun_jax) if lowering_platforms is not None: actual_lowering_platforms = tuple(lowering_platforms) else: actual_lowering_platforms = (default_export_platform(),) # TODO: move to `lower` symbolic_scope: tuple[shape_poly.SymbolicScope, tree_util.KeyPath] | None = None # type: ignore[invalid-annotation,unused-ignore] for k_path, aval in tree_util.tree_flatten_with_path((args_specs, kwargs_specs))[0]: # Static args may have no `shape` attribute. if not hasattr(aval, "shape"): continue for d in aval.shape: if shape_poly.is_symbolic_dim(d): if symbolic_scope is None: symbolic_scope = (d.scope, k_path) continue symbolic_scope[0]._check_same_scope( d, when=f"when exporting {util.fun_name(wrapped_fun_jax)}", self_descr=f"current (from {shape_poly.args_kwargs_path_to_str(symbolic_scope[1])}) ", other_descr=shape_poly.args_kwargs_path_to_str(k_path)) traced = wrapped_fun_jax.trace(*args_specs, **kwargs_specs) lowered = traced.lower( lowering_platforms=actual_lowering_platforms, _private_parameters=mlir.LoweringParameters( for_export=True, export_ignore_forward_compatibility=config.export_ignore_forward_compatibility.value)) return _export_lowered( lowered, traced.jaxpr, traced.fun_name, disabled_checks=disabled_checks, _device_assignment_for_internal_jax2tf_use_only=_device_assignment_for_internal_jax2tf_use_only) return do_export def export( fun_jit: stages.Wrapped, *, platforms: Sequence[str] | None = None, lowering_platforms: Sequence[str] | None = None, disabled_checks: Sequence[DisabledSafetyCheck] = (), ) -> Callable[..., Exported]: """Exports a JAX function for persistent serialization. Args: fun_jit: the function to export. Should be the result of `jax.jit`. platforms: Optional sequence containing a subset of 'tpu', 'cpu', 'cuda', 'rocm'. If more than one platform is specified, then the exported code takes an argument specifying the platform. If None, then use the default JAX backend. The calling convention for multiple platforms is explained at https://jax.readthedocs.io/en/latest/export/export.html#module-calling-convention. lowering_platforms: DEPRECATED, use `platforms`. disabled_checks: the safety checks to disable. See documentation for of `jax.export.DisabledSafetyCheck`. Returns: a function that takes args and kwargs pytrees of {class}`jax.ShapeDtypeStruct`, or values with `.shape` and `.dtype` attributes, and returns an `Exported`. Usage: >>> from jax import export >>> exported: export.Exported = export.export(jnp.sin)( ... np.arange(4, dtype=np.float32)) >>> >>> # You can inspect the Exported object >>> exported.in_avals (ShapedArray(float32[4]),) >>> blob: bytearray = exported.serialize() >>> >>> # The serialized bytes are safe to use in a separate process >>> rehydrated: export.Exported = export.deserialize(blob) >>> rehydrated.fun_name 'sin' >>> rehydrated.call(np.array([.1, .2, .3, .4], dtype=np.float32)) Array([0.09983342, 0.19866933, 0.29552022, 0.38941833], dtype=float32) """ if not isinstance(fun_jit, stages.Wrapped): raise ValueError( f"Function to be exported must be the result of `jit` but is: {fun_jit}") if platforms is not None and lowering_platforms is not None: raise ValueError("Cannot use both `platforms` and `lowering_platforms`") if platforms is None and lowering_platforms is not None: platforms = lowering_platforms if platforms is not None: actual_lowering_platforms = tuple(platforms) else: actual_lowering_platforms = (default_export_platform(),) def do_export(*args_specs, **kwargs_specs) -> Exported: # TODO: move to `lower` symbolic_scope: tuple[shape_poly.SymbolicScope, tree_util.KeyPath] | None = None # type: ignore[invalid-annotation,unused-ignore] for k_path, aval in tree_util.tree_flatten_with_path((args_specs, kwargs_specs))[0]: # Static args may have no `shape` attribute. if not hasattr(aval, "shape"): continue for d in aval.shape: if shape_poly.is_symbolic_dim(d): if symbolic_scope is None: symbolic_scope = (d.scope, k_path) continue symbolic_scope[0]._check_same_scope( d, when=f"when exporting {util.fun_name(fun_jit)}", self_descr=f"current (from {shape_poly.args_kwargs_path_to_str(symbolic_scope[1])}) ", other_descr=shape_poly.args_kwargs_path_to_str(k_path)) traced = fun_jit.trace(*args_specs, **kwargs_specs) lowered = traced.lower( lowering_platforms=actual_lowering_platforms, _private_parameters=mlir.LoweringParameters( for_export=True, export_ignore_forward_compatibility=config.export_ignore_forward_compatibility.value)) return _export_lowered( lowered, traced.jaxpr, traced.fun_name, disabled_checks=disabled_checks) return do_export def _export_lowered( lowered: stages.Lowered, jaxpr: core.ClosedJaxpr, fun_name: str, disabled_checks: Sequence[DisabledSafetyCheck] = (), _device_assignment_for_internal_jax2tf_use_only = None, ) -> Exported: version = config.jax_export_calling_convention_version.value if (version < minimum_supported_calling_convention_version or version > maximum_supported_calling_convention_version): raise ValueError( f"The requested export calling convention version {version} is outside the " f"range of supported versions [{minimum_supported_calling_convention_version}" f"..{maximum_supported_calling_convention_version}]") lowering = lowered._lowering _check_lowering(lowering) mlir_module = lowering.stablehlo() args_avals_flat, _ = tree_util.tree_flatten(lowered.in_avals) if "mut" in lowering.compile_args: if lowering.compile_args["mut"]: raise NotImplementedError if "kept_var_idx" in lowering.compile_args: module_kept_var_idx = tuple(sorted(lowering.compile_args["kept_var_idx"])) else: # For pmap module_kept_var_idx = tuple(range(len(args_avals_flat))) shape_poly_state = lowering.compile_args["shape_poly_state"] if (not all(core.is_constant_shape(a.shape) for a in args_avals_flat) or lowering.compile_args.get("ordered_effects", [])): mlir_module = _wrap_main_func( mlir_module, args_avals_flat, args_kwargs_tree=lowered.in_tree, has_platform_index_argument=shape_poly_state.has_platform_index_argument, module_kept_var_idx=module_kept_var_idx, serialization_version=version) with mlir_module.context: mlir_module_attrs = mlir_module.operation.attributes mlir_module_attrs["jax.uses_shape_polymorphism"] = ( mlir.ir.BoolAttr.get(shape_poly_state.uses_dim_vars)) mlir_module_serialized = _module_to_bytecode(mlir_module) # Figure out the result types and shapes if "global_out_avals" in lowering.compile_args: # This is currently the case for pjit out_avals_flat = lowering.compile_args["global_out_avals"] elif "shards" in lowering.compile_args: # for PmapComputation out_avals_flat = lowering.compile_args["shards"].out_sharded_avals else: out_avals_flat = lowered.compile_args["out_avals"] # type: ignore # Log and then check the module. if logging.vlog_is_on(3): logmsg = (f"fun_name={fun_name} version={version} " f"lowering_platforms={lowering._platforms} " # type: ignore[unused-ignore,attribute-error] f"disabled_checks={disabled_checks}") logging.info("Exported JAX function: %s\n", logmsg) logging.info(mlir.dump_module_message(mlir_module, "export")) _check_module(mlir_module, disabled_checks=disabled_checks) ordered_effects = tuple(lowering.compile_args["ordered_effects"]) unordered_effects = tuple(lowering.compile_args["unordered_effects"]) nr_devices = len(lowering.compile_args["device_assignment"]) def export_sharding(s: LoweringSharding, aval: core.ShapedArray) -> HloSharding | None: if sharding_impls.is_unspecified(s): return None return s._to_xla_hlo_sharding(aval.ndim) # type: ignore[union-attr] all_in_shardings = expand_in_shardings(lowering.compile_args["in_shardings"], module_kept_var_idx, len(args_avals_flat)) in_shardings = tuple( export_sharding(s, aval) for s, aval in zip(all_in_shardings, args_avals_flat)) out_shardings = tuple( export_sharding(s, aval) for s, aval in zip(lowering.compile_args["out_shardings"], out_avals_flat)) device_assignment = lowering.compile_args["device_assignment"] if _device_assignment_for_internal_jax2tf_use_only is not None: _device_assignment_for_internal_jax2tf_use_only[0] = device_assignment def _get_exported_vjp(exp_primal: Exported) -> Exported: # Turn the primal jaxpr into a function, in preparation for exporting # the VJP. Note that jaxpr_as_fun produces a function with flat arguments assert(jaxpr is not None) # None only when the lowered was created outside JAX fun_jax = core.jaxpr_as_fun(jaxpr) fun_vjp_jax, vjp_in_avals = _get_vjp_fun(fun_jax, in_tree=exp_primal.in_tree, in_avals=exp_primal.in_avals, in_shardings_hlo=exp_primal.in_shardings_hlo, out_avals=exp_primal.out_avals, out_shardings_hlo=exp_primal.out_shardings_hlo, device_assignment=device_assignment, apply_jit=True, flat_primal_fun=True) return export(fun_vjp_jax, # type: ignore[arg-type] platforms=exp_primal.platforms, disabled_checks=exp_primal.disabled_safety_checks)(*vjp_in_avals) return Exported( fun_name=fun_name, in_tree=lowered.in_tree, out_tree=lowered.out_tree, in_avals=tuple(args_avals_flat), out_avals=tuple(out_avals_flat), in_shardings_hlo=in_shardings, out_shardings_hlo=out_shardings, nr_devices=nr_devices, platforms=lowering._platforms, # type: ignore ordered_effects=ordered_effects, unordered_effects=unordered_effects, disabled_safety_checks=tuple(disabled_checks), mlir_module_serialized=mlir_module_serialized, module_kept_var_idx=module_kept_var_idx, uses_global_constants=shape_poly_state.uses_dim_vars, calling_convention_version=version, _get_vjp=_get_exported_vjp) def _module_to_bytecode(module: ir.Module) -> bytes: mlir_str = mlir.module_to_bytecode(module) # `target_version` is used to manage situations when a StableHLO producer # (in this case, jax2tf) and a StableHLO consumer were built using # different versions of StableHLO. # # Each StableHLO version `producer_version` has a compatibility window, # i.e. range of versions [`consumer_version_min`, `consumer_version_max`], # where StableHLO portable artifacts serialized by `producer_version` # can be deserialized by `consumer_version` within the window. # See https://github.com/openxla/stablehlo/blob/main/docs/compatibility.md # for the exact extent of these compatibility guarantees. # # `hlo.get_minimum_version()` returns `consumer_version_min` # for the current version of StableHLO. We are using it here to maximize # forward compatibility, i.e. to maximize how far into the past we can go # and still have the payloads produced by `serialize_portable_artifact` # compatible with potential consumers from the past. target_version = hlo.get_minimum_version() module_serialized = xla_client._xla.mlir.serialize_portable_artifact( # type: ignore mlir_str, target_version) return module_serialized def _wrap_main_func( module: ir.Module, args_avals_flat: Sequence[core.ShapedArray], *, args_kwargs_tree: tree_util.PyTreeDef, has_platform_index_argument: bool, module_kept_var_idx: tuple[int, ...], serialization_version: int ) -> ir.Module: """Wraps the lowered module with a new "main" handling dimension arguments. See calling convention documentation https://jax.readthedocs.io/en/latest/export/export.html#module-calling-convention. Args: module: the HLO module as obtained from lowering. args_avals_flat: the avals for all the arguments of the lowered function, which correspond to the array arguments of the `module`. args_kwargs_tree: the PyTreeDef corresponding to `(args, kwargs)`, for error messages. has_platform_index_argument: whether the `module` has a first platform index argument module_kept_var_idx: a sorted tuple of integers with the indices of arguments in `args_avals_flat` that are kept as `module` arguments. serialization_version: the target serialization version Returns the wrapped module, without dimension and token arguments. """ dim_vars = shape_poly.all_dim_vars(args_avals_flat) context = mlir.make_ir_context() with context, ir.Location.unknown(context): # Make a copy, do not mutate because it may be cached wrapped_module = ir.Module.parse(mlir.module_to_bytecode(module)) symbol_table = ir.SymbolTable(wrapped_module.operation) orig_main = symbol_table["main"] orig_main.attributes["sym_visibility"] = ir.StringAttr.get("private") symbol_table.set_symbol_name(orig_main, "_wrapped_jax_export_main") orig_main_name = ir.StringAttr(symbol_table.insert(orig_main)).value def is_token(typ, attrs): return (typ == mlir.token_type()) orig_input_types = orig_main.type.inputs # type: ignore arg_attrs = list(ir.ArrayAttr(orig_main.arg_attrs)) # type: ignore # The order of args: platform_index_arg, dim args, token args, array args. nr_platform_index_args = 1 if has_platform_index_argument else 0 nr_dim_args = len(dim_vars) token_arg_idxs = [i for i, (typ, attrs) in enumerate(zip(orig_input_types, arg_attrs)) if is_token(typ, attrs)] nr_token_args = len(token_arg_idxs) if nr_token_args > 0: assert min(token_arg_idxs) == nr_platform_index_args + nr_dim_args assert token_arg_idxs == list( range(nr_platform_index_args + nr_dim_args, nr_platform_index_args + nr_dim_args + nr_token_args)) nr_array_args = (len(orig_input_types) - nr_platform_index_args - nr_dim_args - nr_token_args) assert nr_array_args >= 0 (platform_input_types, dim_var_input_types, token_input_types, array_input_types) = util.split_list( orig_input_types, [nr_platform_index_args, nr_dim_args, nr_token_args]) # The order of results: tokens, array results orig_output_types = orig_main.type.results # type: ignore result_attrs = list(ir.ArrayAttr(orig_main.result_attrs)) # type: ignore token_result_idxs = [i for i, (typ, attrs) in enumerate(zip(orig_output_types, result_attrs)) if is_token(typ, attrs)] nr_token_results = len(token_result_idxs) assert token_result_idxs == list(range(0, nr_token_results)) nr_array_results = len(orig_output_types) - nr_token_results assert nr_array_results >= 0 new_main_arg_indices = ( *range(nr_platform_index_args), *range(nr_platform_index_args + nr_dim_args, len(orig_input_types))) new_main_result_indices = tuple(range(0, len(orig_output_types))) new_main_input_types = [orig_input_types[idx] for idx in new_main_arg_indices] new_main_output_types = [orig_output_types[idx] for idx in new_main_result_indices] new_main_ftype = ir.FunctionType.get(new_main_input_types, new_main_output_types) new_main_op = func_dialect.FuncOp( "main", new_main_ftype, ip=ir.InsertionPoint.at_block_begin(wrapped_module.body)) new_main_op.attributes["sym_visibility"] = ir.StringAttr.get("public") try: new_arg_attrs = [] for idx in new_main_arg_indices: new_arg_attr = {} for attr in arg_attrs[idx]: if attr.name == "tf.aliasing_output": i = new_main_result_indices.index(attr.attr.value) new_arg_attr[attr.name] = ir.IntegerAttr.get( ir.IntegerType.get_signless(32), i ) else: new_arg_attr[attr.name] = attr.attr new_arg_attrs.append(ir.DictAttr.get(new_arg_attr)) new_main_op.arg_attrs = ir.ArrayAttr.get(new_arg_attrs) except KeyError: pass # TODO: better detection if orig_main.arg_attrs does not exist try: new_main_op.result_attrs = ir.ArrayAttr.get( [result_attrs[idx] for idx in new_main_result_indices]) except KeyError: pass symbol_table.insert(new_main_op) entry_block = new_main_op.add_entry_block() with ir.InsertionPoint(entry_block): # Make a context just for lowering the dimension value computations module_context = mlir.ModuleContext( backend_or_name="cpu", platforms=["cpu"], axis_context=sharding_impls.ShardingContext(0), keepalives=[], channel_iterator=itertools.count(1), host_callbacks=[], module=wrapped_module, context=context, lowering_parameters=mlir.LoweringParameters( global_constant_computation=True, for_export=True, export_ignore_forward_compatibility=config.export_ignore_forward_compatibility.value, )) ctx = mlir.LoweringRuleContext( module_context=module_context, name_stack=source_info_util.new_name_stack(), primitive=None, avals_in=args_avals_flat, avals_out=None, tokens_in=mlir.TokenSet(), tokens_out=None) # We compute dim_values from the array arguments. new_main_op_array_args = new_main_op.arguments[-nr_array_args:] if shape_poly.all_dim_vars(args_avals_flat): # TODO(necula): handle module_kept_var_idx in presence of shape # polymorphism. For now we ensured upstream that we keep all variables. assert len(set(module_kept_var_idx)) == len(args_avals_flat) dim_values = mlir.lower_fun( functools.partial(shape_poly.compute_dim_vars_from_arg_shapes, args_avals_flat, args_kwargs_tree=args_kwargs_tree), multiple_results=True)(ctx, *new_main_op_array_args) else: dim_values = () # The arguments to pass to the call to orig_main orig_main_args: list[ir.Value] = [] # The platform index and the dimension variables for arg, arg_type in zip( list(new_main_op.arguments[0:nr_platform_index_args]) + mlir.flatten_ir_values(dim_values), platform_input_types + dim_var_input_types): if arg.type != arg_type: orig_main_args.append(hlo.convert(arg_type, arg)) else: orig_main_args.append(arg) # Then the token arguments orig_main_args.extend( new_main_op.arguments[nr_platform_index_args: nr_platform_index_args + nr_token_args]) # Then the array arguments. We insert a ConvertOp as the only use of # an input argument. This helps the downstream shape refinement because # it will set the type of input arguments to static shapes, and this # can invalidate the module if the argument is used as the result of a # function, or if it appears as the input to a custom_call with # output_operand_alias attribute. See b/287386268. for arg, arg_type in zip(new_main_op_array_args, array_input_types): if arg.type != arg_type: orig_main_args.append(hlo.convert(arg_type, arg)) else: orig_main_args.append(arg) call = func_dialect.CallOp(orig_output_types, ir.FlatSymbolRefAttr.get(orig_main_name), orig_main_args) func_dialect.ReturnOp([call.results[idx] for idx in new_main_result_indices]) symbol_table.set_symbol_name(new_main_op, "main") return wrapped_module def _check_lowering(lowering) -> None: if not isinstance(lowering, pxla.MeshComputation): raise NotImplementedError(f"serialization is supported only for jit. {lowering}") if lowering.compile_args["host_callbacks"] or lowering.compile_args["keepalive"]: raise NotImplementedError("serialization of host_callbacks is not yet implemented") # Check that we do not see new compile_args. When we add a compile_args it is # safe to add it to the allowed_compile_args if it does not change the semantics # or the calling convention of the lowered module. allowed_compile_args = [ "backend", "platforms", "mesh", "global_in_avals", "global_out_avals", "in_shardings", "out_shardings", "kept_var_idx", "mut", "spmd_lowering", "auto_spmd_lowering", "tuple_args", "ordered_effects", "unordered_effects", "keepalive", "host_callbacks", "pmap_nreps", "committed", "device_assignment", "jaxpr_debug_info", "shape_poly_state", "all_default_mem_kind", "in_layouts", "out_layouts", "all_args_info", "pgle_profiler", "intermediate_shardings"] for compile_arg in lowering.compile_args.keys(): if compile_arg not in allowed_compile_args: raise NotImplementedError(f"Unrecognized lowered.compile_args[{compile_arg}]") # We have not implemented support for some of the compile_args. Check here that # the compile_args have the values that have been implemented. not_implemented_msgs = [] for compile_arg, check_value, err_msg in ( ("spmd_lowering", lambda v: v, "True"), ("auto_spmd_lowering", lambda v: not v, "False"), # tuple_args is a compilation flag, does not affect lowering. ("tuple_args", lambda v: True, "N/A"), # unordered_effects do not change the calling convention. Those from # jax.debug will also result in keepalive being non-empty and unsupported # custom calls. The CallTfEffect is an exception, but we want to allow # that one. ("unordered_effects", lambda v: True, "N/A"), ("ordered_effects", lambda v: True, "N/A"), # used for TPU jax.debug, send/recv. Not supported yet. ("host_callbacks", lambda v: not v, "empty"), # used on all platforms for callbacks. Not supported yet. ("keepalive", lambda v: not v, "empty"), ("pmap_nreps", lambda v: v == 1, "1"), ("shape_poly_state", lambda v: True, "N/A"), ): if compile_arg in lowering.compile_args: if not check_value(lowering.compile_args[compile_arg]): not_implemented_msgs.append( f"{compile_arg} must be {err_msg} and it is {lowering.compile_args[compile_arg]}") if not_implemented_msgs: raise NotImplementedError( "serialization error, unimplemented lowered.compile_args:\n" + "\n".join(not_implemented_msgs)) _CPU_FFI_KERNELS = [ "lapack_spotrf_ffi", "lapack_dpotrf_ffi", "lapack_cpotrf_ffi", "lapack_zpotrf_ffi", ] # These are the JAX custom call target names that are guaranteed to be stable. # Their backwards compatibility is tested by back_compat_test.py. _CUSTOM_CALL_TARGETS_GUARANTEED_STABLE = { *_CPU_FFI_KERNELS, "Sharding", "SPMDFullToShardShape", "SPMDShardToFullShape", "cu_threefry2x32", "cu_threefry2x32_ffi", "__gpu$xla.gpu.triton", # Pallas call on GPU # cholesky on CPU "lapack_spotrf", "lapack_dpotrf", "lapack_cpotrf", "lapack_zpotrf", # eigh on CPU "lapack_ssyevd", "lapack_dsyevd", "lapack_cheevd", "lapack_zheevd", # eigh on GPU "cusolver_syevj", "cusolver_syevd", # eigh on TPU "Eigh", # eig on CPU "lapack_sgeev", "lapack_dgeev", "lapack_cgeev", "lapack_zgeev", # qr on CPU "lapack_sgeqrf", "lapack_dgeqrf", "lapack_cgeqrf", "lapack_zgeqrf", # householder product on CPU "lapack_sorgqr", "lapack_dorgqr", "lapack_cungqr", "lapack_zungqr", # svd on CPU "lapack_sgesdd", "lapack_dgesdd", "lapack_cgesdd", "lapack_zgesdd", # qr on GPU "cusolver_geqrf", "cublas_geqrf_batched", "cusolver_orgqr", # qr and svd on TPU "Qr", "ProductOfElementaryHouseholderReflectors", # triangular_solve on CPU "blas_strsm", "blas_dtrsm", "blas_ctrsm", "blas_ztrsm", # TODO(atondwal, necula): add back_compat tests for lu on CPU/GPU # lu on CPU "lapack_sgetrf", "lapack_dgetrf", "lapack_cgetrf", "lapack_zgetrf", # schur on CPU "lapack_sgees", "lapack_dgees", "lapack_cgees", "lapack_zgees", # # lu on GPU # "cublas_getrf_batched", "cusolver_getrf", # "hipblas_getrf_batched", "hipsolver_getrf", # lu on TPU "LuDecomposition", # ApproxTopK on TPU "ApproxTopK", "stablehlo.dynamic_approx_top_k", "tf.call_tf_function", # From jax2tf.call_tf(func, call_tf_graph=True) "tpu_custom_call", # Pallas/TPU kernels # TODO(burmako): maintain backwards compatibility for these, until they # are upstreamed to StableHLO. # See https://github.com/openxla/stablehlo/issues/8. "stablehlo.dynamic_reduce_window", "stablehlo.dynamic_rng_bit_generator", "stablehlo.dynamic_top_k", "shape_assertion", # Used by shape_poly to evaluate assertions } check_sharding_pattern = re.compile(r"^({replicated}|{unknown shard_as.*}|"")$") def _check_module(mod: ir.Module, *, disabled_checks: Sequence[DisabledSafetyCheck]) -> bool: """Run a number of checks on the module. Args: disabled_checks: the safety checks that are disabled. Returns True if the module uses non-replicated shardings. """ sharding_attr = ir.StringAttr.get("Sharding", mod.context) allowed_custom_call_targets: set[str] = copy.copy(_CUSTOM_CALL_TARGETS_GUARANTEED_STABLE) for dc in disabled_checks: target = dc.is_custom_call() if target is not None: allowed_custom_call_targets.add(target) allowed_custom_call_targets_attrs = { ir.StringAttr.get(target, mod.context) for target in allowed_custom_call_targets} disallowed_custom_call_ops: list[str] = [] module_uses_non_replicated_sharding = False def check_sharding(op: ir.Operation, loc: ir.Location): try: sharding = op.attributes["mhlo.sharding"] except KeyError: pass else: nonlocal module_uses_non_replicated_sharding try: sharding_value = ir.StringAttr(sharding).value except UnicodeDecodeError: # The mhlo.sharding attribute may be in pretty-printed format, or # as an encoding of an HloSharding protobuf in some rare situations. # We handle the latter by conservatively assuming it is non-replicated. module_uses_non_replicated_sharding = True else: if not re.match(check_sharding_pattern, sharding_value): module_uses_non_replicated_sharding = True def check_op(op: ir.Operation): op_name = op.operation.name if op_name == "func.func": check_sharding(op.operation, op.location) elif op_name == "stablehlo.custom_call": call_target_name_attr = op.operation.attributes["call_target_name"] if (call_target_name_attr not in allowed_custom_call_targets_attrs): disallowed_custom_call_ops.append(f"{op} at {op.location}") if call_target_name_attr == sharding_attr: check_sharding(op, op.location) def walk_operations(op): check_op(op) for region in op.operation.regions: for block in region: for op in block: walk_operations(op) walk_operations(mod) if disallowed_custom_call_ops: disallowed_custom_call_ops_str = "\n".join(disallowed_custom_call_ops) msg = ("Cannot serialize code with custom calls whose targets have no " "compatibility guarantees. Examples are:\n" f"{disallowed_custom_call_ops_str}.\n" "See https://jax.readthedocs.io/en/latest/export/export.html#compatibility-guarantees-for-custom-calls") raise ValueError(msg) return module_uses_non_replicated_sharding def expand_in_shardings(in_shardings: Sequence[LoweringSharding], module_kept_var_idx: Sequence[int], nr_inputs: int) -> Sequence[LoweringSharding]: """Expands in_shardings with unspecified shardings for inputs not kept. Assumes in_shardings corresponds to module_kept_var_idx. """ assert len(in_shardings) == len(module_kept_var_idx) assert nr_inputs >= len(module_kept_var_idx) all_in_shardings: list[LoweringSharding] = [sharding_impls.UNSPECIFIED] * nr_inputs for idx, in_s in zip(sorted(module_kept_var_idx), in_shardings): all_in_shardings[idx] = in_s return tuple(all_in_shardings) def _hlo_sharding_to_xla_compatible_sharding( hlo_sharding: HloSharding | None, mesh: sharding.Mesh) -> sharding.Sharding | None: if hlo_sharding is None: return None return sharding_impls._gspmd_to_named_sharding_via_mesh( _hlo_sharding_to_gspmd_sharding(hlo_sharding, tuple(mesh.devices.flat)), # type: ignore[arg-type] mesh) def _hlo_sharding_to_gspmd_sharding( hlo_sharding: HloSharding | None, device_assignment: Sequence[jax.Device]) -> sharding.GSPMDSharding | None: if hlo_sharding is None: return None return sharding.GSPMDSharding(device_assignment, hlo_sharding) def _get_vjp_fun(primal_fun: Callable, *, in_tree: tree_util.PyTreeDef, in_avals: Sequence[core.AbstractValue], out_avals: Sequence[core.AbstractValue], in_shardings_hlo: tuple[HloSharding | None, ...], out_shardings_hlo: tuple[HloSharding | None, ...], device_assignment: Sequence[sharding_impls.Device] | None, apply_jit: bool, flat_primal_fun: bool = False, ) -> tuple[Callable, Sequence[core.AbstractValue]]: # Since jax.vjp does not handle kwargs, it is easier to do all the work # here with flattened functions. # apply_jit=False is only used for backwards compatibility with the graph # graph serialization. When apply_jit=True, we must pass a device assignment. # flat_primal_fun=False is used only from jax2tf, and it means that the # `primal_fun` takes PyTree `*args` and `**kwargs`. def fun_vjp_jax(*args_and_out_cts_flat_jax): # Takes a flat list of primals and output cotangents def flattened_primal_fun_jax(*args_flat): args, kwargs = in_tree.unflatten(args_flat) res = primal_fun(*args, **kwargs) res_flat, _ = tree_util.tree_flatten(res) return res_flat args_flat_jax, out_cts_flat_jax = util.split_list(args_and_out_cts_flat_jax, [len(in_avals)]) _, pullback_jax = jax.vjp(primal_fun if flat_primal_fun else flattened_primal_fun_jax, *args_flat_jax) return pullback_jax(out_cts_flat_jax) vjp_in_avals = list( itertools.chain(in_avals, map(lambda a: a.at_least_vspace(), out_avals))) if apply_jit: assert device_assignment is not None vjp_in_shardings = tuple( _hlo_sharding_to_gspmd_sharding(s, device_assignment) for s in itertools.chain(in_shardings_hlo, out_shardings_hlo)) vjp_out_shardings = tuple( _hlo_sharding_to_gspmd_sharding(s, device_assignment) for s in in_shardings_hlo) return pjit.pjit(fun_vjp_jax, in_shardings=vjp_in_shardings, out_shardings=vjp_out_shardings), vjp_in_avals else: return fun_vjp_jax, vjp_in_avals ### Calling the exported function def call(exported: Exported) -> Callable[..., jax.Array]: if not isinstance(exported, Exported): raise ValueError( "The exported argument must be an export.Exported. " f"Found {exported}.") @jax.custom_vjp def f_flat(*args_flat): return call_exported_p.bind(*args_flat, exported=exported) def f_flat_vjp_fwd(*args_flat): # Return the primal arguments as the residual # TODO: keep as residuals only the arguments that are needed return f_flat(*args_flat), args_flat def f_flat_vjp_bwd(residual, ct_res_flat): args_flat = residual # residual is the primal argument flat tuple exp_vjp = exported.vjp() # ct_res_flat may contain arrays of zeros where exp_vjp expect float0. # We make the proper arrays of float0 to invoke exp_vjp. def fix_float0_ct(ct_res, expected_aval): if expected_aval.dtype != dtypes.float0: return ct_res return ad_util.zeros_like_jaxval(ct_res) ct_res_fixed = map(fix_float0_ct, ct_res_flat, exp_vjp.in_avals[len(args_flat):]) in_ct_flat = call_exported(exp_vjp)(*args_flat, *ct_res_fixed) return in_ct_flat f_flat.defvjp(f_flat_vjp_fwd, f_flat_vjp_bwd) def f_imported(*args, **kwargs): # since custom_vjp does not support kwargs, flatten the function first. args_flat, in_tree = tree_util.tree_flatten((args, kwargs)) if in_tree != exported.in_tree: # Give errors with the precise tree difference; use fake leaves so we can # use tree_util.equality_errors. in_args = in_tree.unflatten([0] * in_tree.num_leaves) exp_in_args = exported.in_tree.unflatten([0] * exported.in_tree.num_leaves) msg = ( "The invocation args and kwargs must have the same pytree structure " f"as when the function '{exported.fun_name}' was exported, but they " "have the following structural differences:\n" + ("\n".join( f" - {shape_poly.args_kwargs_path_to_str(path)} is a {thing1} in the invocation and a " f"{thing2} when exported, so {explanation}.\n" for path, thing1, thing2, explanation in tree_util.equality_errors(in_args, exp_in_args)))) raise ValueError(msg) res_flat = f_flat(*args_flat) return exported.out_tree.unflatten(res_flat) return f_imported call_exported = call # A JAX primitive for invoking a serialized JAX function. call_exported_p = core.Primitive("call_exported") call_exported_p.multiple_results = True @util.cache() def _call_exported_abstract_eval( *in_avals: core.AbstractValue, exported: Exported ) -> tuple[tuple[core.AbstractValue, ...], set[effects.Effect]]: exported_dim_vars = shape_poly.all_dim_vars(exported.in_avals) assert len(in_avals) == len(exported.in_avals) # since the pytrees have the same structure # Check that the expected shapes match the actual ones for arg_idx, (exp_aval, actual_aval) in enumerate(zip(exported.in_avals, in_avals)): if not isinstance(actual_aval, core.ShapedArray): raise ValueError(f"Expected ShapedArray but got: {actual_aval}") def pp_arg_dim(dim_idx: int | None) -> str: return shape_poly.pretty_print_dimension_descriptor(exported.in_tree, arg_idx, dim_idx) if len(exp_aval.shape) != len(actual_aval.shape): raise ValueError( f"Rank mismatch for {pp_arg_dim(None)}: expected {exp_aval.shape} " f"and called with {actual_aval.shape}") if exp_aval.dtype != actual_aval.dtype: raise ValueError( f"Dtype mismatch for {pp_arg_dim(None)}: expected {exp_aval.dtype} " f"and called with {actual_aval.dtype}") for dim_idx, aval_d in enumerate(exp_aval.shape): # If the exp_aval has a constant dimension then the actual argument must have # a matching constant dimension. if core.is_constant_dim(aval_d): if (not core.is_constant_dim(actual_aval.shape[dim_idx]) or aval_d != actual_aval.shape[dim_idx]): raise ValueError( f"Shape mismatch for {pp_arg_dim(dim_idx)} " "(expected same constant): " f"expected {exp_aval.shape} and called with {actual_aval.shape}") # Must express the exported_dim_vars in terms of the shapes in in_avals. solution, shape_constraints, synth_dim_vars = shape_poly.solve_dim_vars( exported.in_avals, args_kwargs_tree=exported.in_tree) synthetic_env = {vname: in_avals[arg_idx].shape[dim_idx] for (vname, arg_idx, dim_idx) in synth_dim_vars} synthetic_eval = shape_poly.CachingShapeEvaluator(**synthetic_env) # We discharge all the constraints statically. This results in much simpler # composability (because we do not have to worry about the constraints of the # Exported called recursively; we only need to worry about entry-point # constraints). This also makes sense from a composability point of view, # because we get the same errors if we invoke the exported module, or if we # trace the exported function. Consider for example, an exported module with # signature `f32[a, a] -> f32[a]`. If we invoke the module with an argument # `f32[c, d]` it is better to fail because `c == d` is inconclusive, than # succeed and add a compile-time check that `c == d`. In the latter case, # it would be ambiguous whether we should continue tracing with a result # of type `f32[c]` or `f32[d]`. shape_constraints.check_statically(synthetic_eval) exported_dim_values = [synthetic_eval.evaluate(solution[var]) for var in exported_dim_vars] out_avals = tuple( core.ShapedArray(core.evaluate_shape(out_aval.shape, exported_dim_vars, *exported_dim_values), dtype=out_aval.dtype, weak_type=out_aval.weak_type) for out_aval in exported.out_avals) return out_avals, set(exported.ordered_effects + exported.unordered_effects) call_exported_p.def_effectful_abstract_eval(_call_exported_abstract_eval) def _call_exported_impl(*args, exported: Exported): return dispatch.apply_primitive(call_exported_p, *args, exported=exported) call_exported_p.def_impl(_call_exported_impl) def _call_exported_lowering(ctx: mlir.LoweringRuleContext, *args, exported: Exported): if exported.uses_global_constants: ctx.module_context.shape_poly_state.uses_dim_vars = True submodule = ir.Module.parse(exported.mlir_module()) axis_context = ctx.module_context.axis_context if isinstance(axis_context, sharding_impls.ShardingContext): num_devices = axis_context.num_devices elif isinstance(axis_context, sharding_impls.SPMDAxisContext): num_devices = axis_context.mesh.size elif isinstance(axis_context, sharding_impls.ReplicaAxisContext): num_devices = axis_context.axis_env.nreps else: raise NotImplementedError(type(axis_context)) if num_devices != exported.nr_devices: # In some special cases we allow running with a different number of devices # than the function was exported for. err_msg = "" if exported.nr_devices != 1: err_msg = "the function was exported for more than 1 device." elif (_check_module(submodule, disabled_checks=()) or any(s is not None and not s.is_replicated() for s in exported.in_shardings_hlo + exported.out_shardings_hlo)): err_msg = "the function contains non-replicated sharding annotations." if err_msg: raise ValueError( f"Function {exported.fun_name} was exported for " f"{exported.nr_devices} devices and is called in a context with " f"{num_devices} devices. This is disallowed because: {err_msg}" ) # Apply in_shardings args = tuple( wrap_with_sharding(ctx, x, x_aval, x_sharding) for x, x_aval, x_sharding in zip(args, ctx.avals_in, exported.in_shardings_hlo)) symtab = ir.SymbolTable(submodule.operation) # The called function may have been exported with polymorphic shapes and called # now with more refined shapes. We insert hlo.ConvertOp to ensure the module # is valid. def convert_shape(x: ir.Value, x_aval: core.AbstractValue, new_aval: core.AbstractValue) -> ir.Value: new_ir_type = mlir.aval_to_ir_type(new_aval) if x.type != new_ir_type: return hlo.convert(mlir.aval_to_ir_type(new_aval), x) else: return x callee_type = symtab["main"].type # TODO: maybe cache multiple calls fn = mlir.merge_mlir_modules(ctx.module_context.module, f"call_exported_{exported.fun_name}", submodule, dst_symtab=ctx.module_context.symbol_table) submodule_args: list[ir.Value] = [] # All the platforms for the current lowering must be among the platforms # for which the callee was lowered. lowering_platforms = ctx.module_context.platforms callee_lowering_platform_index: list[int] = [] for platform in lowering_platforms: if platform in exported.platforms: callee_lowering_platform_index.append( exported.platforms.index(platform)) elif DisabledSafetyCheck.platform() in exported.disabled_safety_checks: callee_lowering_platform_index.append(0) else: raise ValueError( f"Function '{exported.fun_name}' was exported for " f"platforms '{exported.platforms}' but it is used " f"on '{lowering_platforms}'.") if len(exported.platforms) > 1: # The exported module takes a platform index argument if len(lowering_platforms) > 1: current_platform_idx = ctx.dim_var_values[0] else: current_platform_idx = cast(ir.Value, mlir.ir_constant(np.int32(0))) # Compute the rule index based on the current platform i32_type = mlir.aval_to_ir_type(core.ShapedArray((), dtype=np.int32)) if current_platform_idx.type != i32_type: current_platform_idx = hlo.ConvertOp(i32_type, current_platform_idx) callee_platform_idx = hlo.CaseOp([i32_type], index=current_platform_idx, num_branches=len(lowering_platforms)) for i in range(len(lowering_platforms)): branch = callee_platform_idx.regions[i].blocks.append() with ir.InsertionPoint(branch): hlo.return_([mlir.ir_constant( np.int32(callee_lowering_platform_index[i]))]) if callee_platform_idx.result.type != callee_type.inputs[0]: callee_platform_idx = hlo.ConvertOp(callee_type.inputs[0], callee_platform_idx) submodule_args.append(callee_platform_idx) else: assert len(lowering_platforms) == 1 ordered_effects = exported.ordered_effects for eff in ordered_effects: token_in = ctx.tokens_in.get(eff) submodule_args.append(token_in) kept_args = [ convert_shape(a, a_aval, exported_in_aval) for i, (a, a_aval, exported_in_aval) in enumerate(zip(args, ctx.avals_in, exported.in_avals)) if i in exported.module_kept_var_idx] submodule_args = submodule_args + kept_args call = func_dialect.CallOp(callee_type.results, ir.FlatSymbolRefAttr.get(fn), submodule_args) if ordered_effects: tokens_out = {eff: (call.results[effect_idx],) for effect_idx, eff in enumerate(ordered_effects)} ctx.set_tokens_out(mlir.TokenSet(tokens_out)) # The ctx.avals_out already contain the abstract values refined by # _call_exported_abstract_eval. results = tuple( convert_shape(out, out_aval, refined_out_aval) for out, out_aval, refined_out_aval in zip(call.results[len(ordered_effects):], exported.out_avals, ctx.avals_out)) # Apply out_shardings results = tuple( wrap_with_sharding(ctx, x, x_aval, x_sharding) for x, x_aval, x_sharding in zip(results, ctx.avals_out, exported.out_shardings_hlo) ) return results mlir.register_lowering(call_exported_p, _call_exported_lowering) def wrap_with_sharding(ctx: mlir.LoweringRuleContext, x: ir.Value, x_aval: core.AbstractValue, x_sharding: HloSharding | None) -> ir.Value: if x_sharding is None: return x return mlir.wrap_with_sharding_op( ctx, x, x_aval, x_sharding.to_proto())