# Copyright 2021 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. from __future__ import annotations from collections import defaultdict from collections.abc import Callable, Sequence, Iterable import contextlib import dataclasses from functools import partial import inspect import itertools as it import logging import weakref from typing import NamedTuple, Any, Union, cast import warnings import numpy as np from jax._src import api from jax._src import api_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 linear_util as lu from jax._src import mesh as mesh_lib from jax._src import op_shardings from jax._src import profiler from jax._src import sharding_impls from jax._src import source_info_util from jax._src import stages from jax._src import traceback_util from jax._src import tree_util from jax._src import util from jax._src import xla_bridge as xb from jax._src.api_util import ( argnums_partial_except, flatten_axes, flatten_fun, flatten_fun_nokwargs, donation_vector, check_callable, resolve_argnums, argnames_partial_except, debug_info, hoist_obj_attrs, _check_no_aliased_ref_args, _check_no_aliased_closed_over_refs) from jax._src.interpreters import partial_eval as pe from jax._src.partition_spec import PartitionSpec from jax._src.interpreters import xla from jax._src.interpreters import ad from jax._src.interpreters import batching from jax._src.interpreters import mlir from jax._src.interpreters import pxla from jax._src.lib.mlir import ir from jax._src.lib.mlir.dialects import func as func_dialect from jax._src.lib import jax_jit from jax._src.lib import xla_client as xc from jax._src.mesh import AbstractMesh from jax._src.sharding import Sharding from jax._src.sharding_impls import ( NamedSharding, GSPMDSharding, SingleDeviceSharding, PmapSharding, AUTO, UNSPECIFIED, UnspecifiedValue, prepare_axis_resources, parse_flatten_op_sharding, canonicalize_sharding, flatten_spec, _internal_use_concrete_mesh) from jax._src.layout import Format, DeviceLocalLayout, AutoLayout from jax._src.state.types import RefEffect from jax._src.traceback_util import api_boundary from jax._src.tree_util import ( tree_flatten, tree_unflatten, treedef_is_leaf, tree_structure, tree_leaves, treedef_children, broadcast_prefix, all_leaves, prefix_errors, keystr, PyTreeDef, none_leaf_registry as none_lr, tree_map, tree_flatten_with_path) from jax._src.util import ( HashableFunction, safe_map, safe_zip, wraps, distributed_debug_log, split_list, split_list_checked, weakref_lru_cache, merge_lists, subs_list, fun_name, fun_qual_name) from jax._src.attrs import (Box, List, dne_sentinel, jax_setattr, jax_getattr, jax_extendattr) map, unsafe_map = safe_map, map zip, unsafe_zip = safe_zip, zip traceback_util.register_exclusion(__file__) PjitSharding = Union[GSPMDSharding, UnspecifiedValue, AUTO] PjitShardingMinusUnspecified = Union[GSPMDSharding, AUTO] MeshSharding = Union[NamedSharding, UnspecifiedValue, AUTO] MeshShardingMinusUnspecified = Union[NamedSharding, AUTO] logger = logging.getLogger(__name__) class PjitInfo(NamedTuple): """Things that we know about a jit instance before it is called. In other words, this structure contains arguments to jit()/pjit(), preprocessed and validated. """ fun_sourceinfo: str fun_signature: inspect.Signature | None # Shardings, as specified by the user. These can either be UNSPECIFIED or they # can be a tree (prefix) of shardings or None. user_specified_in_shardings: bool in_shardings_treedef: PyTreeDef in_shardings_leaves: tuple[Any, ...] out_shardings_treedef: PyTreeDef out_shardings_leaves: tuple[Any, ...] in_layouts_treedef: PyTreeDef in_layouts_leaves: tuple[Any, ...] out_layouts_treedef: PyTreeDef out_layouts_leaves: tuple[Any, ...] static_argnums: tuple[int, ...] static_argnames: tuple[str, ...] donate_argnums: tuple[int, ...] donate_argnames: tuple[str, ...] device: xc.Device | None backend: str | None keep_unused: bool inline: bool abstracted_axes: Any | None use_resource_env: bool # False for jit, True for pjit compiler_options_kvs: tuple[tuple[str, Any], ...] # Hash and compare PjitInfo by identity when used as a cache key. def __hash__(self): return id(self) def __eq__(self, other): return self is other def _python_pjit_helper(fun: Callable, jit_info: PjitInfo, *args, **kwargs): p, args_flat = _infer_params(fun, jit_info, args, kwargs) for arg in args_flat: dispatch.check_arg(arg) if p.attrs_tracked: init_states = _get_states(p.attrs_tracked) args_flat = [*init_states, *args_flat] try: if (core.trace_state_clean() and not config.debug_key_reuse.value and not p.params['jaxpr'].jaxpr.is_high): args_flat = map(core.full_lower, args_flat) core.check_eval_args(args_flat) out_flat, compiled, profiler = _pjit_call_impl_python(*args_flat, **p.params) else: out_flat = pjit_p.bind(*args_flat, **p.params) compiled = None profiler = None except stages.DeviceAssignmentMismatchError as e: fails, = e.args fun_name = getattr(fun, '__qualname__', getattr(fun, '__name__', str(fun))) msg = stages._device_assignment_mismatch_error( fun_name, fails, args_flat, 'jit', p.arg_names) raise ValueError(msg) from None except xla.InvalidInputException as e: arg_names = [''] * len(args_flat) if p.arg_names is None else p.arg_names # Run canonicalization again to figure out which arg failed. if p.params['jaxpr'].consts: raise TypeError(e.args[0]) from e else: for arg, name, aval in zip(args_flat, arg_names, p.in_avals): try: xla.canonicalize_dtype(arg) except xla.InvalidInputException as _: # Reraise as TypeError with the new message. raise TypeError( f"Argument '{name}' of shape {aval.str_short()} of type" f' {type(arg)} is not a valid JAX type.') from e raise AssertionError("Unreachable") from e except api_util.InternalFloatingPointError as e: if getattr(fun, '_apply_primitive', False): raise FloatingPointError(f"invalid value ({e.ty}) encountered in {fun.__qualname__}") from None api_util.maybe_recursive_nan_check(e, fun, args, kwargs) if p.box_data: box_treedef, out_tree = p.out_tree.children() box_flat, out_flat = split_list_checked(out_flat, [box_treedef.num_leaves, out_tree.num_leaves]) box_out = tree_unflatten(box_treedef, box_flat) leaves = tree_leaves((args, kwargs)) for (i, kind), b in zip(p.box_data, box_out): if kind is pe.BoxAttr: leaves[i].set(tree_unflatten(b.treedef, b.leaves)) elif kind is pe.ListAttr: for item in tree_unflatten(b.treedef, b.leaves): leaves[i].append(item) else: assert False else: out_tree = p.out_tree if p.attrs_tracked: num_states_out = sum(end_tree.num_leaves for _, end_tree, _ in p.attrs_tracked) final_states, out_flat = split_list(out_flat, [num_states_out]) _set_states(p.attrs_tracked, final_states) outs = tree_unflatten(out_tree, out_flat) return (outs, out_flat, out_tree, args_flat, p.params['jaxpr'], p.attrs_tracked, p.box_data, compiled, profiler) def _need_to_rebuild_with_fdo(pgle_profiler): return (pgle_profiler is not None and pgle_profiler.is_enabled() and not pgle_profiler.is_fdo_consumed()) def _get_fastpath_data( executable, out_tree, args_flat, out_flat, attrs_tracked, box_data, effects, consts, abstracted_axes, pgle_profiler ) -> pxla.MeshExecutableFastpathData | None: out_reflattened, out_tree = pxla.reflatten_outputs_for_dispatch(out_tree, out_flat) use_fastpath = ( executable is not None and isinstance(executable, pxla.MeshExecutable) and isinstance(executable.unsafe_call, pxla.ExecuteReplicated) # No effects in computation and not executable.unsafe_call.ordered_effects and not executable.unsafe_call.has_unordered_effects and not executable.unsafe_call.has_host_callbacks and all(isinstance(x, xc.ArrayImpl) for x in out_reflattened) and abstracted_axes is None # no attr state effects and not attrs_tracked and not box_data # no ref state effects and not any(isinstance(e, RefEffect) for e in effects) # no prng reuse checking and not (config.debug_key_reuse.value and any( hasattr(arg, 'dtype') and dtypes.issubdtype(arg.dtype, dtypes.prng_key) for arg in (*args_flat, *out_flat, *consts))) and not _need_to_rebuild_with_fdo(pgle_profiler) ) if use_fastpath: out_avals = [o.aval for o in out_reflattened] out_committed = [o._committed for o in out_reflattened] kept_var_bitvec = [i in executable._kept_var_idx for i in range(len(args_flat))] in_shardings = [ sharding_impls.physical_sharding(a, s) if a is not core.abstract_token and dtypes.issubdtype(a.dtype, dtypes.extended) else s for s, a in zip(executable._in_shardings, executable.in_avals) ] fastpath_data = pxla.MeshExecutableFastpathData( executable.xla_executable, out_tree, in_shardings, executable._out_shardings, out_avals, out_committed, kept_var_bitvec, executable._dispatch_in_layouts) else: fastpath_data = None return fastpath_data # The entries are doubled here from the default 4096 because _pjit_call_impl # also has a cpp dispatch path and that would double the number of entries in # the global shared cache. # This cache is only used for jit's with only fun. For example: jax.jit(f) _cpp_pjit_cache_fun_only = xc._xla.PjitFunctionCache(capacity=8192) # This cache is used for jit where extra arguments are defined other than the # fun. For example: jax.jit(f, donate_argnums=...) OR # jax.jit(f, out_shardings=...), etc. We don't use the same cache because the # capacity might get full very fast because of all the jitted function in JAX # which might evict train_step for example. _cpp_pjit_cache_explicit_attributes = xc._xla.PjitFunctionCache(capacity=8192) def _get_cpp_global_cache(contains_explicit_attributes: bool): if contains_explicit_attributes: return _cpp_pjit_cache_explicit_attributes else: return _cpp_pjit_cache_fun_only def _cpp_pjit(fun: Callable, jit_info: PjitInfo): @api_boundary def cache_miss(*args, **kwargs): if config.no_tracing.value: raise RuntimeError(f"re-tracing function {jit_info.fun_sourceinfo} for " "`jit`, but 'no_tracing' is set") (outs, out_flat, out_tree, args_flat, jaxpr, attrs_tracked, box_data, executable, pgle_profiler) = _python_pjit_helper(fun, jit_info, *args, **kwargs) maybe_fastpath_data = _get_fastpath_data( executable, out_tree, args_flat, out_flat, attrs_tracked, box_data, jaxpr.effects, jaxpr.consts, jit_info.abstracted_axes, pgle_profiler) return outs, maybe_fastpath_data, _need_to_rebuild_with_fdo(pgle_profiler) cache_key = pxla.JitGlobalCppCacheKeys( donate_argnums=jit_info.donate_argnums, donate_argnames=jit_info.donate_argnames, device=jit_info.device, backend=jit_info.backend, in_shardings_treedef=jit_info.in_shardings_treedef, in_shardings_leaves=jit_info.in_shardings_leaves, out_shardings_treedef=jit_info.out_shardings_treedef, out_shardings_leaves=jit_info.out_shardings_leaves, in_layouts_treedef=jit_info.in_layouts_treedef, in_layouts_leaves=jit_info.in_layouts_leaves, out_layouts_treedef=jit_info.out_layouts_treedef, out_layouts_leaves=jit_info.out_layouts_leaves, compiler_options_kvs=jit_info.compiler_options_kvs) cpp_pjit_f = xc._xla.pjit( fun_name(fun), fun, cache_miss, jit_info.static_argnums, jit_info.static_argnames, cache_key, tree_util.dispatch_registry, pxla.cc_shard_arg, _get_cpp_global_cache(cache_key.contains_explicit_attributes)) cpp_pjitted_f = wraps(fun)(cpp_pjit_f) cpp_pjitted_f._fun = fun cpp_pjitted_f._jit_info = jit_info # TODO(necula): move these to top-level; we don't need to do this for # every jit cpp_jitted_f_class = type(cpp_pjitted_f) # TODO(necula): make clear_cache private, no need to have it part of the API cpp_jitted_f_class.clear_cache = jit_evict_fn cpp_jitted_f_class.lower = jit_lower cpp_jitted_f_class.trace = jit_trace cpp_jitted_f_class.eval_shape = jit_eval_shape # We return directly the function produced by _xla.pjit, because we do not # want to have Python in the dispatch path. return cpp_pjitted_f @api_boundary def jit_trace(jit_func, *args, **kwargs) -> stages.Traced: p, args_flat = _infer_params(jit_func._fun, jit_func._jit_info, args, kwargs) donate_argnums = tuple(i for i, d in enumerate(p.donated_invars) if d) args_info = stages.make_args_info(p.in_tree, p.in_avals, donate_argnums) lower_callable = partial(_resolve_and_lower, args_flat, **p.params, pgle_profiler=None) return stages.Traced( p.params['jaxpr'], args_info, p.params["name"], p.out_tree, lower_callable, args_flat, p.arg_names, p.num_consts) @api_boundary def jit_lower(jit_func, *args, **kwargs): return jit_trace(jit_func, *args, **kwargs).lower() @api_boundary def jit_eval_shape(jit_func, *args, **kwargs): p, _ = _infer_params(jit_func._fun, jit_func._jit_info, args, kwargs) out_shardings = [None if isinstance(s, UnspecifiedValue) else s for s in p.params['out_shardings']] out = [] for a, out_s in zip(p.params['jaxpr'].out_avals, out_shardings): if out_s is None: s = a.sharding if a.sharding.mesh._are_all_axes_explicit else out_s else: s = out_s # TODO(yashkatariya): Add `Layout` to SDS. out.append(api.ShapeDtypeStruct(a.shape, a.dtype, sharding=s, weak_type=a.weak_type)) return tree_unflatten(p.out_tree, out) def jit_evict_fn(self): self._clear_cache() _create_pjit_jaxpr.evict_function(self._fun) # pytype: disable=attribute-error _infer_params_cached.cache_clear() def _split_layout_and_sharding(entries): entries_flat, treedef = tree_flatten(entries, is_leaf=lambda x: x is None) layouts, shardings = [], [] for e in entries_flat: if isinstance(e, Format): layouts.append(e.device_local_layout) shardings.append(e.sharding) elif isinstance(e, (DeviceLocalLayout, AutoLayout)): raise ValueError( '`jax.jit` does not accept device-local layouts directly. Create ' 'a `Format` instance wrapping this device-local layout and pass ' f'that to `jit` instead. Got {e}') else: layouts.append(None) shardings.append(e) assert len(layouts) == len(shardings) return tree_unflatten(treedef, layouts), tree_unflatten(treedef, shardings) def _parse_jit_arguments(fun: Callable, *, in_shardings: Any, out_shardings: Any, static_argnums: int | Sequence[int] | None, static_argnames: str | Iterable[str] | None, donate_argnums: int | Sequence[int] | None, donate_argnames: str | Iterable[str] | None, keep_unused: bool, device: xc.Device | None, backend: str | None, inline: bool, abstracted_axes: Any | None, compiler_options: dict[str, Any] | None, use_resource_env: bool) -> PjitInfo: """Parses the arguments to jit/pjit. Performs any preprocessing and validation of the arguments that we can do ahead of time before the jit()-ed function is invoked. """ if abstracted_axes and not config.dynamic_shapes.value: raise ValueError("abstracted_axes must be used with --jax_dynamic_shapes") check_callable(fun) if backend is not None or device is not None: warnings.warn( 'backend and device argument on jit is deprecated. You can use' ' `jax.device_put(..., jax.local_devices(backend="cpu")[0])` on the' ' inputs to the jitted function to get the same behavior.', DeprecationWarning, ) if device is not None and backend is not None: raise ValueError("can't specify both a device and a backend for jit, " f"got {device=} and {backend=}") if in_shardings is not None and not isinstance(in_shardings, UnspecifiedValue): raise ValueError('If backend or device is specified on jit, then ' 'in_shardings should not be specified.') if out_shardings is not None and not isinstance(out_shardings, UnspecifiedValue): raise ValueError('If backend or device is specified on jit, then ' 'out_shardings should not be specified.') if isinstance(in_shardings, list): # To be a tree prefix of the positional args tuple, in_axes can never be a # list: if in_axes is not a leaf, it must be a tuple of trees. However, # in cases like these users expect tuples and lists to be treated # essentially interchangeably, so we canonicalize lists to tuples here # rather than raising an error. https://github.com/jax-ml/jax/issues/2367 in_shardings = tuple(in_shardings) in_layouts, in_shardings = _split_layout_and_sharding(in_shardings) out_layouts, out_shardings = _split_layout_and_sharding(out_shardings) in_shardings = prepare_axis_resources(in_shardings, 'in_shardings') out_shardings = prepare_axis_resources(out_shardings, 'out_shardings', allow_unconstrained_dims=True) user_specified_in_shardings = (in_shardings is not None and not isinstance(in_shardings, UnspecifiedValue)) in_shardings_leaves, in_shardings_treedef = none_lr.flatten(in_shardings) out_shardings_leaves, out_shardings_treedef = none_lr.flatten(out_shardings) in_layouts_leaves, in_layouts_treedef = none_lr.flatten(in_layouts) out_layouts_leaves, out_layouts_treedef = none_lr.flatten(out_layouts) fun_sourceinfo = api_util.fun_sourceinfo(fun) fun_signature = api_util.fun_signature(fun) donate_argnums, donate_argnames, static_argnums, static_argnames = resolve_argnums( fun, fun_signature, donate_argnums, donate_argnames, static_argnums, static_argnames) compiler_options_kvs = (() if compiler_options is None else tuple(compiler_options.items())) return PjitInfo( fun_sourceinfo=fun_sourceinfo, fun_signature=fun_signature, user_specified_in_shardings=user_specified_in_shardings, in_shardings_treedef=in_shardings_treedef, in_shardings_leaves=tuple(in_shardings_leaves), out_shardings_treedef=out_shardings_treedef, out_shardings_leaves=tuple(out_shardings_leaves), in_layouts_treedef=in_layouts_treedef, in_layouts_leaves=tuple(in_layouts_leaves), out_layouts_treedef=out_layouts_treedef, out_layouts_leaves=tuple(out_layouts_leaves), static_argnums=static_argnums, static_argnames=static_argnames, donate_argnums=donate_argnums, donate_argnames=donate_argnames, device=device, backend=backend, keep_unused=keep_unused, inline=inline, abstracted_axes=abstracted_axes, use_resource_env=use_resource_env, compiler_options_kvs=compiler_options_kvs) def make_jit(fun: Callable, *, in_shardings: Any, out_shardings: Any, static_argnums: int | Sequence[int] | None, static_argnames: str | Iterable[str] | None, donate_argnums: int | Sequence[int] | None, donate_argnames: str | Iterable[str] | None, keep_unused: bool, device: xc.Device | None, backend: str | None, inline: bool, abstracted_axes: Any | None, compiler_options: dict[str, Any] | None, use_resource_env: bool) -> Any: """jit() and pjit() are thin wrappers around this function.""" jit_info = _parse_jit_arguments( fun, in_shardings=in_shardings, out_shardings=out_shardings, static_argnums=static_argnums, static_argnames=static_argnames, donate_argnums=donate_argnums, donate_argnames=donate_argnames, keep_unused=keep_unused, device=device, backend=backend, inline=inline, abstracted_axes=abstracted_axes, compiler_options=compiler_options, use_resource_env=use_resource_env) return _cpp_pjit(fun, jit_info) class PjitParams(NamedTuple): consts: list[Any] # Only jaxpr constants, we can't keep other arguments alive params: dict[str, Any] in_avals: tuple[core.AbstractValue, ...] in_tree: PyTreeDef out_tree: PyTreeDef donated_invars: tuple[bool, ...] arg_names: tuple[str, ...] num_consts: int attrs_tracked: list[tuple[PyTreeDef, PyTreeDef, tuple[Any, str, Any]]] box_data: list def _infer_params_impl( fun: Callable, ji: PjitInfo, ctx_mesh: mesh_lib.Mesh | None, dbg: core.DebugInfo, args: tuple[Any, ...], kwargs: dict[str, Any], in_avals: tuple[core.AbstractValue, ...] | None, ) -> tuple[PjitParams, list[Any]]: util.test_event("pjit._infer_params_impl", fun) have_kwargs = bool(kwargs) if have_kwargs and ji.user_specified_in_shardings: raise ValueError( "pjit does not support kwargs when in_shardings is specified.") if ctx_mesh is not None: if (ji.backend or ji.device) and not ctx_mesh.empty: raise ValueError( "Mesh context manager should not be used with jit when backend or " "device is also specified as an argument to jit.") axes_specs = _flat_axes_specs(ji.abstracted_axes, *args, **kwargs) f = lu.wrap_init(fun, debug_info=dbg) f, dyn_args = argnums_partial_except(f, ji.static_argnums, args, allow_invalid=True) del args f, dyn_kwargs = argnames_partial_except(f, ji.static_argnames, kwargs) del kwargs dyn_args, dyn_kwargs, box_data = _flatten_boxes(dbg, dyn_args, dyn_kwargs) if box_data: f = _handle_boxes(f, dbg) explicit_args, in_tree = tree_flatten((dyn_args, dyn_kwargs)) flat_fun, out_tree = flatten_fun(f, in_tree) flat_fun, explicit_args = hoist_obj_attrs(flat_fun, explicit_args) if (ji.donate_argnums or ji.donate_argnames) and not config.debug_nans.value: donated_invars = donation_vector(ji.donate_argnums, ji.donate_argnames, in_tree) else: donated_invars = (False,) * len(explicit_args) # If backend or device is set as an arg on jit, then resolve them to # in_shardings and out_shardings as if user passed in in_shardings # and out_shardings. device_or_backend_set = bool(ji.backend or ji.device) if device_or_backend_set: sharding = _create_sharding_with_device_backend(ji.device, ji.backend) leaves, treedef = tree_flatten(sharding) in_shardings_leaves = out_shardings_leaves = tuple(leaves) in_shardings_treedef = out_shardings_treedef = treedef else: in_shardings_leaves = tuple( _create_sharding_for_array(ctx_mesh, x, 'in_shardings', 'jit') for x in ji.in_shardings_leaves) in_shardings_treedef = ji.in_shardings_treedef out_shardings_leaves = tuple( _create_sharding_for_array(ctx_mesh, x, 'out_shardings', 'jit') for x in ji.out_shardings_leaves) out_shardings_treedef = ji.out_shardings_treedef assert None not in in_shardings_leaves assert None not in out_shardings_leaves in_type: core.InputType | tuple[core.AbstractValue, ...] if config.dynamic_shapes.value: assert in_avals is None in_type = pe.infer_lambda_input_type(axes_specs, explicit_args) in_avals = tuple(a for a, e in in_type if e) elif box_data: in_type = in_avals = tuple(core.shaped_abstractify(x) for x in explicit_args) # type: ignore else: in_type = in_avals # type: ignore in_type = tuple(core.AvalQDD(a, core.cur_qdd(x)) if a.has_qdd # type: ignore else a for a, x in zip(in_type, explicit_args)) assert in_avals is not None in_shardings_flat, in_layouts_flat = _process_in_axis_resources( in_shardings_treedef, in_shardings_leaves, ji.in_layouts_treedef, ji.in_layouts_leaves, in_avals, in_tree, flat_fun.debug_info, device_or_backend_set, have_kwargs) attr_token = _attr_cache_index(flat_fun, in_type) jaxpr, consts, out_avals, attrs_tracked = _create_pjit_jaxpr( flat_fun, in_type, attr_token, IgnoreKey(ji.inline)) if config.mutable_array_checks.value: _check_no_aliased_closed_over_refs(dbg, (*jaxpr.consts, *consts), explicit_args) _attr_cachedata_update(flat_fun, in_type, attr_token, attrs_tracked) out_shardings_flat, out_layouts_flat = _check_and_canonicalize_out_shardings( out_shardings_treedef, out_shardings_leaves, ji.out_layouts_treedef, ji.out_layouts_leaves, HashableFunction(out_tree, closure=()), tuple(out_avals), jaxpr.jaxpr._debug_info, device_or_backend_set) assert len(explicit_args) == len(in_shardings_flat) == len(in_layouts_flat) if config.dynamic_shapes.value: implicit_args = _extract_implicit_args( cast(core.InputType, in_type), explicit_args) else: implicit_args = [] args_flat = [*implicit_args, *explicit_args] num_attrs_in = sum(init_tree.num_leaves for init_tree, _, (_, _, kind) in attrs_tracked if kind in (pe.ReadWrite, pe.BoxAttr)) num_extra_args = len(implicit_args) + num_attrs_in + len(consts) in_shardings_flat = (UNSPECIFIED,) * num_extra_args + in_shardings_flat in_layouts_flat = (None,) * num_extra_args + in_layouts_flat donated_invars = (False,) * num_extra_args + donated_invars assert (len(in_shardings_flat) == len(in_layouts_flat) == len(donated_invars) == num_attrs_in + len(consts) + len(args_flat)) params = dict( jaxpr=jaxpr, in_shardings=in_shardings_flat, out_shardings=out_shardings_flat, in_layouts=in_layouts_flat, out_layouts=out_layouts_flat, donated_invars=donated_invars, ctx_mesh=ctx_mesh, name=fun_qual_name(flat_fun), keep_unused=ji.keep_unused, inline=ji.inline, compiler_options_kvs=ji.compiler_options_kvs, ) return PjitParams(consts, params, in_avals, in_tree, out_tree(), donated_invars, dbg.arg_names, len(consts), attrs_tracked, box_data), args_flat class InferParamsCacheEntry: """Mutable value object for _infer_params_cached.""" __slots__ = ['pjit_params'] pjit_params: PjitParams | None def __init__(self): self.pjit_params = None # We use an outer cache that is keyed on the signature of the arguments, but # when populating a cache entry using _infer_params_impl, we need to provide # actual arguments. In principle, we could refactor _infer_params_impl to look # only at an argument signature instead of args/kwargs in those cases that we # cache, but this was a more minimal change. @util.weakref_lru_cache def _infer_params_cached( fun: Callable, jit_info: PjitInfo, signature: jax_jit.ArgumentSignature, in_avals: tuple[core.AbstractValue, ...], ctx_mesh: mesh_lib.Mesh | None, ) -> InferParamsCacheEntry: return InferParamsCacheEntry() def _infer_params( fun: Callable, ji: PjitInfo, args: tuple[Any, ...], kwargs: dict[str, Any] ) -> tuple[PjitParams, list[Any]]: if ji.use_resource_env: # pjit phys_mesh = mesh_lib.thread_resources.env.physical_mesh with (_internal_use_concrete_mesh(phys_mesh), mesh_lib.use_abstract_mesh(phys_mesh.abstract_mesh)): return _infer_params_internal(fun, ji, args, kwargs) return _infer_params_internal(fun, ji, args, kwargs) def _infer_params_internal( fun: Callable, ji: PjitInfo, args: tuple[Any, ...], kwargs: dict[str, Any] ) -> tuple[PjitParams, list[Any]]: ctx_mesh = mesh_lib.get_concrete_mesh() dbg = debug_info( 'jit', fun, args, kwargs, static_argnums=ji.static_argnums, static_argnames=ji.static_argnames, sourceinfo=ji.fun_sourceinfo, signature=ji.fun_signature) any_boxes = any(isinstance(x, (Box, List)) for x in tree_leaves((args, kwargs))) if config.dynamic_shapes.value or any_boxes: # don't use the cache p, args_flat = _infer_params_impl(fun, ji, ctx_mesh, dbg, args, kwargs, in_avals=None) return p, p.consts + args_flat signature, dynargs = jax_jit.parse_arguments( args, tuple(kwargs.values()), tuple(kwargs.keys()), ji.static_argnums, ji.static_argnames, tree_util.default_registry) avals = _infer_input_type(fun, dbg, dynargs) entry = _infer_params_cached(fun, ji, signature, avals, ctx_mesh) if entry.pjit_params is None: p, args_flat = _infer_params_impl( fun, ji, ctx_mesh, dbg, args, kwargs, in_avals=avals) if p.attrs_tracked or p.box_data or p.params['jaxpr'].jaxpr.is_high: return p, p.consts + args_flat entry.pjit_params = p return entry.pjit_params, entry.pjit_params.consts + dynargs def _infer_input_type(fun: Callable, dbg: core.DebugInfo, explicit_args) -> tuple[core.AbstractValue, ...]: avals = [] try: for i, x in enumerate(explicit_args): avals.append(core.shaped_abstractify(x)) except OverflowError: arg_path = f"argument path is {dbg.arg_names[i]}" # pytype: disable=name-error raise OverflowError( "An overflow was encountered while parsing an argument to a jitted " f"computation, whose {arg_path}." ) from None except TypeError: arg_description = f"path {dbg.arg_names[i]}" # pytype: disable=name-error raise TypeError( f"Error interpreting argument to {fun} as an abstract array." f" The problematic value is of type {type(x)} and was passed to" # pytype: disable=name-error f" the function at {arg_description}.\n" "This typically means that a jit-wrapped function was called with a non-array" " argument, and this argument was not marked as static using the" " static_argnums or static_argnames parameters of jax.jit." ) from None if config.mutable_array_checks.value: _check_no_aliased_ref_args(dbg, avals, explicit_args) return tuple(avals) def _extract_implicit_args( in_type: Sequence[tuple[core.AbstractValue, bool]], explicit_args: Sequence[Any] ) -> Sequence[core.Tracer]: """ Given an input type and explicitly-passed arguments (per the user-facing API calling convention), extract implicit axis size arguments from shapes of explicit arguments (for the trace-time / jaxpr-level calling convention). """ # First, using `in_type` construct a list to represent the full argument list, # leaving the implicit arguments as None placeholders for now. explicit_args_ = iter(explicit_args) args = [next(explicit_args_) if expl else None for _, expl in in_type] assert next(explicit_args_, None) is None del explicit_args, explicit_args_ # Next, populate the implicit arguments using the DBIdxs in `in_type`. for i, (aval, explicit) in enumerate(in_type): if not explicit or not isinstance(aval, core.DShapedArray): continue # can't populate an implicit argument arg = args[i] assert arg is not None for d1, d2 in zip(aval.shape, arg.aval.shape): if isinstance(d1, core.DBIdx): if args[d1.val] is None: args[d1.val] = d2 assert core.same_referent(args[d1.val], d2) assert all(x is not None for x in args) return [x for x, (_, e) in zip(args, in_type) if not e] # type: ignore def _flat_axes_specs(abstracted_axes, *args, **kwargs ) -> list[pe.AbstractedAxesSpec] | None: if abstracted_axes is None: return None if kwargs: raise NotImplementedError def ax_leaf(l): return (isinstance(l, dict) and all_leaves(l.values()) or isinstance(l, tuple) and all_leaves(l, lambda x: x is None)) return broadcast_prefix(abstracted_axes, args, ax_leaf) class JitWrapped(stages.Wrapped): def eval_shape(self, *args, **kwargs): """See ``jax.eval_shape``.""" raise NotImplementedError def trace(self, *args, **kwargs) -> stages.Traced: raise NotImplementedError # in_shardings and out_shardings can't be None as the default value # because `None` means that the input is fully replicated. def pjit( fun: Callable, in_shardings: Any = UNSPECIFIED, out_shardings: Any = UNSPECIFIED, static_argnums: int | Sequence[int] | None = None, static_argnames: str | Iterable[str] | None = None, donate_argnums: int | Sequence[int] | None = None, donate_argnames: str | Iterable[str] | None = None, keep_unused: bool = False, device: xc.Device | None = None, backend: str | None = None, inline: bool = False, abstracted_axes: Any | None = None, compiler_options: dict[str, Any] | None = None, ) -> JitWrapped: """Makes ``fun`` compiled and automatically partitioned across multiple devices. NOTE: This function is now equivalent to jax.jit please use that instead. The returned function has semantics equivalent to those of ``fun``, but is compiled to an XLA computation that runs across multiple devices (e.g. multiple GPUs or multiple TPU cores). This can be useful if the jitted version of ``fun`` would not fit in a single device's memory, or to speed up ``fun`` by running each operation in parallel across multiple devices. The partitioning over devices happens automatically based on the propagation of the input partitioning specified in ``in_shardings`` and the output partitioning specified in ``out_shardings``. The resources specified in those two arguments must refer to mesh axes, as defined by the :py:func:`jax.sharding.Mesh` context manager. Note that the mesh definition at :func:`~pjit` application time is ignored, and the returned function will use the mesh definition available at each call site. Inputs to a :func:`~pjit`'d function will be automatically partitioned across devices if they're not already correctly partitioned based on ``in_shardings``. In some scenarios, ensuring that the inputs are already correctly pre-partitioned can increase performance. For example, if passing the output of one :func:`~pjit`'d function to another :func:`~pjit`’d function (or the same :func:`~pjit`’d function in a loop), make sure the relevant ``out_shardings`` match the corresponding ``in_shardings``. .. note:: **Multi-process platforms:** On multi-process platforms such as TPU pods, :func:`~pjit` can be used to run computations across all available devices across processes. To achieve this, :func:`~pjit` is designed to be used in SPMD Python programs, where every process is running the same Python code such that all processes run the same :func:`~pjit`'d function in the same order. When running in this configuration, the mesh should contain devices across all processes. All inputs arguments must be globally shaped. ``fun`` will still be executed across *all* devices in the mesh, including those from other processes, and will be given a global view of the data spread across multiple processes as a single array. The SPMD model also requires that the same multi-process :func:`~pjit`'d functions must be run in the same order on all processes, but they can be interspersed with arbitrary operations running in a single process. Args: fun: Function to be compiled. Should be a pure function, as side-effects may only be executed once. Its arguments and return value should be arrays, scalars, or (nested) standard Python containers (tuple/list/dict) thereof. Positional arguments indicated by ``static_argnums`` can be anything at all, provided they are hashable and have an equality operation defined. Static arguments are included as part of a compilation cache key, which is why hash and equality operators must be defined. in_shardings: Pytree of structure matching that of arguments to ``fun``, with all actual arguments replaced by resource assignment specifications. It is also valid to specify a pytree prefix (e.g. one value in place of a whole subtree), in which case the leaves get broadcast to all values in that subtree. The ``in_shardings`` argument is optional. JAX will infer the shardings from the input :py:class:`jax.Array`'s, and defaults to replicating the input if the sharding cannot be inferred. The valid resource assignment specifications are: - :py:class:`Sharding`, which will decide how the value will be partitioned. With this, using a mesh context manager is not required. - :py:obj:`None` is a special case whose semantics are: - if the mesh context manager is *not* provided, JAX has the freedom to choose whatever sharding it wants. For in_shardings, JAX will mark is as replicated but this behavior can change in the future. For out_shardings, we will rely on the XLA GSPMD partitioner to determine the output shardings. - If the mesh context manager is provided, None will imply that the value will be replicated on all devices of the mesh. - For backwards compatibility, in_shardings still supports ingesting :py:class:`PartitionSpec`. This option can *only* be used with the mesh context manager. - :py:class:`PartitionSpec`, a tuple of length at most equal to the rank of the partitioned value. Each element can be a :py:obj:`None`, a mesh axis or a tuple of mesh axes, and specifies the set of resources assigned to partition the value's dimension matching its position in the spec. The size of every dimension has to be a multiple of the total number of resources assigned to it. out_shardings: Like ``in_shardings``, but specifies resource assignment for function outputs. The ``out_shardings`` argument is optional. If not specified, :py:func:`jax.jit` will use GSPMD's sharding propagation to determine how to shard the outputs. static_argnums: An optional int or collection of ints that specify which positional arguments to treat as static (compile-time constant). Operations that only depend on static arguments will be constant-folded in Python (during tracing), and so the corresponding argument values can be any Python object. Static arguments should be hashable, meaning both ``__hash__`` and ``__eq__`` are implemented, and immutable. Calling the jitted function with different values for these constants will trigger recompilation. Arguments that are not arrays or containers thereof must be marked as static. If ``static_argnums`` is not provided, no arguments are treated as static. static_argnames: An optional string or collection of strings specifying which named arguments to treat as static (compile-time constant). See the comment on ``static_argnums`` for details. If not provided but ``static_argnums`` is set, the default is based on calling ``inspect.signature(fun)`` to find corresponding named arguments. donate_argnums: Specify which positional argument buffers are "donated" to the computation. It is safe to donate argument buffers if you no longer need them once the computation has finished. In some cases XLA can make use of donated buffers to reduce the amount of memory needed to perform a computation, for example recycling one of your input buffers to store a result. You should not reuse buffers that you donate to a computation, JAX will raise an error if you try to. By default, no argument buffers are donated. If neither ``donate_argnums`` nor ``donate_argnames`` is provided, no arguments are donated. If ``donate_argnums`` is not provided but ``donate_argnames`` is, or vice versa, JAX uses :code:`inspect.signature(fun)` to find any positional arguments that correspond to ``donate_argnames`` (or vice versa). If both ``donate_argnums`` and ``donate_argnames`` are provided, ``inspect.signature`` is not used, and only actual parameters listed in either ``donate_argnums`` or ``donate_argnames`` will be donated. For more details on buffer donation see the `FAQ `_. donate_argnames: An optional string or collection of strings specifying which named arguments are donated to the computation. See the comment on ``donate_argnums`` for details. If not provided but ``donate_argnums`` is set, the default is based on calling ``inspect.signature(fun)`` to find corresponding named arguments. keep_unused: If `False` (the default), arguments that JAX determines to be unused by `fun` *may* be dropped from resulting compiled XLA executables. Such arguments will not be transferred to the device nor provided to the underlying executable. If `True`, unused arguments will not be pruned. device: This argument is deprecated. Please put your arguments on the device you want before passing them to jit. Optional, the Device the jitted function will run on. (Available devices can be retrieved via :py:func:`jax.devices`.) The default is inherited from XLA's DeviceAssignment logic and is usually to use ``jax.devices()[0]``. backend: This argument is deprecated. Please put your arguments on the backend you want before passing them to jit. Optional, a string representing the XLA backend: ``'cpu'``, ``'gpu'``, or ``'tpu'``. Returns: A wrapped version of ``fun``, set up for just-in-time compilation and automatically partitioned by the mesh available at each call site. For example, a convolution operator can be automatically partitioned over an arbitrary set of devices by a single :func:`~pjit` application: >>> import jax >>> import jax.numpy as jnp >>> import numpy as np >>> from jax.sharding import Mesh, PartitionSpec >>> from jax.experimental.pjit import pjit >>> >>> x = jnp.arange(8, dtype=jnp.float32) >>> f = pjit(lambda x: jax.numpy.convolve(x, jnp.asarray([0.5, 1.0, 0.5]), 'same'), ... in_shardings=None, out_shardings=PartitionSpec('devices')) >>> with Mesh(np.array(jax.devices()), ('devices',)): ... print(f(x)) # doctest: +SKIP [ 0.5 2. 4. 6. 8. 10. 12. 10. ] """ return make_jit( fun, in_shardings=in_shardings, out_shardings=out_shardings, static_argnums=static_argnums, static_argnames=static_argnames, donate_argnums=donate_argnums, donate_argnames=donate_argnames, keep_unused=keep_unused, device=device, backend=backend, inline=inline, abstracted_axes=abstracted_axes, compiler_options=compiler_options, use_resource_env=True) def hashable_pytree(pytree): vals, treedef = tree_flatten(pytree) vals = tuple(vals) return HashableFunction(lambda: tree_unflatten(treedef, vals), closure=(treedef, vals)) def _create_sharding_for_array(mesh, x, name, api_name): if x is None and (mesh is None or mesh.empty): return UNSPECIFIED if isinstance(x, (AUTO, UnspecifiedValue, Sharding)): return x if mesh is None: msg = ('jax.jit only supports `Sharding`s being passed to' f' {name}. Looks like you are passing either `PartitionSpec` or `None`' f' which is not allowed in jax.jit.\n') if name == 'in_shardings': msg += (f'Note that {name} argument is optional. JAX will infer the shardings' " from the input jax.Array's and will default to replicating the" ' input if the sharding cannot be inferred.') elif name == 'out_shardings': msg += (f'Note that {name} is optional. If not specified, jax.jit will' " use GSPMD's sharding propagation to figure out what the sharding" ' of the output(s) should be.') raise RuntimeError(msg) if mesh.empty: raise RuntimeError( f'{api_name} requires a non-empty mesh if you are passing' f' `PartitionSpec`s or `None` to {name}! Is a mesh defined at the call' f' site? Alternatively, provide `Sharding`s to {name} and' ' then the mesh context manager is not required.') # A nice user error is raised in prepare_axis_resources. assert x is None or isinstance(x, PartitionSpec), x return sharding_impls.create_mesh_pspec_sharding(mesh, x) def _create_sharding_with_device_backend(device, backend): if device is not None: assert backend is None out = SingleDeviceSharding(device) elif backend is not None: assert device is None out = SingleDeviceSharding(xb.get_backend(backend).local_devices()[0]) else: raise AssertionError('Unreachable!') out._device_backend = True return out def flatten_axis_resources(what, tree, shardings, tupled_args): try: return tuple(flatten_axes(what, tree, shardings, tupled_args=tupled_args)) except ValueError: pass # Raise a tree prefix error below # Tree leaves are always valid prefixes, so if there was a prefix error as # assumed here, axis_resources must not be a leaf. assert not treedef_is_leaf(tree_structure(shardings)) # Check the type directly rather than using isinstance because of namedtuples. if tupled_args and (type(shardings) is not tuple or len(shardings) != len(tree.children())): # We know axis_resources is meant to be a tuple corresponding to the args # tuple, but while it is a non-leaf pytree, either it wasn't a tuple or it # wasn't the right length. msg = (f"{what} specification must be a tree prefix of the positional " f"arguments tuple passed to the `pjit`-decorated function. In " f"particular, {what} must either be a None, a PartitionSpec, or " f"a tuple of length equal to the number of positional arguments.") # If `tree` represents an args tuple, then `axis_resources` must be a tuple. # TODO(mattjj,apaszke): disable implicit list casts, remove 'or list' below if type(shardings) is not tuple: msg += f" But {what} is not a tuple: got {type(shardings)} instead." elif len(shardings) != len(tree.children()): msg += (f" But {what} is the wrong length: got a tuple or list of length " f"{len(shardings)} for an args tuple of length " f"{len(tree.children())}.") # As an extra hint, let's check if the user just forgot to wrap # shardings in a singleton tuple. if len(tree.children()) == 1: try: flatten_axes(what, tree, (shardings,)) except ValueError: pass # That's not the issue. else: msg += (f" Given the corresponding argument being " f"passed, it looks like {what} might need to be wrapped in " f"a singleton tuple.") raise ValueError(msg) axis_tree = shardings # Because we only have the `tree` treedef and not the full pytree here, # we construct a dummy tree to compare against. Revise this in callers? dummy_tree = tree_unflatten(tree, [PytreeLeaf()] * tree.num_leaves) errors = prefix_errors(axis_tree, dummy_tree) if errors: e = errors[0] # Only show information about the first disagreement found. raise e(what) # At this point we've failed to find a tree prefix error. assert False, "Please open a bug report!" # This should be unreachable. class PytreeLeaf: def __repr__(self): return "pytree leaf" @util.cache(max_size=4096, trace_context_in_key=False) def _process_in_axis_resources(in_shardings_treedef, in_shardings_leaves, in_layouts_treedef, in_layouts_leaves, in_avals, in_tree, debug_info: core.DebugInfo, device_or_backend_set, kws): if not kws: in_tree, _ = treedef_children(in_tree) orig_in_shardings = tree_unflatten(in_shardings_treedef, in_shardings_leaves) # Only do this if original in_shardings are unspecified. If it is AUTO, go # via flatten_axis_resources. if isinstance(orig_in_shardings, UnspecifiedValue): in_shardings_flat = (orig_in_shardings,) * len(in_avals) else: in_shardings_flat = flatten_axis_resources( "pjit in_shardings", in_tree, orig_in_shardings, tupled_args=True) in_layouts = tree_unflatten(in_layouts_treedef, in_layouts_leaves) if in_layouts is None: in_layouts_flat = (in_layouts,) * len(in_avals) else: in_layouts_flat = flatten_axis_resources( "pjit in_layouts", in_tree, in_layouts, tupled_args=True) # TODO(dougalm,mattjj): enable debug info with attrs_tracked attrs_tracked = len(debug_info.arg_names) != len(in_avals) if not config.dynamic_shapes.value and not attrs_tracked: pjit_check_aval_sharding(in_shardings_flat, in_avals, debug_info.safe_arg_names(len(in_avals)), "pjit arguments", allow_uneven_sharding=False) check_aval_layout_compatibility( in_layouts_flat, in_avals, debug_info.safe_arg_names(len(in_avals)), "jit arguments") # type: ignore[arg-type] return in_shardings_flat, in_layouts_flat callsites_with_tracing_cache_miss: set[str] = set() def diff_tracing_cache_keys( k: tuple, oldk: tuple, debug_info: lu.DebugInfo) -> tuple[Sequence[str], int]: """Explanations of differences between the cache keys, along with diff sizes. Result: a pair of a list of explanations for differences, and the total size of the differences. The sizes are used to pick the old key with the smallest different size for the explanation that is shown to the user. """ (fun_transforms_k, fun_params_k, fun_in_type_k, (arg_in_type_k, arg_attr_data_k, arg_inline_k), ctx_k) = k (fun_transforms_ok, fun_params_ok, fun_in_type_ok, (arg_in_type_ok, arg_attr_data_ok, arg_inline_ok), ctx_ok) = oldk diffs: list[tuple[str, int]] = [] # each difference with its size def unavailable(key_field: str, what_k, what_ok): diffs.append( (f"different {key_field}:\n now: {what_k}\n != before: {what_ok}.\n" "explanation unavailable! " "please open an issue at https://github.com/jax-ml/jax.", 10)) def list_diff_size(s1: Sequence, s2: Sequence) -> int: min_len = min(len(s1), len(s2)) diff_size = max(len(s1), len(s2)) - min_len diff_size += sum(e1 != e2 for e1, e2 in zip(s1[:min_len], s2[:min_len])) return diff_size different_leaf_count = False def explain_transform_argnums_partial(param_k: tuple, param_ok: tuple): dyn_argnums_k, static_args_k = param_k dyn_argnums_ok, static_args_ok = param_ok if dyn_argnums_k != dyn_argnums_ok: diffs.append( ("different static_argnums:\n" f" dynamic argnums now {dyn_argnums_k} and before {dyn_argnums_ok}", 1)) if static_args_k != static_args_ok: diffs.append( ("different value of static args:\n" f" now {', '.join(repr(a.val) for a in static_args_k)}" f" and before {', '.join(repr(a.val) for a in static_args_ok)}", list_diff_size(static_args_k, static_args_ok))) def explain_transform_argnames_partial(param_k: tuple, param_ok: tuple): static_kwargs_k, = param_k static_kwargs_ok, = param_ok static_kwargs_k = [(k, v.val) for k, v in sorted(static_kwargs_k.val.items())] static_kwargs_ok = [(k, v.val) for k, v in sorted(static_kwargs_ok.val.items())] if static_kwargs_k != static_kwargs_ok: diffs.append( ("different value of static kwargs:\n" f" now {{{', '.join(f'{k}: {repr(v)}' for k, v in static_kwargs_k)}}}" f" and before {{{', '.join(f'{k}: {repr(v)}' for k, v in static_kwargs_ok)}}}", list_diff_size(static_kwargs_k, static_kwargs_ok))) def explain_in_tree_diff(in_tree_k: PyTreeDef, in_tree_ok: PyTreeDef): nonlocal different_leaf_count different_leaf_count = (in_tree_k.num_leaves != in_tree_ok.num_leaves) if not different_leaf_count: # Look for the special case of passing positional args as kwargs or # vice-versa; the common prefix of positional args match. args_tree_k, kwargs_tree_k = treedef_children(in_tree_k) nr_args_k = len(treedef_children(args_tree_k)) args_tree_ok, kwargs_tree_ok = treedef_children(in_tree_ok) nr_args_ok = len(treedef_children(args_tree_k)) if (treedef_children(args_tree_k)[:min(nr_args_k, nr_args_ok)] == treedef_children(args_tree_ok)[:min(nr_args_k, nr_args_ok)]): keys_k = kwargs_tree_k.node_data()[1] # type: ignore[index] keys_ok = kwargs_tree_ok.node_data()[1] # type: ignore[index] diffs.append( (("different number of args and kwargs, but same total number.\n" f" now {nr_args_k} args and kwargs " f"with keys {keys_k}\n" f" before {nr_args_ok} args and kwargs " f"with keys {keys_ok}"), abs(nr_args_ok - nr_args_k))) return in_tree_k_str = str(in_tree_k) in_tree_k_str = (in_tree_k_str if len(in_tree_k_str) < 73 else in_tree_k_str[:73] + "...") in_tree_ok_str = str(in_tree_ok) in_tree_ok_str = (in_tree_ok_str if len(in_tree_ok_str) < 73 else in_tree_ok_str[:73] + "...") diff = [f"different input pytree:\n now: {in_tree_k_str}\n" f" before: {in_tree_ok_str}"] errs = list(tree_util.equality_errors_pytreedef(in_tree_k, in_tree_ok)) for path, thing1, thing2, explanation in errs: fst, *path = path # type: ignore base = ["args", "kwargs"][fst.idx] diff.append( f" * at {base}{keystr(tuple(path))}, now {thing1} and before {thing2}," f" so {explanation}") diffs.append(("\n".join(diff), len(errs))) def explain_args_type_diff(args_k: tuple[core.AbstractValue], args_ok: tuple[core.AbstractValue]): diff_size = 0 arg_names = debug_info.safe_arg_names(len(args_k)) def arg_type_to_str(at): if hasattr(at, "str_short"): return at.str_short(short_dtypes=True) else: return str(at) args_k_str = ", ".join(f"{an}: {arg_type_to_str(at)}" for an, at in zip(arg_names, args_k)) args_k_str = args_k_str if len(args_k_str) < 73 else args_k_str[:73] + "..." diff = [f"different input types:\n types now: {args_k_str}"] add_weak_type_hint = False for name, arg_t_k, arg_t_ok in zip(arg_names, args_k, args_ok): if arg_t_k == arg_t_ok: continue this_arg_diff_size = 0 if type(arg_t_k) == type(arg_t_ok) == core.ShapedArray: s1, s2 = arg_type_to_str(arg_t_k), arg_type_to_str(arg_t_ok) this_arg_diff_size += list_diff_size(arg_t_k.shape, arg_t_ok.shape) # type: ignore if arg_t_k.weak_type != arg_t_ok.weak_type: # type: ignore s1 += f"{{weak_type={arg_t_k.weak_type}}}" # type: ignore s2 += f"{{weak_type={arg_t_ok.weak_type}}}" # type: ignore add_weak_type_hint = True this_arg_diff_size += 1 elif arg_t_k.sharding != arg_t_ok.sharding: # type: ignore s1 = arg_t_k.str_short(short_dtypes=True, mesh_axis_types=True) # type: ignore s2 = arg_t_ok.str_short(short_dtypes=True, mesh_axis_types=True) # type: ignore this_arg_diff_size += 1 else: s1, s2 = str(arg_t_k), str(arg_t_ok) diff_size += max(1, this_arg_diff_size) diff.append(f" * at {name}, now {s1} and before {s2}") if add_weak_type_hint: diff.append( "where weak_type=True often means a Python builtin numeric value, and \n" "weak_type=False means a jax.Array.\n" "See https://docs.jax.dev/en/latest/type_promotion.html#weak-types.") diffs.append(("\n".join(diff), diff_size)) if fun_transforms_k != fun_transforms_ok: if len(fun_transforms_k) != len(fun_transforms_ok): different_leaf_count = True # Skip other more precise checks unavailable("fun_transforms length", fun_transforms_k, fun_transforms_ok) else: for i, (t, ot) in enumerate(zip(fun_transforms_k, fun_transforms_ok)): t_name = t[0].__name__ if t == ot: continue # TODO(mattjj): explain box cache misses if t_name == '_handle_boxes': continue if t[0] != ot[0]: unavailable(f"fun_transforms[{i}] transform", t, ot) continue if t_name == "flatten_fun": explain_in_tree_diff(t[1][0], ot[1][0]) continue if t_name == "_argnums_partial": explain_transform_argnums_partial(t[1], ot[1]) continue if t_name == "_argnames_partial": explain_transform_argnames_partial(t[1], ot[1]) continue unavailable(f"fun_transforms.{t_name} params", t[1:], ot[1:]) continue # If we had different leaf counts, we can discard the _argnums_partial # difference. That transform sometimes occurs before the flatten_fun if different_leaf_count: diffs = [d for d in diffs if "fun_transforms._argnums_partial" not in d[0]] if fun_params_k != fun_params_ok: unavailable("fun_params", fun_params_k, fun_params_ok) if fun_in_type_k != fun_in_type_ok: unavailable("fun_in_type", fun_params_k, fun_params_ok) if arg_in_type_k != arg_in_type_ok and not different_leaf_count: explain_args_type_diff(arg_in_type_k, arg_in_type_ok) if arg_attr_data_k != arg_attr_data_ok: unavailable("arg_attr_data", arg_attr_data_k, arg_attr_data_ok) if arg_inline_k != arg_inline_ok: unavailable("arg_inline", arg_inline_k, arg_inline_ok) if ctx_k != ctx_ok: assert len(ctx_k) == len(ctx_ok) idxs = [f" [{i}]: now {c_k} and before {c_ok}" for i, (c_k, c_ok) in enumerate(zip(ctx_k, ctx_ok)) if c_k != c_ok] diffs.append( ("different tracing context, e.g. due to config or context manager.\n" "found differences at positions\n" + ", and\n".join(idxs) + "\ncompare to tuple returned by " "config.trace_context() in jax/_src/config.py.", len(idxs))) if not diffs: # Should never happen, but let's not crash unavailable("something (unexpected empty diffs)", k, oldk) diffs_and_sizes = util.unzip2(sorted(diffs, key=lambda d: d[1])) return (diffs_and_sizes[0], sum(diffs_and_sizes[1])) def explain_tracing_cache_miss( fun: lu.WrappedFun, unseen_f: bool, cache: dict, key: tuple, elapsed_sec: float): if config.check_tracer_leaks.value: return if key[3][2].val: return # No explanations for "inline" functions debug_info = fun.debug_info func_filename = debug_info.func_filename if func_filename and not source_info_util.is_user_filename(func_filename): return msg: list[str] = [] p = msg.append done = lambda: logger.log(logging.WARNING, "\n".join(msg)) callsite = source_info_util.summarize(source_info_util.current()) p(f"TRACING CACHE MISS at {callsite} costing {elapsed_sec * 1e3:.3f} ms because:") # have we seen this function before at all? src_info = "" if func_filename: src_info += f" defined at {func_filename}" if func_lineno := debug_info.func_lineno: src_info += f":{func_lineno}" func_name = debug_info.func_name if unseen_f or not cache: p(f" never seen function:\n {func_name} id={id(fun.f)}{src_info}") if callsite in callsites_with_tracing_cache_miss: p(" but seen another function defined on the same line; maybe the function is\n" " being re-defined repeatedly, preventing caching?") else: callsites_with_tracing_cache_miss.add(callsite) return done() p(f" for {func_name}{src_info}") diffs = [diff_tracing_cache_keys(key, ok, debug_info) for ok in cache.keys() if key != ok] assert diffs, "we must find some diffs if key differs from all cache keys" min_diff = min(diffs, key=lambda v: v[1]) smallest_diffs: Sequence[Sequence[str]] # the diffs for the closest keys smallest_diffs = [d[0] for d in diffs if d[1] == min_diff[1]] def indent_subsequent_lines(indent: int, msg: str) -> str: return msg.replace("\n", "\n" + " " * indent) def p_one_diff(diff: Sequence[str]): for d in diff: p(" * key with " + indent_subsequent_lines(4, d)) if len(smallest_diffs) == 1: p(" all previously seen cache keys are different. Closest previous key:") p_one_diff(smallest_diffs[0]) else: p(" all previously seen cache keys are different. " "Several previous keys are closest:") for d in smallest_diffs: p_one_diff(d) done() @partial(lu.cache, explain=explain_tracing_cache_miss) def _create_pjit_jaxpr( fun: lu.WrappedFun, in_type: core.InputType | Sequence[core.AbstractValue], attr_data: int, ignored_inline: IgnoreKey ) -> tuple[core.ClosedJaxpr, list[Any], list[core.AbstractValue], list[tuple[PyTreeDef, PyTreeDef, tuple[Any, str, Any]]]]: util.test_event("create_pjit_jaxpr") del ignored_inline # just for explain_cache_miss if config.no_tracing.value: raise RuntimeError(f"re-tracing function {fun.f} for `jit`, but " "'no_tracing' is set") with dispatch.log_elapsed_time( "Finished tracing + transforming {fun_name} for pjit in {elapsed_time:.9f} sec", fun_name=fun.__name__, event=dispatch.JAXPR_TRACE_EVENT): if config.dynamic_shapes.value: jaxpr, global_out_avals, consts = pe.trace_to_jaxpr_dynamic2( lu.annotate(fun, cast(core.InputType, in_type))) attrs_tracked = [] else: jaxpr, global_out_avals, consts, attrs_tracked = pe.trace_to_jaxpr_dynamic(fun, in_type) if config.debug_key_reuse.value: # Import here to avoid circular imports from jax.experimental.key_reuse._core import check_key_reuse_jaxpr # pytype: disable=import-error check_key_reuse_jaxpr(jaxpr) # TODO(mattjj,yashkatariya): if we take the 'true' path then we *must* fall # off the C++ dispatch fast path for correctness. Ensure that happens. if any(isinstance(c, core.Tracer) or core.typeof(c).has_qdd for c in consts): closed_jaxpr = pe.close_jaxpr(pe.convert_constvars_jaxpr(jaxpr)) final_consts = consts else: closed_jaxpr = core.ClosedJaxpr(jaxpr, consts) final_consts = [] return closed_jaxpr, final_consts, global_out_avals, attrs_tracked @util.cache(max_size=4096, trace_context_in_key=False) def _check_and_canonicalize_out_shardings( out_shardings_treedef, out_shardings_leaves, out_layouts_treedef, out_layouts_leaves, out_tree, out_avals, debug_info: core.DebugInfo, device_or_backend_set): orig_out_shardings = tree_unflatten(out_shardings_treedef, out_shardings_leaves) if isinstance(orig_out_shardings, (UnspecifiedValue, Sharding)): out_shardings_flat = (orig_out_shardings,) * len(out_avals) else: out_shardings_flat = flatten_axis_resources( "pjit out_shardings", out_tree(), orig_out_shardings, tupled_args=False) out_layouts = tree_unflatten(out_layouts_treedef, out_layouts_leaves) if out_layouts is None: out_layouts_flat = (out_layouts,) * len(out_avals) else: out_layouts_flat = flatten_axis_resources( "pjit out_layouts", out_tree(), out_layouts, tupled_args=False) if not config.dynamic_shapes.value: pjit_check_aval_sharding( out_shardings_flat, out_avals, debug_info.safe_result_paths(len(out_avals)), "pjit outputs", allow_uneven_sharding=False) check_aval_layout_compatibility( out_layouts_flat, out_avals, debug_info.safe_result_paths(len(out_avals)), "jit outputs") return out_shardings_flat, out_layouts_flat AttrRecord = tuple[object, str, PyTreeDef, list[core.AbstractValue]] _seen_attrs = weakref.WeakKeyDictionary() # type: ignore def seen_attrs_get( fun: lu.WrappedFun, in_type: core.InputType | tuple[core.AbstractValue, ...] ) -> list: cache = _seen_attrs.setdefault(fun.f, defaultdict(list)) assert fun.in_type is None or fun.in_type == in_type return cache[(fun.transforms, fun.params, in_type)] def _attr_cache_index( fun: lu.WrappedFun, in_type: core.InputType | tuple[core.AbstractValue, ...] ) -> int: cases = seen_attrs_get(fun, in_type) for i, records in enumerate(cases): for obj, attr, kind, treedef, avals in records: if kind in (pe.ReadWrite, pe.BoxAttr): val = getattr(obj, attr, dne_sentinel) vals, treedef_ = tree_flatten(val) avals_ = map(core.shaped_abstractify, vals) if treedef != treedef_ or avals != avals_: break else: return i return len(cases) def _attr_cachedata_update(fun, in_type, i, attrs_tracked): leaves = lambda obj, attr: tree_leaves(getattr(obj, attr, dne_sentinel)) records = [(obj, attr, kind, init_tree, map(core.typeof, leaves(obj, attr))) for init_tree, _, (obj, attr, kind) in attrs_tracked] cases = seen_attrs_get(fun, in_type) if i == len(cases): cases.append(records) @dataclasses.dataclass(frozen=True) class IgnoreKey: val: Any def __hash__(self): return hash(self.__class__) def __eq__(self, other): return isinstance(other, IgnoreKey) # ignore self.val! def pjit_check_aval_sharding( shardings, flat_avals, names: Sequence[str], what_aval: str, allow_uneven_sharding: bool): for aval, s, name in zip(flat_avals, shardings, names): if isinstance(s, (UnspecifiedValue, AUTO)): continue name_str = f' with pytree key path {name}' if name else '' shape = aval.shape try: # Sharding interfaces can implement `check_compatible_aval` as an optional # method to raise a more meaningful error. if hasattr(s, 'check_compatible_aval'): s.check_compatible_aval(shape) else: s._to_xla_hlo_sharding(len(shape)) except ValueError as e: raise ValueError( f'One of {what_aval}{name_str} is incompatible with its sharding ' f'annotation {s}: {e}') # Use the `OpSharding` proto to find out how many ways each dimension of # the aval is sharded. This approach will work across all # Sharding. hlo_sharding = s._to_xla_hlo_sharding(len(shape)) assert hlo_sharding is not None num_ways_dim_sharded, _ = op_shardings.get_num_ways_dim_sharded(hlo_sharding) for i, size in enumerate(num_ways_dim_sharded): if not allow_uneven_sharding and shape[i] % size != 0: raise ValueError(f"One of {what_aval}{name_str} was given the sharding " f"of {s}, which implies that " f"the global size of its dimension {i} should be " f"divisible by {size}, but it is equal to {shape[i]} " f"(full shape: {shape})") def check_aval_layout_compatibility( layouts, flat_avals, names: Sequence[str], what_aval: str): for aval, l, name in zip(flat_avals, layouts, names): if l is None or isinstance(l, AutoLayout): continue name_str = f' with pytree key path {name}' if name else '' try: l.check_compatible_aval(aval.shape) except ValueError as e: raise ValueError( f'One of {what_aval}{name_str} is incompatible with its layout ' f'annotation {l}: {e}') # -------------------- pjit rules -------------------- pjit_p = core.Primitive("pjit") pjit_p.multiple_results = True pjit_p.skip_canonicalization = True def _is_high(jaxpr, **_) -> bool: return jaxpr.jaxpr.is_high pjit_p.is_high = _is_high # type: ignore def _to_lojax(*hi_args, jaxpr, **params): # convert closed-over boxes to explicit args jaxpr, closed_over_himutables = pe.convert_const_himutables(jaxpr) hi_args = [*closed_over_himutables, *hi_args] params = _converted_mutables_add_params(len(closed_over_himutables), **params) # expand pjit params that must match number of lo inputs/outputs lo_nums_in = [len(aval.lo_ty()) for aval in jaxpr.in_aval_qdds] lo_nums_out = [len(t.lo_ty()) for t in jaxpr.out_avals] lo_muts_out = sum(len(aval.lo_ty()) for aval in jaxpr.final_aval_qdds if aval.has_qdd) params = _lojax_expand_params(lo_nums_in, lo_nums_out, lo_muts_out, **params) # collect lo input values lo_args = [lo_val for aval, x in zip(jaxpr.in_aval_qdds, hi_args) for lo_val in (aval.read_loval(x) if aval.has_qdd else aval.lower_val(x))] # lower the jaxpr and bind it using lo input values lo_jaxpr = pe.lower_jaxpr(jaxpr) all_outs = pjit_p.bind(*lo_args, jaxpr=lo_jaxpr, **params) out_mut, lo_outs = split_list(all_outs, [lo_muts_out]) # collect and apply mutations out_mut_ = iter(out_mut) in_idx = {v: i for i, v in enumerate(jaxpr.jaxpr.invars)} for v in jaxpr.jaxpr.invars: if v.final_qdd is not None: qdd = v.final_qdd lo_vals = it.islice(out_mut_, len(v.aval.lo_ty_qdd(qdd))) v.aval.update_from_loval(qdd, hi_args[in_idx[v]], *lo_vals) assert next(out_mut_, None) is None # collect output values into hi types lo_outs_ = iter(lo_outs) hi_outs = [t.raise_val(*it.islice(lo_outs_, len(t.lo_ty()))) for t in jaxpr.out_avals] assert next(lo_outs_, None) is None return hi_outs pjit_p.to_lojax = _to_lojax def _converted_mutables_add_params( n, *, donated_invars, in_shardings, in_layouts, **params): donated_invars = (False,) * n + donated_invars in_shardings = (UNSPECIFIED,) * n + in_shardings in_layouts = (None,) * n + in_layouts return dict(params, donated_invars=donated_invars, in_shardings=in_shardings, in_layouts=in_layouts) def _lojax_expand_params( nums_in, nums_out, muts_out, *, donated_invars, in_shardings, in_layouts, out_shardings, out_layouts, **params): # some pjit params match the length of hi_jaxpr.invars/outvars, so when # lowering we must expand them to match their number of lojax types def expand(ns, xs): return tuple(y for n, x in zip(ns, xs) for y in (x,) * n) donated_invars = expand(nums_in , donated_invars) in_shardings = expand(nums_in , in_shardings ) in_layouts = expand(nums_in , in_layouts ) out_shardings = expand(nums_out, out_shardings ) out_layouts = expand(nums_out, out_layouts ) # also, the lo_jaxpr has pure outputs corresponding to mutable hi_jaxpr types out_shardings = (UNSPECIFIED,) * muts_out + out_shardings out_layouts = (None,) * muts_out + out_layouts new_params = dict(params, donated_invars=donated_invars, in_shardings=in_shardings, in_layouts=in_layouts, out_shardings=out_shardings, out_layouts=out_layouts) return new_params def _resolve_in_layouts(args, jit_in_layouts, resolved_in_shardings, in_avals): # If device or backend is set, return the default layout. This is because you # can pass arrays on cpu (with untiled layouts) to jit with backend='tpu' # which causes error checks to fail. Returning the default layout allows # this to exist. It's the same for handling shardings. if pxla.check_device_backend_on_shardings(resolved_in_shardings): return (None,) * len(jit_in_layouts) resolved_in_layouts = [] for arg, jit_in_l, rs, aval in safe_zip( args, jit_in_layouts, resolved_in_shardings, in_avals): committed = getattr(arg, '_committed', True) # `arg_layout` is only used for checking purposes in the `else` branch # below. We cannot replace default layout with None to raise nicer errors. # `dispatch_arg_layout` replaces default layouts with `None` to simplify # dispatch and lowering logic downstream. if hasattr(arg, 'format'): arg_layout = arg.format.device_local_layout dispatch_arg_layout = (None if pxla.is_default_layout(arg_layout, rs, aval) else arg_layout) else: arg_layout, dispatch_arg_layout = None, None # Sharding can be unspecified when array is committed if it's a PmapSharding. is_pmap_sharding = (isinstance(rs, UnspecifiedValue) or isinstance(getattr(arg, 'sharding', None), PmapSharding)) if jit_in_l is None: if committed: if is_pmap_sharding: resolved_in_layouts.append(None) else: resolved_in_layouts.append(dispatch_arg_layout) else: resolved_in_layouts.append(None) else: # arg_layout can be None because some backends don't implement the # required layout methods. Hence `arr.format` can return # `Format(None, sharding)` if (committed and not is_pmap_sharding and arg_layout is not None and not pxla.is_user_xla_layout_equal(jit_in_l, arg_layout)): extra_msg = '' if isinstance(jit_in_l, AutoLayout): extra_msg = ( ' The layout given to `jax.jit` is `DeviceLocalLayout.AUTO` but' ' the corresponding argument passed is a `jax.Array` with a' ' concrete layout. Consider passing a `jax.ShapeDtypeStruct`' ' instead of `jax.Array` as an argument to the jitted function ' ' when using `DeviceLocalLayout.AUTO`.' ) raise ValueError('Layout passed to jit does not match the layout ' 'on the respective arg. ' f'Got pjit layout: {jit_in_l},\n' f'arg layout: {arg_layout} for ' f'arg shape: {core.shaped_abstractify(arg).str_short()}.' f'{extra_msg}') jit_in_l = (None if isinstance(jit_in_l, DeviceLocalLayout) and pxla.is_default_layout(jit_in_l, rs, aval) else jit_in_l) resolved_in_layouts.append(jit_in_l) return tuple(resolved_in_layouts) def _resolve_out_layouts(out_layouts, out_shardings, out_avals): new_out_layouts = [] for out_l, out_s, out_aval in safe_zip(out_layouts, out_shardings, out_avals): if out_l is None: new_out_layouts.append(None) elif (isinstance(out_l, DeviceLocalLayout) and pxla.is_default_layout(out_l, out_s, out_aval)): new_out_layouts.append(None) else: new_out_layouts.append(out_l) return tuple(new_out_layouts) def finalize_arg_sharding(arg_s, committed): if isinstance(arg_s, UnspecifiedValue): return arg_s else: if committed: # If the arg has a PmapSharding, then reshard it unconditionally. return UNSPECIFIED if isinstance(arg_s, PmapSharding) else arg_s else: assert isinstance(arg_s, Sharding) if dispatch.is_single_device_sharding(arg_s): return UNSPECIFIED raise NotImplementedError('Having uncommitted Array sharded on ' 'multiple devices is not supported.') def _resolve_in_shardings(args, pjit_in_shardings: Sequence[PjitSharding] ) -> Sequence[PjitSharding]: # If True, means that device or backend is set by the user on pjit and it # has the same semantics as device_put i.e. doesn't matter which device the # arg is on, reshard it to the device mentioned. So don't do any of the # checks and just return the pjit_in_shardings directly. `shard_args` will # handle the resharding. if pxla.check_device_backend_on_shardings(pjit_in_shardings): return pjit_in_shardings committed_arg_shardings = [] for a in args: arg_s = getattr(a, 'sharding', None) # arg sharding can be None in case of ShapeDtypeStruct. jax.Array does # not allow None as the sharding. if arg_s is None: continue # Don't consider PmapSharding inputs as committed. They will get resharded # unconditionally. if isinstance(arg_s, PmapSharding): continue if getattr(a, '_committed', True): committed_arg_shardings.append((arg_s, stages.MismatchType.ARG_SHARDING, None)) resolved_in_shardings: list[PjitSharding] = [] for arg, pjit_in_s in zip(args, pjit_in_shardings): # arg sharding can be None in case of ShapeDtypeStruct. jax.Array does # not allow None as the sharding. arg_s, committed = ((arg.sharding, getattr(arg, '_committed', True)) if hasattr(arg, 'sharding') and arg.sharding is not None else (UNSPECIFIED, False)) if isinstance(arg_s, NamedSharding) and arg_s.mesh.empty: arg_s, committed = UNSPECIFIED, False if isinstance(pjit_in_s, UnspecifiedValue): resolved_in_shardings.append(finalize_arg_sharding(arg_s, committed)) else: if (isinstance(arg, np.ndarray) and not pjit_in_s.is_fully_replicated and # type: ignore[union-attr] xb.process_count() > 1): raise ValueError( 'Passing non-trivial shardings for numpy ' 'inputs is not allowed. To fix this error, either specify a ' 'replicated sharding explicitly or use ' '`jax.make_array_from_process_local_data(...)` ' 'to convert your host local numpy inputs to a jax.Array which you ' 'can pass to jit. ' 'If the numpy input is the same on each process, then you can use ' '`jax.make_array_from_callback(...) to create a `jax.Array` which ' 'you can pass to jit. ' f'Got arg shape: {arg.shape}, arg value: {arg}') if not isinstance(arg_s, UnspecifiedValue) and arg_s._is_concrete: # jax.jit does not allow resharding across different memory kinds even # if the argument is uncommitted. Use jax.device_put for those cases, # either outside or inside jax.jit. if pjit_in_s.memory_kind != arg_s.memory_kind: # type: ignore[union-attr] raise ValueError( 'Memory kinds passed to jax.jit does not match memory kind on the' f' respective arg. Got pjit memory kind: {pjit_in_s.memory_kind}, ' # type: ignore[union-attr] f'arg memory kind: {arg_s.memory_kind} for ' f'arg shape: {core.shaped_abstractify(arg).str_short()}') if (committed and not isinstance(arg_s, PmapSharding) and not op_shardings.are_op_shardings_equal( pjit_in_s._to_xla_hlo_sharding(arg.ndim), # type: ignore[union-attr] arg_s._to_xla_hlo_sharding(arg.ndim))): raise ValueError('Sharding passed to pjit does not match the sharding ' 'on the respective arg. ' f'Got pjit sharding: {pjit_in_s},\n' f'arg sharding: {arg_s} for ' f'arg shape: {core.shaped_abstractify(arg).str_short()}') resolved_in_shardings.append(pjit_in_s) return tuple(resolved_in_shardings) def _resolve_and_lower( args, jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, lowering_platforms, lowering_parameters, pgle_profiler, compiler_options_kvs): in_shardings = _resolve_in_shardings(args, in_shardings) in_layouts = _resolve_in_layouts(args, in_layouts, in_shardings, jaxpr.in_avals) out_layouts = _resolve_out_layouts(out_layouts, out_shardings, jaxpr.out_avals) return _pjit_lower( jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs, lowering_platforms=lowering_platforms, lowering_parameters=lowering_parameters, pgle_profiler=pgle_profiler) _pgle_profiler_dict = weakref.WeakKeyDictionary() # type: ignore def _pjit_call_impl_python( *args, jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): pgle_compile_options, pgle_profiler = {}, None if config.enable_pgle.value and config.pgle_profiling_runs.value > 0: compilation_target_key = jaxpr pgle_profiler = _pgle_profiler_dict.get(compilation_target_key) if pgle_profiler is None: pgle_profiler = profiler.PGLEProfiler( config.pgle_profiling_runs.value, config.pgle_aggregation_percentile.value) _pgle_profiler_dict[compilation_target_key] = pgle_profiler # The method below will return FDO profile when module was profiled # config.jax_pgle_profiling_runs amount of times, otherwise the result will # be None. fdo_profile = pgle_profiler.consume_fdo_profile() if fdo_profile is not None: pgle_compile_options['fdo_profile'] = fdo_profile compiler_options_kvs = compiler_options_kvs + tuple(pgle_compile_options.items()) # Passing mutable PGLE profile here since it should be extracted by JAXPR to # initialize the fdo_profile compile option. compiled = _resolve_and_lower( args, jaxpr=jaxpr, in_shardings=in_shardings, out_shardings=out_shardings, in_layouts=in_layouts, out_layouts=out_layouts, donated_invars=donated_invars, ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, lowering_platforms=None, lowering_parameters=mlir.LoweringParameters(), pgle_profiler=pgle_profiler, compiler_options_kvs=compiler_options_kvs, ).compile() # This check is expensive so only do it if enable_checks is on. if compiled._auto_spmd_lowering and config.enable_checks.value: pxla.check_array_xla_sharding_layout_match( args, compiled._in_shardings, compiled._in_layouts, jaxpr.jaxpr._debug_info, compiled._kept_var_idx) if config.distributed_debug.value: # Defensively only perform fingerprint logic if debug logging is enabled # NOTE(skyewm): I didn't benchmark this fingerprint = None if hasattr(compiled.runtime_executable(), "fingerprint"): fingerprint = compiled.runtime_executable().fingerprint if fingerprint is not None: fingerprint = fingerprint.hex() distributed_debug_log(("Running pjit'd function", name), ("in_shardings", in_shardings), ("out_shardings", out_shardings), ("in_layouts", in_layouts), ("out_layouts", out_layouts), ("abstract args", map(core.abstractify, args)), ("fingerprint", fingerprint)) return compiled.unsafe_call(*args), compiled, pgle_profiler @weakref_lru_cache def _get_jaxpr_as_fun(jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): # The input jaxpr to `_get_jaxpr_as_fun` is under a weakref_lru_cache so # returning `core.jaxpr_as_fun(jaxpr)` directly creates a strong reference to # the jaxpr defeating the purpose of weakref_lru_cache. So return a function # that closes over a weakrefed jaxpr and gets called inside that function. # This way there won't be a strong reference to the jaxpr from the output # function. jaxpr = weakref.ref(jaxpr) return lambda *args: core.jaxpr_as_fun(jaxpr())(*args) # pylint: disable=unnecessary-lambda def _pjit_call_impl(*args, jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): def call_impl_cache_miss(*args_, **kwargs_): out_flat, compiled, pgle_profiler = _pjit_call_impl_python( *args, jaxpr=jaxpr, in_shardings=in_shardings, out_shardings=out_shardings, in_layouts=in_layouts, out_layouts=out_layouts, donated_invars=donated_invars, ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) fastpath_data = _get_fastpath_data( compiled, tree_structure(out_flat), args, out_flat, [], [], jaxpr.effects, jaxpr.consts, None, pgle_profiler) return out_flat, fastpath_data, _need_to_rebuild_with_fdo(pgle_profiler) f = _get_jaxpr_as_fun( jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs) donated_argnums = tuple(i for i, d in enumerate(donated_invars) if d) cache_key = pxla.JitGlobalCppCacheKeys( donate_argnums=donated_argnums, donate_argnames=None, device=None, backend=None, in_shardings_treedef=None, in_shardings_leaves=in_shardings, out_shardings_treedef=None, out_shardings_leaves=out_shardings, in_layouts_treedef=None, in_layouts_leaves=in_layouts, out_layouts_treedef=None, out_layouts_leaves=out_layouts) return xc._xla.pjit( name, f, call_impl_cache_miss, [], [], cache_key, tree_util.dispatch_registry, pxla.cc_shard_arg, _get_cpp_global_cache(cache_key.contains_explicit_attributes))(*args) pjit_p.def_impl(_pjit_call_impl) def _pjit_lower(*args, **kwargs): util.test_event("pjit_lower") return _pjit_lower_cached(*args, **kwargs) # This cache is important for python dispatch performance. @weakref_lru_cache def _pjit_lower_cached( jaxpr: core.ClosedJaxpr, in_shardings, out_shardings, in_layouts: pxla.MaybeLayout, out_layouts: pxla.MaybeLayout, donated_invars, ctx_mesh, name: str, keep_unused: bool, inline: bool, compiler_options_kvs: tuple[tuple[str, Any], ...], *, lowering_platforms: tuple[str, ...] | None, lowering_parameters: mlir.LoweringParameters, pgle_profiler: profiler.PGLEProfiler | None): return pxla.lower_sharding_computation( jaxpr, 'jit', name, in_shardings, out_shardings, in_layouts, out_layouts, tuple(donated_invars), keep_unused=keep_unused, context_mesh=ctx_mesh, compiler_options_kvs=compiler_options_kvs, lowering_platforms=lowering_platforms, lowering_parameters=lowering_parameters, pgle_profiler=pgle_profiler) def pjit_staging_rule(trace, source_info, *args, **params): # If we're inlining, no need to compute forwarding information; the inlined # computation will in effect forward things. if (params["inline"] and all(isinstance(i, UnspecifiedValue) for i in params["in_shardings"]) and all(isinstance(o, UnspecifiedValue) for o in params["out_shardings"]) and all(i is None for i in params["in_layouts"]) and all(o is None for o in params["out_layouts"])): jaxpr = params["jaxpr"] if config.dynamic_shapes.value: # Inline jaxpr doesn't handle dynamic shapes when inlining. If dynamic # shapes are enabled, use eval_jaxpr, which uses the tracing machinery, # but redundantly performs abstract evaluation again. with core.set_current_trace(trace): out = core.eval_jaxpr(jaxpr.jaxpr, jaxpr.consts, *args, propagate_source_info=False) else: out = pe.inline_jaxpr_into_trace( trace, source_info, jaxpr.jaxpr, jaxpr.consts, *args) return [trace.to_jaxpr_tracer(x, source_info) for x in out] jaxpr = params['jaxpr'] if config.dynamic_shapes.value: jaxpr, in_fwd, out_shardings, out_layouts = _pjit_forwarding( jaxpr, params['out_shardings'], params['out_layouts']) params = dict(params, jaxpr=jaxpr, out_shardings=out_shardings, out_layouts=out_layouts) out_tracers = [] for aval in _out_type(jaxpr): if type(aval) is core.DShapedArray: shape = [args[d.val] if type(d) is core.InDBIdx else out_tracers[d.val] if type(d) is core.OutDBIdx else d for d in aval.shape] aval = aval.update(shape=tuple(core.get_referent(d) for d in shape)) out_tracers.append(pe.DynamicJaxprTracer(trace, aval, source_info)) eqn = core.new_jaxpr_eqn( map(trace.getvar, args), map(trace.makevar, out_tracers), pjit_p, params, jaxpr.effects, source_info) trace.frame.add_eqn(eqn) out_tracers_ = iter(out_tracers) out_tracers = [args[f] if type(f) is int else next(out_tracers_) for f in in_fwd] assert next(out_tracers_, None) is None elif any(isinstance(c, core.MutableArray) for c in jaxpr.consts): jaxpr, consts = pxla._move_mutable_consts(jaxpr) consts = [trace.new_const(c, source_info) for c in consts] in_shardings = (*params['in_shardings'],) + (UNSPECIFIED,) * len(consts) in_layouts = (*params['in_layouts'],) + (None,) * len(consts) donated_invars = (*params['donated_invars'],) + (False,) * len(consts) new_params = dict(params, jaxpr=jaxpr, in_shardings=in_shardings, in_layouts=in_layouts, donated_invars=donated_invars) out_tracers = trace.default_process_primitive( pjit_p, (*args, *consts), new_params, source_info=source_info) else: out_tracers = trace.default_process_primitive( pjit_p, args, params, source_info=source_info) return out_tracers pe.custom_staging_rules[pjit_p] = pjit_staging_rule def _pjit_forwarding(jaxpr, out_shardings, out_layouts): in_fwd: list[int | None] = pe._jaxpr_forwarding(jaxpr.jaxpr) in_fwd = [fwd if isinstance(os, UnspecifiedValue) and ol is None else None for fwd, os, ol in zip(in_fwd, out_shardings, out_layouts)] keep = [f is None for f in in_fwd] jaxpr = pe.prune_closed_jaxpr_outputs(jaxpr, keep) out_shardings = tuple(o for o, k in zip(out_shardings, keep) if k) out_layouts = tuple(o for o, k in zip(out_layouts , keep) if k) return jaxpr, in_fwd, out_shardings, out_layouts def pjit_forwarding_rule(eqn): if not config.dynamic_shapes.value: return [None] * len(eqn.outvars), eqn jaxpr, in_fwd, out_shardings, out_layouts = _pjit_forwarding( eqn.params['jaxpr'], eqn.params['out_shardings'], eqn.params['out_layouts']) new_outvars = [v for v, f in zip(eqn.outvars, in_fwd) if f is None] new_params = dict(eqn.params, jaxpr=jaxpr, out_shardings=out_shardings, out_layouts=out_layouts) new_eqn = eqn.replace(params=new_params, outvars=new_outvars) return in_fwd, new_eqn # TODO(mattjj): Remove pjit_forwarding_rule and also in staging rule. pe.forwarding_rules[pjit_p] = pjit_forwarding_rule # TODO(mattjj): remove/trivialize this when jaxprs have type annotation on them, # since it's actually not possible in general to infer the type from the term def _out_type(jaxpr: core.ClosedJaxpr) -> list[core.AbstractValue]: out = [] in_idx = {v: i for i, v in enumerate(jaxpr.jaxpr.invars)} out_idx = {x: i for i, x in enumerate(jaxpr.jaxpr.invars) if type(x) is core.Var} for x in jaxpr.jaxpr.outvars: aval = x.aval if type(aval) is core.DShapedArray: shape = [core.InDBIdx(in_idx[d]) if d in in_idx else core.OutDBIdx(out_idx[d]) if d in out_idx else d for d in x.aval.shape] aval = aval.update(shape=tuple(shape)) out.append(aval) return out def _pjit_typecheck(ctx_factory, *in_atoms, jaxpr, **params): return core._check_call(ctx_factory, pjit_p, in_atoms, dict(params, call_jaxpr=jaxpr.jaxpr)) core.custom_typechecks[pjit_p] = _pjit_typecheck def _pjit_abstract_eval(*args, jaxpr, out_shardings, **_): return jaxpr.out_avals, jaxpr.effects pjit_p.def_effectful_abstract_eval(_pjit_abstract_eval) def _pjit_cached_lower_jaxpr_to_fun(ctx: mlir.LoweringRuleContext, name: str, jaxpr: core.ClosedJaxpr, effects, in_shardings, out_shardings, in_layouts, out_layouts, api_name): mod_ctx = ctx.module_context axis_ctx = ctx.module_context.axis_context num_devices = None if isinstance(axis_ctx, sharding_impls.ShardingContext): num_devices = axis_ctx.num_devices elif isinstance(axis_ctx, sharding_impls.SPMDAxisContext): num_devices = axis_ctx.mesh.size key = (pjit_p, name, jaxpr, effects, num_devices, pxla.SemanticallyEqualShardings(in_shardings, jaxpr.in_avals), # pytype: disable=wrong-arg-types pxla.SemanticallyEqualShardings(out_shardings, jaxpr.out_avals), # pytype: disable=wrong-arg-types in_layouts, out_layouts, api_name) func = mod_ctx.cached_primitive_lowerings.get(key, None) if func is None: arg_shardings = [None if isinstance(i, UnspecifiedValue) else i for i in in_shardings] result_shardings = [None if isinstance(o, UnspecifiedValue) else o for o in out_shardings] # TODO(b/228598865): inlined calls cannot have shardings set directly on the # inputs or outputs because they are lost during MLIR->HLO conversion. # using_sharding_annotation=False means we add an identity operation instead. num_callbacks = len(mod_ctx.host_callbacks) func = mlir.lower_jaxpr_to_fun( mod_ctx, name, jaxpr, effects, ctx.name_stack, arg_shardings=arg_shardings, result_shardings=result_shardings, use_sharding_annotations=False, api_name=api_name, arg_layouts=in_layouts, result_layouts=out_layouts) # If this Jaxpr includes callbacks, we can't cache the lowering because # on TPU every callback must have a globally unique channel, but the # channel gets assigned during lowering. has_callbacks = len(mod_ctx.host_callbacks) > num_callbacks if not has_callbacks or "tpu" not in mod_ctx.platforms: mod_ctx.cached_primitive_lowerings[key] = func return func def _pjit_lowering(ctx: mlir.LoweringRuleContext, *args, name: str, jaxpr: core.ClosedJaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, keep_unused, inline, compiler_options_kvs): effects = list(ctx.tokens_in.effects()) output_types = map(mlir.aval_to_ir_type, ctx.avals_out) output_types = [mlir.token_type()] * len(effects) + output_types flat_output_types = mlir.flatten_ir_types(output_types) func = _pjit_cached_lower_jaxpr_to_fun( ctx, name, jaxpr, tuple(effects), in_shardings, out_shardings, in_layouts, out_layouts, api_name='jit') tokens_in = [ctx.tokens_in.get(eff) for eff in effects] args = (*ctx.dim_var_values, *tokens_in, *args) call = func_dialect.CallOp(flat_output_types, ir.FlatSymbolRefAttr.get(func.name.value), mlir.flatten_ir_values(args)) mlir.wrap_compute_type_in_place(ctx, call) out_nodes = mlir.unflatten_ir_values_like_types(call.results, output_types) tokens, out_nodes = split_list(out_nodes, [len(effects)]) tokens_out = ctx.tokens_in.update_tokens(mlir.TokenSet(zip(effects, tokens))) ctx.set_tokens_out(tokens_out) return out_nodes mlir.register_lowering(pjit_p, _pjit_lowering) def _pjit_batcher(axis_data, vals_in, dims_in: tuple[int, ...], jaxpr: core.ClosedJaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): segment_lens, dims_in = batching.indirectify_ragged_axes(dims_in) new_jaxpr, axes_out = batching.batch_jaxpr2(jaxpr, axis_data, dims_in) # TODO(axch): prepend with Nones (?) to account for new segment_lens inputs in_shardings = tuple( _pjit_batcher_for_sharding(i, axis_in, axis_data.spmd_name, ctx_mesh, aval.ndim) if axis_in is not None else i for axis_in, i, aval in zip(dims_in, in_shardings, new_jaxpr.in_avals)) out_shardings = tuple( _pjit_batcher_for_sharding(o, axis_out, axis_data.spmd_name, ctx_mesh, aval.ndim) if axis_out is not None else o for axis_out, o, aval in zip(axes_out, out_shardings, new_jaxpr.out_avals)) # TODO(yashkatariya): Figure out layouts should change under vmap. if not (all(l is None for l in in_layouts) and all(l is None for l in out_layouts)): raise NotImplementedError( 'Concrete layouts are not supported for vmap(jit).') vals_out = pjit_p.bind( *vals_in, jaxpr=new_jaxpr, in_shardings=in_shardings, out_shardings=out_shardings, in_layouts=in_layouts, out_layouts=out_layouts, donated_invars=donated_invars, ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) resolved_axes_out = batching.resolve_ragged_axes_against_inputs_outputs( vals_in, vals_out, axes_out) return vals_out, resolved_axes_out batching.fancy_primitive_batchers[pjit_p] = _pjit_batcher batching.ragged_prop_rules[pjit_p] = batching.ragged_mask_no_op_rule def _pjit_batcher_for_sharding( s: Sharding | UnspecifiedValue, dim: int | batching.RaggedAxis, spmd_axis_name: tuple[str, ...] | None, mesh, ndim: int): if isinstance(s, UnspecifiedValue): return s hlo_s = s._to_xla_hlo_sharding(ndim) if spmd_axis_name is None: if sharding_impls.is_op_sharding_replicated(hlo_s): return s if isinstance(s, NamedSharding) and isinstance(s.mesh, AbstractMesh): return NamedSharding( s.mesh, pxla.batch_spec(s.spec, dim, PartitionSpec.UNCONSTRAINED)) new_op = hlo_s.to_proto().clone() tad = list(new_op.tile_assignment_dimensions) tad.insert(dim, 1) # type: ignore new_op.tile_assignment_dimensions = tad new_gs = GSPMDSharding( s._device_assignment, new_op, _device_list=getattr(s, '_internal_device_list', None)) return pxla._get_out_sharding_from_orig_sharding([new_gs], [None], s, None)[0] else: if isinstance(s, NamedSharding) and isinstance(s.mesh, AbstractMesh): return NamedSharding( s.mesh, pxla.batch_spec(s.spec, dim, spmd_axis_name)) if isinstance(s, NamedSharding): mesh = s.mesh if mesh is None or mesh.empty: raise ValueError( 'If you are using spmd_axis_name parameter of jax.vmap,' ' please make sure to run your jitted function inside the mesh' ' context manager. Only `jax.lax.with_sharding_constraint` with' ' `jax.sharding.NamedSharding` as an input can be transformed with' ' spmd_axis_name batching rules outside of an explicit mesh context' f' manager scope{s!r}') spec = parse_flatten_op_sharding(hlo_s, mesh)[0] return NamedSharding( mesh, pxla.batch_spec(spec, dim, spmd_axis_name)) def _pjit_jvp(primals_in, tangents_in, jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): is_nz_tangents_in = [type(t) is not ad.Zero for t in tangents_in] jaxpr_jvp, is_nz_tangents_out = ad.jvp_jaxpr( jaxpr, is_nz_tangents_in, instantiate=False) def _filter_zeros(is_nz_l, l): return (x for nz, x in zip(is_nz_l, l) if nz) _filter_zeros_in = partial(_filter_zeros, is_nz_tangents_in) _filter_zeros_out = partial(_filter_zeros, is_nz_tangents_out) outputs = pjit_p.bind( *primals_in, *_filter_zeros_in(tangents_in), jaxpr=jaxpr_jvp, in_shardings=(*in_shardings, *_filter_zeros_in(in_shardings)), out_shardings=(*out_shardings, *_filter_zeros_out(out_shardings)), in_layouts=(*in_layouts, *_filter_zeros_in(in_layouts)), out_layouts=(*out_layouts, *_filter_zeros_out(out_layouts)), donated_invars=(*donated_invars, *_filter_zeros_in(donated_invars)), ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) primals_out, tangents_out = split_list(outputs, [len(jaxpr.jaxpr.outvars)]) assert len(primals_out) == len(jaxpr.jaxpr.outvars) tangents_out_it = iter(tangents_out) return primals_out, [next(tangents_out_it) if nz else ad.Zero(aval) for nz, aval in zip(is_nz_tangents_out, jaxpr.out_avals)] ad.primitive_jvps[pjit_p] = _pjit_jvp def _pjit_linearization(nzs, *primals_in, jaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): primal_jaxpr, num_residuals, nzs_out, tangent_jaxpr = ad.linearize_jaxpr(jaxpr, nzs) res_shardings = (UNSPECIFIED,) * num_residuals res_layouts = (None,) * num_residuals res_donated = (False,) * num_residuals primal_out_shardings = res_shardings + tuple(out_shardings) primal_out_layouts = res_layouts + tuple(out_layouts) def keep_where(l, should_keep): return tuple(x for x, keep in zip(l, should_keep) if keep) # Input-to-output forwarding. in_fwd = pe._jaxpr_forwarding(primal_jaxpr.jaxpr) in_fwd_res, in_fwd_primal = split_list(in_fwd, [num_residuals]) in_fwd = in_fwd_res + [ fwd if isinstance(os, UnspecifiedValue) and ol is None else None for os, ol, fwd in zip(out_shardings, out_layouts, in_fwd_primal) ] del in_fwd_res, in_fwd_primal keep = [f is None for f in in_fwd] primal_jaxpr = pe.prune_closed_jaxpr_outputs(primal_jaxpr, keep) primal_out_shardings = keep_where(primal_out_shardings, keep) primal_out_layouts = keep_where(primal_out_layouts, keep) kept_res, _ = split_list(keep, [num_residuals]) num_kept_residuals = sum(kept_res) del keep, kept_res # Output-to-output forwarding. num_out_primals = len(primal_jaxpr.jaxpr.outvars) - num_kept_residuals res_vars, out_vars = split_list(primal_jaxpr.jaxpr.outvars, [num_kept_residuals]) idx_map = {id(v): i for i, v in enumerate(out_vars)} offset = sum(id(v) not in idx_map for v in res_vars) idx_map = {k: v + offset for k, v in idx_map.items()} out_fwd = [idx_map.get(id(v)) for v in res_vars] + [None] * num_out_primals keep = [f is None for f in out_fwd] primal_jaxpr = pe.prune_closed_jaxpr_outputs(primal_jaxpr, keep) primal_out_shardings = keep_where(primal_out_shardings, keep) primal_out_layouts = keep_where(primal_out_layouts, keep) del keep def tangent_fun(consts_, *tangents): tangents_nz = _filter_zeros(nzs, tangents) nz_tangents_out = pjit_p.bind(*tangents_nz, *consts_, jaxpr=tangent_jaxpr, in_shardings=_filter_zeros(nzs, in_shardings) + res_shardings, out_shardings=_filter_zeros(nzs_out, out_shardings), in_layouts=_filter_zeros(nzs, in_layouts) + res_layouts, out_layouts=_filter_zeros(nzs_out, out_layouts), donated_invars=_filter_zeros(nzs, donated_invars) + res_donated, ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) tangent_avals_out = [v.aval.to_tangent_aval() for v in jaxpr.jaxpr.outvars] nz_tangents_out_ = iter(nz_tangents_out) tangents_out = [next(nz_tangents_out_) if nz else ad.Zero(aval) for (aval, nz) in zip(tangent_avals_out, nzs_out)] return tangents_out def _filter_zeros(is_nz_l, l): return tuple(x for nz, x in zip(is_nz_l, l) if nz) ans = pjit_p.bind(*primals_in, jaxpr=primal_jaxpr, in_shardings=in_shardings, out_shardings=primal_out_shardings, in_layouts=in_layouts, out_layouts=primal_out_layouts, donated_invars=donated_invars, ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) ans = subs_list(out_fwd, ans, ans) ans = subs_list(in_fwd, primals_in, ans) residuals_ans, primal_ans = split_list(ans, [num_residuals]) return primal_ans, nzs_out, residuals_ans, tangent_fun ad.primitive_linearizations[pjit_p] = _pjit_linearization def _pjit_partial_eval(trace: pe.JaxprTrace, *in_tracers, jaxpr: core.ClosedJaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): in_pvals = [t.pval for t in in_tracers] known_ins = tuple(pv.is_known() for pv in in_pvals) unknown_ins = tuple(not k for k in known_ins) known_jaxpr, unknown_jaxpr, unknown_outs, res_out_avals, in_fwd_res = \ pe.partial_eval_jaxpr_nounits_fwd(jaxpr, unknown_ins, instantiate=False) unknown_outs = tuple(unknown_outs) # type: ignore[assignment] known_outs = tuple(not uk for uk in unknown_outs) # out_shardings and out_layouts for residual values output by known_jaxpr def keep_where(l, should_keep): return tuple(x for x, keep in zip(l, should_keep) if keep) known_out_shardings = (keep_where(out_shardings, known_outs) + (UNSPECIFIED,) * len(res_out_avals)) known_out_layouts = (keep_where(out_layouts, known_outs) + (None,) * len(res_out_avals)) # Input-to-output forwarding: compute which outputs are just forwarded inputs. num_out_primals = len(known_jaxpr.out_avals) - len(res_out_avals) in_fwd: list[int | None] = pe._jaxpr_forwarding(known_jaxpr.jaxpr) in_fwd_primal, in_fwd_res_ = split_list(in_fwd, [num_out_primals]) assert all(f is None for f in in_fwd_res_) in_fwd = [ fwd if isinstance(os, UnspecifiedValue) and ol is None else None for os, ol, fwd in zip( keep_where(out_shardings, known_outs), keep_where(out_layouts, known_outs), in_fwd_primal) ] + in_fwd_res_ del in_fwd_primal, in_fwd_res_ # Prune jaxpr outputs and out_shardings by removing the input-forwards. keep = [f is None for f in in_fwd] known_jaxpr = pe.prune_closed_jaxpr_outputs(known_jaxpr, keep) known_out_shardings = keep_where(known_out_shardings, keep) known_out_layouts = keep_where(known_out_layouts, keep) # Update num_out_primals to reflect pruning. kept_primals, kept_res = split_list(keep, [num_out_primals]) num_out_primals = sum(kept_primals) del keep, kept_primals, kept_res # Output-to-output forwarding: compute which residuals are just primal outputs out_vars, res_vars = split_list(known_jaxpr.jaxpr.outvars, [num_out_primals]) idx_map = {id(v): i for i, v in enumerate(out_vars)} out_fwd = [None] * num_out_primals + [idx_map.get(id(v)) for v in res_vars] # Prune jaxpr outputs and out_shardings by removing forwarded residuals. keep = [f is None for f in out_fwd] known_jaxpr = pe.prune_closed_jaxpr_outputs(known_jaxpr, keep) known_out_shardings = keep_where(known_out_shardings, keep) known_out_layouts = keep_where(known_out_layouts, keep) del keep known_params = dict( jaxpr=known_jaxpr, in_shardings=keep_where(in_shardings, known_ins), out_shardings=known_out_shardings, in_layouts=keep_where(in_layouts, known_ins), out_layouts=known_out_layouts, donated_invars=keep_where(donated_invars, known_ins), ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) assert len(known_params['out_shardings']) == len(known_params['jaxpr'].out_avals) assert len(known_params['out_layouts']) == len(known_params['jaxpr'].out_avals) # Bind known things to pjit_p. known_inputs = [pv.get_known() for pv in in_pvals if pv.is_known()] all_known_outs = pjit_p.bind(*known_inputs, **known_params) # Add back in the output fwds. all_known_outs = subs_list(out_fwd, all_known_outs, all_known_outs) # Add back in the input fwds. all_known_outs = subs_list(in_fwd, known_inputs, all_known_outs) known_out_vals, residual_vals = \ split_list(all_known_outs, [len(all_known_outs) - len(res_out_avals)]) residual_vals_ = iter(residual_vals) residual_vals = [next(residual_vals_) if f is None else [*jaxpr.consts, *known_inputs][f] for f in in_fwd_res] assert next(residual_vals_, None) is None residual_tracers = map(trace.new_instantiated_const, residual_vals) # The convention of partial_eval_jaxpr_nounits is to place residual binders at # the front of the jaxpr produced, so we move them to the back since both the # jaxpr equation built below and the pjit transpose rule assume a # residual-inputs-last convention. unknown_jaxpr = pe.move_binders_to_back( unknown_jaxpr, [True] * len(residual_vals) + [False] * sum(unknown_ins)) # Set up staged-out 'unknown' eqn unknown_in_shardings = (keep_where(in_shardings, unknown_ins) + (UNSPECIFIED,) * len(residual_tracers)) unknown_in_layouts = (keep_where(in_layouts, unknown_ins) + (None,) * len(residual_tracers)) unknown_donated_invars = (keep_where(donated_invars, unknown_ins) + (False,) * len(residual_tracers)) unknown_params = dict( jaxpr=unknown_jaxpr, in_shardings=unknown_in_shardings, in_layouts=unknown_in_layouts, out_shardings=keep_where(out_shardings, unknown_outs), out_layouts=keep_where(out_layouts, unknown_outs), donated_invars=unknown_donated_invars, ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) unknown_tracers_in = [t for t in in_tracers if not t.pval.is_known()] unknown_out_avals = unknown_jaxpr.out_avals unknown_tracers_out = [ pe.JaxprTracer(trace, pe.PartialVal.unknown(aval), None) for aval in unknown_out_avals ] eqn = pe.new_eqn_recipe(trace, (*unknown_tracers_in, *residual_tracers), unknown_tracers_out, pjit_p, unknown_params, unknown_jaxpr.effects, source_info_util.current()) for t in unknown_tracers_out: t.recipe = eqn return merge_lists(unknown_outs, known_out_vals, unknown_tracers_out) pe.custom_partial_eval_rules[pjit_p] = _pjit_partial_eval def _pjit_partial_eval_custom_params_updater( unks_in: Sequence[bool], inst_in: Sequence[bool], kept_outs_known: Sequence[bool], kept_outs_staged: Sequence[bool], num_res_out: int, num_res_in: int, params_known: dict, params_staged: dict ) -> tuple[dict, dict]: # prune inputs to jaxpr_known according to unks_in donated_invars_known, _ = pe.partition_list(unks_in, params_known['donated_invars']) in_shardings_known, _ = pe.partition_list(unks_in, params_known['in_shardings']) _, out_shardings_known = pe.partition_list(kept_outs_known, params_known['out_shardings']) in_layouts_known, _ = pe.partition_list(unks_in, params_known['in_layouts']) _, out_layouts_known = pe.partition_list(kept_outs_known, params_known['out_layouts']) new_params_known = dict(params_known, in_shardings=tuple(in_shardings_known), out_shardings=(*out_shardings_known, *[UNSPECIFIED] * num_res_out), in_layouts=tuple(in_layouts_known), out_layouts=(*out_layouts_known, *[None] * num_res_out), donated_invars=tuple(donated_invars_known)) assert len(new_params_known['in_shardings']) == len(params_known['jaxpr'].in_avals) assert len(new_params_known['out_shardings']) == len(params_known['jaxpr'].out_avals) assert len(new_params_known['in_layouts']) == len(params_known['jaxpr'].in_avals) assert len(new_params_known['out_layouts']) == len(params_known['jaxpr'].out_avals) # added num_res new inputs to jaxpr_staged, and pruning according to inst_in _, donated_invars_staged = pe.partition_list(inst_in, params_staged['donated_invars']) donated_invars_staged = [False] * num_res_in + donated_invars_staged _, in_shardings_staged = pe.partition_list(inst_in, params_staged['in_shardings']) in_shardings_staged = [*[UNSPECIFIED] * num_res_in, *in_shardings_staged] _, out_shardings_staged = pe.partition_list(kept_outs_staged, params_staged['out_shardings']) _, in_layouts_staged = pe.partition_list(inst_in, params_staged['in_layouts']) in_layouts_staged = [*[None] * num_res_in, *in_layouts_staged] _, out_layouts_staged = pe.partition_list(kept_outs_staged, params_staged['out_layouts']) new_params_staged = dict(params_staged, in_shardings=tuple(in_shardings_staged), out_shardings=tuple(out_shardings_staged), in_layouts=tuple(in_layouts_staged), out_layouts=tuple(out_layouts_staged), donated_invars=tuple(donated_invars_staged)) assert len(new_params_staged['in_shardings']) == len(params_staged['jaxpr'].in_avals) assert len(new_params_staged['out_shardings']) == len(params_staged['jaxpr'].out_avals) assert len(new_params_staged['in_layouts']) == len(params_staged['jaxpr'].in_avals) assert len(new_params_staged['out_layouts']) == len(params_staged['jaxpr'].out_avals) return new_params_known, new_params_staged pe.partial_eval_jaxpr_custom_rules[pjit_p] = \ partial(pe.closed_call_partial_eval_custom_rule, 'jaxpr', _pjit_partial_eval_custom_params_updater) @lu.cache def _pjit_transpose_trace(fun: lu.WrappedFun, in_avals: Sequence[core.AbstractValue]): transpose_jaxpr, _, consts, attrs_tracked = pe.trace_to_jaxpr_dynamic( fun, in_avals) transpose_jaxpr = core.ClosedJaxpr(transpose_jaxpr, consts) return transpose_jaxpr, attrs_tracked def _pjit_transpose(cts_in, *primals_in, jaxpr: core.ClosedJaxpr, in_shardings, out_shardings, in_layouts, out_layouts, donated_invars, ctx_mesh, name, keep_unused, inline, compiler_options_kvs): def prune_type(ty, xs, maybe_zeros): return tuple(x for x, mz in zip(xs, maybe_zeros) if type(mz) is not ty) body = lu.wrap_init(ad.closed_backward_pass, debug_info=jaxpr.jaxpr._debug_info) body = lu.hashable_partial(body, jaxpr, False) primals_and_nz_cts_in, in_treedef = tree_flatten((primals_in, cts_in)) body, cts_out_treedef_thunk = flatten_fun_nokwargs(body, in_treedef) transpose_in_shardings = ( *prune_type(ad.UndefinedPrimal, in_shardings, primals_in), *prune_type(ad.Zero, out_shardings, cts_in) ) transpose_in_layouts = ( *prune_type(ad.UndefinedPrimal, in_layouts, primals_in), *prune_type(ad.Zero, out_layouts, cts_in) ) global_cts_in_avals = tuple(core.get_aval(ct) for ct in primals_and_nz_cts_in) transpose_jaxpr, attrs_tracked = _pjit_transpose_trace( body, global_cts_in_avals) cts_out_treedef = cts_out_treedef_thunk() transpose_out_shardings = prune_type( ad.Zero, in_shardings, tree_unflatten(cts_out_treedef, [object()] * cts_out_treedef.num_leaves)) transpose_out_layouts = prune_type( ad.Zero, in_layouts, tree_unflatten(cts_out_treedef, [object()] * cts_out_treedef.num_leaves)) if attrs_tracked: init_states = _get_states(attrs_tracked) num_attr_outs = sum(final_tree.num_leaves for _, final_tree, _ in attrs_tracked) primals_and_nz_cts_in = [*init_states, *primals_and_nz_cts_in] transpose_in_shardings = (UNSPECIFIED,) * len(init_states) + transpose_in_shardings transpose_out_shardings = (UNSPECIFIED,) * num_attr_outs + transpose_out_shardings transpose_in_layouts = (None,) * len(init_states) + transpose_in_layouts transpose_out_layouts = (None,) * num_attr_outs + transpose_out_layouts try: nz_cts_out = pjit_p.bind( *primals_and_nz_cts_in, jaxpr=transpose_jaxpr, in_shardings=transpose_in_shardings, out_shardings=transpose_out_shardings, in_layouts=transpose_in_layouts, out_layouts=transpose_out_layouts, donated_invars=(False,) * len(primals_and_nz_cts_in), ctx_mesh=ctx_mesh, name=name, keep_unused=keep_unused, inline=inline, compiler_options_kvs=compiler_options_kvs) except api_util.InternalFloatingPointError as e: print("Invalid nan value encountered in the backward pass of a jax.jit " "function. Calling the de-optimized backward pass.") try: _ = ad.closed_backward_pass(jaxpr, None, primals_in, cts_in) except (FloatingPointError, ZeroDivisionError) as e2: raise e2 from None # great else: # If control reaches this line, we got a NaN on the output of `compiled` # but not `fun.call_wrapped` on the same arguments. Let's tell the user. api_util._raise_no_nan_in_deoptimized(e) if attrs_tracked: final_states, nz_cts_out = split_list(nz_cts_out, [num_attr_outs]) _set_states(attrs_tracked, final_states) return tree_unflatten(cts_out_treedef, nz_cts_out) ad.primitive_transposes[pjit_p] = _pjit_transpose @weakref_lru_cache def _dce_jaxpr_pjit( jaxpr: core.ClosedJaxpr, used_outputs: tuple[bool, ...] ) -> tuple[core.ClosedJaxpr, list[bool]]: new_jaxpr, used_inputs = pe.dce_jaxpr(jaxpr.jaxpr, used_outputs) return core.ClosedJaxpr(new_jaxpr, jaxpr.consts), used_inputs def dce_jaxpr_pjit_rule(used_outputs: list[bool], eqn: core.JaxprEqn ) -> tuple[list[bool], core.JaxprEqn | None]: if not any(used_outputs) and not pe.has_effects(eqn): return [False] * len(eqn.invars), None dced_jaxpr, used_inputs = _dce_jaxpr_pjit( eqn.params['jaxpr'], tuple(used_outputs)) def keep_where(xs, keeps): return tuple(x for x, keep in zip(xs, keeps) if keep) eqn_params = eqn.params new_params = dict( eqn_params, jaxpr=dced_jaxpr, in_shardings=keep_where(eqn_params["in_shardings"], used_inputs), out_shardings=keep_where(eqn_params["out_shardings"], used_outputs), in_layouts=keep_where(eqn_params["in_layouts"], used_inputs), out_layouts=keep_where(eqn_params["out_layouts"], used_outputs), donated_invars=keep_where(eqn_params["donated_invars"], used_inputs), ) if not any(used_inputs) and not any(used_outputs) and not dced_jaxpr.effects: return used_inputs, None else: new_eqn = core.new_jaxpr_eqn( [v for v, used in zip(eqn.invars, used_inputs) if used], [v for v, used in zip(eqn.outvars, used_outputs) if used], eqn.primitive, new_params, dced_jaxpr.effects, eqn.source_info, eqn.ctx) return used_inputs, new_eqn pe.dce_rules[pjit_p] = dce_jaxpr_pjit_rule def _pjit_pp_rule(eqn: core.JaxprEqn, context: core.JaxprPpContext, settings: core.JaxprPpSettings) -> core.pp.Doc: params = dict(eqn.params) del params['inline'] if not any(params['donated_invars']): del params['donated_invars'] if all(isinstance(s, UnspecifiedValue) for s in params['in_shardings']): del params['in_shardings'] if all(isinstance(s, UnspecifiedValue) for s in params['out_shardings']): del params['out_shardings'] if all(l is None for l in params['in_layouts']): del params['in_layouts'] if all(l is None for l in params['out_layouts']): del params['out_layouts'] if not params['keep_unused']: del params['keep_unused'] if params['ctx_mesh'] is None or params['ctx_mesh'].empty: del params['ctx_mesh'] if not params['compiler_options_kvs']: del params['compiler_options_kvs'] if params['jaxpr'].jaxpr not in context.shared_jaxprs: context.suggest_same_var_names(params['jaxpr'].jaxpr.invars, eqn.invars) context.suggest_same_var_names(params['jaxpr'].jaxpr.outvars, eqn.outvars) # Move name= to the front to make the resulting equation easier to scan. del params["name"] return core._pp_eqn(eqn, context, settings, params=["name"] + sorted(params)) core.pp_eqn_rules[pjit_p] = _pjit_pp_rule # -------------------- with_sharding_constraint -------------------- def check_shardings_are_auto(shardings_flat): for s in shardings_flat: if not isinstance(s, NamedSharding): continue mesh = s.mesh.abstract_mesh if not all(mesh._name_to_type[i] == mesh_lib.AxisType.Auto for axes in s.spec if axes is not PartitionSpec.UNCONSTRAINED and axes is not None for i in (axes if isinstance(axes, tuple) else (axes,))): raise ValueError( 'The spec of NamedSharding passed to with_sharding_constraint can' f' only refer to Auto axes of the mesh. Got spec={s.spec} and' f' mesh={mesh}. You probably meant to use `reshard` API?') cur_mesh = mesh_lib.get_abstract_mesh() if cur_mesh._are_all_axes_explicit: raise ValueError( 'with_sharding_constraint cannot be used when all axes of the mesh are' ' of type `Explicit`. Please use the `reshard` API.') def with_sharding_constraint(x, shardings): """Mechanism to constrain the sharding of an Array inside a jitted computation This is a strict constraint for the GSPMD partitioner and not a hint. For examples of how to use this function, see `Distributed arrays and automatic parallelization`_. Inside of a jitted computation, with_sharding_constraint makes it possible to constrain intermediate values to an uneven sharding. However, if such an unevenly sharded value is output by the jitted computation, it will come out as fully replicated, no matter the sharding annotation given. Args: x: PyTree of jax.Arrays which will have their shardings constrained shardings: PyTree of sharding specifications. Valid values are the same as for the ``in_shardings`` argument of :func:`jax.experimental.pjit`. Returns: x_with_shardings: PyTree of jax.Arrays with specified sharding constraints. .. _Distributed arrays and automatic parallelization: https://docs.jax.dev/en/latest/notebooks/Distributed_arrays_and_automatic_parallelization.html """ x_flat, tree = tree_flatten(x) x_avals_flat = [core.shaped_abstractify(x) for x in x_flat] layouts, shardings = _split_layout_and_sharding(shardings) user_shardings = prepare_axis_resources( shardings, "shardings", allow_unconstrained_dims=True) del shardings user_shardings_flat = tuple( flatten_axes("with_sharding_constraint shardings", tree, user_shardings)) del user_shardings user_layouts_flat = tuple( flatten_axes("with_sharding_constraint layouts", tree, layouts)) del layouts context_mesh = ( mesh_lib.get_abstract_mesh() if mesh_lib.get_concrete_mesh() is not None else mesh_lib.thread_resources.env.physical_mesh) shardings_flat = [_create_sharding_for_array(context_mesh, a, 'shardings', 'with_sharding_constraint') for a in user_shardings_flat] for s, u in zip(shardings_flat, user_shardings_flat): if isinstance(s, (UnspecifiedValue, AUTO)): raise ValueError( f'One of with_sharding_constraint arguments got sharding {u} which is' ' not allowed. Please only pass `jax.sharding.Sharding` instances.') del user_shardings_flat # TODO(bartchr): remove `unconstrained_dims` after migrating to Shardy. It's # already part of the shardings. unconstrained_dims = [get_unconstrained_dims(s) if isinstance(s, NamedSharding) else frozenset() for s in shardings_flat] pjit_check_aval_sharding( shardings_flat, x_avals_flat, ("",) * len(shardings_flat), "with_sharding_constraint arguments", allow_uneven_sharding=True) check_shardings_are_auto(shardings_flat) check_aval_layout_compatibility(user_layouts_flat, x_avals_flat, ("",) * len(user_layouts_flat), "with_sharding_constraint arguments") outs = [sharding_constraint_p.bind(xf, sharding=s, layout=l, context_mesh=context_mesh, unconstrained_dims=ud) for xf, s, l, ud in zip(x_flat, shardings_flat, user_layouts_flat, unconstrained_dims)] return tree_unflatten(tree, outs) def _identity_fn(x): return x def _sharding_constraint_impl(x, sharding, layout, context_mesh, unconstrained_dims): if (isinstance(sharding, NamedSharding) and isinstance(sharding.mesh, AbstractMesh)): if (not context_mesh.empty and isinstance(context_mesh, AbstractMesh) and not hasattr(x, 'sharding')): concrete_mesh = mesh_lib.get_concrete_mesh() assert concrete_mesh is not None sharding = NamedSharding(concrete_mesh, sharding.spec) else: aval = core.shaped_abstractify(x) if not hasattr(x, 'sharding'): raise ValueError( 'Target sharding contains a `jax.sharding.AbstractMesh` which' ' requires the input passed should be a `jax.Array`. Got' f' {type(x)} with shape {aval.str_short()}') if not isinstance(x.sharding, NamedSharding): raise TypeError( 'The sharding on the input must be a `NamedSharding` since the' ' target sharding has an `AbstractMesh` in it. Got sharding type' f' {type(x.sharding)} for shape {aval.str_short()}') if x.sharding.mesh.shape_tuple != sharding.mesh.shape_tuple: raise ValueError( f'Mesh shape of the input {x.sharding.mesh.shape_tuple} does not' ' match the mesh shape of the target sharding' f' {sharding.mesh.shape_tuple} for shape {aval.str_short()}') sharding = NamedSharding(x.sharding.mesh, sharding.spec) if layout is None: if hasattr(x, 'sharding') and x.sharding.is_equivalent_to(sharding, x.ndim): return x # Run a jit here to raise good errors when device assignment don't match. return api.jit(_identity_fn, out_shardings=sharding)(x) else: if (hasattr(x, 'format') and x.format.device_local_layout == layout and x.sharding.is_equivalent_to(sharding, x.ndim)): return x return api.jit(_identity_fn, out_shardings=Format(layout, sharding))(x) sharding_constraint_p = core.Primitive("sharding_constraint") sharding_constraint_p.def_impl(_sharding_constraint_impl) sharding_constraint_p.def_abstract_eval(lambda x, **_: x) ad.deflinear2(sharding_constraint_p, lambda ct, _, **params: (sharding_constraint_p.bind(ct, **params),)) def _sharding_constraint_hlo_lowering(ctx, x_node, *, sharding, layout, context_mesh, unconstrained_dims): in_aval, = ctx.avals_in out_aval, = ctx.avals_out axis_ctx = ctx.module_context.axis_context if dtypes.issubdtype(in_aval.dtype, dtypes.extended): in_aval = core.physical_aval(in_aval) if (isinstance(axis_ctx, sharding_impls.SPMDAxisContext) and axis_ctx.manual_axes): sharding = mlir.add_manual_axes(axis_ctx, sharding, in_aval.ndim) if config.use_shardy_partitioner.value: sharding = sharding._to_sdy_sharding(in_aval.ndim) else: sharding = sharding._to_xla_hlo_sharding(in_aval.ndim).to_proto() out = mlir.wrap_with_sharding_op( ctx, x_node, out_aval, sharding, unspecified_dims=unconstrained_dims) if layout is not None: out = mlir.wrap_with_layout_op(ctx, out, out_aval, layout, in_aval) return [out] mlir.register_lowering(sharding_constraint_p, _sharding_constraint_hlo_lowering) def _sharding_constraint_batcher( axis_data, vals_in, dims_in, sharding, layout, context_mesh, unconstrained_dims): if axis_data.spmd_name is not None and isinstance(sharding, NamedSharding): used = {n for ns in sharding.spec for n in (ns if isinstance(ns, tuple) else (ns,))} if set(axis_data.spmd_name) & used: raise ValueError(f"vmap spmd_axis_name {axis_data.spmd_name} cannot appear in " "with_sharding_constraint spec, but got spec " f"{sharding.spec}") x, = vals_in d, = dims_in unconstrained_dims = {ud + (d <= ud) for ud in unconstrained_dims} if axis_data.spmd_name is None: unconstrained_dims.add(d) vmapped_sharding = _pjit_batcher_for_sharding( sharding, d, axis_data.spmd_name, context_mesh, x.ndim) if unconstrained_dims and isinstance(vmapped_sharding, NamedSharding): new_spec = list(vmapped_sharding.spec) + [None] * (x.ndim - len(vmapped_sharding.spec)) for u in unconstrained_dims: new_spec[u] = PartitionSpec.UNCONSTRAINED vmapped_sharding = NamedSharding( vmapped_sharding.mesh, PartitionSpec(*new_spec)) # TODO(yashkatariya): Figure out layouts should change under vmap. if layout is not None: raise NotImplementedError( 'Concrete layout is not supported for vmap(with_sharding_constraint). ' f'Got layout {layout}') y = sharding_constraint_p.bind( x, sharding=vmapped_sharding, layout=layout, context_mesh=context_mesh, unconstrained_dims=frozenset(unconstrained_dims)) return y, d batching.fancy_primitive_batchers[sharding_constraint_p] = _sharding_constraint_batcher batching.skippable_batchers[sharding_constraint_p] = lambda _: () # -------------------- mesh_cast --------------------------- # TODO(yashkatariya): Make shardings optional. def mesh_cast(xs, out_shardings): x_flat, treedef = tree_flatten(xs) shardings_flat = flatten_axes("mesh_cast shardings", treedef, out_shardings) out_flat = [ mesh_cast_p.bind( x, dst_sharding=canonicalize_sharding( s, 'mesh_cast', check_mesh_consistency=False)) for x, s in safe_zip(x_flat, shardings_flat) ] return tree_unflatten(treedef, out_flat) mesh_cast_p = core.Primitive('mesh_cast') mesh_cast_p.skip_canonicalization = True def _mesh_cast_abstract_eval(aval, dst_sharding): src_sharding = aval.sharding if src_sharding == dst_sharding: return aval if src_sharding.mesh.empty or dst_sharding.mesh.empty: return aval.update(sharding=dst_sharding) if src_sharding.mesh.shape_tuple != dst_sharding.mesh.shape_tuple: raise ValueError( f'Mesh shape of the input {src_sharding.mesh.shape_tuple} does not' ' match the mesh shape of the target sharding' f' {dst_sharding.mesh.shape_tuple} for shape {aval.str_short()}') if (src_sharding.mesh._axis_types_dict == dst_sharding.mesh._axis_types_dict and src_sharding.spec != dst_sharding.spec): raise ValueError( 'mesh_cast should only be used when AxisType changes between the' ' input mesh and the target mesh. Got src' f' axis_types={src_sharding.mesh._axis_types_dict} and dst' f' axis_types={dst_sharding.mesh._axis_types_dict}. To reshard between' ' the same mesh, use `jax.sharding.reshard` instead?') if src_sharding.mesh._any_axis_explicit and dst_sharding.mesh._any_axis_explicit: for s, d in safe_zip(flatten_spec(src_sharding.spec), flatten_spec(dst_sharding.spec)): if s is None and d is None: continue if s is None and d is not None: assert (src_sharding.mesh._name_to_type[d] == mesh_lib.AxisType.Auto and dst_sharding.mesh._name_to_type[d] == mesh_lib.AxisType.Explicit) continue if s is not None and d is None: assert (src_sharding.mesh._name_to_type[s] == mesh_lib.AxisType.Explicit and dst_sharding.mesh._name_to_type[s] == mesh_lib.AxisType.Auto) continue if d != s: raise ValueError( 'Explicit data movement in mesh_cast is not allowed. Got src spec:' f' {s} and dst spec: {d}') return aval.update(sharding=dst_sharding) mesh_cast_p.def_abstract_eval(_mesh_cast_abstract_eval) def _mesh_cast_impl(x, dst_sharding): return dispatch.apply_primitive(mesh_cast_p, x, dst_sharding=dst_sharding) mesh_cast_p.def_impl(_mesh_cast_impl) def _mesh_cast_transpose_rule(ct, x, dst_sharding): return [mesh_cast_p.bind(ct, dst_sharding=x.aval.sharding)] ad.deflinear2(mesh_cast_p, _mesh_cast_transpose_rule) def _mesh_cast_hlo_lowering(ctx, x_node, *, dst_sharding): aval, = ctx.avals_in aval_out, = ctx.avals_out proto = (dst_sharding._to_sdy_sharding(aval.ndim) if config.use_shardy_partitioner.value else dst_sharding._to_xla_hlo_sharding(aval.ndim).to_proto()) return [mlir.lower_with_sharding_in_types(ctx, x_node, aval_out, proto)] mlir.register_lowering(mesh_cast_p, _mesh_cast_hlo_lowering) def _mesh_cast_batcher(axis_data, vals_in, dims_in, dst_sharding): x, = vals_in d, = dims_in vmapped_dst_sharding = batching.get_sharding_for_vmap( axis_data, dst_sharding, d) y = mesh_cast_p.bind(x, dst_sharding=vmapped_dst_sharding) return y, d batching.fancy_primitive_batchers[mesh_cast_p] = _mesh_cast_batcher batching.skippable_batchers[mesh_cast_p] = lambda _: () # -------------------- reshard ------------------------------------ def reshard(xs, out_shardings): x_flat, treedef = tree_flatten(xs) shardings_flat = flatten_axes("reshard shardings", treedef, out_shardings) x_avals_flat = [core.shaped_abstractify(x) for x in x_flat] out_flat = [] for x, x_aval, s in safe_zip(x_flat, x_avals_flat, shardings_flat): ds = canonicalize_sharding(s, 'reshard') if ds is None: raise ValueError( 'Reshard should only be used with out_shardings which are non-None ' f'and have a nonempty mesh. Got sharding {s}.' ) ds = ds.update(spec=ds.spec._normalized_spec_for_aval(x_aval.ndim)) # pytype: disable=attribute-error out_flat.append(reshard_p.bind(x, dst_sharding=ds)) return tree_unflatten(treedef, out_flat) reshard_p = core.Primitive('reshard') def _reshard_abstract_eval(aval, dst_sharding): src_sharding = aval.sharding if (not src_sharding.mesh.empty and src_sharding.mesh.abstract_mesh != dst_sharding.mesh.abstract_mesh): raise ValueError( f'Mesh of the input {src_sharding.mesh.abstract_mesh} does not' ' equal the mesh of the target sharding' f' {dst_sharding.mesh.abstract_mesh} for shape {aval.str_short()}') return aval.update(sharding=dst_sharding) reshard_p.def_abstract_eval(_reshard_abstract_eval) def _reshard_impl(x, dst_sharding): return dispatch.apply_primitive(reshard_p, x, dst_sharding=dst_sharding) reshard_p.def_impl(_reshard_impl) def _reshard_transpose_rule(ct, x, dst_sharding): return [reshard_p.bind(ct, dst_sharding=x.aval.to_cotangent_aval().sharding)] ad.deflinear2(reshard_p, _reshard_transpose_rule) def _reshard_hlo_lowering(ctx, x_node, *, dst_sharding): aval, = ctx.avals_in aval_out, = ctx.avals_out proto = (dst_sharding._to_sdy_sharding(aval.ndim) if config.use_shardy_partitioner.value else dst_sharding._to_xla_hlo_sharding(aval.ndim).to_proto()) return [mlir.lower_with_sharding_in_types(ctx, x_node, aval_out, proto)] mlir.register_lowering(reshard_p, _reshard_hlo_lowering) def _reshard_batcher(axis_data, vals_in, dims_in, dst_sharding): assert axis_data.spmd_name is None x, = vals_in d, = dims_in vmapped_dst_sharding = batching.get_sharding_for_vmap( axis_data, dst_sharding, d) y = reshard_p.bind(x, dst_sharding=vmapped_dst_sharding) return y, d batching.fancy_primitive_batchers[reshard_p] = _reshard_batcher batching.skippable_batchers[reshard_p] = lambda _: () # -------------------- auto and user mode ------------------------- def _get_new_mesh(axes: str | tuple[str, ...] | None, axis_type: mesh_lib.AxisType, name: str, shardings=None, error_on_manual_to_auto_explicit=False): cur_mesh = mesh_lib.get_abstract_mesh() flat_shardings, _ = tree_flatten(shardings) sharding_mesh = mesh_lib.empty_abstract_mesh for i in flat_shardings: if isinstance(i, NamedSharding): if not sharding_mesh.empty and sharding_mesh != i.mesh: raise ValueError( f'Shardings passed to {name} should have the same mesh. Got one' f' mesh {sharding_mesh} and another {i.mesh}') sharding_mesh = i.mesh.abstract_mesh if sharding_mesh.empty and cur_mesh.empty: raise ValueError( f'Context mesh {cur_mesh} cannot be empty. Please use' ' `jax.sharding.use_mesh` API to enter into a mesh context when using' f' `{name}` API.') if not sharding_mesh.empty and not cur_mesh.empty: if sharding_mesh != cur_mesh: raise ValueError( f'Context mesh {cur_mesh} must match the mesh passed to shardings' f' {sharding_mesh}. Recommended approach is to use' ' `jax.sharding.use_mesh` context manager.') mesh_to_use = cur_mesh elif sharding_mesh.empty and not cur_mesh.empty: mesh_to_use = cur_mesh else: assert not sharding_mesh.empty and cur_mesh.empty mesh_to_use = sharding_mesh if axes is None: axes = mesh_to_use.axis_names if not isinstance(axes, tuple): axes = (axes,) for a in axes: if (error_on_manual_to_auto_explicit and mesh_to_use._name_to_type[a] == mesh_lib.AxisType.Manual and axis_type in {mesh_lib.AxisType.Auto, mesh_lib.AxisType.Explicit}): raise NotImplementedError( 'Going from `Manual` AxisType to `Auto` or `Explicit` AxisType is not' ' allowed. Please file a bug at https://github.com/jax-ml/jax/issues' ' with your use case') return mesh_to_use.update_axis_types({a: axis_type for a in axes}) def auto_axes(fun, *, axes: str | tuple[str, ...] | None = None, out_sharding=None): def decorator(*args, **kwargs): if out_sharding is None: if "out_sharding" in kwargs: _out_sharding = kwargs.pop("out_sharding") else: raise TypeError("Missing required keyword argument: 'out_sharding'") else: _out_sharding = out_sharding new_mesh = _get_new_mesh( axes, mesh_lib.AxisType.Auto, 'auto_axes', shardings=_out_sharding, error_on_manual_to_auto_explicit=True) with mesh_lib.use_abstract_mesh(new_mesh): in_specs = tree_map(lambda a: core.modify_spec_for_auto_manual( core.get_aval(a).sharding.spec, new_mesh), args) args = mesh_cast(args, in_specs) out = fun(*args, **kwargs) return mesh_cast(out, _out_sharding) return decorator @contextlib.contextmanager def use_auto_axes(*axes): new_mesh = _get_new_mesh(axes, mesh_lib.AxisType.Auto, 'use_auto_axes') with mesh_lib.use_abstract_mesh(new_mesh): yield def explicit_axes(fun, *, axes: str | tuple[str, ...] | None = None, in_sharding=None): def decorator(*args, **kwargs): if in_sharding is None: if "in_sharding" in kwargs: _in_sharding = kwargs.pop("in_sharding") else: raise TypeError("Missing required keyword argument: 'in_sharding'") else: _in_sharding = in_sharding new_mesh = _get_new_mesh(axes, mesh_lib.AxisType.Explicit, 'explicit_axes', error_on_manual_to_auto_explicit=True) with mesh_lib.use_abstract_mesh(new_mesh): args = mesh_cast(args, _in_sharding) out = fun(*args, **kwargs) out_specs = tree_map(lambda o: core.modify_spec_for_auto_manual( core.get_aval(o).sharding.spec, mesh_lib.get_abstract_mesh()), out) return mesh_cast(out, out_specs) return decorator @contextlib.contextmanager def use_explicit_axes(*axes): new_mesh = _get_new_mesh(axes, mesh_lib.AxisType.Explicit, 'use_explicit_axes') with mesh_lib.use_abstract_mesh(new_mesh): yield # -------------------- with_layout_constraint -------------------- def with_layout_constraint(x, layouts): x_flat, tree = tree_flatten(x) x_avals_flat = [core.shaped_abstractify(x) for x in x_flat] layouts_flat = tuple(flatten_axes("with_layout_constraint layouts", tree, layouts)) if any(not isinstance(l, DeviceLocalLayout) for l in layouts_flat): raise ValueError( 'layouts passed to `with_layout_constraint` must be of type' f' `DeviceLocalLayout`. Got {[type(l) for l in layouts_flat]}') check_aval_layout_compatibility( layouts_flat, x_avals_flat, ("",) * len(layouts_flat), "with_layout_constraint arguments") outs = [layout_constraint_p.bind(xf, layout=l) for xf, l in zip(x_flat, layouts_flat)] return tree_unflatten(tree, outs) layout_constraint_p = core.Primitive('layout_constraint') layout_constraint_p.def_abstract_eval(lambda x, **_: x) ad.deflinear2(layout_constraint_p, lambda ct, _, **params: (layout_constraint_p.bind(ct, **params),)) def _layout_constraint_impl(x, *, layout): if not isinstance(x, xc.ArrayImpl): raise ValueError( 'with_layout_constraint in eager mode can only be applied to' f' jax.Arrays. Got {type(x)}') if x.format.device_local_layout == layout: # type: ignore return x return api.jit(_identity_fn, out_shardings=Format(layout, x.sharding))(x) layout_constraint_p.def_impl(_layout_constraint_impl) def _layout_constraint_hlo_lowering(ctx, x_node, *, layout): aval, = ctx.avals_in out_aval, = ctx.avals_out return [mlir.wrap_with_layout_op(ctx, x_node, out_aval, layout, aval)] mlir.register_lowering(layout_constraint_p, _layout_constraint_hlo_lowering) def _layout_constraint_batcher(axis_data, vals_in, dims_in, layout): raise NotImplementedError batching.fancy_primitive_batchers[layout_constraint_p] = _layout_constraint_batcher batching.skippable_batchers[layout_constraint_p] = lambda _: () # -------------------- helpers -------------------- def get_unconstrained_dims(sharding: NamedSharding): assert sharding.spec is not None return frozenset(i for i, axes in enumerate(sharding.spec) if axes is PartitionSpec.UNCONSTRAINED) # -------------------- attrs etc -------------------- def _set_states(attrs_tracked, vals): valss = split_list(vals, [td.num_leaves for _, td, _ in attrs_tracked[:-1]]) for ((_, treedef, (obj, attr, kind)), leaves) in zip(attrs_tracked, valss): if kind is pe.ReadWrite: val = tree_unflatten(treedef, leaves) jax_setattr(obj, attr, val) elif kind is pe.Append: del treedef val, = leaves jax_extendattr(obj, attr, val) elif kind is pe.BoxAttr: val = tree_unflatten(treedef, leaves) obj.set(val) elif kind is pe.ListAttr: for item in tree_unflatten(treedef, leaves): obj.append(item) else: assert False def _get_states(attrs_tracked): vals = [] for treedef, _, (obj, attr, kind) in attrs_tracked: if kind is pe.ReadWrite: tree = jax_getattr(obj, attr) if hasattr(obj, attr) else dne_sentinel leaves, treedef_ = tree_flatten(tree) assert treedef == treedef_ vals.extend(leaves) elif kind is pe.Append: pass elif kind is pe.BoxAttr: tree = obj.get() # not getattr! leaves, treedef_ = tree_flatten(tree) assert treedef == treedef_ vals.extend(leaves) elif kind is pe.ListAttr: pass else: assert False return vals def static(): return dataclasses.field(metadata=dict(static=True)) @tree_util.register_dataclass @dataclasses.dataclass class BoxTree: leaves: list treedef: PyTreeDef = static() @tree_util.register_dataclass @dataclasses.dataclass class ListTree: leaves: list treedef: PyTreeDef | None = static() def _flatten_boxes(dbg, args, kwargs): # TODO(mattjj,dougalm): refine this implementation of box-handling... if all(not isinstance(x, (Box, List)) for x in tree_leaves((args, kwargs))): return args, kwargs, [] box_data = [] id_first_occurrences = {} idxs = it.count() def visit(x): i = next(idxs) if (isinstance(x, (Box, List)) and (dup_idx := id_first_occurrences.setdefault(id(x), i)) != i): type_name = type(x).__name__ raise ValueError( f"a {type_name} instance can't be passed as an argument more than " f"once, but when tracing {dbg.func_src_info} for {dbg.traced_for}, " f"the object {x} appeared at both arguments " f"{dbg.arg_names[dup_idx]} and {dbg.arg_names[i]}" if dbg else f"at both flat index {dup_idx} and flat index {i}") if type(x) is Box: leaves, treedef = tree_flatten(x._val) ty = tuple(core.shaped_abstractify(l) for l in leaves) box_data.append((i, pe.BoxAttr)) return BoxTree(leaves, treedef) elif type(x) is List: box_data.append((i, pe.ListAttr)) return ListTree([], None) else: return x args, kwargs = tree_map(visit, (args, kwargs)) return args, kwargs, box_data # TODO(mattjj): because _handle_boxes's caller passes arguments splatted, the # names of its first two parameters must not collide with user-suppliedkwargs. # Using obscure names is a temporary workaround; revise! @lu.transformation2 def _handle_boxes(__f, __dbg, *args, **kwargs): f, dbg = __f, __dbg new_args = [] arg_mutables = [] def visit(x): if type(x) is BoxTree: box = Box(tree_unflatten(x.treedef, x.leaves)) arg_mutables.append(box) return box elif type(x) is ListTree: lst = List() lst._is_arg = True arg_mutables.append(lst) return lst else: return x args, kwargs = tree_map(visit, (args, kwargs), is_leaf=lambda x: isinstance(x, (BoxTree, ListTree))) out = f(*args, **kwargs) for path, leaf in tree_flatten_with_path(out)[0]: if isinstance(leaf, (Box, List)): type_name = type(leaf).__name__ raise ValueError( f"a {type_name} instance can't be returned from a transformed " f"function, but when tracing {dbg.func_src_info} for {dbg.traced_for} " f"the object {leaf} appeared at result{keystr(path)}") if not arg_mutables: return out extra_outs = [] for mutable in arg_mutables: if type(mutable) is Box: leaves, treedef = tree_flatten(mutable._val) extra_outs.append(BoxTree(leaves, treedef)) elif type(mutable) is List: leaves, treedef = tree_flatten(mutable._val) extra_outs.append(ListTree(leaves, treedef)) else: assert False return extra_outs, out