Source code for jax.experimental.pjit

# Copyright 2021 Google LLC
#
# 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 enum import IntEnum
import numpy as np
from collections import OrderedDict, Counter
from typing import Callable, Sequence, Tuple, Union
from warnings import warn
import itertools as it
from functools import partial

from jax.experimental import maps
from jax.experimental.global_device_array import GlobalDeviceArray as GDA
from jax import core
from jax import linear_util as lu
from jax._src.api import _check_callable, _check_arg, Lowered
from jax._src import source_info_util
from jax._src.api_util import (argnums_partial_except, flatten_axes,
                               flatten_fun_nokwargs, _ensure_index_tuple,
                               donation_vector, rebase_donate_argnums,
                               shaped_abstractify)
from jax.errors import JAXTypeError
from jax.interpreters import ad
from jax.interpreters import mlir
from jax.interpreters import pxla
from jax.interpreters import xla
from jax.interpreters import batching
from jax.interpreters import partial_eval as pe
from jax.interpreters.sharded_jit import PartitionSpec
from jax._src.lib import xla_client as xc
from jax.tree_util import tree_map, tree_flatten, tree_unflatten, tree_leaves
from jax._src.util import (extend_name_stack, HashableFunction, safe_zip,
                         wrap_name, wraps, distributed_debug_log,
                         split_list, cache, tuple_insert)
xops = xc._xla.ops

class _FromGsdaSingleton:
  pass
FROM_GDA = _FromGsdaSingleton()

# TODO(yashkatariya): Add pjit microbenchmarks.
[docs]def pjit(fun: Callable, in_axis_resources, out_axis_resources, static_argnums: Union[int, Sequence[int]] = (), donate_argnums: Union[int, Sequence[int]] = ()): """Makes ``fun`` compiled and automatically partitioned across multiple devices. 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_axis_resources`` and the output partitioning specified in ``out_axis_resources``. The resources specified in those two arguments must refer to mesh axes, as defined by the :py:func:`jax.experimental.maps.mesh` context manager. Note that the mesh definition at ``pjit`` application time is ignored, and the returned function will use the mesh definition available at each call site. Inputs to a pjit'd function will be automatically partitioned across devices if they're not already correctly partitioned based on ``in_axis_resources``. In some scenarios, ensuring that the inputs are already correctly pre-partitioned can increase performance. For example, if passing the output of one pjit'd function to another pjit’d function (or the same pjit’d function in a loop), make sure the relevant ``out_axis_resources`` match the corresponding ``in_axis_resources``. .. note:: **Multi-process platforms:** On multi-process platforms such as TPU pods, ``pjit`` can be used to run computations across all available devices across processes. To achieve this, ``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 pjit'd function in the same order. When running in this configuration, the mesh should contain devices across all processes. However, any input argument dimensions partitioned over multi-process mesh axes should be of size equal to the corresponding *local* mesh axis size, and outputs will be similarly sized according to the local mesh. ``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. However, outside of ``pjit`` every process only "sees" its local piece of the input and output, corresponding to its local sub-mesh. The SPMD model requires that the same multi-process ``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_axis_resources: 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 valid resource assignment specifications are: - :py:obj:`None`, in which case the value will be replicated on all devices - :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_axis_resources: Like ``in_axis_resources``, but specifies resource assignment for function 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. donate_argnums: Specify which arguments are "donated" to the computation. It is safe to donate arguments 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. Returns: A wrapped version of ``fun``, set up for just-in-time compilation and automaticly 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 ```pjit`` application: >>> import jax >>> import jax.numpy as jnp >>> from jax.experimental.maps import mesh >>> from jax.experimental.pjit import PartitionSpec, 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_axis_resources=None, out_axis_resources=PartitionSpec('devices')) >>> with mesh(jax.devices(), ('devices',)): ... print(f(x)) # doctest: +SKIP [ 0.5 2. 4. 6. 8. 10. 12. 10. ] """ warn("pjit is an experimental feature and probably has bugs!") _check_callable(fun) if isinstance(in_axis_resources, 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/google/jax/issues/2367 in_axis_resources = tuple(in_axis_resources) in_axis_resources, _, _ = _prepare_axis_resources( in_axis_resources, "in_axis_resources") out_axis_resources, _, _ = _prepare_axis_resources( out_axis_resources, "out_axis_resources") static_argnums = _ensure_index_tuple(static_argnums) donate_argnums = _ensure_index_tuple(donate_argnums) donate_argnums = rebase_donate_argnums(donate_argnums, static_argnums) def infer_params(*args, **kwargs): if kwargs: raise NotImplementedError("pjit does not support kwargs") if max(static_argnums + donate_argnums, default=-1) >= len(args): raise ValueError(f"jitted function has static_argnums={static_argnums}, " f"donate_argnums={donate_argnums} but " f"was called with only {len(args)} positional arguments.") # Putting this outside of wrapped would make resources lexically scoped resource_env = maps.thread_resources.env mesh = resource_env.physical_mesh if mesh.empty: raise RuntimeError("pjit requires a non-empty mesh! Are you sure that " "it's defined at the call site?") if any(d.platform not in {'gpu', 'tpu'} for d in mesh.devices.flat): raise RuntimeError("pjit only supports GPU and TPU devices") f = lu.wrap_init(fun) if static_argnums: f, dyn_args = argnums_partial_except( f, static_argnums, args, allow_invalid=False) else: dyn_args = args args_flat, in_tree = tree_flatten(args) flat_fun, out_tree = flatten_fun_nokwargs(f, in_tree) if donate_argnums: donated_invars = donation_vector(donate_argnums, dyn_args, ()) else: donated_invars = (False,) * len(args_flat) local_in_avals = tuple(shaped_abstractify(a) for a in args_flat) # TODO(yashkatariya): This is a hack. This should go away when avals have # is_global attribute. in_positional_semantics = tuple( maps._PositionalSemantics.GLOBAL if type(a) is GDA else maps ._positional_semantics for a in args_flat) out_positional_semantics = maps._positional_semantics jaxpr, in_axis_resources_flat, out_axis_resources_flat = _pjit_jaxpr( flat_fun, mesh, local_in_avals, in_tree, hashable_pytree(in_axis_resources), HashableFunction(out_tree, closure=()), hashable_pytree(out_axis_resources), in_positional_semantics, out_positional_semantics, tuple(isinstance(a, GDA) for a in args_flat)) in_axis_resources_flat = tree_map(_canonicalize_spec, in_axis_resources_flat, tuple(args_flat)) params = dict( jaxpr=jaxpr, in_axis_resources=in_axis_resources_flat, out_axis_resources=out_axis_resources_flat, resource_env=resource_env, donated_invars=donated_invars, name=flat_fun.__name__, in_positional_semantics=in_positional_semantics, out_positional_semantics=out_positional_semantics) return args_flat, params, in_tree, out_tree(), donate_argnums @wraps(fun) def wrapped(*args, **kwargs): for arg in tree_leaves(args): _check_arg(arg) args_flat, params, _, out_tree, _ = infer_params(*args, **kwargs) out = pjit_p.bind(*args_flat, **params) return tree_unflatten(out_tree, out) def lower(*args, **kwargs): args_flat, params, in_tree, out_tree, donate_argnums = \ infer_params(*args, **kwargs) lowering = _pjit_lower( params['jaxpr'], params['in_axis_resources'], params['out_axis_resources'], params['resource_env'], params['donated_invars'], params['name'], params['in_positional_semantics'], params['out_positional_semantics']) return Lowered(lowering, in_tree, out_tree, donate_argnums, no_kwargs=True) wrapped.lower = lower return wrapped
class _ListWithW(list): __slots__ = ('__weakref__',) def hashable_pytree(pytree): vals, treedef = tree_flatten(pytree) vals = tuple(vals) return HashableFunction(lambda: tree_unflatten(treedef, vals), closure=(treedef, vals)) def flatten_axis_resources(what, tree, axis_resources, tupled_args): try: return tuple(flatten_axes(what, tree, axis_resources, tupled_args=tupled_args)) except ValueError: pass # Replace axis_resources with unparsed versions to avoid revealing internal details flatten_axes(what, tree, tree_map(lambda parsed: parsed.user_spec, axis_resources), tupled_args=tupled_args) raise AssertionError("Please open a bug request!") # This should be unreachable @lu.cache def _pjit_jaxpr(fun, mesh, local_in_avals, in_tree, in_axis_resources_thunk, out_tree, out_axis_resources_thunk, in_positional_semantics, out_positional_semantics, is_gda): # TODO(yashkatariya): Make this work with FROM_GDA special value. in_axis_resources_flat = flatten_axis_resources( "pjit in_axis_resources", in_tree, in_axis_resources_thunk(), tupled_args=True) # This check should be above local_to_global call below otherwise if # `FROM_GDA` is passed to any input other than GDA, a ugly error message # will be raised because get_array_mapping (in local_to_global) of a # FROM_GDA cannot happen. tree_map(_check_resources_mismatch, in_axis_resources_flat, is_gda) # If all inputs are GDAs, then the avals are global and the mesh should also # be global. This split is because non-contiguous mesh can only be used if all # inputs are GDAs. if all(is_gda): _check_shapes_against_resources( "pjit arguments", mesh.is_multi_process, mesh.shape, local_in_avals, in_axis_resources_flat) else: _check_shapes_against_resources("pjit arguments", False, mesh.local_mesh.shape, local_in_avals, in_axis_resources_flat) global_in_avals = local_to_global(in_positional_semantics, mesh, local_in_avals, in_axis_resources_flat) prev_positional = maps._positional_semantics try: maps._positional_semantics = maps._PositionalSemantics.GLOBAL jaxpr, global_out_avals, consts = pe.trace_to_jaxpr_dynamic(fun, global_in_avals) finally: maps._positional_semantics = prev_positional jaxpr = core.ClosedJaxpr(jaxpr, consts) out_axis_resources_flat = flatten_axis_resources( "pjit out_axis_resources", out_tree(), out_axis_resources_thunk(), tupled_args=False) _check_shapes_against_resources("pjit outputs", mesh.is_multi_process, mesh.shape, global_out_avals, out_axis_resources_flat) # lu.cache needs to be able to create weakrefs to outputs, so we can't return a plain tuple return _ListWithW([jaxpr, in_axis_resources_flat, out_axis_resources_flat]) class SpecSync(IntEnum): """Encodes how much out of sync the real value of partitions is compared to the user specified one. We use this to make sure we don't show garbage modified values while claiming that the users have specified them like that. """ OUT_OF_SYNC = 0 # Arbitrary changes, including new axes inserted DIM_PERMUTE = 1 # Dimensions permuted, but no new sharding axes IN_SYNC = 2 # Entirely in sync class ParsedPartitionSpec: __slots__ = ('partitions', 'unsafe_user_spec', 'sync') def __init__(self, user_spec, partitions, sync=SpecSync.IN_SYNC): self.partitions = tuple(partitions) self.unsafe_user_spec = user_spec self.sync = sync @property def user_spec(self): return self.unsynced_user_spec(SpecSync.IN_SYNC) def unsynced_user_spec(self, min_sync): if self.sync < min_sync: raise AssertionError(f"Please open a bug report! ({self.sync} >= {min_sync})") return self.unsafe_user_spec def insert_axis_partitions(self, dim, val): parts = self.partitions too_short = dim - len(parts) if too_short > 0: parts += ((),) * too_short new_partitions = tuple_insert(parts, dim, val) new_sync = SpecSync.DIM_PERMUTE if val == () else SpecSync.OUT_OF_SYNC return ParsedPartitionSpec(self.unsafe_user_spec, new_partitions, sync=new_sync) @classmethod def from_user_input(cls, entry, arg_name): if entry is None: return cls(entry, ()) if not isinstance(entry, PartitionSpec): raise TypeError(f"{arg_name} are expected to be " f"PartitionSpec instances or None, but got {entry}") axis_specs = [] for axis_spec in entry: if axis_spec is None: axis_spec = () elif isinstance(axis_spec, (list, tuple)): axis_spec = tuple(axis_spec) else: axis_spec = (axis_spec,) axis_specs.append(axis_spec) return cls(entry, axis_specs) def __hash__(self): return hash((self.partitions, self.sync)) def __eq__(self, other): return (self.partitions == other.partitions and self.unsafe_user_spec == other.unsafe_user_spec and self.sync == other.sync) def __len__(self): return len(self.partitions) def __getitem__(self, i): return self.partitions[i] def __iter__(self): return iter(self.partitions) def __repr__(self): return f"<partitions={self.partitions} sync={self.sync}>" REPLICATED = ParsedPartitionSpec(None, ()) def _prepare_axis_resources(axis_resources, arg_name): # PyTrees don't treat None values as leaves, so we explicitly need # to explicitly declare them as such entries, treedef = tree_flatten(axis_resources, is_leaf=lambda x: x is None) what = f"{arg_name} leaf specifications" entries = [ entry if entry is FROM_GDA else ParsedPartitionSpec.from_user_input( entry, what) for entry in entries ] _check_unique_resources(entries, arg_name) return tree_unflatten(treedef, entries), entries, treedef def _check_resources_mismatch(in_axis_resources_flat, is_gda): if not is_gda and in_axis_resources_flat is FROM_GDA: raise ValueError('For a non-GDA input, the corresponding resource in ' 'in_axis_resources cannot be `pjit.FROM_GDA`.') def _check_unique_resources(axis_resources, arg_name): for arg_axis_resources in axis_resources: if not arg_axis_resources: continue if arg_axis_resources is FROM_GDA: continue resource_counts = Counter(it.chain.from_iterable(arg_axis_resources)) if not resource_counts: continue if resource_counts.most_common(1)[0][1] > 1: multiple_uses = [r for r, c in resource_counts.items() if c > 1] if multiple_uses: raise ValueError(f"A single {arg_name} specification can map every mesh axis " f"to at most one positional dimension, but {arg_axis_resources.user_spec} " f"has duplicate entries for {maps.show_axes(multiple_uses)}") def _check_shapes_against_resources(what: str, is_global_shape: bool, mesh_shape, flat_avals, flat_axis_resources): global_str = " global" if is_global_shape else "" for aval, aval_axis_resources in zip(flat_avals, flat_axis_resources): if aval_axis_resources is FROM_GDA: continue shape = aval.shape if len(shape) < len(aval_axis_resources): raise ValueError(f"One of {what} was given the resource assignment " f"of {aval_axis_resources.user_spec}, which implies that " f"it has a rank of at least {len(aval_axis_resources)}, " f"but it is {len(shape)}") for i, axis_resources in enumerate(aval_axis_resources): try: size = int(np.prod([mesh_shape[resource] for resource in axis_resources], dtype=np.int64)) except KeyError as e: raise ValueError(f"One of {what} was given the resource assignment " f"of {aval_axis_resources.user_spec}, but resource axis " f"{e.args[0]} is undefined. Did you forget to declare the mesh?") from None if shape[i] % size != 0: raise ValueError(f"One of {what} was given the resource assignment " f"of {aval_axis_resources.user_spec}, which implies that " f"the{global_str} size of its dimension {i} should be " f"divisible by {size}, but it is equal to {shape[i]}") # -------------------- pjit rules -------------------- pjit_p = core.Primitive("pjit") pjit_p.multiple_results = True def _pjit_call_impl(*args, jaxpr, in_axis_resources, out_axis_resources, resource_env, donated_invars, name, in_positional_semantics, out_positional_semantics): compiled = _pjit_lower( jaxpr, in_axis_resources, out_axis_resources, resource_env, donated_invars, name, in_positional_semantics, out_positional_semantics).compile() distributed_debug_log(("Running pjit'd function", name), ("mesh", resource_env.physical_mesh)) return compiled.unsafe_call(*args) pjit_p.def_impl(_pjit_call_impl) @cache() def _pjit_lower( jaxpr: core.ClosedJaxpr, in_axis_resources: Tuple[ParsedPartitionSpec, ...], out_axis_resources: Tuple[ParsedPartitionSpec, ...], resource_env, donated_invars, name: str, in_positional_semantics, out_positional_semantics): in_axes = [get_array_mapping(axes) for axes in in_axis_resources] out_axes = [get_array_mapping(axes) for axes in out_axis_resources] pxla.resource_typecheck(jaxpr, resource_env, {}, lambda: "pjit") f = core.jaxpr_as_fun(jaxpr) f.__name__ = name fun = lu.wrap_init(f) return pxla.lower_mesh_computation( fun, name, resource_env.physical_mesh, in_axes, out_axes, donated_invars, True, jaxpr.in_avals, tile_by_mesh_axes=False) def _pjit_abstract_eval(*args, jaxpr, out_axis_resources, resource_env, out_positional_semantics, **_): return global_to_local(out_positional_semantics, resource_env.physical_mesh, jaxpr.out_avals, out_axis_resources) pjit_p.def_abstract_eval(_pjit_abstract_eval) def _pjit_translation_rule(ctx, avals_in, avals_out, *in_nodes, name, jaxpr, in_axis_resources, out_axis_resources, resource_env, donated_invars, in_positional_semantics, out_positional_semantics): mesh = resource_env.physical_mesh subc = xc.XlaBuilder(f"pjit_{name}") args = [] for i, (n, aval, axis_resources) in enumerate( safe_zip(in_nodes, avals_in, in_axis_resources)): # N.B. inlined calls shouldn't have shardings set directly on the inputs or # outputs (set_sharding_proto adds an identity operation). arg = xla.parameter(subc, i, ctx.builder.GetShape(n)) args.append( xla.set_sharding_proto( subc, arg, get_aval_sharding_proto(aval, axis_resources, mesh))) # TODO: Think about how to avoid duplicating constants with the outer jaxpr sub_ctx = ctx.replace( builder=subc, name_stack=extend_name_stack(ctx.name_stack, wrap_name(name, "pjit"))) out_nodes = xla.jaxpr_subcomp( sub_ctx, jaxpr.jaxpr, xla._xla_consts(subc, jaxpr.consts), *args) out_nodes = [ xla.set_sharding_proto(subc, out, get_aval_sharding_proto(aval, axis_resources, mesh)) for out, aval, axis_resources in safe_zip( out_nodes, avals_out, out_axis_resources) ] subc = subc.build(xops.Tuple(subc, out_nodes)) return xla.xla_destructure(ctx.builder, xops.Call(ctx.builder, subc, list(in_nodes))) xla.register_translation(pjit_p, _pjit_translation_rule) def _pjit_batcher(insert_axis, axis_size, axis_name, main_type, vals_in, dims_in, jaxpr, in_axis_resources, out_axis_resources, resource_env, donated_invars, name, in_positional_semantics, out_positional_semantics): # batch_jaxpr expects all batching dimensions to be equal to 0 vals_in = [batching.moveaxis(x, d, 0) if d is not batching.not_mapped and d != 0 else x for x, d in zip(vals_in, dims_in)] is_mapped_in = [d is not batching.not_mapped for d in dims_in] new_jaxpr, is_mapped_out = batching.batch_jaxpr( jaxpr, axis_size, is_mapped_in, instantiate=False, axis_name=axis_name, main_type=main_type) new_parts = (axis_name,) if insert_axis else () in_axis_resources = tuple( spec.insert_axis_partitions(0, new_parts) if is_mapped else spec for is_mapped, spec in zip(is_mapped_in, in_axis_resources)) out_axis_resources = tuple( spec.insert_axis_partitions(0, new_parts) if is_mapped else spec for is_mapped, spec in zip(is_mapped_out, out_axis_resources)) vals_out = pjit_p.bind( *vals_in, jaxpr=new_jaxpr, in_axis_resources=in_axis_resources, out_axis_resources=out_axis_resources, resource_env=resource_env, donated_invars=donated_invars, name=name, in_positional_semantics=in_positional_semantics, out_positional_semantics=out_positional_semantics) dims_out = [0 if batched else batching.not_mapped for batched in is_mapped_out] return vals_out, dims_out batching.axis_primitive_batchers[pjit_p] = partial(_pjit_batcher, False) pxla.spmd_primitive_batchers[pjit_p] = partial(_pjit_batcher, True) def _pjit_jvp(primals_in, tangents_in, jaxpr, in_axis_resources, out_axis_resources, resource_env, donated_invars, name, in_positional_semantics, out_positional_semantics): 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_axis_resources=(*in_axis_resources, *_filter_zeros_in(in_axis_resources)), out_axis_resources=(*out_axis_resources, *_filter_zeros_out(out_axis_resources)), resource_env=resource_env, donated_invars=(*donated_invars, *_filter_zeros_in(donated_invars)), name=wrap_name(name, 'jvp'), in_positional_semantics=(*in_positional_semantics, *_filter_zeros_in(in_positional_semantics)), out_positional_semantics=out_positional_semantics) 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_partial_eval(trace, *in_tracers, jaxpr, in_axis_resources, out_axis_resources, resource_env, donated_invars, name, in_positional_semantics, out_positional_semantics): # XXX: At the moment all residuals get fully replicated, which is extremely # wasteful and might quickly lead to OOM errors. mesh = resource_env.physical_mesh 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) raw_known_jaxpr, raw_unknown_jaxpr, unknown_outs = pe.partial_eval_jaxpr( jaxpr, unknown_ins, instantiate=False) unknown_outs = tuple(unknown_outs) known_outs = tuple(not uk for uk in unknown_outs) num_residuals = len(raw_known_jaxpr.jaxpr.outvars) - len(unknown_outs) def keep_where(l, should_keep): return tuple(x for x, keep in zip(l, should_keep) if keep) # Prepare the known jaxpr # TODO(apaszke): map_jaxpr will break caching! known_jaxpr = raw_known_jaxpr.map_jaxpr(lambda jaxpr: pe._drop_vars( jaxpr, drop_ins=unknown_ins, drop_outs=unknown_outs + (False,) * num_residuals)) # Compute the known outputs known_params = dict( jaxpr=known_jaxpr, in_axis_resources=keep_where(in_axis_resources, known_ins), out_axis_resources=(keep_where(out_axis_resources, known_outs) + (REPLICATED,) * num_residuals), resource_env=resource_env, donated_invars=keep_where(donated_invars, known_ins), name=name, in_positional_semantics=keep_where(in_positional_semantics, known_ins), out_positional_semantics=out_positional_semantics) if num_residuals: executable = _pjit_lower(**known_params).compile( _allow_propagation_to_outputs=True, _allow_compile_replicated=False) output_op_sharding = \ executable.xla_executable.hlo_modules()[0].spmd_output_sharding output_sharding_specs = parse_op_sharding(output_op_sharding, mesh) residual_specs = tuple(output_sharding_specs[-num_residuals:]) else: residual_specs = () known_params['out_axis_resources'] = ( keep_where(out_axis_resources, known_outs) + residual_specs) all_known_outs = pjit_p.bind( *(pv.get_known() for pv in in_pvals if pv.is_known()), **known_params) if num_residuals: known_out_vals, residual_vals = split_list(all_known_outs, [-num_residuals]) else: known_out_vals, residual_vals = all_known_outs, () known_tracers_out = [trace.new_const(known_out) for known_out in known_out_vals] residual_tracers = [trace.new_instantiated_const(residual) for residual in residual_vals] # Prepare the unknown jaxpr # TODO(apaszke): map_jaxpr will break caching! unknown_jaxpr = raw_unknown_jaxpr.map_jaxpr(lambda jaxpr: pe._drop_vars( jaxpr, drop_ins=known_ins + (False,) * num_residuals, drop_outs=known_outs)) # Prepare unknown tracers unknown_params = dict( jaxpr=unknown_jaxpr, in_axis_resources=(keep_where(in_axis_resources, unknown_ins) + residual_specs), out_axis_resources=keep_where(out_axis_resources, unknown_outs), resource_env=resource_env, donated_invars=(keep_where(donated_invars, unknown_ins) + (False,) * num_residuals), name=name, in_positional_semantics=(keep_where( in_positional_semantics, unknown_ins) + (out_positional_semantics,) * num_residuals), out_positional_semantics=out_positional_semantics) unknown_tracers_in = [t for t in in_tracers if not t.pval.is_known()] unknown_tracers_out = [ pe.JaxprTracer(trace, pe.PartialVal.unknown(aval), None) for aval in global_to_local(unknown_params["out_positional_semantics"], mesh, unknown_jaxpr.out_avals, unknown_params["out_axis_resources"]) ] eqn = pe.new_eqn_recipe((*unknown_tracers_in, *residual_tracers), unknown_tracers_out, pjit_p, unknown_params, source_info_util.current()) for t in unknown_tracers_out: t.recipe = eqn return pe._zip_knowns(known_tracers_out, unknown_tracers_out, unknown_outs) pe.custom_partial_eval_rules[pjit_p] = _pjit_partial_eval def _pjit_transpose(reduce_axes, cts_in, *primals_in, jaxpr, in_axis_resources, out_axis_resources, resource_env, donated_invars, name, in_positional_semantics, out_positional_semantics): mesh = resource_env.physical_mesh def prune_type(ty, xs, maybe_zeros): return tuple(x for x, mz in zip(xs, maybe_zeros) if not type(mz) is ty) body = lu.wrap_init(ad.closed_backward_pass) body = lu.hashable_partial(body, jaxpr, reduce_axes) 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_axis_resources = ( *prune_type(ad.UndefinedPrimal, in_axis_resources, primals_in), *prune_type(ad.Zero, out_axis_resources, cts_in) ) transpose_in_positional_semantics = ( *prune_type(ad.UndefinedPrimal, in_positional_semantics, primals_in), *prune_type(ad.Zero, (out_positional_semantics,) * len(cts_in), cts_in) ) global_cts_in_avals = local_to_global( transpose_in_positional_semantics, mesh, [core.raise_to_shaped(core.get_aval(ct)) for ct in primals_and_nz_cts_in], transpose_in_axis_resources) transpose_jaxpr, global_cts_out_avals, consts = pe.trace_to_jaxpr_dynamic( body, global_cts_in_avals) # TODO(apaszke): Creating ClosedJaxpr by hand will break compilation cache! transpose_jaxpr = core.ClosedJaxpr(transpose_jaxpr, consts) del consts cts_out_treedef = cts_out_treedef_thunk() transpose_out_axis_resources = prune_type( ad.Zero, in_axis_resources, tree_unflatten(cts_out_treedef, [object()] * cts_out_treedef.num_leaves)) nz_cts_out = pjit_p.bind( *primals_and_nz_cts_in, jaxpr=transpose_jaxpr, in_axis_resources=transpose_in_axis_resources, out_axis_resources=transpose_out_axis_resources, resource_env=resource_env, donated_invars=(False,) * len(primals_and_nz_cts_in), name=name, in_positional_semantics=transpose_in_positional_semantics, out_positional_semantics=out_positional_semantics) return tree_unflatten(cts_out_treedef, nz_cts_out) ad.reducing_transposes[pjit_p] = _pjit_transpose def _check_resources_against_named_axes(what, aval, pos_axis_resources, named_axis_resources): pjit_resources = set(it.chain.from_iterable(pos_axis_resources)) aval_resources = set(it.chain.from_iterable( named_axis_resources[a] for a in aval.named_shape)) overlap = pjit_resources & aval_resources if overlap: raise JAXTypeError( f"{what} has an axis resources specification of " f"{pos_axis_resources.unsynced_user_spec(SpecSync.DIM_PERMUTE)} " f"that uses one or more mesh axes already used by xmap to partition " f"a named axis appearing in its named_shape (both use mesh axes " f"{maps.show_axes(overlap)})") def _resource_typing_pjit(avals, params, source_info, resource_env, named_axis_resources): jaxpr = params["jaxpr"] what = "pjit input" if resource_env.physical_mesh != params['resource_env'].physical_mesh: raise RuntimeError("Changing the physical mesh is not allowed inside pjit.") for aval, pos_axis_resources in zip(jaxpr.in_avals, params['in_axis_resources']): _check_resources_against_named_axes(what, aval, pos_axis_resources, named_axis_resources) pxla.resource_typecheck( jaxpr.jaxpr, resource_env, named_axis_resources, lambda: (f"a pjit'ed function {params['name']} " f"(pjit called at {source_info_util.summarize(source_info)})")) what = "pjit output" for aval, pos_axis_resources in zip(jaxpr.out_avals, params['out_axis_resources']): _check_resources_against_named_axes(what, aval, pos_axis_resources, named_axis_resources) pxla.custom_resource_typing_rules[pjit_p] = _resource_typing_pjit # -------------------- with_sharding_constraint -------------------- def with_sharding_constraint(x, axis_resources): x_flat, tree = tree_flatten(x) parsed_axis_resources, entries, _ = _prepare_axis_resources(axis_resources, "axis_resources") axis_resources_flat = tuple( flatten_axes("with_sharding_constraint axis_resources", tree, parsed_axis_resources)) resource_env = maps.thread_resources.env mesh = resource_env.physical_mesh _check_shapes_against_resources( "with_sharding_constraint arguments", mesh.is_multi_process, mesh.shape, x_flat, axis_resources_flat) outs = [sharding_constraint_p.bind(y, axis_resources=r, resource_env=resource_env) for y, r in safe_zip(x_flat, axis_resources_flat)] return tree_unflatten(tree, outs) def _sharding_constraint_impl(x, axis_resources, resource_env): # TODO(skye): can we also prevent this from being called in other # non-pjit contexts? (e.g. pmap, control flow) raise NotImplementedError( "with_sharding_constraint() should only be called inside pjit()") 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, _, axis_resources, resource_env: ( sharding_constraint_p.bind( ct, axis_resources=axis_resources, resource_env=resource_env),)) def _sharding_constraint_translation_rule(ctx, avals_in, avals_out, x_node, *, axis_resources, resource_env): aval, = avals_in mesh = resource_env.physical_mesh return [xla.set_sharding_proto( ctx.builder, x_node, get_aval_sharding_proto(aval, axis_resources, mesh))] xla.register_translation(sharding_constraint_p, _sharding_constraint_translation_rule) def _sharding_constraint_mhlo_lowering(ctx, avals_in, avals_out, x_node, *, axis_resources, resource_env): aval, = avals_in mesh = resource_env.physical_mesh return [mlir.wrap_with_sharding_op( x_node, get_aval_sharding_proto(aval, axis_resources, mesh))] mlir.register_lowering(sharding_constraint_p, _sharding_constraint_mhlo_lowering) def _sharding_constraint_batcher(insert_axis, axis_size, axis_name, main_type, vals_in, dims_in, axis_resources, resource_env): x, = vals_in d, = dims_in new_parts = (axis_name,) if insert_axis else () y = sharding_constraint_p.bind( x, axis_resources=axis_resources.insert_axis_partitions(d, new_parts), resource_env=resource_env) return y, d batching.axis_primitive_batchers[sharding_constraint_p] = partial(_sharding_constraint_batcher, False) pxla.spmd_primitive_batchers[sharding_constraint_p] = partial(_sharding_constraint_batcher, True) def _resource_typing_sharding_constraint(avals, params, source_info, resource_env, named_axis_resources): aval, = avals _check_resources_against_named_axes( "with_sharding_constraint input", aval, params['axis_resources'], named_axis_resources) pxla.custom_resource_typing_rules[sharding_constraint_p] = \ _resource_typing_sharding_constraint # -------------------- helpers -------------------- def get_array_mapping(axis_resources: ParsedPartitionSpec) -> pxla.ArrayMapping: return OrderedDict((axis, i) for i, axes in enumerate(axis_resources) for axis in axes) def get_aval_sharding_proto(aval: core.AbstractValue, axis_resources: ParsedPartitionSpec, mesh: maps.Mesh) -> xc.OpSharding: array_mapping = get_array_mapping(axis_resources) sharding_spec = pxla.mesh_sharding_specs(mesh.shape, mesh.axis_names)( aval, array_mapping) return sharding_spec.sharding_proto() def global_to_local(positional_semantics, mesh, avals, axes): if isinstance(positional_semantics, maps._PositionalSemantics): positional_semantics = [positional_semantics] * len(axes) return [ aval if ps == maps._PositionalSemantics.GLOBAL else mesh.global_to_local( get_array_mapping(aval_axes), aval) for aval, aval_axes, ps in safe_zip(avals, axes, positional_semantics) ] def local_to_global(positional_semantics, mesh, avals, axes): return [ aval if ps == maps._PositionalSemantics.GLOBAL else mesh.local_to_global( get_array_mapping(aval_axes), aval) for aval, aval_axes, ps in safe_zip(avals, axes, positional_semantics) ] def _canonicalize_spec(in_axis_resources_flat: ParsedPartitionSpec, arg): if isinstance(arg, GDA): gsda_ppspec = gsda_mesh_axes_to_parsed_pspec(arg._mesh_axes) if in_axis_resources_flat is not FROM_GDA and in_axis_resources_flat != gsda_ppspec: raise ValueError( 'Got an input GDA to pjit with different partitioning than specified in ' 'the in_axis_resources argument to pjit. The paritioning must match, or ' "use `jax.experimental.pjit.FROM_GDA` in `in_axis_resources`. " f'Got GDA spec: {gsda_ppspec}, pjit spec: {in_axis_resources_flat}') return gsda_ppspec return in_axis_resources_flat def gsda_mesh_axes_to_parsed_pspec(mesh_axes) -> ParsedPartitionSpec: if not isinstance(mesh_axes, PartitionSpec): pspec = PartitionSpec(*mesh_axes) else: pspec = mesh_axes return ParsedPartitionSpec.from_user_input(pspec, arg_name='GDA mesh_axes') # -------------------- XLA OpSharding to PartitionSpec -------------------- # Note that OpSharding is more expressive than PartitionSpecs, so it's not # always possible to convert them, but the code below should at least # support handle all cases when this is possible. def strides_for_sizes(sizes): """Returns an array of strides for major-to-minor sizes.""" return np.cumprod(sizes[::-1])[::-1] // np.asarray(sizes) def unflatten_array(named_sizes, assignment): """Recovers the ordering of axis names based on a device assignment. The device assignments that this function can convert into axis orders are of the form:: np.arange(np.prod(named_sizes.values())).transpose(...).flatten() for some transposition ``...``. This is satisfied by all OpSharding assignments generated from partition specs. Arguments: named_sizes: A dictionary mapping axis names to their sizes. assignment: A permutation of integers between 0 and the product of all named sizes. Returns: A major-to-minor list of axis names that corresponds to the given assignment. """ named_sizes = {name: size for name, size in named_sizes.items() if size != 1} sizes = np.fromiter(named_sizes.values(), dtype=np.int64) strides = strides_for_sizes(sizes) dims = explode_superdims(sizes, unflatten_superdims(assignment)) dim_to_name = {(size, stride): name for size, stride, name in zip(sizes, strides, named_sizes)} return [dim_to_name[d] for d in dims] def unflatten_superdims(assignment): """Unflatten a list of dimension sizes and their strides that generates assignment. If this function succeeds for a given ``assignment``, then the following property should be satisfied:: dims_with_strides = unflatten_superdims(assignment) base_array = np.arange(map(fst, sorted(dims_with_strides, key=snd, reverse=True))) assignment == base_array.transpose(argsort(dims_with_strides, key=snd, reverse=True)).flatten() That is, the returned dimensions list all sizes of the base array (with strides indicating their initial order). The order of dimensions in the list corresponds to the permutation that applied to the base array generates the assignment. """ def check(cond): if cond: return raise NotImplementedError("Failed to convert OpSharding into a ShardingSpec. " "Please open a bug report!") flat_assignment = np.asarray(assignment, dtype=np.int64) check(flat_assignment[0] == 0) dims = [] while flat_assignment.size > 1: stride = flat_assignment[1] for i in range(len(flat_assignment)): if flat_assignment[i] != i * stride: break else: # After this loop i should point to an "element after the sequence", so # we have to increment it if the whole array is a strided sequence. i += 1 size = i dims.append((size, stride)) assert size > 1 # Ensure progress flat_assignment = flat_assignment[::size] return dims def explode_superdims(sizes, dims): """Explode superdims to fit a known shape. The unflattening process might mistakenly generate too few too large dimensions. For example, ``unflatten_superdims(np.arange(n))`` always returns ``[(n, 1)]``. This function takes a list of such contiguous super-dimensions and splits them into smaller dimensions such that:: set(map(fst, explode_superdims(sizes, dims))) == set(sizes) """ strides_to_sizes = {stride: size for size, stride in zip(sizes, strides_for_sizes(sizes))} dims = list(reversed(dims)) final_dims = [] for size, stride in dims: target_size = strides_to_sizes[stride] new_dims = [] while size > target_size: assert target_size > 1 # Ensure progress assert size % target_size == 0 new_dims.append((target_size, stride)) size //= target_size stride *= target_size target_size = strides_to_sizes[stride] assert size == target_size new_dims.append((size, stride)) final_dims += reversed(new_dims) return final_dims def parse_op_sharding(op_sharding, mesh): if op_sharding.type == xc.OpSharding.Type.TUPLE: return [parse_op_sharding(s, mesh) for s in op_sharding.tuple_shardings] elif op_sharding.type == xc.OpSharding.Type.REPLICATED: return REPLICATED elif op_sharding.type == xc.OpSharding.Type.OTHER: mesh_shape = mesh.shape mesh_axis_order = unflatten_array(mesh.shape, op_sharding.tile_assignment_devices) mesh_axis = iter(mesh_axis_order) shape = op_sharding.tile_assignment_dimensions partitions = [] for dim_size in shape: dim_partitions = [] while dim_size > 1: axis = next(mesh_axis) axis_size = mesh_shape[axis] assert dim_size % axis_size == 0 dim_size //= axis_size dim_partitions.append(axis) partitions.append(tuple(dim_partitions)) if op_sharding.replicate_on_last_tile_dim: partitions = partitions[:-1] return ParsedPartitionSpec('<internally generated spec>', partitions) else: raise AssertionError("Unhandled OpSharding type. Please open a bug report!")