rocm_jax/jax/experimental/multihost_utils.py

# 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.
"""Utilities for synchronizing and communication across multiple hosts."""

from __future__ import annotations

from functools import partial, lru_cache
import zlib

from typing import Any
import jax
import jax.numpy as jnp
from jax.tree_util import tree_flatten, tree_unflatten
from jax._src import core
from jax._src.interpreters import ad
from jax._src.interpreters import batching
from jax._src.interpreters import mlir
from jax._src import array
from jax._src import sharding_impls
from jax._src.interpreters import pxla
from jax.interpreters import xla
from jax._src import pjit as pjit_lib
from jax.sharding import PartitionSpec as P
from jax._src import distributed
from jax._src.util import safe_zip
from jax._src import xla_bridge
from jax._src.lib import xla_client
import numpy as np


def _psum(x: Any) -> Any:
  return jax.tree.map(partial(jnp.sum, axis=0), x)


def broadcast_one_to_all(in_tree: Any, is_source: bool | None = None) -> Any:
  """Broadcast data from a source host (host 0 by default) to all other hosts.

  Args:
    in_tree: pytree of arrays - each array *must* have the same shape across the
      hosts.
    is_source: optional bool denoting whether the caller is the source. Only
      'source host' will contribute the data for the broadcast. If None, then
      host 0 is used.

  Returns:
    A pytree matching in_tree where the leaves now all contain the data from the
    first host.
  """
  if jax.process_count() == 1:
    return jax.tree.map(np.asarray, in_tree)

  if is_source is None:
    is_source = jax.process_index() == 0

  devices: np.ndarray = np.array(
      jax.devices()).reshape(jax.process_count(), jax.local_device_count())
  global_mesh = jax.sharding.Mesh(devices, ('processes', 'local_devices'))
  pspec = P('processes')

  def pre_jit(x):
    if is_source:
      inp = x
    else:
      inp = np.zeros_like(x)
    inp = np.expand_dims(inp, axis=0)
    return host_local_array_to_global_array(inp, global_mesh, pspec)

  def post_jit(x):
    return jax.device_get(x.addressable_data(0))

  in_tree = jax.tree.map(pre_jit, in_tree)
  out_tree = jax.jit(_psum, out_shardings=jax.sharding.NamedSharding(
      global_mesh, P()))(in_tree)
  return jax.tree.map(post_jit, out_tree)


def sync_global_devices(name: str):
  """Creates a barrier across all hosts/devices."""
  h = np.uint32(zlib.crc32(name.encode()))
  assert_equal(h, f"sync_global_devices name mismatch ('{name}')")


# Identity function is at the top level so that `process_allgather` doesn't
# recompile on every invocation.
def _identity_fn(x):
  return x


def _handle_array_process_allgather(inp, tiled):
  if isinstance(inp, array.ArrayImpl) and not inp.is_fully_addressable:
    if isinstance(inp.sharding, sharding_impls.NamedSharding):
      reps = inp.sharding.with_spec(P())
    else:
      reps = sharding_impls.GSPMDSharding.get_replicated(
          inp.sharding._device_assignment, memory_kind=inp.sharding.memory_kind)
    out = jax.jit(_identity_fn, out_shardings=reps)(inp)
  else:
    # All inputs here will be fully addressable.
    if jax.process_count() == 1:
      out = np.asarray(inp)
      return np.expand_dims(out, axis=0) if not tiled else out

    devices = np.array(jax.devices()).reshape(jax.process_count(),
                                              jax.local_device_count())
    global_mesh = jax.sharding.Mesh(devices, ('processes', 'local_devices'))
    pspec = P('processes')
    s = jax.sharding.NamedSharding(global_mesh, pspec)

    host_np_arr = np.asarray(inp)
    if host_np_arr.ndim == 0 or not tiled:
      host_np_arr = np.expand_dims(host_np_arr, axis=0)

    aval = core.ShapedArray(host_np_arr.shape, host_np_arr.dtype)
    global_aval = pxla.mesh_local_to_global(
        global_mesh, pxla.get_array_mapping(pspec), aval)

    bufs = [jax.device_put(host_np_arr, d) for d in jax.local_devices()]
    global_arr = array.make_array_from_single_device_arrays(
        global_aval.shape, s, bufs)
    out = jax.jit(_identity_fn,
                  out_shardings=jax.NamedSharding(global_mesh, P()))(global_arr)

  return np.asarray(out.addressable_data(0))


def process_allgather(in_tree: Any, tiled: bool = False) -> Any:
  """Gather data from across processes.

  Args:
    in_tree: pytree of arrays - each array _must_ have the same shape across the
      hosts.
    tiled: Whether to stack or concat the output. Defaults to False i.e. stack
      into a new positional axis at index 0.

  Returns:
    Pytrees of numpy arrays.
      * If the input is a non-fully addressable jax.Array, then the data is
        fully replicated.
      * If the input is numpy array or fully addressable jax.Array, then the
        output shape is dependent on the `tiled` argument.
        If its False, then the output will be stacked else concatenated.
      * If the input is a scalar, then the output will be stacked.
  """

  def _pjit(inp):
    return _handle_array_process_allgather(inp, tiled)
  return jax.tree.map(_pjit, in_tree)


def assert_equal(in_tree, fail_message: str = ''):
  """Verifies that all the hosts have the same tree of values."""
  expected = broadcast_one_to_all(in_tree)
  if not jax.tree_util.tree_all(
      jax.tree_util.tree_map(lambda *x: np.all(np.equal(*x)), in_tree, expected)):
    raise AssertionError(
        f'{fail_message} Expected: {expected}; got: {in_tree}.')


def reached_preemption_sync_point(step_id: int) -> bool:
  """Determine whether all hosts have reached a preemption sync step.

  When any host receives a preemption notice, the notice is propagated to all
  hosts and triggers a synchronization protocol in the background. The
  synchronization protocol calculates the maximum step ids from all hosts, and
  uses the next step id (i.e., max + 1) as the safe step to save a checkpoint.
  All hosts should continue training more steps until this method returns True,
  indicating that the `step_id` is equal to the safe step and the hosts should
  start saving a checkpoint.

  To use this API, all hosts must start training from the same step and call it
  at every training step. Example usage:

  ```
  def should_save(step_id: int) -> bool:

    # Should save an on-demand checkpoint for preemption
    if multihost_utils.reached_preemption_sync_point(step_id):
      return True

    # Should save a regular checkpoint
    return step_id - last_saved_checkpoint_step >= save_interval_steps
  ```

  Preemption notice is provided by the cluster scheduler to notify the
  application in advance before it gets evicted. By default, we use SIGTERM as
  the signal for preemption notice.

  TODO(b/230630494): Add instructions for customized preemption notice.

  Returns:
    A boolean indicating whether all hosts have reached a synchronization step
    after some hosts are preempted.

  Raises:
    RuntimeError: if preemption sync manager has not been inititialized.
  """
  if distributed.global_state.client is None:
    return False
  sync_manager = distributed.global_state.preemption_sync_manager
  if sync_manager is None:
    raise RuntimeError("Preemption sync manager has not been initialized.")
  return sync_manager.reached_sync_point(step_id)


@lru_cache
def _flatten_pspecs(name, in_tree, pspecs_thunk):
  return pjit_lib.flatten_axis_resources(
      name, in_tree, pspecs_thunk(), tupled_args=True)

@lru_cache
def _local_to_global_aval(local_aval, mesh, pspec):
  return pxla.mesh_local_to_global(mesh, pxla.get_array_mapping(pspec),
                                   local_aval)

@lru_cache
def _global_to_local_aval(global_aval, mesh, pspec):
  return pxla.mesh_global_to_local(mesh, pxla.get_array_mapping(pspec),
                                   global_aval)


def host_local_array_to_global_array_impl(
    arr: Any, *, global_mesh: jax.sharding.Mesh, pspec: Any):
  if pspec is None:
    raise ValueError(
        '`None` is not a valid input to the pspecs argument. Please use '
        'jax.sharding.PartitionSpec() if you wanted to replicate your input.')
  # If the Array is not fully addressable i.e. not host local, return it.
  if isinstance(arr, array.ArrayImpl) and not arr.is_fully_addressable:
    return arr
  if isinstance(arr, array.ArrayImpl) and isinstance(
      arr.sharding, jax.sharding.PmapSharding):
    arr = np.array(arr)

  local_sharding = jax.sharding.NamedSharding(global_mesh.local_mesh, pspec)

  # If the input is a concrete jax.Array and the input array sharding
  # matches the `local_sharding`, then there's no need to reshard and create
  # copies.
  if (isinstance(arr, array.ArrayImpl) and
      arr.sharding.is_equivalent_to(local_sharding, arr.ndim)):
    arrays = [x.data for x in arr.addressable_shards]
  else:
    arr = xla.canonicalize_dtype(arr)
    arrays = [
        arr[index]
        for d, index in local_sharding.devices_indices_map(arr.shape).items()]

  global_aval = _local_to_global_aval(
      core.ShapedArray(arr.shape, arr.dtype), global_mesh, pspec)

  return pxla.batched_device_put(
      global_aval, jax.sharding.NamedSharding(global_mesh, pspec),
      arrays, list(global_mesh.local_mesh.devices.flat))


def host_local_array_to_global_array(
    local_inputs: Any, global_mesh: jax.sharding.Mesh, pspecs: Any):
  r"""Converts a host local value to a globally sharded jax.Array.

  This function takes host-local data (which might be different
  across hosts), and populates a global array with this data, where each
  device on each host, get the appropriate slice of the data according to
  sharding defined by the global_mesh/pspects.

  For example:

  >>> global_mesh = jax.sharding.Mesh(jax.devices(), 'x')
  >>> pspecs = jax.sharding.PartitionSpec('x')
  >>> host_id = jax.process_index()
  >>> arr = host_local_array_to_global_array(np.arange(4) * host_id, mesh, pspecs)  # NB: assumes jax.local_device_count() divides 4.   # doctest: +SKIP

  The resulting array will have the shape (4 * num_processes) and will
  have distributed value of: (0, 1, 2, 3, 0, 2, 4, 6, 0, 3, 6, 9, ... ),
  where each slice np.arange(4) * host_id will be partitioned across the
  corresponding host's devices.

  Similarly:

  >>> mesh = jax.sharding.Mesh(np.array(jax.devices()).reshape(jax.process_count(), jax.local_device_count()), ['host', 'dev'])
  >>> pspecs = jax.sharding.PartitionSpec('host')
  >>> host_id = jax.process_index()
  >>> arr = host_local_array_to_global_array(np.arange(4) * host_id, mesh, pspecs)  # doctest: +SKIP

  will create the same distributed value (0, 1, 2, 3, 0, 2, 4, 6, ...),
  however each slice np.arange(4) * i will be *replicated* across corresponding
  host devices.

  On the other hand, if pspecs = PartitionSpec(), which means
  replication across all axes, then this snippet:

  >>> pspecs = jax.sharding.PartitionSpec()
  >>> arr = host_local_array_to_global_array(np.arange(4), mesh, pspecs)  # doctest: +SKIP

  will have the shape (4,) and the value (0, 1, 2, 3) will be replicated
  across all hosts and devices.

  It is an undefined behavior to have not identical local_inputs with pspec
  indicating data replication.

  You can use this function to transition to jax.Array. Using jax.Array with
  pjit has the same semantics of using GDA with pjit i.e. all jax.Array
  inputs to pjit should be globally shaped.

  If you are currently passing host local values to pjit, you can use this
  function to convert your host local values to global Arrays and then pass that
  to pjit.


  Example usage.

  >>> from jax.experimental import multihost_utils # doctest: +SKIP
  >>>
  >>> global_inputs = multihost_utils.host_local_array_to_global_array(host_local_inputs, global_mesh, in_pspecs) # doctest: +SKIP
  >>>
  >>> with mesh: # doctest: +SKIP
  >>>   global_out = pjitted_fun(global_inputs) # doctest: +SKIP
  >>>
  >>> host_local_output = multihost_utils.global_array_to_host_local_array(global_out, mesh, out_pspecs) # doctest: +SKIP

  Please note ths function requires global mesh to be a continuous mesh, meaning
  that  devices that belong to each host should form a subcube in this mesh.
  To move local data to global array with non-continuous mesh use
  jax.make_array_from_callback or jax.make_array_from_single_device_arrays
  instead.

  Args:
    local_inputs: A Pytree of host local values.
    global_mesh: A jax.sharding.Mesh object. The mesh must be a contiguous mesh,
    that is all hosts' devices must form a subcube in this mesh.
    pspecs: A Pytree of jax.sharding.PartitionSpec's.

  Returns:
    A pytree of global arrays.
  """
  flat_inps, in_tree = tree_flatten(local_inputs)
  in_pspecs = _flatten_pspecs('input pspecs', in_tree,
                              pjit_lib.hashable_pytree(pspecs))
  out_flat = [
      host_local_array_to_global_array_p.bind(inp, global_mesh=global_mesh,
                                              pspec=in_spec)
      for inp, in_spec in safe_zip(flat_inps, in_pspecs)
  ]
  return tree_unflatten(in_tree, out_flat)

host_local_array_to_global_array_p = core.Primitive('host_local_array_to_global_array')
host_local_array_to_global_array_p.def_impl(host_local_array_to_global_array_impl)

def ltg_abstract_eval(arr, *, global_mesh, pspec):
  return _local_to_global_aval(
      core.ShapedArray(arr.shape, arr.dtype), global_mesh, pspec)
host_local_array_to_global_array_p.def_abstract_eval(ltg_abstract_eval)

ad.deflinear2(host_local_array_to_global_array_p,
              lambda ct, _, **params: (
                  host_local_array_to_global_array_p.bind(ct, **params),))

def ltg_batcher(insert_axis, axis_data, vals_in, dims_in, global_mesh, pspec):
  x, = vals_in
  d, = dims_in
  new_parts = None if axis_data.spmd_name is None else axis_data.spmd_name
  new_pspec = list(pspec)
  new_pspec.insert(d, new_parts)
  new_pspec = P(*new_pspec)
  y = host_local_array_to_global_array_p.bind(
      x, global_mesh=global_mesh, pspec=new_pspec)
  return y, d
batching.fancy_primitive_batchers[host_local_array_to_global_array_p] = partial(
    ltg_batcher, False)

def _ltg_lowering(ctx, x, *, global_mesh, pspec):
  return [x]
mlir.register_lowering(host_local_array_to_global_array_p, _ltg_lowering)


def global_array_to_host_local_array_impl(
    arr: Any, *, global_mesh: jax.sharding.Mesh, pspec: Any):
  if pspec is None:
    raise ValueError(
        '`None` is not a valid input to the pspecs argument. Please use '
        'jax.sharding.PartitionSpec() if you wanted to replicate your input.')
  # If the Array is already fully addressable i.e. host local, return it.
  if isinstance(arr, array.ArrayImpl) and arr.is_fully_addressable:
    return arr

  global_sharding = jax.sharding.NamedSharding(global_mesh, pspec)
  local_sharding = jax.sharding.NamedSharding(global_mesh.local_mesh, pspec)
  local_aval = _global_to_local_aval(
      core.ShapedArray(arr.shape, arr.dtype), global_mesh, pspec)

  if isinstance(arr, array.ArrayImpl):
    if arr.sharding.is_equivalent_to(global_sharding, arr.ndim):
      arrays = arr._arrays
    else:
      resharded_array = jax.device_put(arr, global_sharding)
      arrays = resharded_array._arrays
    return array.ArrayImpl(local_aval, local_sharding, arrays, committed=True)
  else:
    # numpy array can show up here during AD.
    arr = xla.canonicalize_dtype(arr)
    arrays = [
        arr[index]
        for d, index in local_sharding.devices_indices_map(arr.shape).items()]
    return pxla.batched_device_put(
        local_aval, local_sharding, arrays,
        list(global_mesh.local_mesh.devices.flat))


def global_array_to_host_local_array(
    global_inputs: Any, global_mesh: jax.sharding.Mesh, pspecs: Any):
  r"""Converts a global `jax.Array` to a host local `jax.Array`.

  You can use this function to transition to `jax.Array`. Using `jax.Array` with
  pjit has the same semantics of using GDA with pjit i.e. all `jax.Array`
  inputs to pjit should be globally shaped and the output from pjit will also
  be globally shaped jax.Array's

  You can use this function to convert the globally shaped `jax.Array` output
  from pjit to host local values again so that the transition to jax.Array can
  be a mechanical change.

  Example usage:

  >>> from jax.experimental import multihost_utils # doctest: +SKIP
  >>>
  >>> global_inputs = multihost_utils.host_local_array_to_global_array(host_local_inputs, global_mesh, in_pspecs) # doctest: +SKIP
  >>>
  >>> with mesh: # doctest: +SKIP
  ...   global_out = pjitted_fun(global_inputs) # doctest: +SKIP
  >>>
  >>> host_local_output = multihost_utils.global_array_to_host_local_array(global_out, mesh, out_pspecs) # doctest: +SKIP

  Args:
    global_inputs: A Pytree of global jax.Array's.
    global_mesh: A :class:`jax.sharding.Mesh` object. The mesh must be contiguous
      meaning all local devices of the host must form a subcube.
    pspecs: A Pytree of :class:`jax.sharding.PartitionSpec` objects.

  Returns:
    A Pytree of host local arrays.
  """
  flat_inps, out_tree = tree_flatten(global_inputs)
  out_pspecs = _flatten_pspecs('output pspecs', out_tree,
                               pjit_lib.hashable_pytree(pspecs))
  out_flat = [
      global_array_to_host_local_array_p.bind(inp, global_mesh=global_mesh,
                                              pspec=o)
      for inp, o in safe_zip(flat_inps, out_pspecs)
  ]
  return tree_unflatten(out_tree, out_flat)

global_array_to_host_local_array_p = core.Primitive('global_array_to_host_local_array')
global_array_to_host_local_array_p.def_impl(global_array_to_host_local_array_impl)

def gtl_abstract_eval(arr, *, global_mesh, pspec):
  return _global_to_local_aval(
      core.ShapedArray(arr.shape, arr.dtype), global_mesh, pspec)
global_array_to_host_local_array_p.def_abstract_eval(gtl_abstract_eval)

ad.deflinear2(global_array_to_host_local_array_p,
              lambda ct, _, **params: (
                  global_array_to_host_local_array_p.bind(ct, **params),))
batching.defvectorized(global_array_to_host_local_array_p)

def _gtl_lowering(ctx, x, *, global_mesh, pspec):
  return [x]
mlir.register_lowering(global_array_to_host_local_array_p, _gtl_lowering)


def live_devices(devices: list[xla_client.Device]) -> list[xla_client.Device]:
  """Returns the subset of the provided devices that are live and healthy.

  This API is under active development and is not stable.

  `live_devices` is a low-level fault tolerance primitive that can be used to
  implement fault tolerant multi-process JAX programs.

  Barrier Semantics

  It's important that every process agrees on which devices are live to avoid
  the processes' behavior from diverging. For example, imagine a set of
  processes trying to run an AllGather, but they all disagree on which devices
  should be participating in the AllGather. This is buggy.

  To ensure that every process agrees on the set of live devices, the
  `live_devices` function has barrier-like semantics. Consider an invocation
  `live_devices(devices)` where `devices` includes devices across a set of
  processes P. The invocation acts as a barrier, waiting for every process in P
  to call `live_devices(devices)`. Afterwards, `live_devices` returns the same
  set of live devices `A` to all the processes in P. This ensures that every
  process agrees on the set of live devices.

  `live_devices` does not actually act as a barrier for *every* process in P
  because some processes in P might have failed. Instead, the `live_devices`
  function waits only for the processes with a device in the returned set of
  live devices A.

  An Example

  Imagine we have four processes, each with two devices:

    Process A: Devices 1 and 2
    Process B: Devices 3 and 4
    Process C: Devices 5 and 6
    Process D: Devices 7 and 8

  Further imagine that process D fails and that every process calls
  `live_devices(jax.devices())`. The invocation returns devices 1, 2, 3, 4, 5,
  and 6. Because these devices are hosted by processes A, B, and C, the call to
  `live_devices` acts as a barrier across processes A, B, and C. Process D,
  which failed, is ignored.

  Args:
    devices: A list of devices. The provided devices must include at least one
    local device.

  Returns:
    The subset of the provided devices that are live and healthy.

  Raises:
    RuntimeError: If the distributed runtime was not initialized.
    ValueError: If no local devices are provided.
  """
  client = distributed.global_state.client
  if client is None:
    raise RuntimeError('Distributed JAX not initialized.')

  if not devices:
    # TODO(mwhittaker): Make devices optional. If it's not provided, use
    # jax.devices() as a default.
    raise ValueError('No devices provided.')

  process_ids = {d.process_index for d in devices}
  if xla_bridge.process_index() not in process_ids:
    # A process can only participate in an live_devices call if it hosts some
    # of the provided devices.
    raise ValueError('Provided devices do not have any local devices.')

  if len(process_ids) == 1:
    # If the provided devices are hosted by a single process (this one), then we
    # don't have to perform any distributed computation. We know our local
    # devices are all live.
    return devices

  live_process_ids = client.get_live_nodes(list(process_ids))
  return [d for d in devices if d.process_index in live_process_ids]