def _sharded_jit_translation_rule(ctx, avals_in, avals_out, *in_nodes,
in_parts, out_parts_thunk, nparts,
name, call_jaxpr, local_in_parts,
local_out_parts_thunk, local_nparts):
subc = xc.XlaBuilder(f"sharded_jit_{name}")
# We assume any extra leading in_nodes are constants and replicate them.
num_extra_nodes = len(in_nodes) - len(in_parts)
assert num_extra_nodes >= 0
in_parts = (None,) * num_extra_nodes + in_parts
args = []
for i, (n, sharding) in enumerate(safe_zip(in_nodes, in_parts)):
# We use xla.set_sharding instead of xla.with_sharding because inlined calls
# shouldn't have shardings set directly on the inputs or outputs.
arg = xla.parameter(subc, i, ctx.builder.GetShape(n))
args.append(xla.set_sharding(subc, arg, sharding))
sub_ctx = ctx.replace(
builder=subc,
name_stack=new_name_stack(wrap_name(name, "sharded_jit")))
out_nodes = xla.jaxpr_subcomp(sub_ctx, call_jaxpr, (), *args)
out_parts = out_parts_thunk()
assert len(out_parts) == len(out_nodes)
out_nodes = [xla.set_sharding(subc, out, sharding)
for out, sharding in safe_zip(out_nodes, out_parts)]
subc = subc.build(xops.Tuple(subc, out_nodes))
return xla.xla_destructure(ctx.builder,
xops.Call(ctx.builder, subc, list(in_nodes)))
def primitive_subcomputation(platform: str, axis_env: 'AxisEnv',
prim: core.Primitive,
avals_in: Sequence[core.AbstractValue],
avals_out: Sequence[core.AbstractValue],
**params):
c = xc.XlaBuilder(f"primitive_computation_{prim.name}")
counts = it.count()
xla_args = [
parameter(c, next(counts), xla_shape) for a in avals_in
for xla_shape in aval_to_xla_shapes(a)
]
if (platform is not None
and prim in _backend_specific_translations[platform]):
rule = _backend_specific_translations[platform][prim]
elif prim in _translations:
rule = _translations[prim]
ctx = TranslationContext(builder=c,
platform=platform,
axis_env=axis_env,
name_stack=new_name_stack())
ans = rule(ctx, avals_in, avals_out, *xla_args, **params)
if prim.multiple_results:
return c.build(xops.Tuple(c, ans))
else:
x, = ans
return c.build(x)
def f_with_avals(c, avals, xla_args, params):
# parallelism is only supported via the new-style API.
axis_env = AxisEnv(1, (), ())
wrapped_fun = lu.wrap_init(fun, params)
if not multiple_results:
wrapped_fun = _tuple_output(wrapped_fun)
with core.extend_axis_env_nd(zip(axis_env.names, axis_env.sizes)):
jaxpr, _, consts = pe.trace_to_jaxpr_dynamic(wrapped_fun, avals)
ctx = TranslationContext(c, backend, axis_env, new_name_stack())
outs = jaxpr_subcomp(ctx, jaxpr, _xla_consts(c, consts), *xla_args)
if (multiple_results or any(
len(aval_to_xla_shapes(v.aval)) > 1 for v in jaxpr.outvars)):
return xops.Tuple(c, outs)
else:
assert len(outs) == 1, outs
return outs[0]
def primitive_subcomputation(platform: str, axis_env: 'AxisEnv',
prim: core.Primitive,
*avals: core.AbstractValue, **params):
c = xc.XlaBuilder(f"primitive_computation_{prim.name}")
f = lower_fun(prim.bind, multiple_results=prim.multiple_results,
new_style=True)
xla_args, _ = _xla_callable_args(c, avals, tuple_args=False,
filter_tokens=False)
ctx = TranslationContext(builder=c, platform=platform, axis_env=axis_env,
name_stack=new_name_stack())
ans = f(ctx.replace(builder=c), avals, None, *xla_args, **params)
if prim.multiple_results:
ans = xops.Tuple(c, ans)
else:
ans, = ans
return c.build(ans)
def _sharded_jit_lowering(ctx, *in_nodes, in_parts, out_parts_thunk, nparts,
name, call_jaxpr, local_in_parts,
local_out_parts_thunk, local_nparts):
# We assume any extra leading in_nodes are constants and replicate them.
num_extra_nodes = len(in_nodes) - len(in_parts)
assert num_extra_nodes >= 0
in_parts = (None, ) * num_extra_nodes + in_parts
args = []
for ns, sharding in safe_zip(
safe_map(mlir.wrap_singleton_ir_values, in_nodes), in_parts):
if sharding is not None:
args.append([
mlir.wrap_with_sharding_op(n, xla.sharding_to_proto(sharding))
for n in ns
])
else:
args.append(ns)
sub_ctx = ctx.module_context.replace(
name_stack=new_name_stack(wrap_name(name, "sharded_jit")))
fn = mlir.lower_jaxpr_to_fun(sub_ctx, f"sharded_jit_{name}",
core.ClosedJaxpr(call_jaxpr, ()))
output_types = safe_map(mlir.aval_to_ir_types, ctx.avals_out)
flat_output_types = util.flatten(output_types)
call = func_dialect.CallOp(flat_output_types,
ir.FlatSymbolRefAttr.get(fn.name.value),
mlir.flatten_lowering_ir_args(args))
out_nodes = util.unflatten(call.results, safe_map(len, output_types))
out_parts = out_parts_thunk()
outputs = []
for ns, sharding in safe_zip(out_nodes, out_parts):
if sharding is not None:
outputs.append([
mlir.wrap_with_sharding_op(n, xla.sharding_to_proto(sharding))
for n in ns
])
else:
outputs.append(ns)
return outputs
def _sharded_callable(
fun: lu.WrappedFun, nparts: Optional[int],
in_parts: Tuple[pxla.PartitionsOrReplicated, ...],
out_parts_thunk: Callable[[], Tuple[pxla.PartitionsOrReplicated, ...]],
local_in_parts: Optional[Tuple[pxla.PartitionsOrReplicated, ...]],
local_out_parts_thunk: Callable[[], Optional[Tuple[
pxla.PartitionsOrReplicated,
...]]], local_nparts: Optional[int], name: str, *abstract_args):
nrep = 1
if local_in_parts is None:
local_in_parts = in_parts
global_abstract_args = [
pxla.get_global_aval(arg, parts,
lparts) for arg, parts, lparts in safe_zip(
abstract_args, in_parts, local_in_parts)
]
if logging.vlog_is_on(2):
logging.vlog(2, "abstract_args: %s", abstract_args)
logging.vlog(2, "global_abstract_args: %s", global_abstract_args)
logging.vlog(2, "in_parts: %s", in_parts)
logging.vlog(2, "local_in_parts: %s", local_in_parts)
jaxpr, global_out_avals, consts = pe.trace_to_jaxpr_final(
fun, global_abstract_args)
platform = xb.get_backend().platform
nparts = pxla.reconcile_num_partitions(jaxpr, nparts)
assert nparts is not None
if nparts > xb.device_count():
raise ValueError(
f"sharded_jit computation requires {nparts} devices, "
f"but only {xb.device_count()} devices are available.")
if xb.local_device_count() < nparts < xb.device_count():
raise NotImplementedError(
f"sharded_jit across multiple hosts must use all available devices. "
f"Got {nparts} out of {xb.device_count()} requested devices "
f"(local device count: {xb.local_device_count()})")
if local_nparts is None:
if nparts > xb.local_device_count():
raise ValueError(
"Specify 'local_nparts' when using cross-process sharded_jit "
"and all inputs and outputs are replicated.")
else:
local_nparts = nparts
if local_nparts > xb.local_device_count():
raise ValueError(
f"sharded_jit computation requires {local_nparts} local devices, "
f"but only {xb.local_device_count()} local devices are available.")
if logging.vlog_is_on(2):
logging.vlog(2, "nparts: %d local_nparts: %d", nparts, local_nparts)
out_parts = out_parts_thunk()
local_out_parts = local_out_parts_thunk()
if local_out_parts is None:
local_out_parts = out_parts
if logging.vlog_is_on(2):
logging.vlog(2, "out_parts: %s", out_parts)
logging.vlog(2, "local_out_parts: %s", local_out_parts)
local_out_avals = [
pxla.get_local_aval(out, parts,
lparts) for out, parts, lparts in safe_zip(
global_out_avals, out_parts, local_out_parts)
]
log_priority = logging.WARNING if config.jax_log_compiles else logging.DEBUG
logging.log(log_priority, "Compiling %s for %d devices with args %s.",
fun.__name__, nparts, global_abstract_args)
axis_env = xla.AxisEnv(nrep, (), ())
unordered_effects = [
eff for eff in jaxpr.effects if eff not in core.ordered_effects
]
ordered_effects = [
eff for eff in jaxpr.effects if eff in core.ordered_effects
]
module, _ = mlir.lower_jaxpr_to_module(
f"spjit_{fun.__name__}",
core.ClosedJaxpr(jaxpr, consts),
unordered_effects,
ordered_effects,
platform=platform,
axis_context=mlir.ReplicaAxisContext(axis_env),
name_stack=new_name_stack(wrap_name(name, "sharded_jit")),
donated_args=[False] * len(in_parts),
arg_shardings=safe_map(xla.sharding_to_proto, in_parts),
result_shardings=safe_map(xla.sharding_to_proto, out_parts))
built = xc._xla.mlir.mlir_module_to_xla_computation(
mlir.module_to_string(module), use_tuple_args=False, return_tuple=True)
if nparts <= xb.local_device_count():
devices = xb.local_devices()[:nparts]
else:
assert nparts == xb.device_count()
devices = xb.devices()
device_assignment = np.array([[d for d in devices]])
device_assignment = np.reshape(device_assignment, (-1, nparts))
# device_assignment = None # TODO(skye): replace with default device assignment?
compiled = dispatch.backend_compile(
xb.get_backend(), built,
xb.get_compile_options(nrep, nparts, device_assignment))
input_specs = [
pxla.partitioned_sharding_spec(local_nparts, parts, aval)
for parts, aval in zip(local_in_parts, abstract_args)
]
input_indices = [
pxla.spec_to_indices(aval.shape, spec) if spec is not None else None
for aval, spec in zip(abstract_args, input_specs)
]
handle_args = partial(pxla.shard_args, compiled.local_devices(),
input_indices)
handle_outs = _avals_to_results_handler(
nrep,
local_nparts, # type: ignore
local_out_parts,
local_out_avals)
return partial(_execute_spatially_partitioned, compiled, handle_args,
handle_outs)
def lower_xla_callable(fun: lu.WrappedFun, device, backend, name,
donated_invars, always_lower: bool, keep_unused: bool,
*arg_specs):
"""Lower into XLA.
Args:
always_lower: If `True`, even trivial programs (not doing any computation
such as lambda x: x) will be lowered into an XLA program.
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.
"""
if device is not None and backend is not None:
raise ValueError("can't specify both a device and a backend for jit, "
"got device={} and backend={}".format(
device, backend))
abstract_args, arg_devices = util.unzip2(arg_specs)
if fun.in_type is not None:
abstract_args, which_explicit = util.unzip2(fun.in_type)
else:
which_explicit = None
with log_elapsed_time(f"Finished tracing + transforming {fun.__name__} "
"for jit in {elapsed_time} sec"):
jaxpr, out_avals, consts = pe.trace_to_jaxpr_final(
fun, abstract_args, pe.debug_info_final(fun, "jit"),
which_explicit)
if any(isinstance(c, core.Tracer) for c in consts):
raise UnexpectedTracerError("Encountered an unexpected tracer.")
# TODO(mattjj): handle argument pruning w/ dynamic shapes
if fun.in_type is None and not keep_unused:
jaxpr, kept_const_idx, kept_var_idx = _prune_unused_inputs(jaxpr)
consts = [c for i, c in enumerate(consts) if i in kept_const_idx]
abstract_args, arg_devices = util.unzip2(
[a for i, a in enumerate(arg_specs) if i in kept_var_idx])
donated_invars = [
x for i, x in enumerate(donated_invars) if i in kept_var_idx
]
del kept_const_idx
else:
kept_var_idx = set(range(len(abstract_args)))
map(prefetch, itertools.chain(consts, jaxpr_literals(jaxpr)))
jaxpr = apply_outfeed_rewriter(jaxpr)
nreps = jaxpr_replicas(jaxpr)
device = _xla_callable_device(nreps, backend, device, arg_devices)
backend = xb.get_device_backend(device) if device else xb.get_backend(
backend)
if (config.jax_dynamic_shapes and jaxpr_has_bints(jaxpr)
and not _backend_supports_unbounded_dynamic_shapes(backend)):
jaxpr, consts = pe.pad_jaxpr(jaxpr, consts)
# Computations that only produce constants and/or only rearrange their inputs,
# which are often produced from partial evaluation, don't need compilation,
# and don't need to evaluate their arguments.
if not jaxpr.eqns and not always_lower:
return XlaComputation(name,
None,
True,
None,
None,
jaxpr=jaxpr,
consts=consts,
device=device,
in_avals=abstract_args,
out_avals=out_avals,
has_unordered_effects=False,
ordered_effects=[],
kept_var_idx=kept_var_idx,
keepalive=None)
if not _on_exit:
log_priority = logging.WARNING if config.jax_log_compiles else logging.DEBUG
if len(abstract_args) > 10:
msg = f"Compiling {fun.__name__} ({id(fun)}) for {len(abstract_args)} args."
else:
msg = f"Compiling {fun.__name__} ({id(fun)} for args {abstract_args}."
logging.log(log_priority, msg)
if nreps > 1:
warnings.warn(
f"The jitted function {name} includes a pmap. Using "
"jit-of-pmap can lead to inefficient data movement, as the outer jit "
"does not preserve sharded data representations and instead collects "
"input and output arrays onto a single device. "
"Consider removing the outer jit unless you know what you're doing. "
"See https://github.com/google/jax/issues/2926.")
if nreps > xb.device_count(backend):
raise ValueError(
f"compiling computation `{name}` that requires {nreps} replicas, but "
f"only {xb.device_count(backend)} XLA devices are available.")
if xb.process_count() > 1 and (nreps > 1 or jaxpr_has_pmap(jaxpr)):
raise NotImplementedError(
"jit of multi-host pmap not implemented (and jit-of-pmap can cause "
"extra data movement anyway, so maybe you don't want it after all)."
)
# pass long arg lists as tuple for TPU
tuple_args = len(abstract_args) > 100
axis_env = xla.AxisEnv(nreps, (), ())
name_stack = util.new_name_stack(util.wrap_name(name, 'jit'))
closed_jaxpr = core.ClosedJaxpr(jaxpr, consts)
module_name = f"jit_{fun.__name__}"
unordered_effects = [
eff for eff in closed_jaxpr.effects if eff not in core.ordered_effects
]
ordered_effects = [
eff for eff in closed_jaxpr.effects if eff in core.ordered_effects
]
module, keepalive = mlir.lower_jaxpr_to_module(
module_name, closed_jaxpr,
unordered_effects, ordered_effects, backend.platform,
mlir.ReplicaAxisContext(axis_env), name_stack, donated_invars)
return XlaComputation(name,
module,
False,
donated_invars,
which_explicit,
nreps=nreps,
device=device,
backend=backend,
tuple_args=tuple_args,
in_avals=abstract_args,
out_avals=out_avals,
has_unordered_effects=bool(unordered_effects),
ordered_effects=ordered_effects,
kept_var_idx=kept_var_idx,
keepalive=keepalive)
def _sharded_callable(
fun: lu.WrappedFun, nparts: Optional[int],
in_parts: Tuple[pxla.PartitionsOrReplicated, ...],
out_parts_thunk: Callable[[], Tuple[pxla.PartitionsOrReplicated, ...]],
local_in_parts: Optional[Tuple[pxla.PartitionsOrReplicated, ...]],
local_out_parts_thunk: Callable[[], Optional[Tuple[
pxla.PartitionsOrReplicated,
...]]], local_nparts: Optional[int], name: str, *abstract_args):
nrep = 1
if local_in_parts is None:
local_in_parts = in_parts
global_abstract_args = [
pxla.get_global_aval(arg, parts,
lparts) for arg, parts, lparts in safe_zip(
abstract_args, in_parts, local_in_parts)
]
if logging.vlog_is_on(2):
logging.vlog(2, "abstract_args: %s", abstract_args)
logging.vlog(2, "global_abstract_args: %s", global_abstract_args)
logging.vlog(2, "in_parts: %s", in_parts)
logging.vlog(2, "local_in_parts: %s", local_in_parts)
jaxpr, global_out_avals, consts = pe.trace_to_jaxpr_final(
fun, global_abstract_args)
platform = xb.get_backend().platform
if platform not in ["tpu", "gpu"]:
# TODO(skye): fall back to regular jit?
raise ValueError(f"sharded_jit not supported for {platform}")
nparts = pxla.reconcile_num_partitions(jaxpr, nparts)
assert nparts is not None
if nparts > xb.device_count():
raise ValueError(
f"sharded_jit computation requires {nparts} devices, "
f"but only {xb.device_count()} devices are available.")
if xb.local_device_count() < nparts < xb.device_count():
raise NotImplementedError(
f"sharded_jit across multiple hosts must use all available devices. "
f"Got {nparts} out of {xb.device_count()} requested devices "
f"(local device count: {xb.local_device_count()})")
if local_nparts is None:
if nparts > xb.local_device_count():
raise ValueError(
"Specify 'local_nparts' when using cross-process sharded_jit "
"and all inputs and outputs are replicated.")
else:
local_nparts = nparts
if local_nparts > xb.local_device_count():
raise ValueError(
f"sharded_jit computation requires {local_nparts} local devices, "
f"but only {xb.local_device_count()} local devices are available.")
if logging.vlog_is_on(2):
logging.vlog(2, "nparts: %d local_nparts: %d", nparts, local_nparts)
out_parts = out_parts_thunk()
local_out_parts = local_out_parts_thunk()
if local_out_parts is None:
local_out_parts = out_parts
if logging.vlog_is_on(2):
logging.vlog(2, "out_parts: %s", out_parts)
logging.vlog(2, "local_out_parts: %s", local_out_parts)
local_out_avals = [
pxla.get_local_aval(out, parts,
lparts) for out, parts, lparts in safe_zip(
global_out_avals, out_parts, local_out_parts)
]
log_priority = logging.WARNING if config.jax_log_compiles else logging.DEBUG
logging.log(log_priority, "Compiling %s for %d devices with args %s.",
fun.__name__, nparts, global_abstract_args)
c = xc.XlaBuilder("spjit_{}".format(fun.__name__))
xla_consts = _map(partial(xla.pyval_to_ir_constant, c), consts)
xla_args = _xla_sharded_args(c, global_abstract_args, in_parts)
axis_env = xla.AxisEnv(nrep, (), ())
ctx = xla.TranslationContext(
c, platform, axis_env, new_name_stack(wrap_name(name, "sharded_jit")))
out_nodes = xla.jaxpr_subcomp(ctx, jaxpr, xla_consts, *xla_args)
out_tuple = xla.with_sharding(c, out_parts, xops.Tuple, c, out_nodes)
built = c.Build(out_tuple)
if nparts <= xb.local_device_count():
devices = xb.local_devices()[:nparts]
else:
assert nparts == xb.device_count()
devices = xb.devices()
device_assignment = np.array([[d for d in devices]])
device_assignment = np.reshape(device_assignment, (-1, nparts))
# device_assignment = None # TODO(skye): replace with default device assignment?
compiled = dispatch.backend_compile(
xb.get_backend(), built,
xb.get_compile_options(nrep, nparts, device_assignment))
input_specs = [
pxla.partitioned_sharding_spec(local_nparts, parts, aval)
for parts, aval in zip(local_in_parts, abstract_args)
]
input_indices = [
pxla.spec_to_indices(aval.shape, spec) if spec is not None else None
for aval, spec in zip(abstract_args, input_specs)
]
handle_args = partial(pxla.shard_args, compiled.local_devices(),
input_indices)
handle_outs = _avals_to_results_handler(
nrep,
local_nparts, # type: ignore
local_out_parts,
local_out_avals)
return partial(_execute_spatially_partitioned, compiled, handle_args,
handle_outs)
