Source code for jax.core

# Copyright 2018 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.


import collections
from collections import namedtuple
from contextlib import contextmanager
from functools import partial, partialmethod, total_ordering
import gc
import itertools as it
import operator
from operator import attrgetter
import threading
import types
from typing import (Any, Callable, ClassVar, DefaultDict, Dict, Generator,
                    Iterator, List, NamedTuple, Optional, Sequence, Set, Tuple,
                    Type, Union, cast, Iterable, Hashable)
import warnings
from weakref import ref

import numpy as np

from jax._src import dtypes
from jax._src import config as jax_config
from jax._src.config import FLAGS, config
from jax.errors import (ConcretizationTypeError, TracerArrayConversionError,
                        TracerIntegerConversionError, UnexpectedTracerError)
from jax import linear_util as lu

from jax._src import source_info_util
from jax._src.util import (safe_zip, safe_map, curry, prod, tuple_insert,
                        tuple_delete, cache, as_hashable_function,
                        HashableFunction)
import jax._src.pretty_printer as pp

from jax._src import traceback_util
traceback_util.register_exclusion(__file__)

zip, unsafe_zip = safe_zip, zip
map, unsafe_map = safe_map, map


# -------------------- jaxprs --------------------

[docs]class Jaxpr: constvars: List['Var'] invars: List['Var'] outvars: List['Atom'] eqns: List['JaxprEqn']
[docs] def __init__(self, constvars: Sequence['Var'], invars: Sequence['Var'], outvars: Sequence['Atom'], eqns: Sequence['JaxprEqn']): """ Args: constvars: list of variables introduced for constants. Array constants are replaced with such variables while scalar constants are kept inline. invars: list of input variables. Together, `constvars` and `invars` are the inputs to the Jaxpr. outvars: list of output variables. eqns: list of equations. """ self.constvars = list(constvars) self.invars = list(invars) self.outvars = list(outvars) self.eqns = list(eqns)
def __str__(self): return str(pp_jaxpr(self, JaxprPpContext(), custom_pp_eqn_rules=True)) __repr__ = __str__ def pretty_print(self, *, source_info=False, print_shapes=True, custom_pp_eqn_rules=True, **kw): doc = pp_jaxpr(self, JaxprPpContext(), source_info=source_info, print_shapes=print_shapes, custom_pp_eqn_rules=custom_pp_eqn_rules) return doc.format(**kw)
def jaxprs_in_params(params) -> Iterator[Jaxpr]: for val in params.values(): vals = val if isinstance(val, tuple) else (val,) for v in vals: if isinstance(v, Jaxpr): yield v elif isinstance(v, ClosedJaxpr): yield v.jaxpr def subjaxprs(jaxpr: Jaxpr) -> Iterator[Jaxpr]: """Generator for all subjaxprs found in the params of jaxpr.eqns. Does not descend recursively into the found subjaxprs. """ for eqn in jaxpr.eqns: yield from jaxprs_in_params(eqn.params)
[docs]class ClosedJaxpr: jaxpr: Jaxpr consts: List['Any']
[docs] def __init__(self, jaxpr: Jaxpr, consts: Sequence): assert len(consts) == len(jaxpr.constvars) self.jaxpr = jaxpr self.consts = list(consts)
@property def in_avals(self): return [v.aval for v in self.jaxpr.invars] @property def out_avals(self): return [v.aval for v in self.jaxpr.outvars] @property def literals(self): return self.consts # backwards compatible alias @property def eqns(self): return self.jaxpr.eqns def map_jaxpr(self, f): return ClosedJaxpr(f(self.jaxpr), self.consts) def __str__(self): return str(self.jaxpr) def __repr__(self): return repr(self.jaxpr) def pretty_print(self, *, source_info=False, print_shapes=True, **kw): return pp_jaxpr(self.jaxpr, JaxprPpContext(), source_info=source_info, print_shapes=print_shapes).format(**kw)
@curry def jaxpr_as_fun(closed_jaxpr: ClosedJaxpr, *args): return eval_jaxpr(closed_jaxpr.jaxpr, closed_jaxpr.consts, *args) class JaxprEqn(NamedTuple): invars: List['Atom'] outvars: List['Var'] primitive: 'Primitive' params: Dict[str, Any] source_info: source_info_util.SourceInfo def __repr__(self): return str(pp_eqn(self, JaxprPpContext(), custom_pp_eqn_rules=False) ).rstrip() def new_jaxpr_eqn(invars, outvars, primitive, params, source_info=None): if primitive.call_primitive: assert len(outvars) == len(params["call_jaxpr"].outvars) source_info = source_info or source_info_util.new_source_info() return JaxprEqn(invars, outvars, primitive, params, source_info) @total_ordering class Var: # TODO(frostig,mattjj): We don't override __eq__ or __hash__, so comparison is # by object id, but pretty printing might collide. count: int suffix: str aval: 'AbstractValue' def __init__(self, count: int, suffix: str, aval: 'AbstractValue'): self.count = count self.suffix = suffix self.aval = raise_to_shaped(aval) def __lt__(self, other): if not isinstance(other, Var): return NotImplemented else: return (self.count, self.suffix) < (other.count, other.suffix) def __repr__(self): return _encode_digits_alphabetic(self.count) + self.suffix def _encode_digits_alphabetic(n): s = '' while len(s) == 0 or n: n, i = n // 26, n % 26 s = chr(97 + i % 26) + s return s def _jaxpr_vars(jaxpr): return it.chain( jaxpr.invars, jaxpr.constvars, (v for eqn in jaxpr.eqns for v in eqn.outvars)) def gensym(jaxprs: Optional[Sequence[Jaxpr]] = None, suffix: str = '') -> Callable[['AbstractValue'], Var]: """Produce distinct variables, printed with the optional suffix. If `jaxprs` is provided, the variables produced will be distinct from those in any of the given jaxprs. """ if jaxprs is None: start = 0 else: all_vars = it.chain.from_iterable(_jaxpr_vars(j) for j in jaxprs) start = 1 + max((v.count for v in all_vars), default=-1) counter = it.count(start=start) return lambda aval: Var(next(counter), suffix, aval) # In a jaxpr, `dropvar` can appear in place of a bound variable to indicate that # the assignment is dropped, i.e. that an expression's output value will never # be read. In that sense, `dropvar` is not a variable, but it is convenient to # treat it as a special case of one. Its `aval` is similarly inexact. class DropVar(Var): def __init__(self, aval: 'AbstractValue'): super().__init__(-1, '', aval) def __repr__(self): return '_' class Literal: __slots__ = ["val", "hash"] val: Any hash: Optional[int] def __init__(self, val): self.val = val try: self.hash = hash(val) except TypeError: if type(val) in literalable_types: try: self.hash = hash((val.item(), val.dtype)) except (TypeError, AttributeError, ValueError): self.hash = None @property def aval(self): return raise_to_shaped(get_aval(self.val)) __hash__ = None # type: ignore def __repr__(self): if hasattr(self, 'hash'): return '{}'.format(self.val) else: return 'Literal(val={})'.format(self.val) literalable_types: Set[type] = set() Atom = Union[Var, Literal] class Primitive: name: str multiple_results = False # set for multi-output primitives call_primitive = False # set for call primitives processed in final style map_primitive = False # set for map primitives processed in final style _dispatch_on_params = False # whether to include axis names from params in dispatch def __init__(self, name: str): self.name = name def __repr__(self): return '{}'.format(self.name) def bind(self, *args, **params): assert (not config.jax_enable_checks or all(isinstance(arg, Tracer) or valid_jaxtype(arg) for arg in args)), args top_trace = find_top_trace( args, used_axis_names(self, params) if self._dispatch_on_params else None) tracers = map(top_trace.full_raise, args) out = top_trace.process_primitive(self, tracers, params) return map(full_lower, out) if self.multiple_results else full_lower(out) def def_impl(self, impl): self.impl = impl return impl def def_abstract_eval(self, abstract_eval): self.abstract_eval = abstract_eval return abstract_eval def def_custom_bind(self, bind): self.bind = bind return bind def impl(self, *args, **params): raise NotImplementedError("Evaluation rule for '{}' not implemented" .format(self.name)) def abstract_eval(self, *args, **params): raise NotImplementedError("Abstract evaluation for '{}' not implemented" .format(self.name)) # -------------------- lifting -------------------- # TODO(necula): this belongs next to pe.new_eqn_recipe, but is needed in # core.py. Plan to move all these utilities to jaxpr.py. def extract_call_jaxpr( primitive: Primitive, params: Dict[str, Any]) -> Tuple[Optional[Jaxpr], Dict[str, Any]]: """Extract the call primitive subjaxpr from the params. Returns the subjaxpr and the params without the "call_jaxpr" value. If this is not a call primitive then returns (None, params). """ if not (primitive.call_primitive or primitive.map_primitive): return (None, params) else: assert "call_jaxpr" in params new_params = dict(params) del new_params["call_jaxpr"] return (params["call_jaxpr"], new_params) # TODO(mattjj): replace this approach with a primitive-keyed table of rules def traverse_jaxpr_params(f, params): """Applies f to each jaxpr parameter and returns a tuple of returned values.""" return {name: f(p) for name, param in params.items() for p in (param if isinstance(param, (tuple, list)) else [param]) if type(p) in (Jaxpr, ClosedJaxpr)} def eval_jaxpr_eqn(eqn, in_vals): """Evaluates the jaxpr equation with the provided input values.""" call_jaxpr, params = extract_call_jaxpr(eqn.primitive, eqn.params) if call_jaxpr: subfuns = [lu.wrap_init(partial(eval_jaxpr, call_jaxpr, ()))] else: subfuns = [] if eqn.primitive in initial_to_final_param_rules: bind_params = initial_to_final_param_rules[eqn.primitive](params) elif eqn.primitive.map_primitive: out_axes_thunk = HashableFunction(lambda: params['out_axes'], closure=params['out_axes']) bind_params = dict(params, out_axes_thunk=out_axes_thunk) del bind_params['out_axes'] else: bind_params = params with source_info_util.user_context(eqn.source_info.traceback): return eqn.primitive.bind(*(subfuns + in_vals), **bind_params) def eval_jaxpr(jaxpr: Jaxpr, consts, *args): def read(v): if type(v) is Literal: return v.val else: return env[v] def write(v, val): env[v] = val env: Dict[Var, Any] = {} write(unitvar, unit) map(write, jaxpr.constvars, consts) map(write, jaxpr.invars, args) for eqn in jaxpr.eqns: ans = eval_jaxpr_eqn(eqn, map(read, eqn.invars)) if eqn.primitive.multiple_results: map(write, eqn.outvars, ans) else: write(eqn.outvars[0], ans) return map(read, jaxpr.outvars) initial_to_final_param_rules: Dict[Primitive, Callable] = {} # -------------------- tracing -------------------- class Trace: __slots__ = ['main', 'level', 'sublevel'] main: 'MainTrace' level: int sublevel: 'Sublevel' def __init__(self, main: 'MainTrace', sublevel: 'Sublevel') -> None: self.main = main self.level = main.level self.sublevel = sublevel def full_raise(self, val) -> 'Tracer': if not isinstance(val, Tracer): return self.pure(val) val._assert_live() level = self.level sublevel = self.sublevel if val._trace.main is self.main: if val._trace.sublevel == sublevel: return val elif val._trace.sublevel < sublevel: return self.sublift(val) else: raise escaped_tracer_error( val, f"Can't lift sublevels {val._trace.sublevel} to {sublevel}") elif val._trace.level < level: if val._trace.sublevel > sublevel: raise escaped_tracer_error( val, f"Incompatible sublevel: {val._trace}, {(level, sublevel)}") return self.lift(val) elif val._trace.level > level: raise escaped_tracer_error( val, f"Can't lift level {val} to {self}") else: # val._trace.level == self.level: raise escaped_tracer_error( val, f"Different traces at same level: {val}, {self}") def pure(self, val): raise NotImplementedError("must override") def lift(self, tracer): raise NotImplementedError("must override") def sublift(self, tracer): raise NotImplementedError("must override") def process_primitive(self, primitive, tracers, params): raise NotImplementedError("must override") def __repr__(self): return '{}(level={}/{})'.format( self.__class__.__name__, self.level, self.sublevel) def process_call(self, call_primitive, f, tracers, params): msg = (f"{type(self)} must override process_call to handle call-like " "primitives") raise NotImplementedError(msg) def process_map(self, map_primitive, f, tracers, params): msg = (f"{type(self)} must override process_map to handle map-like " "primitives") raise NotImplementedError(msg) def process_custom_jvp_call(self, primitive, fun, jvp, tracers): msg = (f"{type(self)} must override process_custom_jvp_call " "to handle custom_jvp primitives") raise NotImplementedError(msg) def process_custom_vjp_call(self, primitive, fun, fwd, bwd, tracers, out_trees): msg = (f"{type(self)} must override process_custom_vjp_call " "to handle custom_vjp primitives") raise NotImplementedError(msg) def escaped_tracer_error(tracer, detail=None): num_frames = FLAGS.jax_tracer_error_num_traceback_frames msg = ('Encountered an unexpected tracer. A function transformed by JAX ' 'had a side effect, allowing for a reference to an intermediate value ' f'with shape {tracer.shape} and dtype {tracer.dtype} to escape.\n' 'JAX transformations require that functions explicitly return their ' 'outputs, and disallow saving intermediate values to global state.') dbg = getattr(tracer._trace.main, 'debug_info', None) if dbg is not None: msg += ('\nThe function being traced when the value leaked was ' f'{dbg.func_src_info} traced for {dbg.traced_for}.') line_info = getattr(tracer, '_line_info', None) if line_info is not None: divider = '\n' + '-'*30 + '\n' msg += divider msg += ('The leaked intermediate value was created on line ' f'{source_info_util.summarize(line_info)}. ') msg += divider if num_frames > 0: msg += (f'When the value was created, the final {num_frames} stack ' 'frames (most recent last) excluding JAX-internal frames were:') msg += divider + source_info_util.summarize( line_info, num_frames=num_frames) + divider msg += ('\nTo catch the leak earlier, try setting the environment variable ' 'JAX_CHECK_TRACER_LEAKS or using the `jax.checking_leaks` context ' 'manager.') if detail: msg += f'Detail: {detail}' return UnexpectedTracerError(msg) class Tracer: __array_priority__ = 1000 __slots__ = ['_trace', '__weakref__', '_line_info'] def __array__(self, *args, **kw): raise TracerArrayConversionError(self) def __index__(self): raise TracerIntegerConversionError(self) def __init__(self, trace: Trace): self._trace = trace def __iter__(self): return iter(self.aval._iter(self)) def __len__(self): return self.aval._len(self) @property def aval(self): raise NotImplementedError("must override") def _assert_live(self) -> None: pass # Override for liveness checking # Python looks up special methods only on classes, not instances. This means # these methods needs to be defined explicitly rather than relying on # __getattr__. def __neg__(self): return self.aval._neg(self) def __pos__(self): return self.aval._pos(self) def __eq__(self, other): return self.aval._eq(self, other) def __ne__(self, other): return self.aval._ne(self, other) def __lt__(self, other): return self.aval._lt(self, other) def __le__(self, other): return self.aval._le(self, other) def __gt__(self, other): return self.aval._gt(self, other) def __ge__(self, other): return self.aval._ge(self, other) def __abs__(self): return self.aval._abs(self) def __add__(self, other): return self.aval._add(self, other) def __radd__(self, other): return self.aval._radd(self, other) def __sub__(self, other): return self.aval._sub(self, other) def __rsub__(self, other): return self.aval._rsub(self, other) def __mul__(self, other): return self.aval._mul(self, other) def __rmul__(self, other): return self.aval._rmul(self, other) def __div__(self, other): return self.aval._div(self, other) def __rdiv__(self, other): return self.aval._rdiv(self, other) def __truediv__(self, other): return self.aval._truediv(self, other) def __rtruediv__(self, other): return self.aval._rtruediv(self, other) def __floordiv__(self, other): return self.aval._floordiv(self, other) def __rfloordiv__(self, other): return self.aval._rfloordiv(self, other) def __divmod__(self, other): return self.aval._divmod(self, other) def __rdivmod__(self, other): return self.aval._rdivmod(self, other) def __mod__(self, other): return self.aval._mod(self, other) def __rmod__(self, other): return self.aval._rmod(self, other) def __pow__(self, other): return self.aval._pow(self, other) def __rpow__(self, other): return self.aval._rpow(self, other) def __matmul__(self, other): return self.aval._matmul(self, other) def __rmatmul__(self, other): return self.aval._rmatmul(self, other) def __and__(self, other): return self.aval._and(self, other) def __rand__(self, other): return self.aval._rand(self, other) def __or__(self, other): return self.aval._or(self, other) def __ror__(self, other): return self.aval._ror(self, other) def __xor__(self, other): return self.aval._xor(self, other) def __rxor__(self, other): return self.aval._rxor(self, other) def __invert__(self): return self.aval._invert(self) def __lshift__(self, other): return self.aval._lshift(self, other) def __rlshift__(self, other): return self.aval._rlshift(self, other) def __rshift__(self, other): return self.aval._rshift(self, other) def __rrshift__(self, other): return self.aval._rrshift(self, other) def __getitem__(self, idx): return self.aval._getitem(self, idx) def __nonzero__(self): return self.aval._nonzero(self) def __bool__(self): return self.aval._bool(self) def __int__(self): return self.aval._int(self) def __long__(self): return self.aval._long(self) def __hex__(self): return self.aval._hex(self) def __oct__(self): return self.aval._oct(self) def __float__(self): return self.aval._float(self) def __complex__(self): return self.aval._complex(self) # raises the better error message from ShapedArray def __setitem__(self, idx, val): return self.aval._setitem(self, idx, val) # NumPy also only looks up special methods on classes. def __array_module__(self, types): return self.aval._array_module(self, types) def __getattr__(self, name): # if the aval property raises an AttributeError, gets caught here assert not config.jax_enable_checks or name != "aval" try: attr = getattr(self.aval, name) except KeyError as err: raise AttributeError( "{} has no attribute {}".format(self.__class__.__name__, name) ) from err else: t = type(attr) if t is aval_property: return attr.fget(self) elif t is aval_method: return types.MethodType(attr.fun, self) else: return attr def _pretty_print(self): base = pp.text(f'Traced<{self.aval}>with<{self._trace}>') contents = [(name, attr._pretty_print() if isinstance(attr, Tracer) else pp.text(repr(attr))) for name, attr in self._contents()] if contents: base = pp.group(pp.nest(2, pp.concat([ base, pp.text(' with'), pp.brk(), pp.join(pp.brk(), [ pp.text('{} = '.format(name)) + pp_payload for name, pp_payload in contents]) ]))) return base def __repr__(self): return self._pretty_print().format() def _contents(self): try: return [(name, getattr(self, name)) for name in self.__slots__] except AttributeError: return () def __copy__(self): return self def __deepcopy__(self, unused_memo): return self def _origin_msg(self) -> str: return "" # these can be used to set up forwarding of properties and instance methods from # Tracer instances to the underlying avals aval_property = namedtuple("aval_property", ["fget"]) aval_method = namedtuple("aval_method", ["fun"]) class EvalTrace(Trace): # See comments in https://github.com/google/jax/pull/3370 def pure(self, x): return x lift = sublift = pure def process_primitive(self, primitive, tracers, params): return primitive.impl(*tracers, **params) def process_call(self, primitive, f, tracers, params): return primitive.impl(f, *tracers, **params) process_map = process_call def process_custom_jvp_call(self, primitive, fun, jvp, tracers): del primitive, jvp # Unused. with new_sublevel(): return fun.call_wrapped(*tracers) def process_custom_vjp_call(self, primitive, fun, fwd, bwd, tracers, out_trees): del primitive, fwd, bwd, out_trees # Unused. with new_sublevel(): return fun.call_wrapped(*tracers) class MainTrace: level: int trace_type: Type[Trace] payload: Dict[str, Any] def __init__(self, level, trace_type, **payload) -> None: self.level = level self.trace_type = trace_type self.payload = payload def __repr__(self) -> str: return "MainTrace({},{})".format(self.level, self.trace_type.__name__) def __hash__(self) -> int: return hash((self.level, self.trace_type)) def __eq__(self, other: object) -> bool: return (isinstance(other, MainTrace) and self.level == other.level and self.trace_type == other.trace_type and self.payload == other.payload) def with_cur_sublevel(self): return self.trace_type(self, cur_sublevel(), **self.payload) class TraceStack: # See comments in https://github.com/google/jax/pull/3370 stack: List[MainTrace] dynamic: MainTrace def __init__(self): eval_trace = MainTrace(0, EvalTrace) self.stack = [eval_trace] self.dynamic = eval_trace def next_level(self) -> int: return len(self.stack) def push(self, main_trace: MainTrace) -> None: self.stack.append(main_trace) def pop(self) -> None: self.stack.pop() def __repr__(self) -> str: stack_str = map(' {}\n'.format, self.stack[::-1]) return f'Trace stack\n{stack_str}\n{self.dynamic}' def copy(self): new = self.__new__(TraceStack) new.stack = self.stack[:] new.dynamic = self.dynamic return new @total_ordering class Sublevel: def __init__(self, level: int): self.level = level def __repr__(self): return str(self.level) def __eq__(self, other): return type(other) is Sublevel and self.level == other.level def __lt__(self, other): return type(other) is Sublevel and self.level < other.level AxisEnvFrame = namedtuple('AxisEnvFrame', ['name', 'size', 'main_trace']) AxisName = Hashable no_axis_name = object() class TraceState: trace_stack: TraceStack substack: List[Sublevel] axis_env: List[AxisEnvFrame] def __init__(self) -> None: self.trace_stack = TraceStack() self.substack = [Sublevel(0)] self.axis_env = [] def copy(self): new = self.__new__(TraceState) new.trace_stack = self.trace_stack.copy() new.substack = self.substack[:] new.axis_env = self.axis_env[:] return new def _update_thread_local_jit_state(dynamic): # Copies the MainTrace instance, removing any .debug_info or .jaxpr_stack # fields that should not be kept alive as part of a cache key. # TODO(mattjj): split debug_info and jaxpr_stack out of MainTrace. # TODO(mattjj): add a test that verifies that JIT-ted functions are not kept # alive by the JIT cache, particularly for nested JIT-ted functions. copy = MainTrace(dynamic.level, dynamic.trace_type, **dynamic.payload) jax_config.update_thread_local_jit_state(dynamic_trace_state=copy) # The global state of the tracer is accessed by a thread-local object. # This allows concurrent tracing in separate threads; passing traced objects # between threads is forbidden. class ThreadLocalState(threading.local): def __init__(self): self.trace_state = TraceState() _update_thread_local_jit_state(self.trace_state.trace_stack.dynamic) thread_local_state = ThreadLocalState() def trace_state_clean() -> bool: trace_state = thread_local_state.trace_state return (trace_state.substack == [Sublevel(0)] and trace_state.axis_env == [] and trace_state.trace_stack.stack == [MainTrace(0, EvalTrace)] and trace_state.trace_stack.dynamic == MainTrace(0, EvalTrace)) def reset_trace_state() -> bool: "Reset the global trace state and return True if it was already clean." if not trace_state_clean(): thread_local_state.trace_state.__init__() # type: ignore return False else: return True def cur_sublevel() -> Sublevel: return thread_local_state.trace_state.substack[-1] TRACER_LEAK_DEBUGGER_WARNING = """\ JAX check_tracer_leaks behavior can trigger false positives when used with a debugger. To avoid false positives and silence this warning, you can disable thread tracing using the following: import threading threading.current_thread().pydev_do_not_trace = True """ def maybe_find_leaked_tracers(x: Optional[Union[MainTrace, Sublevel]]): """Find the leaked tracers holding a reference to the MainTrace or SubLevel. It's possible there's none! eg. there's some cases where JAX itself holds a reference to `x` inside of a lambda closure, and no tracers were leaked by the user. In this case an empty list is returned. """ if not getattr(threading.current_thread(), 'pydev_do_not_trace', True): warnings.warn(TRACER_LEAK_DEBUGGER_WARNING) # Trigger garbage collection to filter out cyclical dependency false positives gc.collect() traces = list(filter(lambda x: isinstance(x, Trace), gc.get_referrers(x))) tracers = list(filter(lambda x: isinstance(x, Tracer), gc.get_referrers(*traces))) return tracers @contextmanager def new_main(trace_type: Type[Trace], dynamic: bool = False, **payload) -> Generator[MainTrace, None, None]: # See comments in https://github.com/google/jax/pull/3370 stack = thread_local_state.trace_state.trace_stack level = stack.next_level() main = MainTrace(level, trace_type, **payload) stack.push(main) if dynamic: prev_dynamic, stack.dynamic = stack.dynamic, main _update_thread_local_jit_state(stack.dynamic) try: yield main finally: stack.pop() if dynamic: stack.dynamic = prev_dynamic _update_thread_local_jit_state(stack.dynamic) if config.jax_check_tracer_leaks: t = ref(main) del main if t() is not None: leaked_tracers = maybe_find_leaked_tracers(t()) if leaked_tracers: raise Exception(f'Leaked level {t()}. Leaked tracer(s): {leaked_tracers}.') @contextmanager def new_base_main(trace_type: Type[Trace]) -> Generator[MainTrace, None, None]: # See comments in https://github.com/google/jax/pull/3370 stack = thread_local_state.trace_state.trace_stack main = MainTrace(0, trace_type) prev_dynamic, stack.dynamic = stack.dynamic, main prev_base, stack.stack[0] = stack.stack[0], main _update_thread_local_jit_state(stack.dynamic) try: yield main finally: stack.dynamic = prev_dynamic stack.stack[0] = prev_base _update_thread_local_jit_state(stack.dynamic) if config.jax_check_tracer_leaks: t = ref(main) del main if t() is not None: leaked_tracers = maybe_find_leaked_tracers(t()) if leaked_tracers: raise Exception(f'Leaked level {t()}. Leaked tracer(s): {leaked_tracers}.') @contextmanager def eval_context(): with new_base_main(EvalTrace): yield @contextmanager def new_sublevel() -> Generator[None, None, None]: sublevel = Sublevel(len(thread_local_state.trace_state.substack)) thread_local_state.trace_state.substack.append(sublevel) try: yield finally: thread_local_state.trace_state.substack.pop() if config.jax_check_tracer_leaks: t = ref(sublevel) del sublevel if t() is not None: leaked_tracers = maybe_find_leaked_tracers(t()) if leaked_tracers: raise Exception(f'Leaked sublevel {t()}. Leaked tracer(s): {leaked_tracers}.') def full_lower(val): if isinstance(val, Tracer): return val.full_lower() else: return val def find_top_trace(xs, axis_names=None) -> Trace: top_main: Optional[MainTrace] = None if axis_names: top_main = max((axis_frame(a).main_trace for a in axis_names), default=None, key=lambda t: getattr(t, 'level', -1)) top_tracer = max((x for x in xs if isinstance(x, Tracer)), default=None, key=attrgetter('_trace.level')) if top_tracer is not None: top_tracer._assert_live() if top_tracer._trace.main.level > getattr(top_main, 'level', -1): top_main = top_tracer._trace.main dynamic = thread_local_state.trace_state.trace_stack.dynamic top_main = (dynamic if top_main is None or dynamic.level > top_main.level else top_main) return top_main and top_main.with_cur_sublevel() # type: ignore # -------------------- abstract values -------------------- class AbstractValue: __slots__: List[str] = [] def at_least_vspace(self): raise NotImplementedError("must override") def __repr__(self): try: kv_pairs = ('{}={}'.format(k, v) for k, v in self.__dict__.items()) return '{}({})'.format(self.__class__.__name__, ','.join(kv_pairs)) except AttributeError: return self.__class__.__name__ def strip_weak_type(self) -> 'AbstractValue': return self def strip_named_shape(self) -> 'AbstractValue': return self def join(self, other): raise NotImplementedError("must override") def update(self, **kwargs): raise NotImplementedError("must override") def str_short(self, short_dtypes=False): raise NotImplementedError("must override") class Bot(AbstractValue): pass bot = Bot() class AbstractUnit(AbstractValue): def at_least_vspace(self): return self def join(self, other): if config.jax_enable_checks: assert other is abstract_unit, other return self def _eq(self, self_traced, other): return get_aval(other) is self def str_short(self, short_dtypes=False): return '*' abstract_unit = AbstractUnit() def lattice_join(x: Optional[AbstractValue], y: Optional[AbstractValue]) -> AbstractValue: if x is None: return cast(AbstractValue, y) elif y is None: return cast(AbstractValue, x) elif isinstance(x, type(y)): return y.join(x) elif isinstance(y, type(x)): return x.join(y) else: raise TypeError(x, y) # For use in typing annotations to denote either a Tracer or a `valid_jaxtype`. Value = Any def valid_jaxtype(x): try: concrete_aval(x) except TypeError: return False else: return True def check_valid_jaxtype(x): if not valid_jaxtype(x): raise TypeError( f"Value {repr(x)} of type {type(x)} is not a valid JAX type") def concrete_aval(x): for typ in type(x).mro(): handler = pytype_aval_mappings.get(typ) if handler: return handler(x) if hasattr(x, '__jax_array__'): return concrete_aval(x.__jax_array__()) raise TypeError(f"Value {repr(x)} with type {type(x)} is not a valid JAX " "type") def get_aval(x): if isinstance(x, Tracer): return x.aval else: return concrete_aval(x) pytype_aval_mappings: Dict[type, Callable[[Any], AbstractValue]] = {} class Unit: def __repr__(self): return '*' unit: Unit = Unit() literalable_types.add(Unit) class UnitVar(Var): count = -1 suffix = '' def __init__(self): pass @property def aval(self): return abstract_unit def __repr__(self): return '*' unitvar = UnitVar() pytype_aval_mappings[Unit] = lambda _: abstract_unit def concretization_function_error(fun, suggest_astype=False): fname = getattr(fun, "__name__", fun) fname_context = f"The problem arose with the `{fname}` function. " if suggest_astype: fname_context += ("If trying to convert the data type of a value, " f"try using `x.astype({fun.__name__})` " f"or `jnp.array(x, {fun.__name__})` instead.") def error(self, arg): raise ConcretizationTypeError(arg, fname_context) return error def concrete_or_error(force: Any, val: Any, context=""): """Like force(val), but gives the context in the error message.""" if force is None: force = lambda x: x if isinstance(val, Tracer): if isinstance(val.aval, ConcreteArray): return force(val.aval.val) else: raise ConcretizationTypeError(val, context) else: return force(val) def _short_dtype_name(dtype): return (dtype.name.replace('float', 'f').replace('uint', 'u') .replace('int', 'i').replace('complex', 'c')) class UnshapedArray(AbstractValue): __slots__ = ['dtype', 'weak_type'] array_abstraction_level = 2 def __init__(self, dtype, weak_type=False): self.dtype = np.dtype(dtypes.canonicalize_dtype(dtype)) self.weak_type = weak_type def update(self, dtype=None, weak_type=None): if dtype is None: dtype = self.dtype if weak_type is None: weak_type = self.weak_type return UnshapedArray(dtype, weak_type) def __eq__(self, other): return (type(self) is type(other) and self.dtype == other.dtype and self.weak_type == other.weak_type) def __ne__(self, other): return not self == other def __hash__(self): # can use hash(self.dtype) and rely on the fact that numpy reuses base dtype # objects, e.g. `np.zeros(3).dtype is np.zeros(4).dtype`, or we can use # the unique character code via hash(self.dtype.char) return hash((self.dtype, self.weak_type)) def __repr__(self): return '{}({}{})'.format(self.__class__.__name__, self.str_short(), ", weak_type=True" if self.weak_type else "") _bool = _nonzero = concretization_function_error(bool) _float = concretization_function_error(float, True) _int = concretization_function_error(int, True) _complex = concretization_function_error(complex, True) _hex = concretization_function_error(hex) _oct = concretization_function_error(oct) def at_least_vspace(self) -> AbstractValue: return UnshapedArray(primal_dtype_to_tangent_dtype(self.dtype), self.weak_type) def join(self, other): if self.dtype == other.dtype: if self.weak_type == other.weak_type: return self else: return UnshapedArray(self.dtype, weak_type=False) else: raise TypeError(self, other) def str_short(self, short_dtypes=False) -> str: return _short_dtype_name(self.dtype) if short_dtypes else self.dtype.name def strip_weak_type(self): """Returns a copy of the aval with weak_type=False.""" return self.update(weak_type=False) @property def shape(self): msg = ("UnshapedArray has no shape. Please open an issue at " "https://github.com/google/jax/issues because it's unexpected for " "UnshapedArray instances to ever be produced.") raise TypeError(msg) class ShapedArray(UnshapedArray): __slots__ = ['shape', 'named_shape'] array_abstraction_level = 1 def __init__(self, shape, dtype, weak_type=False, named_shape=None): super().__init__(dtype, weak_type=weak_type) self.shape = canonicalize_shape(shape) self.named_shape = {} if named_shape is None else dict(named_shape) def update(self, shape=None, dtype=None, weak_type=None, named_shape=None): if shape is None: shape = self.shape if dtype is None: dtype = self.dtype if weak_type is None: weak_type = self.weak_type if named_shape is None: named_shape = self.named_shape return ShapedArray(shape, dtype, weak_type, named_shape) ndim = property(lambda self: len(self.shape)) size = property(lambda self: prod(self.shape)) broadcast: ClassVar[Optional[aval_method]] = None transpose: ClassVar[Optional[aval_method]] = None reshape: ClassVar[Optional[aval_method]] = None _iter: ClassVar[Optional[staticmethod]] = None def __eq__(self, other): return (type(self) is type(other) and self.dtype == other.dtype and self.shape == other.shape and self.weak_type == other.weak_type and self.named_shape == other.named_shape) def __hash__(self): # can use hash(self.dtype) and rely on the fact that numpy reuses base dtype # objects, e.g. `np.zeros(3).dtype is np.zeros(4).dtype`, or we can use # the unique character code via hash(self.dtype.char) return hash((self.shape, self.dtype, self.weak_type, tuple(self.named_shape.items()))) def at_least_vspace(self): return ShapedArray(self.shape, primal_dtype_to_tangent_dtype(self.dtype), self.weak_type, self.named_shape) def join(self, other): if symbolic_equal_shape(self.shape, other.shape) and self.dtype == other.dtype: weak_type = self.weak_type and other.weak_type named_shape = join_named_shapes(self.named_shape, other.named_shape) return self.update(weak_type=weak_type, named_shape=named_shape) elif self.dtype == other.dtype: return UnshapedArray(self.dtype) else: raise TypeError(self, other) def str_short(self, short_dtypes=False): dt_str = _short_dtype_name(self.dtype) if short_dtypes else self.dtype.name shapestr = ','.join(map(str, self.shape)) if self.named_shape: named_shapestr = ','.join(f'{k}:{v}' for k, v in self.named_shape.items()) return f'{dt_str}[{shapestr};{named_shapestr}]' else: return f'{dt_str}[{shapestr}]' def strip_named_shape(self): return self.update(named_shape={}) def _len(self, ignored_tracer): try: return self.shape[0] except IndexError as err: raise TypeError("len() of unsized object") from err # same as numpy error def _forward_to_value(self, fun, ignored_tracer, *args): return fun(self.val, *args) class ConcreteArray(ShapedArray): __slots__ = ['val'] array_abstraction_level = 0 def __init__(self, val, weak_type=None): super().__init__(np.shape(val), np.result_type(val), weak_type=dtypes.is_weakly_typed(val) if weak_type is None else weak_type) # Note: canonicalized self.dtype doesn't necessarily match self.val self.val = val assert self.dtype != np.dtype('O'), val def update(self, val=None, weak_type=None): if val is None: val = self.val if weak_type is None: weak_type = self.weak_type return ConcreteArray(val, weak_type) def __eq__(self, other): if (type(self) is type(other) and self.dtype == other.dtype and self.shape == other.shape and self.weak_type == other.weak_type): with eval_context(): # in case self.val is a DeviceArray return (self.val == other.val).all() else: return False def __hash__(self): return id(self.val) def join(self, other) -> AbstractValue: if self == other: return self elif self.shape == other.shape and self.dtype == other.dtype: weak_type = self.weak_type and other.weak_type named_shape = join_named_shapes(self.named_shape, other.named_shape) return ShapedArray( self.shape, self.dtype, weak_type=weak_type, named_shape=named_shape) elif self.dtype == other.dtype: return UnshapedArray(self.dtype, weak_type=self.weak_type and other.weak_type) else: raise TypeError(self, other) def str_short(self, short_dtypes=False) -> str: dt_str = _short_dtype_name(self.dtype) if short_dtypes else self.dtype.name return f'{self.val}, dtype={dt_str}' _bool = _nonzero = partialmethod(_forward_to_value, bool) _int = partialmethod(_forward_to_value, int) _hex = partialmethod(_forward_to_value, hex) _oct = partialmethod(_forward_to_value, oct) _float = concretization_function_error(float, True) _complex = concretization_function_error(complex, True) def primal_dtype_to_tangent_dtype(primal_dtype): if not dtypes.issubdtype(primal_dtype, np.inexact): return dtypes.float0 else: return primal_dtype class AbstractToken(AbstractValue): def join(self, other): if isinstance(other, AbstractToken): return self else: assert False, f"Cannot join {self} with {other}" def str_short(self, short_dtypes=False): return 'Tok' def at_least_vspace(self): return self abstract_token: AbstractToken = AbstractToken() # Concrete token object class Token(object): pass token: Token = Token() pytype_aval_mappings[Token] = lambda _: abstract_token def raise_to_shaped(aval: AbstractValue, weak_type=None): if weak_type is None: weak_type = getattr(aval, 'weak_type', False) for typ in type(aval).mro(): handler = raise_to_shaped_mappings.get(typ) if handler: return handler(aval, weak_type) raise TypeError(type(aval)) raise_to_shaped_mappings : Dict[type, Callable] = { AbstractUnit: lambda aval, _: aval, AbstractToken: lambda aval, _: aval, Bot: lambda aval, _: aval, UnshapedArray: lambda aval, _: aval, ShapedArray: lambda aval, weak_type: ShapedArray( aval.shape, aval.dtype, weak_type, aval.named_shape) } ### Operations on shapes and dimension sizes. # Shapes are tuples of dimension sizes, which are normally integers. We allow # modules to extend the set of dimension sizes to contain other types, e.g., # symbolic dimensions in jax2tf.shape_poly.DimVar and masking.Poly. DimSize = Union[int, Any] # extensible Shape = Sequence[DimSize] class InconclusiveDimensionOperation(Exception): """Raised when we cannot conclusively compute with symbolic dimensions.""" pass class DimensionHandler: """Operations on dimension sizes. Dimension sizes are normally integer constants, but can also be symbolic, e.g., masking.Poly or jax2tf.shape_poly.DimVar. The base class works for integers only. Subclasses are invoked when at least one of the operands has a type registered in _SPECIAL_DIMENSION_HANDLERS. In that case, all operands are guaranteed to be either the special dimension type, or Python integer scalars. Subclasses should raise InconclusiveDimensionOperation if the result cannot be computed in some contexts. """ def is_constant(self, d: DimSize) -> bool: """The dimension is a constant.""" return True def symbolic_equal(self, d1: DimSize, d2: DimSize) -> bool: """True iff the dimension sizes are equal in all contexts; False otherwise. Unlike `d1 == d2` this never raises InconclusiveDimensionOperation. """ return d1 == d2 def greater_equal(self, d1: DimSize, d2: DimSize) -> bool: """Computes `d1 >= d2`. Raise InconclusiveDimensionOperation if the result is different in different contexts. """ return d1 >= d2 def sum(self, *ds: DimSize) -> DimSize: """Sum of dimensions. Raises InconclusiveDimensionOperation if the result cannot be represented by the same DimSize in all contexts. """ return sum(ds) def diff(self, d1: DimSize, d2: DimSize) -> DimSize: """Difference of dimensions. Raises InconclusiveDimensionOperation if the result cannot be represented by the same DimSize in all contexts. """ return d1 - d2 def divide_shape_sizes(self, s1: Shape, s2: Shape) -> DimSize: """Computes integer "i" such that i * size(s2) == size(s1). Raise InconclusiveDimensionOperation if there is no such integer for all contexts, """ sz1 = int(np.prod(s1)) sz2 = int(np.prod(s2)) if sz1 == 0 and sz2 == 0: return 1 if sz1 % sz2: raise InconclusiveDimensionOperation(f"Cannot divide evenly the sizes of shapes {tuple(s1)} and {tuple(s2)}") return sz1 // sz2 def stride(self, d: DimSize, window_size: DimSize, window_stride: DimSize) -> DimSize: """(d - window_size) // window_stride + 1""" return (d - window_size) // window_stride + 1 def dilate(self, d: DimSize, dilation: int) -> DimSize: """Implements `0 if d == 0 else 1 + dilation * (d - 1))`""" return 0 if d == 0 else 1 + dilation * (d - 1) def as_value(self, d: DimSize): """Turns a dimension size into a JAX value that we can compute with.""" return d _dimension_handler_int = DimensionHandler() _SPECIAL_DIMENSION_HANDLERS: Dict[type, DimensionHandler] = {} def _dim_handler_and_canonical(*dlist: DimSize) -> Tuple[DimensionHandler, Tuple[DimSize, ...]]: """Finds the handler for the given dimensions; also returns the canonical dimensions. A dimension is canonical if it is a Python integer scalar, or has a type registered in _SPECIAL_DIMENSION_HANDLERS. """ special_handlers = set() canonical = [] for d in dlist: handler = _SPECIAL_DIMENSION_HANDLERS.get(type(d)) if handler: special_handlers.add(handler) canonical.append(d) else: try: canonical.append(operator.index(d)) except TypeError: raise _invalid_shape_error(dlist) if len(special_handlers) > 1: msg = (f"Dimension size operation involves multiple special dimension types {dlist}") raise ValueError(msg) return next(iter(special_handlers), _dimension_handler_int), tuple(canonical) def is_special_dim_size(v: Any) -> bool: """Checks if a value is a special DimSize.""" handler = _SPECIAL_DIMENSION_HANDLERS.get(type(v)) return (handler is not None) def is_constant_dim(d: DimSize) -> bool: handler, ds = _dim_handler_and_canonical(d) return handler.is_constant(*ds) def symbolic_equal_dim(d1: DimSize, d2: DimSize) -> bool: handler, ds = _dim_handler_and_canonical(d1, d2) return handler.symbolic_equal(*ds) def symbolic_equal_one_of_dim(d1: DimSize, dlist: Sequence[DimSize]) -> bool: handler, ds = _dim_handler_and_canonical(d1, *dlist) return any([handler.symbolic_equal(ds[0], d) for d in ds[1:]]) def symbolic_equal_shape(s1: Shape, s2: Shape) -> bool: return (len(s1) == len(s2) and all(unsafe_map(symbolic_equal_dim, s1, s2))) def greater_equal_dim(d1: DimSize, d2: DimSize) -> bool: handler, ds = _dim_handler_and_canonical(d1, d2) return handler.greater_equal(*ds) def greater_equal_shape(s1: Shape, s2: Shape) -> bool: return all(map(greater_equal_dim, s1, s2)) def sum_dim(*ds: DimSize) -> DimSize: handler, ds = _dim_handler_and_canonical(*ds) return handler.sum(*ds) def sum_shapes(*ss: Shape) -> Shape: return tuple(map(sum_dim, *ss)) def diff_dim(d1: DimSize, d2: DimSize) -> DimSize: handler, ds = _dim_handler_and_canonical(d1, d2) return handler.diff(*ds) def diff_shape(s1: Shape, s2: Shape) -> Shape: return tuple(map(diff_dim, s1, s2)) def divide_shape_sizes(s1: Shape, s2: Shape) -> DimSize: """Returns an integer "i" s.t., i * size(s2) == size(s1). Raises if there is no such integer.""" s1 = s1 or (1,) s2 = s2 or (1,) handler, ds = _dim_handler_and_canonical(*s1, *s2) return handler.divide_shape_sizes(ds[:len(s1)], ds[len(s1):]) def same_shape_sizes(s1: Shape, s2: Shape) -> bool: return 1 == divide_shape_sizes(s1, s2) def is_empty_shape(s: Shape) -> bool: return any(symbolic_equal_dim(d, 0) for d in s) def dilate_dim(d: DimSize, dilation: DimSize) -> DimSize: """Implements `0 if d == 0 else 1 + dilation * (d - 1))`""" handler, ds = _dim_handler_and_canonical(d, dilation) return handler.dilate(*ds) def dilate_shape(s: Shape, dilations: Sequence[int]) -> Shape: return tuple(map(dilate_dim, s, dilations)) def stride_dim(d: DimSize, window_size: DimSize, window_stride: DimSize) -> DimSize: handler, ds = _dim_handler_and_canonical(d, window_size, window_stride) return handler.stride(*ds) def stride_shape(s: Shape, window_size: Shape, window_stride: Shape) -> Shape: """(s - window_size) // window_stride + 1""" return tuple(map(stride_dim, s, window_size, window_stride)) def dimension_as_value(d: DimSize): """Turns a dimension size into a JAX value that we can compute with. This is the identity function for constant dimensions.""" handler, ds = _dim_handler_and_canonical(d) return handler.as_value(*ds) def _canonicalize_dimension(dim: DimSize) -> DimSize: if type(dim) in _SPECIAL_DIMENSION_HANDLERS: return dim else: return operator.index(dim) def canonicalize_shape(shape: Shape, context: str="") -> Shape: """Canonicalizes and checks for errors in a user-provided shape value. Args: shape: a Python value that represents a shape. Returns: A tuple of canonical dimension values. """ try: return tuple(map(_canonicalize_dimension, shape)) except TypeError: pass raise _invalid_shape_error(shape, context) def canonicalize_dim(d: DimSize, context: str="") -> DimSize: """Canonicalizes and checks for errors in a user-provided shape dimension value. Args: f: a Python value that represents a dimension. Returns: A canonical dimension value. """ return canonicalize_shape((d,), context)[0] def _invalid_shape_error(shape: Shape, context: str=""): msg = ("Shapes must be 1D sequences of concrete values of integer type, " f"got {shape}.") if context: msg += f" {context}." if any(isinstance(x, Tracer) and isinstance(get_aval(x), ShapedArray) and not isinstance(get_aval(x), ConcreteArray) for x in shape): msg += ("\nIf using `jit`, try using `static_argnums` or applying `jit` to " "smaller subfunctions.") return TypeError(msg) # ------------------- Named shapes ------------------- class NamedShape: def __init__(self, *args, **kwargs): self.__positional = canonicalize_shape(args) # TODO: Assert that kwargs match axis env? self.__named = dict(kwargs) @property def rank(self): return len(self.__positional) + len(self.__named) @property def positional_rank(self): return len(self.__positional) @property def named_rank(self): return len(self.__named) @property def positional(self): return self.__positional @property def names(self): return self.__named.keys() @property def named_sizes(self): return self.__named.values() @property def named_items(self): return self.__named.items() def __getitem__(self, idx): try: idx = operator.index(idx) return self.__positional[idx] except TypeError: pass return self.__named[idx] @property def total(self): total = 1 for s in self.__positional: total *= s for s in self.__named.values(): total *= s return total def __str__(self): return (f"({', '.join(map(str, self.__positional))}{', ' if self.__named else ''}" f"{', '.join(f'{k}={v}' for k, v in self.__named.items())})") def __eq__(self, other): if isinstance(other, NamedShape): return (self.__positional, self.__named) == (other.__positional, other.__named) if isinstance(other, tuple): return not self.__named and self.__positional == other raise TypeError(f"NamedShape doesn't support comparisons with {type(other)}") def __hash__(self): named = frozenset(self.__named.items()) return hash((self.__positional, named)) def join_named_shapes(*named_shapes): result = {} for named_shape in named_shapes: for name, size in named_shape.items(): if result.setdefault(name, size) != size: raise TypeError( f"Axis name {name} used with inconsistent sizes: {result[name]} != {size}") return result # TODO: Make canonicalize_shape return named shapes? def as_named_shape(shape) -> NamedShape: if isinstance(shape, NamedShape): return shape return NamedShape(*shape) # ------------------- Call ------------------- def apply_todos(todos, outs): todos_list = list(todos) while todos_list: outs = map(full_lower, todos_list.pop()(outs)) return outs class _IgnoreElemList(list): """Compares equal to all other _ignore_elem_lists.""" def __hash__(self): return 0 def __eq__(self, other): return type(other) is _IgnoreElemList @lu.transformation_with_aux def process_env_traces(primitive: Union['CallPrimitive', 'MapPrimitive'], level: int, params_tuple: tuple, out_axes_transforms, *args): outs = yield args, {} params = dict(params_tuple) todo = [] assert not out_axes_transforms while True: tracers = [x for x in outs if isinstance(x, Tracer) and (level is None or x._trace.level > level)] if tracers: ans = max(tracers, key=lambda x: x._trace.level) else: break trace = ans._trace.main.with_cur_sublevel() outs = map(trace.full_raise, outs) outs, cur_todo = primitive.post_process(trace, outs, params) if isinstance(primitive, MapPrimitive): cur_todo, out_axes_transform = cur_todo out_axes_transforms.append(out_axes_transform) todo.append(cur_todo) yield outs, tuple(todo) # Ensure the aux output is immutable def call_bind(primitive: Union['CallPrimitive', 'MapPrimitive'], fun, *args, **params): out_axes_transforms = _IgnoreElemList() if primitive.map_primitive: out_axes_thunk = params['out_axes_thunk'] # The new thunk depends deterministically on the old thunk and the wrapped # function. Any caching already has to include the wrapped function as part # of the key, so we only use the previous thunk for equality checks. @as_hashable_function(closure=out_axes_thunk) def new_out_axes_thunk(): out_axes = out_axes_thunk() for t in out_axes_transforms: out_axes = t(out_axes) return out_axes params = dict(params, out_axes_thunk=new_out_axes_thunk) params_tuple = tuple(params.items()) top_trace = find_top_trace(args) fun, env_trace_todo = process_env_traces( fun, primitive, top_trace and top_trace.level, params_tuple, out_axes_transforms) tracers = map(top_trace.full_raise, args) outs = primitive.process(top_trace, fun, tracers, params) return map(full_lower, apply_todos(env_trace_todo(), outs)) class CallPrimitive(Primitive): multiple_results = True call_primitive = True def bind(self, fun, *args, **params): return call_bind(self, fun, *args, **params) def process(self, trace, fun, tracers, params): return trace.process_call(self, fun, tracers, params) def post_process(self, trace, out_tracers, params): return trace.post_process_call(self, out_tracers, params) def call_impl(f: lu.WrappedFun, *args, **params): del params # params parameterize the call primitive, not the function with new_sublevel(): return f.call_wrapped(*args) call_p: CallPrimitive = CallPrimitive('call') call = call_p.bind call_p.def_impl(call_impl) named_call_p: CallPrimitive = CallPrimitive('named_call') named_call_p.def_impl(call_impl) # ------------------- Map ------------------- def mapped_aval(size: int, axis: int, aval: AbstractValue) -> AbstractValue: handler, _ = aval_mapping_handlers.get(type(aval), (None, None)) if handler is not None: return handler(size, axis, aval) else: raise TypeError(f"no mapping handler for {aval} of type {type(aval)}") def unmapped_aval(size: int, axis_name, axis: int, aval: AbstractValue) -> AbstractValue: _, handler = aval_mapping_handlers.get(type(aval), (None, None)) if handler is not None: return handler(size, axis_name, axis, aval) else: raise TypeError(f"no unmapping handler for {aval} of type {type(aval)}") def _map_unit(*_) -> AbstractUnit: return abstract_unit def _map_shaped_array(size: int, axis: int, aval: ShapedArray) -> ShapedArray: assert aval.shape[axis] == size # TODO: Extend the named shape return ShapedArray(tuple_delete(aval.shape, axis), aval.dtype, named_shape=aval.named_shape) def _unmap_shaped_array(size: int, axis_name, axis: int, aval: ShapedArray) -> ShapedArray: named_shape = dict(aval.named_shape) # TODO: Make this mandatory named_shape.pop(axis_name, None) return ShapedArray(tuple_insert(aval.shape, axis, size), aval.dtype, named_shape=named_shape) AvalMapHandlerPair = Tuple[Callable, Callable] aval_mapping_handlers: Dict[Type, AvalMapHandlerPair] = { AbstractUnit: (_map_unit, _map_unit), ShapedArray: (_map_shaped_array, _unmap_shaped_array), ConcreteArray: (_map_shaped_array, _unmap_shaped_array), } class MapPrimitive(Primitive): multiple_results = True map_primitive = True def bind(self, fun, *args, **params): assert len(params['in_axes']) == len(args) return call_bind(self, fun, *args, **params) def process(self, trace, fun, tracers, params): return trace.process_map(self, fun, tracers, params) def post_process(self, trace, out_tracers, params): return trace.post_process_map(self, out_tracers, params) @contextmanager def extend_axis_env(axis_name: AxisName, size: int, tag: Any): frame = AxisEnvFrame(axis_name, size, tag) thread_local_state.trace_state.axis_env.append(frame) try: yield finally: thread_local_state.trace_state.axis_env.pop() @contextmanager def extend_axis_env_nd(axes: Iterable[Tuple[AxisName, int]]): frames = [AxisEnvFrame(axis_name, size, None) for axis_name, size in axes] thread_local_state.trace_state.axis_env.extend(frames) try: yield finally: for _ in frames: thread_local_state.trace_state.axis_env.pop() # When a mapped function is given no axis name, we generate a name object based # on the id of the function object. Collisions aren't important because this # name can't be used in collectives, as user code never gets a ref to this # object. We don't want to use the function object itself because that might # persist references to the function object. # TODO(mattjj): revisit this unique axis name strategy @total_ordering class _TempAxisName: def __init__(self, obj): self.id = id(obj) def __repr__(self): return f'<axis {hex(self.id)}>' def __hash__(self): return hash(self.id) def __eq__(self, other): return type(other) is _TempAxisName and self.id == other.id def __lt__(self, other): return type(other) is _TempAxisName and self.id < other.id def axis_frame(axis_name): frames = thread_local_state.trace_state.axis_env for frame in reversed(frames): if frame.name == axis_name: return frame named_axes = [frame.name for frame in reversed(frames) if not isinstance(frame.name, _TempAxisName)] raise NameError( f'unbound axis name: {axis_name}. The following axis names (e.g. defined ' f'by pmap) are available to collective operations: {named_axes}') ParamDict = Dict[str, Any] AxisSubst = Callable[[AxisName], Tuple[AxisName, ...]] class NameGatheringSubst: def __init__(self): self.axis_names = set() def __call__(self, axis_name): self.axis_names.add(axis_name) return (axis_name,) def used_axis_names(primitive: Primitive, params: ParamDict) -> Set[AxisName]: subst = NameGatheringSubst() subst_axis_names(primitive, params, subst) return subst.axis_names def subst_axis_names(primitive: Primitive, params: ParamDict, subst: AxisSubst, traverse: bool = True) -> ParamDict: if primitive in axis_substitution_rules: return axis_substitution_rules[primitive](params, subst, traverse) if not traverse: return params # Default implementation: substitute names in all jaxpr parameters if isinstance(primitive, MapPrimitive): def shadowed_subst(name): return (name,) if name == params['axis_name'] else subst(name) else: shadowed_subst = subst jaxpr_params = [(n, v) for n, v in params.items() if isinstance(v, (Jaxpr, ClosedJaxpr))] if not jaxpr_params: return params new_params = dict(params) for name, jaxpr in jaxpr_params: new_params[name] = subst_axis_names_jaxpr(jaxpr, shadowed_subst) return new_params class DuplicateAxisNameError(Exception): def __init__(self, var): self.var = var self.eqn = None def subst_axis_names_var(v: Var, subst: AxisSubst, var_map: Dict[Var, Var]) -> Var: # Var identity is load-bearing, so we can't have duplicates! if v is unitvar: return v if isinstance(v, DropVar): return v assert v not in var_map if not hasattr(v.aval, 'named_shape'): var_map[v] = v return v names = tuple(it.chain.from_iterable(subst(name) for name in v.aval.named_shape)) named_shape = {name: axis_frame(name).size for name in names} if len(named_shape) != len(names): raise DuplicateAxisNameError(v) new_v = Var(v.count, v.suffix, v.aval.update(named_shape=named_shape)) var_map[v] = new_v return new_v def subst_axis_names_eqn(eqn: JaxprEqn, subst: AxisSubst, var_map: Dict[Var, Var]) -> JaxprEqn: invars: List[Atom] = [v if isinstance(v, Literal) else var_map[v] for v in eqn.invars] try: outvars = [subst_axis_names_var(v, subst, var_map) for v in eqn.outvars] except DuplicateAxisNameError as e: e.eqn = eqn raise params = subst_axis_names(eqn.primitive, eqn.params, subst) return new_jaxpr_eqn(invars, outvars, eqn.primitive, params, eqn.source_info) def do_subst_axis_names_jaxpr(jaxpr: Union[Jaxpr, ClosedJaxpr], subst: AxisSubst): consts = None if isinstance(jaxpr, ClosedJaxpr): consts = jaxpr.consts jaxpr = jaxpr.jaxpr var_map: Dict[Var, Var] = {unitvar: unitvar} invars = [subst_axis_names_var(v, subst, var_map) for v in jaxpr.invars] constvars = [subst_axis_names_var(v, subst, var_map) for v in jaxpr.constvars] eqns = [subst_axis_names_eqn(eqn, subst, var_map) for eqn in jaxpr.eqns] outvars: List[Atom] = [v if isinstance(v, Literal) else var_map[v] for v in jaxpr.outvars] new_jaxpr = Jaxpr(constvars, invars, outvars, eqns) if consts is not None: return ClosedJaxpr(new_jaxpr, consts) return new_jaxpr @cache() def used_axis_names_jaxpr(jaxpr: Union[Jaxpr, ClosedJaxpr]): subst = NameGatheringSubst() do_subst_axis_names_jaxpr(jaxpr, subst) return frozenset(subst.axis_names) def subst_axis_names_jaxpr(jaxpr: Union[Jaxpr, ClosedJaxpr], subst: AxisSubst): if isinstance(subst, NameGatheringSubst): # This is a common case, so we optimize it! subst.axis_names |= used_axis_names_jaxpr(jaxpr) return jaxpr return do_subst_axis_names_jaxpr(jaxpr, subst) axis_substitution_rules: Dict[Primitive, Callable[[ParamDict, AxisSubst, bool], ParamDict]] = {} # ------------------- AxisPrimitive ------------------- # Primitives that store axis names in params and want those axis names to # participate in dispatch should subclass AxisPrimitive. class AxisPrimitive(Primitive): _dispatch_on_params = True # ------------------- Jaxpr checking ------------------- def typecheck(aval: AbstractValue, x) -> bool: return typecompat(aval, get_aval(x)) def typecompat(aval_ref: AbstractValue, aval: AbstractValue) -> bool: """Determine whether `aval` conforms to `aval_ref`. Ignores weak_type and named_shape, other than to check that an axis name isn't used with different sizes. """ try: return typematch(aval_ref, lattice_join(aval_ref, aval)) except TypeError: return False def typematch(aval1: AbstractValue, aval2: AbstractValue) -> bool: """Determine whether `aval1` and `aval2` are equivalent. Ignores weak_type and named_shape, other than to check that an axis name isn't used with different sizes. """ if aval1 == aval2: return True # unequal avals may still represent the same type, because type is represented # by avals at the shaped level, and because weak type tags and (for now) named # shape components aren't considered part of the type if isinstance(aval1, ShapedArray) and isinstance(aval2, ShapedArray): # a bonus check for whether any named axes have inconsistent sizes join_named_shapes(aval1.named_shape, aval2.named_shape) return (raise_to_shaped(aval1, weak_type=False).strip_named_shape() == raise_to_shaped(aval2, weak_type=False).strip_named_shape()) class JaxprTypeError(TypeError): pass def typecheck_assert(pred, msg): if not pred: raise JaxprTypeError(msg) custom_typechecks: Dict[Primitive, Callable] = {} def check_jaxpr(jaxpr: Jaxpr): """Checks well-formedness of a jaxpr. Specifically, check that: - variables that are read are bound beforehand - variables are typed equally throughout a jaxpr - variable type annotations are compatible with their binding expression Raises `JaxprTypeError` if `jaxpr` is determined invalid. Returns `None` otherwise. """ try: _check_jaxpr(jaxpr, [v.aval for v in jaxpr.invars]) except JaxprTypeError as e: if len(e.args) == 2: msg, eqn_idx = e.args jaxpr_str = str(pp_jaxpr_eqn_range(jaxpr, eqn_idx - 10, eqn_idx + 10, JaxprPpContext())) else: msg, = e.args jaxpr_str = str(pp_jaxpr_eqn_range(jaxpr, 0, 20, JaxprPpContext())) msg = "\n\n".join([msg, "while checking jaxpr:", jaxpr_str]) raise JaxprTypeError(msg) from None def _check_jaxpr(jaxpr: Jaxpr, in_avals: Sequence[AbstractValue]): def read(v: Atom) -> AbstractValue: if isinstance(v, Literal): return raise_to_shaped(get_aval(v.val)) else: typecheck_assert(v in env, f"Variable '{v}' not defined") return env[v] def write(v: Var, a: AbstractValue) -> None: typecheck_assert(v not in env, f"Variable '{v}' already bound") if not isinstance(v, DropVar): typecheck_assert(typecompat(v.aval, a), f"Variable '{v}' inconsistently typed as {a}, " f"bound as {v.aval}") env[v] = a env : Dict[Var, AbstractValue] = {} write(unitvar, abstract_unit) map(write, jaxpr.constvars, [v.aval for v in jaxpr.constvars]) map(write, jaxpr.invars, in_avals) for eqn_idx, eqn in enumerate(jaxpr.eqns): prim = eqn.primitive try: in_avals = map(read, eqn.invars) typecheck_assert(all(not isinstance(ina, ConcreteArray) for ina in in_avals), "Equation given ConcreteArray type inputs") if prim in custom_typechecks: out_avals = custom_typechecks[prim](*in_avals, **eqn.params) if out_avals is None: out_avals = [v.aval for v in eqn.outvars] elif prim.call_primitive: out_avals = check_call(prim, in_avals, eqn.params) elif prim.map_primitive: out_avals = check_map(prim, in_avals, eqn.params) else: out_avals = check_eqn(prim, in_avals, eqn.params) map(write, eqn.outvars, out_avals) except JaxprTypeError as e: msg, = e.args src = source_info_util.summarize(eqn.source_info) msg = "\n\n".join([msg, "in equation:", str(pp.nest(2, pp_eqn(eqn, JaxprPpContext()))), f"from source: {src}"]) raise JaxprTypeError(msg, eqn_idx) from None map(read, jaxpr.outvars) def check_eqn(prim, in_avals, params): for jaxpr in jaxprs_in_params(params): check_jaxpr(jaxpr) out_avals = prim.abstract_eval(*in_avals, **params) if not prim.multiple_results: out_avals = [out_avals] return out_avals def check_call(prim, in_avals, params): typecheck_assert("call_jaxpr" in params, f"Call primitive {prim} missing 'call_jaxpr' parameter") call_jaxpr = params["call_jaxpr"] # These checks also happen in recursive call, but give better errors here. typecheck_assert(len(in_avals) == len(call_jaxpr.invars), f"Call primitive {prim} with {len(call_jaxpr.invars)} " f"operands cannot call jaxpr with {len(call_jaxpr.invars)} " f"inputs") binder_avals = [v.aval for v in call_jaxpr.invars] for binder_aval, in_aval in zip(binder_avals, in_avals): typecheck_assert(typecompat(binder_aval, in_aval), f"Call primitive {prim} passes operand {in_aval} " f"to jaxpr expecting {binder_aval}") _check_jaxpr(call_jaxpr, in_avals) out_avals = [v.aval for v in call_jaxpr.outvars] return out_avals def check_map(prim, in_avals, params): typecheck_assert("call_jaxpr" in params, f"Map primitive {prim} missing 'call_jaxpr' parameter") call_jaxpr = params["call_jaxpr"] typecheck_assert("axis_size" in params, f"Map primitive {prim} missing 'axis_size' parameter") axis_size = params["axis_size"] typecheck_assert("axis_name" in params, f"Map primitive {prim} missing 'axis_name' parameter") axis_name = params["axis_name"] typecheck_assert("in_axes" in params, f"Map primitive {prim} missing 'in_axes' parameter") in_axes = params["in_axes"] typecheck_assert("out_axes" in params, f"Map primitive {prim} missing 'out_axes' parameter") out_axes = params["out_axes"] binder_avals = [unmapped_aval(axis_size, axis_name, in_axis, v.aval) if in_axis is not None else v.aval for v, in_axis in zip(call_jaxpr.invars, in_axes)] for binder_aval, in_aval in zip(binder_avals, in_avals): typecheck_assert(typecompat(binder_aval, in_aval), f"Call primitive {prim} passes operand {in_aval} " f"to jaxpr expecting {binder_aval}") mapped_avals = [mapped_aval(axis_size, in_axis, aval) if in_axis is not None else aval for aval, in_axis in zip(in_avals, in_axes)] with extend_axis_env(params['axis_name'], axis_size, None): _check_jaxpr(call_jaxpr, mapped_avals) mapped_out_avals = [v.aval for v in call_jaxpr.outvars] out_avals = [unmapped_aval(axis_size, axis_name, out_axis, aval) if out_axis is not None else aval for aval, out_axis in zip(mapped_out_avals, out_axes)] return out_avals # ------------------- Jaxpr printed representation ------------------- # A JaxprPpContext allows us to globally uniquify variable names within nested # Jaxprs. class JaxprPpContext: var_ids: DefaultDict[Var, int] def __init__(self): self.var_ids = collections.defaultdict(it.count().__next__) def pp_var(v: Var, context: JaxprPpContext): if isinstance(v, (Literal, DropVar)): return str(v) return f"{_encode_digits_alphabetic(context.var_ids[v])}{v.suffix}" def pp_vars(vs: Sequence[Any], context: JaxprPpContext, *, separator="", print_shapes: bool = False) -> pp.Doc: if print_shapes: return pp.nest(2, pp.group( pp.join(pp.text(separator) + pp.group(pp.brk()), [ pp.text(pp_var(v, context)) + pp.dim(pp.text(":" + v.aval.str_short(short_dtypes=True))) for v in vs ]) )) else: return pp.nest(2, pp.group( pp.join(pp.text(separator) + pp.group(pp.brk()), [pp.text(pp_var(v, context)) for v in vs]) )) def pp_kv_pair(k:str, v: Any, context: JaxprPpContext) -> pp.Doc: if type(v) is tuple and all(isinstance(j, (Jaxpr, ClosedJaxpr)) for j in v): pp_v = pp_jaxprs(v, context) elif isinstance(v, Jaxpr): pp_v = pp_jaxpr(v, context) elif isinstance(v, ClosedJaxpr): pp_v = pp_jaxpr(v.jaxpr, context) else: pp_v = pp.text(str(v)) return pp.text(f'{k}=') + pp_v def pp_kv_pairs(kv_pairs, context: JaxprPpContext) -> pp.Doc: if not kv_pairs: return pp.nil() return pp.group( pp.nest(2, pp.concat([ pp.text("["), pp.brk(""), pp.join(pp.brk(), [pp_kv_pair(k, v, context) for k, v in kv_pairs]) ])) + pp.brk("") + pp.text("]") ) def pp_eqn(eqn, context: JaxprPpContext, *, print_shapes=True, source_info=False, custom_pp_eqn_rules=True) -> pp.Doc: lhs = pp_vars(eqn.outvars, context, print_shapes=print_shapes) annotation = (source_info_util.summarize(eqn.source_info) if source_info else None) rule = pp_eqn_rules.get(eqn.primitive) if rule and custom_pp_eqn_rules: rhs = rule(eqn, context) else: rhs = [pp.text(eqn.primitive.name), pp_kv_pairs(sorted(eqn.params.items()), context), pp.text(" ") + pp_vars(eqn.invars, context)] return pp.concat([lhs, pp.text(" = ", annotation=annotation), *rhs]) CustomPpEqnRule = Callable[[JaxprEqn, JaxprPpContext], Sequence[pp.Doc]] pp_eqn_rules: Dict[Primitive, CustomPpEqnRule] = {} def pp_eqns(eqns, context: JaxprPpContext, *, print_shapes=True, source_info=False, custom_pp_eqn_rules=True ) -> pp.Doc: return pp.join( pp.brk("; "), [pp_eqn(e, context, print_shapes=print_shapes, source_info=source_info, custom_pp_eqn_rules=custom_pp_eqn_rules) for e in eqns]) def pp_eqn_compact(primitive_name: str, params: Dict, context: JaxprPpContext ) -> pp.Doc: filtered_params = {k: v for k, v in params.items() if (k != 'branches' and not isinstance(v, (Jaxpr, ClosedJaxpr)))} return (pp.text(primitive_name) + pp_kv_pairs(sorted(filtered_params.items()), context)) def pp_jaxpr_skeleton(jaxpr, eqns_fn, context: JaxprPpContext, *, print_shapes=True) -> pp.Doc: constvars = pp_vars(jaxpr.constvars, context, print_shapes=print_shapes) invars = pp_vars(jaxpr.invars, context, print_shapes=print_shapes) eqns = eqns_fn() outvars = pp.concat([ pp.text("("), pp_vars(jaxpr.outvars, context, separator=","), pp.text(")" if len(jaxpr.outvars) != 1 else ",)")]) return pp.group(pp.nest(2, pp.concat([ pp.text("{ "), pp.bright(pp.text("lambda ")), constvars, pp.text("; "), invars, pp.text(". "), pp.bright(pp.text("let")), pp.nest(2, pp.brk() + eqns), pp.brk(), pp.bright(pp.text("in ")), outvars ])) + pp.text(" }")) def pp_jaxpr(jaxpr, context: JaxprPpContext, *, print_shapes=True, source_info=False, custom_pp_eqn_rules=True) -> pp.Doc: eqns_fn = lambda: pp_eqns(jaxpr.eqns, context, print_shapes=print_shapes, source_info=source_info, custom_pp_eqn_rules=custom_pp_eqn_rules) return pp_jaxpr_skeleton(jaxpr, eqns_fn, context, print_shapes=print_shapes) def pp_jaxprs(jaxprs, context: JaxprPpContext) -> pp.Doc: jaxprs = [j.jaxpr if isinstance(j, ClosedJaxpr) else j for j in jaxprs] return pp.group(pp.nest(2, pp.concat([ pp.text('('), pp.brk(""), pp.join(pp.brk(), map(lambda x: pp_jaxpr(x, context), jaxprs))] )) + pp.brk("") + pp.text(')') ) def pp_jaxpr_eqn_range(jaxpr: Jaxpr, lo: int, hi: int, context: JaxprPpContext, print_shapes=True, source_info: bool = False) -> pp.Doc: lo = max(lo, 0) hi = max(lo, min(hi, len(jaxpr.eqns))) eqns = jaxpr.eqns[lo:hi] def eqns_fn(): pps = [] if len(eqns) == 0 and len(jaxpr.eqns) != 0: pps.append(pp.text('...')) else: if lo != 0: pps.append(pp.text('...')) pps.extend(map((lambda e: pp_eqn(e, context, print_shapes=print_shapes, source_info=source_info)), eqns)) if hi != len(jaxpr.eqns): pps.append(pp.text('...')) return pp.join(pp.brk("; "), pps) return pp_jaxpr_skeleton(jaxpr, eqns_fn, context, print_shapes=print_shapes)