allow for recursive uses of custom_transpose

Co-authored-by: Matthew Johnson <mattjj@google.com>

jax/_src/custom_transpose.py

@@ -19,6 +19,7 @@
 from jax import linear_util as lu
 from jax.interpreters import ad
 from jax.interpreters import mlir
+from jax.interpreters import partial_eval as pe
 from jax.interpreters import xla
 from jax.tree_util import (tree_flatten, tree_leaves, tree_map,
                            tree_structure, treedef_tuple, tree_unflatten)
@@ -133,6 +134,16 @@ def check_transpose_rule_trees(rule, lin_tree, rule_out_tree):
           f'Transpose rule output: {rule_out_tree}\n'
           f'Linear primal inputs: {lin_tree}')
 
+def make_transpose_from_thunk(thunk, lin_tree):
+  transpose_jaxpr, transpose_consts = thunk()
+  transpose_jaxpr = core.ClosedJaxpr(
+      pe.convert_constvars_jaxpr(transpose_jaxpr), ())
+  def transpose(res_arg, ct_out):
+    args_flat = tree_leaves((res_arg, ct_out))
+    ct_ins = core.jaxpr_as_fun(transpose_jaxpr)(*transpose_consts, *args_flat)
+    return tree_unflatten(lin_tree, ct_ins)
+  return transpose
+
 
 ### custom_transpose primitive and rules
 
@@ -157,8 +168,14 @@ def bind(self, call, *args, **params):
   # TODO(frostig,mattjj): consider keeping `call` as a named parameter
   # instead of following this "call primitive" convention.
   def get_bind_params(self, params):
+    assert 'call_jaxpr' in params
+    assert 'transpose_jaxpr_thunk' in params
     new_params = dict(params)
-    return [new_params.pop('call')], new_params
+    new_params['transpose'] = make_transpose_from_thunk(
+        new_params.pop('transpose_jaxpr_thunk'),
+        new_params['lin_tree'])
+    call = lu.wrap_init(core.jaxpr_as_fun(new_params.pop('call_jaxpr')))
+    return [call], new_params
 
 
 # TODO(frostig,mattjj): reinstate checks
@@ -167,7 +184,16 @@ def custom_transpose_typecheck(*avals, **params):
 
 
 def custom_transpose_transpose_rule(
-    cts, *args, call, transpose, out_types, res_tree, lin_tree, out_tree):
+    cts, *args, out_types, res_tree, lin_tree, out_tree, **params):
+
+  if 'transpose_jaxpr_thunk' in params:
+    assert 'call_jaxpr' in params
+    transpose = make_transpose_from_thunk(
+        params['transpose_jaxpr_thunk'], lin_tree)
+  else:
+    assert 'call' in params
+    transpose = params['transpose']
+
   call_in_tree = treedef_tuple((res_tree, lin_tree))
 
   # TODO(frostig,mattjj): `lin_arg` indicates the inputs with respect

jax/_src/lax/control_flow.py

@@ -805,7 +805,7 @@ def switch(index, branches, operand):
 
 
 def _cond(pred, true_fun: Callable, false_fun: Callable, *operands,
-          operand=_no_operand_sentinel):
+          operand=_no_operand_sentinel, linear=None):
   """Conditionally apply ``true_fun`` or ``false_fun``.
 
   ``cond()`` has equivalent semantics to this Python implementation::
@@ -865,6 +865,12 @@ def cond(pred, true_fun, false_fun, *operands):
       return false_fun(*operands)
 
   ops, ops_tree = tree_flatten(operands)
+  if linear is None:
+    linear_ops = [False] * len(ops)
+  else:
+    linear_ops, ops_tree2 = tree_flatten(linear)
+    if ops_tree != ops_tree2:
+      raise TypeError('linear tree and operand tree mismatch')
   ops_avals = tuple(_map(_abstractify, ops))
 
   jaxprs, consts, out_trees = _initial_style_jaxprs_with_common_consts(
@@ -878,10 +884,10 @@ def cond(pred, true_fun, false_fun, *operands):
 
   index = lax.convert_element_type(pred, np.int32)
 
-  linear = (False,) * (len(consts) + len(ops))
+  linear = [False] * len(consts) + linear_ops
   out = cond_p.bind(
       index, *consts, *ops,
-      branches=(false_jaxpr, true_jaxpr), linear=linear)
+      branches=(false_jaxpr, true_jaxpr), linear=tuple(linear))
   return tree_unflatten(out_tree, out)
 
 @api_boundary

jax/interpreters/ad.py

@@ -164,8 +164,10 @@ def recast_to_float0(primal, tangent):
   else:
     return tangent
 
-# NOTE: The FIXMEs below are caused by primal/tangent mixups (type errors if you will)
-def backward_pass(jaxpr: core.Jaxpr, reduce_axes, transform_stack, consts, primals_in, cotangents_in):
+# NOTE: The FIXMEs below are caused by primal/tangent mixups (type
+# errors if you will)
+def backward_pass(jaxpr: core.Jaxpr, reduce_axes, transform_stack,
+                  consts, primals_in, cotangents_in):
   if all(type(ct) is Zero for ct in cotangents_in):
     return map(lambda v: Zero(v.aval), jaxpr.invars)
 

jax/interpreters/partial_eval.py

@@ -1626,20 +1626,22 @@ def process_custom_transpose(self, prim, call, tracers,
 
     transpose_flat, in_tree2 = flatten_fun_nokwargs(
         lu.wrap_init(transpose), treedef_tuple((res_tree, out_tree)))
-    transpose_jaxpr, in_avals2, transpose_consts = trace_to_subjaxpr_dynamic(
-        transpose_flat, self.main, in_avals_t)
-    closed_transpose_jaxpr = core.ClosedJaxpr(
-        convert_constvars_jaxpr(transpose_jaxpr), ())
+
+    main_ = ref(self.main)
+    # the following thunk evaluates to a pair: transpose_jaxpr, transpose_consts
+    transpose_jaxpr_thunk = _memoize(
+        lambda: trace_to_subjaxpr_dynamic(
+            transpose_flat, main_(), in_avals_t)[::2])
 
     out_tracers = [DynamicJaxprTracer(self, a) for a in out_avals]
     invars = map(self.getvar, tracers)
     constvars = map(self.getvar, map(self.instantiate_const, call_consts))
     outvars = map(self.makevar, out_tracers)
     eqn = new_jaxpr_eqn([*constvars, *invars], outvars, prim,
                         dict(call_jaxpr=closed_call_jaxpr,
-                             transpose_jaxpr=(closed_transpose_jaxpr,
-                                              transpose_consts),
-                             num_consts=len(call_consts)),
+                             transpose_jaxpr_thunk=transpose_jaxpr_thunk,
+                             out_types=out_types, res_tree=res_tree,
+                             lin_tree=lin_tree, out_tree=out_tree),
                         source_info_util.current())
     self.frame.eqns.append(eqn)
     return out_tracers

tests/api_test.py

@@ -6891,7 +6891,6 @@ def tp(r, t): return 2 * t / r
     self.assertAllClose(f_t(x), g_t(x))
 
   def test_jit_recursive(self):
-    raise unittest.SkipTest('unimplemented')  # TODO(frostig,mattjj)
     def f(x, y):
       @custom_transpose(jnp.ones(2))
       def fn(r, x): return x / r
@@ -6913,6 +6912,56 @@ def tp(r, t): return 2 * fn(r, t)
     self.assertAllClose(f_(x), g_(x))
     self.assertAllClose(f_t(x), g_t(x))
 
+  def test_cond(self):
+    def f(x, y):
+      @custom_transpose(jnp.ones(2))
+      def fn(r, x): return x / r
+      @fn.def_transpose
+      def tp(r, t): return 2 * t / r
+
+      return x + fn(y, x)
+
+    def cond_wrap(f):
+      return lambda i, x: lax.cond(i > 0, f, lambda x: x, x,
+                                   linear=(True,))
+
+    i = 7.
+    x = jnp.ones(2) * 6.
+    y = jnp.ones(2) * 3.
+
+    f_ = lambda x: f(x, y)
+    f_t = transpose_unary(f_, x)
+    g_ = partial(cond_wrap(f_), i)
+    g_t = transpose_unary(g_, x)
+
+    self.assertAllClose(f_(x), g_(x))
+    self.assertAllClose(f_t(x), g_t(x))
+
+  def test_cond_recursive(self):
+    def f(x, y):
+      @custom_transpose(jnp.ones(2))
+      def fn(r, x): return x / r
+      @fn.def_transpose
+      def tp(r, t): return 2 * fn(r, t)
+
+      return x + fn(y, x)
+
+    def cond_wrap(f):
+      return lambda i, x: lax.cond(i > 0, f, lambda x: x, x,
+                                   linear=(True,))
+
+    i = 7.
+    x = jnp.ones(2) * 6.
+    y = jnp.ones(2) * 3.
+
+    f_ = lambda x: f(x, y)
+    f_t = transpose_unary(f_, x)
+    g_ = partial(cond_wrap(f_), i)
+    g_t = transpose_unary(g_, x)
+
+    self.assertAllClose(f_(x), g_(x))
+    self.assertAllClose(f_t(x), g_t(x))
+
 
 class CustomVmapTest(jtu.JaxTestCase):