Merge pull request #11656 from mattjj:while-loop-new-remat

PiperOrigin-RevId: 463973829

jax/_src/lax/control_flow/conditionals.py

@@ -446,7 +446,7 @@ def _cond_partial_eval_custom(saveable, unks_in, inst_in, eqn):
   unks_out: List[bool] = [False] * len(eqn.outvars)
   for jaxpr in branches:
     _, _, unks_out_, _, _ = pe.partial_eval_jaxpr_custom(
-        jaxpr.jaxpr, in_unknowns=ops_uk, in_inst=[True] * len(ops_uk),
+        jaxpr.jaxpr, in_unknowns=ops_uk, in_inst=True,
         ensure_out_unknowns=False, ensure_out_inst=True, saveable=saveable)
     unks_out = map(operator.or_, unks_out, unks_out_)
 
@@ -458,7 +458,7 @@ def _cond_partial_eval_custom(saveable, unks_in, inst_in, eqn):
   for jaxpr in branches:
     jaxpr_known, jaxpr_staged, _, inst_out, num_res = \
         pe.partial_eval_jaxpr_custom(
-            jaxpr.jaxpr, in_unknowns=ops_uk, in_inst=[True] * len(ops_uk),
+            jaxpr.jaxpr, in_unknowns=ops_uk, in_inst=True,
             ensure_out_unknowns=unks_out, ensure_out_inst=True,
             saveable=saveable)
     branches_known_.append( core.ClosedJaxpr(jaxpr_known,  jaxpr.consts))
@@ -481,7 +481,7 @@ def _cond_partial_eval_custom(saveable, unks_in, inst_in, eqn):
   # passing in_inst argument to partial_eval_jaxpr_custom above).
   new_inst = [x for x, inst in zip(eqn.invars, inst_in)
               if type(x) is core.Var and not inst]
-  inst_in = [True] * len(inst_in)
+  del inst_in
 
   # Create residual variables.
   newvar = core.gensym()

jax/_src/lax/control_flow/loops.py

@@ -828,14 +828,14 @@ def _scan_partial_eval_custom(saveable, unks_in, inst_in, eqn):
     unks_in = const_uk   + carry_uk   + xs_uk
     jaxpr_known_, jaxpr_staged_, unks_out, inst_out, num_res = \
         pe.partial_eval_jaxpr_custom(
-            jaxpr.jaxpr, in_unknowns=unks_in, in_inst=[True] * len(unks_in),
+            jaxpr.jaxpr, in_unknowns=unks_in, in_inst=True,
             ensure_out_unknowns=carry_uk + [False] * num_ys,
             ensure_out_inst=True, saveable=saveable)
     carry_uk_out, ys_uk = split_list(unks_out, [num_carry])
     if carry_uk_out == carry_uk:
       break
     else:
-      carry_uk = _map(operator.or_, carry_uk  , carry_uk_out  )
+      carry_uk = _map(operator.or_, carry_uk, carry_uk_out)
   else:
     assert False, "Fixpoint not reached"
   jaxpr_known  = core.ClosedJaxpr(jaxpr_known_ , jaxpr.consts)
@@ -1309,6 +1309,70 @@ def _while_partial_eval(trace: pe.JaxprTrace, *tracers: pe.Tracer, cond_nconsts:
   out_tracers = [t for t, uk in zip(out_tracers_, carry_uk) if uk]
   return util.merge_lists(carry_uk, out_known, out_tracers)
 
+# TODO(mattjj): de-duplicate code with _while_partial_eval
+def _while_partial_eval_custom(saveable, unks_in, inst_in, eqn):
+  del saveable  # We can't save any residuals anyway (w/o dynamic shapes)!
+  cond_jaxpr = eqn.params['cond_jaxpr']
+  cond_nconsts = eqn.params['cond_nconsts']
+  body_jaxpr = eqn.params['body_jaxpr']
+  body_nconsts = eqn.params['body_nconsts']
+
+  cond_consts_uk, body_consts_uk, carry_init_uk = \
+      split_list(unks_in, [cond_nconsts, body_nconsts])
+
+  # Fixpoint to compute known part of the body (trivial on 'inst_in', since we
+  # make all inputs available as DCE can subsequently prune any unused ones)
+  carry_uk = carry_init_uk
+  for _ in range(1 + len(carry_uk)):
+    body_unks_in = body_consts_uk + carry_uk
+    jaxpr_known_, _, carry_uk_out, _, num_res = \
+        pe.partial_eval_jaxpr_custom(
+            body_jaxpr.jaxpr, in_unknowns=body_unks_in, in_inst=True,
+            ensure_out_unknowns=carry_uk, ensure_out_inst=True,
+            saveable=ad_checkpoint.nothing_saveable)
+    if carry_uk_out == carry_uk:
+      break
+    else:
+      carry_uk = _map(operator.or_, carry_uk, carry_uk_out)
+  else:
+    assert False, "Fixpoint not reached"
+  assert not num_res
+  body_jaxpr_known = core.ClosedJaxpr(jaxpr_known_, body_jaxpr.consts)
+  del jaxpr_known_, carry_uk_out, num_res
+
+  # Compute the known part of cond_fun (basically pruning inputs on known side).
+  cond_unks_in = cond_consts_uk + carry_uk
+  cond_jaxpr_known_, _, [cond_uk], _, _ = \
+      pe.partial_eval_jaxpr_custom(
+          cond_jaxpr.jaxpr, cond_unks_in, in_inst=True,
+          ensure_out_unknowns=False, ensure_out_inst=True,
+          saveable=ad_checkpoint.nothing_saveable)
+  assert not cond_uk  # only possible with old-style remat
+  cond_jaxpr_known = core.ClosedJaxpr(cond_jaxpr_known_, cond_jaxpr.consts)
+  del cond_uk
+
+  # Build the known eqn.
+  ins_known, _ = partition_list(unks_in, eqn.invars)
+  out_binders_known, _ = partition_list(carry_uk, eqn.outvars)
+  params_known = dict(cond_jaxpr=cond_jaxpr_known, body_jaxpr=body_jaxpr_known,
+                      cond_nconsts=len(cond_consts_uk) - sum(cond_consts_uk),
+                      body_nconsts=len(body_consts_uk) - sum(body_consts_uk))
+  effects_known = core.join_effects(cond_jaxpr_known.effects,
+                                    body_jaxpr_known.effects)
+  eqn_known = pe.new_jaxpr_eqn(ins_known, out_binders_known, while_p,
+                               params_known, effects_known, eqn.source_info)
+
+  # Staged eqn is same as input eqn.
+  eqn_staged = eqn
+
+  # Instantiate all inputs (b/c jaxpr_staged takes all inputs).
+  new_inst = [x for x, inst in zip(eqn.invars, inst_in)
+              if type(x) is core.Var and not inst]
+
+  unks_out = carry_uk
+  inst_out = [True] * len(unks_out)
+  return eqn_known, eqn_staged, unks_out, inst_out, new_inst
+
 def _while_transpose_error(*_, **kwargs):
   raise ValueError("Reverse-mode differentiation does not work for "
                    "lax.while_loop or lax.fori_loop. "
@@ -1323,8 +1387,7 @@ pe.custom_partial_eval_rules[while_p] = _while_partial_eval
 xla.register_initial_style_primitive(while_p)
 ad.primitive_transposes[while_p] = _while_transpose_error
 batching.axis_primitive_batchers[while_p] = _while_loop_batching_rule
-pe.partial_eval_jaxpr_custom_rules[while_p] = \
-    partial(pe.partial_eval_jaxpr_custom_rule_not_implemented, 'while_loop')
+pe.partial_eval_jaxpr_custom_rules[while_p] = _while_partial_eval_custom
 
 
 def _pred_bcast_select_mhlo(

jax/interpreters/partial_eval.py

@@ -1241,11 +1241,13 @@ call_partial_eval_rules[remat_call_p] = _remat_partial_eval
 def partial_eval_jaxpr_custom(
     jaxpr: Jaxpr,
     in_unknowns: Sequence[bool],
-    in_inst: Sequence[bool],
+    in_inst: Union[bool, Sequence[bool]],
     ensure_out_unknowns: Union[bool, Sequence[bool]],
     ensure_out_inst: Union[bool, Sequence[bool]],
     saveable: Callable[..., bool],
   ) -> Tuple[Jaxpr, Jaxpr, List[bool], List[bool], int]:
+  if type(in_inst) is bool:
+    in_inst = (in_inst,) * len(jaxpr.invars)
   if type(ensure_out_unknowns) is bool:
     ensure_out_unknowns = (ensure_out_unknowns,) * len(jaxpr.outvars)
   if type(ensure_out_inst) is bool:

tests/api_test.py

@@ -5207,6 +5207,50 @@ class RematTest(jtu.JaxTestCase):
     self.assertEqual(jaxpr_text.count(' sin '), 1)
     self.assertEqual(jaxpr_text.count(' cos '), 2)
 
+  @parameterized.named_parameters(
+      {"testcase_name": f"{suffix}", "remat": remat}
+      for suffix, remat in [
+          ('', api.remat),
+          ('_new', new_checkpoint),
+      ])
+  def test_remat_of_while_loop(self, remat):
+    def cond_fn(carry):
+      i, _ = carry
+      return i < 3
+    def body_fn(carry):
+      i, x = carry
+      return i + 1, jnp.sin(x)
+    def f(x):
+      _, y = lax.while_loop(cond_fn, body_fn, (0, x))
+      return y
+
+    _, f_lin = jax.linearize(remat(f), 3.)
+    y_dot = f_lin(1.0)
+    expected = jax.grad(lambda x: jnp.sin(jnp.sin(jnp.sin(x))))(3.)
+    self.assertArraysAllClose(y_dot, expected, check_dtypes=False)
+
+    jaxpr = api.make_jaxpr(jax.linearize(remat(f), 4.)[1])(1.)
+    self.assertIn(' sin ', str(jaxpr))
+    self.assertIn(' cos ', str(jaxpr))
+
+  def test_remat_of_while_loop_policy(self):
+    def cond_fn(carry):
+      i, _ = carry
+      return i < 3
+    def body_fn(carry):
+      i, x = carry
+      return i + 1, jnp.sin(x)
+    def f(x):
+      _, y = lax.while_loop(cond_fn, body_fn, (0, x))
+      return y
+
+    # even with a policy, we can't save residuals (w/o dynamic shapes)!
+    save_cos = lambda prim, *_, **__: str(prim) == 'cos'
+    g = new_checkpoint(f, policy=save_cos)
+    jaxpr = api.make_jaxpr(jax.linearize(g, 4.)[1])(1.)
+    self.assertIn(' sin ', str(jaxpr))
+    self.assertIn(' cos ', str(jaxpr))
+
 
 class JaxprTest(jtu.JaxTestCase):
 

tests/lax_control_flow_test.py

@@ -105,6 +105,15 @@ SCAN_IMPLS_WITH_FOR = [
 ]
 
 
+def while_loop_new_checkpoint(cond_fun, body_fun, init_val):
+  return new_checkpoint(partial(lax.while_loop, cond_fun, body_fun))(init_val)
+
+WHILE_LOOP_IMPLS = [
+    (lax.while_loop, 'while_loop'),
+    (while_loop_new_checkpoint, 'new_checkpoint'),
+]
+
+
 def while_loop_reference(cond, body, carry):
   while cond(carry):
     carry = body(carry)
@@ -2007,13 +2016,16 @@ class LaxControlFlowTest(jtu.JaxTestCase):
     jtu.check_grads(loop, (x,), order=2, modes=["fwd"])
 
   @parameterized.named_parameters(
-      {"testcase_name": "_jit_loop={}_jit_body={}_jit_cond={}".format(
-          jit_loop, jit_body, jit_cond),
-       "jit_loop": jit_loop, "jit_body": jit_body, "jit_cond": jit_cond}
+      {"testcase_name": "_jit_loop={}_jit_body={}_jit_cond={}_impl={}".format(
+          jit_loop, jit_body, jit_cond, while_name),
+       "jit_loop": jit_loop, "jit_body": jit_body, "jit_cond": jit_cond,
+       "while_loop": while_impl}
       for jit_loop in [False, True]
       for jit_body in [False, True]
-      for jit_cond in [False, True])
-  def testWhileLinearize(self, jit_loop=True, jit_body=False, jit_cond=True):
+      for jit_cond in [False, True]
+      for while_impl, while_name in WHILE_LOOP_IMPLS)
+  def testWhileLinearize(self, while_loop, jit_loop=True, jit_body=False,
+                         jit_cond=True):
     cond = lambda x: x[0, 2] <= 8
     body = lambda x: x * x
 
@@ -2022,7 +2034,7 @@ class LaxControlFlowTest(jtu.JaxTestCase):
     if jit_body:
       body = jax.jit(body)
 
-    loop = partial(lax.while_loop, cond, body)
+    loop = partial(while_loop, cond, body)
     if jit_loop:
       loop = jax.jit(loop)