jax-ml/jax · error · NotImplementedError
non-bool prevent_cse together with…
Error message
non-bool prevent_cse together with static_argnums/static_argnames
What it means
Raised by _remat3 when checkpoint/remat is combined with custom_vjp whose fwd result is rematerialized as an opaque unit; fires when a non-boolean prevent_cse is passed together with static_argnums/static_argnames, a combination the remat implementation does not support.
Solutions
- Nest the input to match the treedef structure
- Validate structure with tree_structure comparisons before calling
- Use a treedef derived from the input itself
Example fix
# before out = treedef.from_iterable_tree([a, b]) # treedef expects ((a,b),(c,d)) # after out = treedef.from_iterable_tree([(a, b), (c, d)])
Defensive patterns
Strategy: type-guard
Validate before calling
def from_iterable_safe(treedef, xs):
try:
return treedef.from_iterable_tree(xs)
except ValueError:
return None
out = from_iterable_safe(treedef, xs)
if out is None:
_, treedef = jax.tree_util.tree_flatten(xs_template_correctly_nested) Type guard
def shape_ok(treedef, xs) -> bool:
try:
treedef.from_iterable_tree(xs)
return True
except ValueError:
return False Try / catch
try:
treedef.from_iterable_tree(xs)
except ValueError as e:
if 'Tree structures did not match' in str(e):
xs = _add_missing_nesting(xs)
else:
raise Prevention
- Validate nesting depth before rebuilding
- Use treedefs derived from the same data shape
- Wrap from_iterable_tree in structure checks in library code
When it happens
Trigger: treedef.FromIterableTree(xs) with xs less nested (or smaller) than the treedef expects.
Common situations: Feeding scalars or flat lists where nested tuples are required; template/output-shape mismatches.
Related errors
- the `static_argnums` argument to `jax.checkpoint` /…
- But the tree structures differ:
- cond_fun must return a boolean scalar, but got pytree
- COO.tree_unflatten: invalid
- cotangent tree does not match function output, expected
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/08bea57c6a345d76.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/ad_checkpoint.py:1008
custom_vjp's fwd rule* is rematerialized as an opaque unit. For the
conventional idiom of a fwd rule calling its own custom_vjp-decorated
function to compute the primal output, that is exactly the intended
semantics. Values and gradients are unaffected at every order of
differentiation. The one observable consequence arises only under
higher-order AD: differentiating a second time runs the inner application's
own fwd rule (at first order it never runs, since the outer bwd rule
discharges the derivative), and values inside it (e.g.
``checkpoint_name``-tagged intermediates) cannot be marked saveable by the
checkpoint ``policy`` there; they are always recomputed.
"""
kwargs = dict(policy=policy, static_argnums=static_argnums,
static_argnames=static_argnames, prevent_cse=prevent_cse)
if f is None: return lambda g: _remat3(g, **kwargs)
return _remat3(f, **kwargs)
def _remat3(f, *, policy, static_argnums, static_argnames, prevent_cse=True):
if not isinstance(prevent_cse, bool) and (static_argnums or static_argnames):
raise NotImplementedError(
"non-bool prevent_cse together with static_argnums/static_argnames")
@wraps(f)
def decorator(*args, **kwargs):
if static_argnums or static_argnames:
# Like classic remat (and custom_vjp3), support unhashable static
# values by closing over them instead of threading them through the
# tracing machinery, which hashes them.
args_ = api_util.resolve_kwargs(f, args, kwargs)
argnums_ = (static_argnums,) if type(static_argnums) is int else static_argnums
argnums = frozenset(i % len(args_) for i in _static_argnums(
f, argnums_, static_argnames))
if not all(api_util.is_hashable(args_[i]) for i in argnums):
which_static = [i in argnums for i in range(len(args_))]
dyn_args, static_args = partition_list(which_static, args_)
f2 = _dyn_args_fun(f, argnums, tuple(map(WrapHashably, static_args)),
len(args_))
return _remat3(f2, policy=policy, static_argnums=(),
static_argnames=())(*dyn_args)View on GitHub (pinned to 1e1c6a8fc0)