jax-ml/jax · error · InternalFloatingPointError
inf
Error message
inf
What it means
Raised when jax_debug_infs (debug_infs config) is enabled and a computed buffer contains infinity. Like the NaN check, it turns silent Inf values into an early InternalFloatingPointError at the op that produced them, at the cost of performance.
Solutions
- Locate the overflowing op via traceback/jax.debug.print and guard it (clip logits, use logsumexp, epsilon in denominators)
- Switch to a wider dtype (float32) or enable loss scaling for mixed precision
- Turn off jax_debug_infs once resolved, as it disables caching and adds overhead
Example fix
# before p = jnp.exp(logits) # overflow with debug_infs # after p = jax.nn.softmax(logits) # uses stable logsumexp internally
Defensive patterns
Strategy: validation
Validate before calling
import jax.numpy as jnp def finite(x): return bool(jnp.isfinite(jnp.asarray(x, dtype=jnp.float32)).all()) assert all(finite(a) for a in jax.tree_util.tree_leaves(inputs))
Try / catch
try:
with jax.debug_infs(True):
out = f(x)
except jax._src.dispatch.InternalFloatingPointError as e:
print('Inf produced by op:', e); raise Prevention
- Clip exponents: jnp.clip(x, -30, 30) before exp
- Use float32 or loss scaling for mixed precision
- Add epsilons to denominators
When it happens
Trigger: Enabling jax_debug_infs while a computation overflows to ±Inf, e.g. exp(large), division by zero, or fp16 overflow; check_special inspects every op output buffer.
Common situations: Mixed-precision (bfloat16/fp16) training overflow; exp on large logits; division by near-zero denominators; enabled during numerical debugging sessions.
Related errors
- Invalid value ( ) encountered in sharded computation.
- 0th dimension of all xs should be replicated. Got
- 0th dimension of leaf passed to `jax.lax.map` should be…
- a and b must be 2D, but got
- a and b must have the same element type, but got
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/972c7bf88c158cd2.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/dispatch.py:313
if needs_check_special():
for buf in bufs:
_check_special(name, buf.dtype, buf)
def check_special_array(name: str, arr: array.ArrayImpl) -> array.ArrayImpl:
if needs_check_special():
if dtypes.issubdtype(arr.dtype, np.inexact):
for buf in arr._arrays:
_check_special(name, buf.dtype, buf)
return arr
def _check_special(name: str, dtype: np.dtype, buf: basearray.Array) -> None:
if dtypes.issubdtype(dtype, np.inexact):
if config.debug_nans.value and np.any(np.isnan(np.asarray(buf))):
raise InternalFloatingPointError(name, "nan")
if config.debug_infs.value and np.any(np.isinf(np.asarray(buf))):
raise InternalFloatingPointError(name, "inf")
def _device_put_reshard(x): return x
@util.cache(max_size=2048, trace_context_in_key=False)
def _cached_logical_device_ids(
inp_device_list: xc.DeviceList,
target_device_list: xc.DeviceList
) -> tuple[int, ...]:
device_to_index = {d: i for i, d in enumerate(target_device_list)}
return tuple(device_to_index[d] for d in inp_device_list)
def _different_device_order_reshard(
x: array.ArrayImpl, target_sharding: NamedSharding, copy: ArrayCopySemantics
) -> array.ArrayImpl:
x._check_if_deleted()
inp_sharding = x.shardingView on GitHub (pinned to 1e1c6a8fc0)