jax-ml/jax · error · NotImplementedError
+
Error message
{uninitialized_memory} + {str(dtype)} What it means
get_uninitialized_value could not produce an uninitialized sentinel for the combination of uninitialized_memory mode and dtype. Only NaN (floating) and zero modes are supported; NaN for non-float or unrecognized dtype combos hits the NotImplementedError.
Solutions
- Use uninitialized_memory='zero' if the dtype has no NaN sentinel
- Avoid depending on OOB/padded reads for non-float dtypes; mask reads instead
- File/patch support for the dtype in get_uninitialized_value if it should be supported
Example fix
// before params = interpret.InterpretParams(uninitialized_memory='nan') # with int32 buffers // after params = interpret.InterpretParams(uninitialized_memory='zero')
Defensive patterns
Strategy: fallback
Validate before calling
import jax.numpy as jnp mode = 'nan' if jnp.issubdtype(dtype, jnp.floating) else 'zero'
Try / catch
try:
val = get_uninitialized_value(dtype, 'nan')
except NotImplementedError:
val = 0 # fall back to zero for non-float dtypes Prevention
- Mask OOB/padded reads instead of consuming uninitialized sentinel values
- Choose 'nan' mode only when all relevant buffers are floating point
When it happens
Trigger: Requesting uninitialized (padded/out-of-bounds) memory with uninitialized_memory='nan' for an integer/bool/bfloat16-unsupported dtype, or an unrecognized dtype, in interpret-mode reads.
Common situations: Kernels reading padded regions of integer accumulators in interpret mode; new/extended dtypes not yet covered by the sentinel table.
Related errors
- Attempting to deallocate allocation with key
- Attempting to get contents of allocation with key
- Attempting to store into allocation with key
- mixes JAX mesh and Pallas mesh grid axes
- Batching over dynamic grid values is not supported yet.
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/d7ea9f7e7441c44f.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/pallas/mosaic/interpret/utils.py:43
from jax._src.pallas import primitives
from jax._src.util import safe_map
import jax.numpy as jnp
import numpy as np
def get_uninitialized_value(
dtype, uninitialized_memory: Literal["nan", "zero"]
):
if uninitialized_memory == "nan":
if jnp.issubdtype(dtype, jnp.floating):
return np.nan
elif jnp.issubdtype(dtype, jnp.integer):
return jnp.iinfo(dtype).max
elif jnp.issubdtype(dtype, jnp.bool):
return True
if uninitialized_memory == "zero":
return 0
raise NotImplementedError(uninitialized_memory + " + " + str(dtype))
def get_uninitialized_array(
shape, dtype, uninitialized_memory: Literal["nan", "zero"]
):
return jnp.full(
shape,
get_uninitialized_value(dtype, uninitialized_memory),
dtype,
)
def pad_to_block_dimension(
value, block_shape, uninitialized_memory: Literal["nan", "zero"]
):
"""Pads values so the shape evenly divides into block dimensions.
For example, if values has a shape of (33, 2, 5) with a block_shape ofView on GitHub (pinned to 1e1c6a8fc0)