jax-ml/jax · error · NotImplementedError
Encountered non-leading UntilingTransform or UnswizzleRef…
Error message
Encountered non-leading UntilingTransform or UnswizzleRef transforms: {transforms} What it means
During transform handling for references, UntilingTransform/UnswizzleRef transforms must appear at the front of the transform list. If they occur after other transforms, this NotImplementedError fires because undoing them out of order is not supported.
Solutions
- Upgrade JAX — transform ordering bugs are fixed upstream over time
- Simplify the kernel: avoid combining swizzled layouts with discharge/untile on the same ref
- Report a minimal repro to the JAX team
Defensive patterns
Strategy: retry
Prevention
- Keep JAX current — transform-ordering fixes land upstream
- Avoid mixing swizzle and discharge transforms on one ref
When it happens
Trigger: A ref carrying a mixed transform stack where un-tiling/un-swizzling is not leading, e.g. discharge followed by unswizzle ordering produced by nested transforms on buffers.
Common situations: Using swizzled buffers combined with multiple transforms (scatter/discharge/untile) in complex pallas pipelines; often a JAX-internal ordering issue rather than user-controllable.
Understand the failure class
Background: UnsupportedOperationException and "is not supported" errors: when a library deliberately refuses a call — this error's family across 30 libraries.
Related errors
- Non-indexing transforms on GMEM refs are not implemented.
- Not all transforms could be handled. Remaining transforms
- Unsupported transform
- Barrier does not support arbitrary transforms
- Commuting a `UntilingTransform` with a `ReshapeTransform`…
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/48c0c7b29e1156cc.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/pallas/mosaic_gpu/lowering.py:4773
ref, transforms = ref
ref_aval, transform_avals = ref_aval
# We ignore other transforms here, because they are already embedded
# in the jaxpr.
assert isinstance(ref_aval, state_types.AbstractRef)
ref, ref_aval, _ = _handle_transforms(
ctx, ref_aval, ref, transform_avals, transforms,
handle_reshapes=False, handle_transposes=False
)
if ctx.module_ctx.lowering_semantics == mgpu.LoweringSemantics.Warpgroup:
# In warpgroup semantics, we must reapply the transforms that were on
# the `BlockSpec` here, as the below expects the transformed value to be
# fed in.
spec_transforms = tuple(
t for t in transforms
if isinstance(t, (gpu_core.UntilingTransform, gpu_core.UnswizzleRef))
)
if spec_transforms != transforms[:len(spec_transforms)]:
raise NotImplementedError(
"Encountered non-leading UntilingTransform or UnswizzleRef "
f"transforms: {transforms}"
)
for t in pallas_core.undo_transforms(ref_aval, spec_transforms):
ref_aval = cast(state_types.AbstractRef, t.transform_type(ref_aval))
ref = _reinterpret_cast(ref, ref_aval)
args.append(ref)
program_ids = program_ids_treedef.unflatten(flat_program_ids)
for axis, pid in enumerate(program_ids):
if pid is not None:
continue
program_ids[axis] = _program_id(
axis, ctx.module_ctx.squashed_dims, len(program_ids)
)
new_module_ctx = dataclasses.replace(ctx.module_ctx, program_ids=program_ids)
return lower_jaxpr_to_mosaic_gpu(
new_module_ctx, ctx.launch_ctx, jaxpr, args
)View on GitHub (pinned to 1e1c6a8fc0)