jax-ml/jax · error · ValueError
Unknown action
Error message
Unknown action: {action} What it means
The helper _copy_start_or_wait dispatches on an action string that must be exactly 'start' or 'wait' for async copy descriptors. Any other string (typo, wrong case, new action like 'commit') raises ValueError.
Solutions
- Use exactly 'start' or 'wait' as the action string
- If you need a new action, implement it in the descriptor API rather than the dispatch string
- Avoid relying on this private helper; use pltpu async_copy primitives instead
Example fix
// before _copy_start_or_wait(src, dst, action='Start') # wrong case // after _copy_start_or_wait(src, dst, action='start')
Defensive patterns
Strategy: validation
Validate before calling
assert action in ('start', 'wait'), action Type guard
def is_copy_action(a: str) -> TypeGuard[Literal['start','wait']]: return a in ('start','wait') Prevention
- Don't call private helpers; use pltpu async_copy primitives
- Constants for action strings instead of literals
When it happens
Trigger: Calling the async copy helper (used by Mosaic/TensorCore copy plumbing) with an action argument other than 'start'/'wait' — typically only reachable from internal code or monkey-patched wrappers passing a custom action.
Common situations: Extending JAX internals with new async-copy actions; typos in patched versions; calling private helpers directly from user code.
Understand the failure class
Background: "Must be a positive integer", "Invalid value", "Unsupported": the invalid-argument-value error family, when a library rejects the value you pass — this error's family across 35 libraries.
Related errors
- Loads and stores are only allowed on VMEM and SMEM…
- Loads are only allowed on VMEM and SMEM references.
- Acc ref dtype must be float32 or int32, got
- Acc ref must be at least 2D, got shape
- Accumulators are not available on TPU
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/e98b972c0b48c645.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/pallas/mosaic/helpers.py:42
def sync_copy(src_ref, dst_ref, *, add: bool = False) -> None:
"""Synchronously copies a PyTree of refs to another PyTree of refs."""
if not jax.tree.leaves(src_ref):
# No buffers to copy so skip the function.
return
@functools.partial(
pl_primitives.run_scoped, sem=tpu_core.SemaphoreType.DMA(())
)
def _(sem):
def _copy_start_or_wait(action, src_ref, dst_ref):
descriptor = plm_primitives.make_async_copy(src_ref, dst_ref, sem)
if action == "start":
descriptor.start(add=add)
elif action == "wait":
descriptor.wait()
else:
raise ValueError(f"Unknown action: {action}")
jax.tree.map(
functools.partial(_copy_start_or_wait, "start"),
src_ref,
dst_ref,
)
jax.tree.map(
functools.partial(_copy_start_or_wait, "wait"),
src_ref,
dst_ref,
)
def run_on_first_core(core_axis_name: str):
"""Runs a function on the first core in a given axis."""
num_cores = jax.lax.axis_size(core_axis_name)
if num_cores == 1:
return lambda f: f()View on GitHub (pinned to 1e1c6a8fc0)