jax-ml/jax · error · NotImplementedError
Transforms not supported for matmul_acc_lhs.
Error message
Transforms not supported for matmul_acc_lhs.
What it means
Raised during Mosaic/TPU lowering of a Pallas kernel that uses the tpu.matmul_acc_lhs primitive when accumulator transforms are supplied. The lowering path for this matmul variant simply does not implement transforms (e.g. scaling/offset accumulations), so it refuses to compile rather than producing wrong results.
Solutions
- Remove accumulator transforms from the kernel using matmul_acc_lhs
- Use the regular tpu.matmul primitive instead, which supports transforms
- Check the JAX version release notes for matmul_acc_lhs transform support before upgrading
Example fix
// before tpu.matmul_acc_lhs(acc, lhs, acc_transforms=transforms) // after tpu.matmul(lhs, rhs, acc=acc) # no transforms on acc
Defensive patterns
Strategy: validation
Validate before calling
from jax.tree_util import tree_flatten flat, treedef = tree_flatten(acc_transforms) assert not flat, 'matmul_acc_lhs does not support acc transforms'
Prevention
- Keep accumulator transforms off matmul_acc_lhs/matmul_pop kernels
- Pin supported JAX versions when using exotic matmul primitives
When it happens
Trigger: Calling tpu.matmul_acc_lhs(...) (directly or via a Pallas TPU kernel) where the accumulator carries transform metadata — typically when jit transformations or accumulation transforms propagate into the primitive's lowering.
Common situations: Writing custom TPU Pallas matmul kernels against jax._src.pallas.mosaic.tpu and composing them with transformed accumulators; version upgrades where transforms started being threaded through matmul primitives.
Related errors
- Transforms not supported for matmul_pop.
- Batching over dynamic grid values is not supported yet.
- Cannot split a Pallas key. Use fold_in instead to generate…
- {ctx.avals_out[0].dtype}
- dma_start not implemented in LoJAX yet.
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/c34ddb9728795853.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/pallas/mosaic/lowering.py:5761
mxu_id: int
def __post_init__(self):
tpu_core.check_accumulator_ref(self.shape, self.dtype, self.mxu_id)
@register_lowering_rule(tpu_primitives.matmul_acc_lhs_p)
def _matmul_acc_lhs_lowering_rule(
ctx: LoweringRuleContext,
acc: AccRef,
lhs: ir.Value,
*flat_acc_transforms,
load_staged_rhs: int | None,
acc_transforms_tree,
):
del ctx
acc_transforms = jax.tree.unflatten(acc_transforms_tree, flat_acc_transforms)
if acc_transforms:
raise NotImplementedError("Transforms not supported for matmul_acc_lhs.")
staged_rhs_kwarg: dict[str, Any] = {}
if load_staged_rhs is not None:
staged_rhs_kwarg = {"load_staged_rhs": load_staged_rhs}
tpu.matmul_acc_lhs(acc.base_address, lhs, acc.mxu_id, **staged_rhs_kwarg)
return []
@register_lowering_rule(tpu_primitives.matmul_pop_p)
def _matmul_pop_lowering_rule(
ctx: LoweringRuleContext,
acc: AccRef,
*flat_acc_transforms,
acc_transforms_tree,
):
acc_transforms = jax.tree.unflatten(acc_transforms_tree, flat_acc_transforms)
if acc_transforms:
raise NotImplementedError("Transforms not supported for matmul_pop.")
return tpu.matmul_pop(View on GitHub (pinned to 1e1c6a8fc0)