jax-ml/jax · error · ValueError

Unsupported constant type: {x.type}

Error message

Unsupported constant type: {x.type}

What it means

The Mosaic constant-folding helper only understands arith.constant ops with IntegerType or FloatType results. A constant of any other MLIR type (bool/i1 handled elsewhere, complex, etc.) raises ValueError when the folder tries to extract its value.

Source

Thrown at jax/_src/pallas/mosaic/lowering.py:3494

class FoldingError(Exception):
  pass


def _fold(x, fuel):
  if fuel <= 0:
    raise FoldingError()
  op_name = getattr(x.owner, "name", None)
  binop_folds = {
      "arith.maxsi": max,
      "arith.minsi": min,
  }
  if op_name == "arith.constant":
    if isinstance(x.type, ir.IntegerType):
      return ir.IntegerAttr(x.owner.attributes["value"]).value
    elif isinstance(x.type, ir.FloatType):
      return ir.FloatAttr(x.owner.attributes["value"]).value
    else:
      raise ValueError(f"Unsupported constant type: {x.type}")
  if op_name in binop_folds:
    return binop_folds[op_name](_fold(v, fuel - 1) for v in x.owner.operands)
  raise FoldingError()


def _fold_and_get_constant_value(x):
  try:
    return _fold(x, 10)
  except FoldingError:
    return None


@register_lowering_rule(lax.stop_gradient_p)
def _stop_gradient_lowering_rule(_: LoweringRuleContext, x):
  return x

@register_lowering_rule(
    lax.max_p, ensure_mlir_values=False, kernel_types=[*tpu_core.CoreType]

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

1. Update jaxlib to match your JAX version — this is often an internal version-mismatch bug
2. Simplify the kernel to avoid constants of exotic types reaching the folder (materialize them as kernel inputs)
3. Replace bool constants with int32 0/1
4. Report upstream at github.com/google/jax with a minimal repro

Example fix

// before
flag = True  # traced bool constant used in arithmetic
// after
flag = jnp.int32(1)  # passed as kernel argument

catch ValueError from constant folding, simplify kernel constants, and retry; report if persistent

When it happens

Trigger: Constant folding encountering an arith.constant with a non-integer/non-float MLIR type while lowering a Pallas kernel, typically triggered indirectly by an operation whose operands fold to unusual constants.

Common situations: Kernels with boolean constants feeding into arithmetic; complex or index-type constants; version skew between JAX and jaxlib where new constant types appear in the pipeline.

Related errors

• Unsupported block shape type: {type(dim)}
• BlockMapping for {self.origin} has captured constants: {self
• Unsupported aval type: {aval}, {type(aval)}, {t}
• cannot cast {src} to {dst_type}
• Unexpected dim_block_size: {dim_block_size}

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/2c775d09c3bcb66e. Report an issue: GitHub.