jax-ml/jax · error · ValueError

Empty array

Error message

Empty array

What it means

Raised by pack_array, which packs a list of SSA values into a stack-allocated LLVM array (used e.g. to fill TMA descriptors). An empty list has no element type to allocate for, so it's rejected immediately rather than producing an ambiguous alloca.

Source

Thrown at jax/experimental/mosaic/gpu/utils.py:130

      desc, llvm.ConstantOp(i64, ir.IntegerAttr.get(i64, 0)).result, [2]
  ).result
  if rank > 0:
    for i, s in enumerate(memref_ty.shape):
      desc = llvm.InsertValueOp(
          desc, llvm.ConstantOp(i64, ir.IntegerAttr.get(i64, s)).result, [3, i]
      ).result
    for i, s in enumerate(strides):
      desc = llvm.InsertValueOp(
          desc, llvm.ConstantOp(i64, ir.IntegerAttr.get(i64, s)).result, [4, i]
      ).result
  result = builtin.unrealized_conversion_cast([memref_ty], [desc])
  assert isinstance(result, ir.Value)
  return result


def pack_array(values):
  if not values:
    raise ValueError("Empty array")
  elem_ty = values[0].type
  i64 = ir.IntegerType.get_signless(64)
  ptr_ty = ir.Type.parse("!llvm.ptr")
  arr_ptr = llvm.alloca(ptr_ty, c(len(values), i64), elem_ty)
  for i, v in enumerate(values):
    elem_ptr = getelementptr(arr_ptr, [i], elem_ty)
    llvm.store(v, elem_ptr)
  return arr_ptr


def get_contiguous_strides(xs):
  strides_ret = []
  stride = 1
  for x in xs[::-1]:
    strides_ret.append(stride)
    stride *= x
  return strides_ret[::-1]

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

1. Guard callers: skip pack_array/init_tma_desc when the values list is empty
2. Ensure tensor shapes fed into TMA descriptor creation are non-empty
3. Pass at least one value (e.g. a zero constant) if the API requires a placeholder

Example fix

# before
vals = compute_dims(t)  # t has 0-sized dim -> []
packed = utils.pack_array(vals)
# after
if not vals:
  return None  # or raise a clearer upstream error
packed = utils.pack_array(vals)

if not values:
    raise ValueError('cannot build TMA descriptor from empty dimensions')
packed = utils.pack_array(values)

def is_packable(values):
    return len(values) > 0

When it happens

Trigger: Calling utils.pack_array([]) — in practice via init_tma_desc with an empty tensor of box dimensions, e.g. a TMA descriptor for a 0-rank/empty tensor shape.

Common situations: Building TMA descriptors from dynamically-sized tensors that can be empty at trace time (e.g. 0-sized dimension), or computing box dims that degenerate to an empty list.

Related errors

• Transfer of {total_bits} bits is not divisible by {8 * utils
• Expected both or neither of scales to be specified.
• {iterations=} must be positive
• Unsupported trace scope: {trace_scope}
• Scope {spec.trace_scope} not supported

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/73b8eaab7983f649. Report an issue: GitHub.