xai-org/x-algorithm · error · ValueError

Input must not be scalar.

Error message

Input must not be scalar.

What it means

Linear.__call__ requires inputs to have at least one dimension because it reads inputs.shape[-1] to infer input_size. A scalar (shape ()) has no last axis, so the call is rejected with ValueError before any weights are created.

Source

Thrown at phoenix/xrex/models/linear_layer.py:56

        self.input_size = None
        self.output_size = output_size
        self.with_bias = with_bias
        self.w_init = w_init
        self.b_init = b_init or jnp.zeros

        self.config = config
        self.rms_clip_axes = rms_clip_axes

        self.sharding_context = sharding_context
        self.pspec = pspec

    def __call__(
        self,
        inputs: jax.Array,
    ) -> jax.Array:
        fprop_dtype = inputs.dtype
        if not inputs.shape:
            raise ValueError("Input must not be scalar.")

        input_size = self.input_size = inputs.shape[-1]
        output_size = self.output_size

        w_init = self.w_init
        if w_init is None:
            stddev = self.config.init_scale / math.sqrt(self.input_size)
            w_init = hk.initializers.TruncatedNormal(stddev=stddev)

        w = get_parameter(
            "w",
            [input_size, output_size],
            jnp.float32,
            init=w_init,
            pspec=self.pspec,
            lr_multiplier=self.config.lr_multiplier,
            rms_clip_axes=self.rms_clip_axes,
        )

View on GitHub (pinned to 24c60942c5)

Solutions

1. Reshape the scalar to at least 1D before the layer: x[None] or jnp.atleast_1d(x).
2. Keep dims during reductions: x.sum(axis=-1, keepdims=True).
3. Batch per-example scalars into a [B, 1] tensor upstream.

Example fix

# before
y = linear(jnp.float32(0.5))

# after
y = linear(jnp.atleast_1d(jnp.float32(0.5)))

import jax.numpy as jnp
x = jnp.atleast_1d(x) if getattr(x, "ndim", 1) == 0 else x

def is_non_scalar(x) -> bool:
    return hasattr(x, "shape") and len(x.shape) > 0

When it happens

Trigger: Passing a JAX scalar (e.g. jnp.float32(1.0), the output of .sum() or .mean() without keepdims) into a Haiku Linear layer.

Common situations: Reduced tensors losing their dims in feature pipelines; indexing that yields 0-d arrays; feeding per-scalar summaries into a projection layer.

Understand the failure class

Background: Tensor shape mismatch errors ("must have shape", "expected shape ... got ..."): when tensor dimensions disagree with what an op or layer was told to expect — this error's family across 6 libraries.

Related errors

• Only 1D arrays are supported for unique.
• q, k, and v should all be 4D, got: {q.ndim=}, {k.ndim=}, {v.
• Expected {k.shape=} to be {kv_shape} (inferred from q)
• Expected {v.shape=} to be {kv_shape} (inferred from q)
• {kv_seq_len=} must be a multiple of {block_kv=}

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/92ad9f67ae8b78c8. Report an issue: GitHub.