sgl-project/sglang · error · ValueError

Expected x.shape[-1] to be even for split rotary, got

Error message

Expected x.shape[-1] to be even for split rotary, got {last}.

What it means

ValueError from apply_split_rotary_emb: the 'split' rotary embedding variant requires the last dimension of x to be even so it can be reshaped into (2, r). An odd last dim means head_dim is odd, which is incompatible with split rotary.

Solutions

1. Use an even head_dim in the model config
2. Switch rope_type to 'interleaved' if the checkpoint supports it
3. Check upstream reshapes: ensure x is (b, h, t, head_dim) with even head_dim before this call

Example fix

// before
attn = LTXConnector(..., dim_head=57, rope_type="split")
// after
attn = LTXConnector(..., dim_head=64, rope_type="split")

assert head_dim % 2 == 0, f"split rope needs even head_dim, got {head_dim}"

When it happens

Trigger: Calling the LTX-2 connector attention forward with rope_type='split' and a head_dim (x.shape[-1]) that is odd, e.g. head_dim=57 after projection.

Common situations: Loading an LTX-2 checkpoint with unusual head dims; manual model configs where dim_head is odd; accidental transpose/reshape upstream leaving an odd trailing dim.

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

• head_dim must be a multiple of 8, got
• audio_num_frames must be provided for RoPE coordinate…
• Cannot pad RoPE freqs of length
• Cannot repeat tensor with batch=
• cos and sin must have matching [S, D/2] shapes

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/10398a9c16e89d74. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/models/adapter/ltx_2_connector.py:46

def apply_split_rotary_emb(
    x: torch.Tensor, freqs: Tuple[torch.Tensor, torch.Tensor]
) -> torch.Tensor:
    cos, sin = freqs

    x_dtype = x.dtype
    needs_reshape = False
    if x.ndim != 4 and cos.ndim == 4:
        # cos is (#b, h, t, r) -> reshape x to (b, h, t, dim_per_head)
        # The cos/sin batch dim may only be broadcastable, so take batch size from x
        b = x.shape[0]
        _, h, t, _ = cos.shape
        x = x.reshape(b, t, h, -1).transpose(1, 2)
        needs_reshape = True

    # Split last dim (2*r) into (d=2, r)
    last = x.shape[-1]
    if last % 2 != 0:
        raise ValueError(
            f"Expected x.shape[-1] to be even for split rotary, got {last}."
        )
    r = last // 2

    # (..., 2, r)
    split_x = x.reshape(*x.shape[:-1], 2, r)
    first_x = split_x[..., :1, :]  # (..., 1, r)
    second_x = split_x[..., 1:, :]  # (..., 1, r)

    cos_u = cos.unsqueeze(-2)  # broadcast to (..., 1, r) against (..., 2, r)
    sin_u = sin.unsqueeze(-2)

    out = split_x * cos_u
    first_out = out[..., :1, :]
    second_out = out[..., 1:, :]

    first_out.addcmul_(-sin_u, second_x)
    second_out.addcmul_(sin_u, first_x)

View on GitHub (pinned to 0132848349)