sgl-project/sglang · error · ValueError

Cannot pad RoPE freqs of length {cos.shape[0]} to shorter ta

Error message

Cannot pad RoPE freqs of length {cos.shape[0]} to shorter target {target_len}

What it means

Raised by ZImage._pad_freqs_cis_to_length when the precomputed RoPE cos/sin table is longer than the requested target length. Padding only extends tables; truncation is refused because silently dropping positional frequencies would misalign position ids.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/zimage.py:1132

                size=(1, 1, 1),
                start=(0, 0, 0),
                device=device,
            )
            .flatten(0, 2)
            .repeat(image_padding_len, 1)
        )
        image_pos_ids = torch.cat([image_ori_pos_ids, image_padding_pos_ids], dim=0)

        return self.rotary_emb(cap_pos_ids), self.rotary_emb(image_pos_ids)

    @staticmethod
    def _pad_freqs_cis_to_length(
        freqs_cis: Tuple[torch.Tensor, torch.Tensor], target_len: int
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        cos, sin = freqs_cis
        pad_len = target_len - cos.shape[0]
        if pad_len < 0:
            raise ValueError(
                f"Cannot pad RoPE freqs of length {cos.shape[0]} to shorter target {target_len}"
            )
        if pad_len == 0:
            return cos, sin
        return (
            torch.cat([cos, cos.new_zeros(pad_len, cos.shape[-1])], dim=0),
            torch.cat([sin, sin.new_zeros(pad_len, sin.shape[-1])], dim=0),
        )

    def _build_batched_freqs_cis(
        self,
        images: list[torch.Tensor],
        cap_feats: list[torch.Tensor],
        patch_size: int,
        f_patch_size: int,
        image_target_len: int,
        cap_target_len: int,
    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]:

View on GitHub (pinned to 0132848349)

Solutions

1. Compute target_len as max(cos.shape[0] for all sequences in the batch)
2. Fix the caller computing image_seq_len_target so it is never below the actual token count
3. Truncate explicitly before calling if truncation is genuinely intended

Example fix

# before
target_len = min(seq.shape[0] for seq in seqs)
# after
target_len = max(seq.shape[0] for seq in seqs)

target_len = max(s.shape[0] for s in seqs)
assert target_len >= cos.shape[0]

def can_pad_to(cos: torch.Tensor, target: int) -> bool:
    return target >= cos.shape[0]

When it happens

Trigger: The batched RoPE builder targets a sequence length smaller than an individual sequence's cos table length — e.g. a longer image sequence batched with a max-length computed from a shorter member.

Common situations: Mixed-resolution image batches where the target length (e.g. SP-local padded length) is computed from the wrong tensor; off-by-one in seq_len calculation.

Related errors

• rope_pool_fused expects q/k/v to be 3-D
• rope_pool_fused expects positions/slots to be 1-D
• rope_pool_fused expects pool tensors to be 3-D
• q shape must be [num_tokens, num_qo_heads, head_dim], got {q
• k shape must be [num_tokens, num_kv_heads, head_dim], got {k

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5cb18175757ea10d. Report an issue: GitHub.