sgl-project/sglang · error · RuntimeError

Multiview attention was not initialized.

Error message

Multiview attention was not initialized.

What it means

When more than one view is generated (num_views > 1) and use_multiview_attention is enabled, the block needs self.attn_multiview. If it is None because multiview attention was never initialized for this block, forward raises this RuntimeError.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/hunyuan3d_paint.py:131

        if mode is not None and "r" in mode and self.use_reference_attention:
            if self.attn_refview is None:
                raise RuntimeError("Reference attention was not initialized.")
            reference = condition_embeddings[self.layer_name]
            reference = reference.unsqueeze(1).repeat(1, num_views, 1, 1)
            reference = rearrange(reference, "b n l c -> (b n) l c")
            reference_output = self.attn_refview(
                normalized, encoder_hidden_states=reference
            )
            reference_scale = self._broadcast_scale(
                1.0 if self.is_turbo else options.get("ref_scale", 1.0),
                reference_output,
                num_views,
            )
            hidden_states = hidden_states + reference_scale * reference_output

        if num_views > 1 and self.use_multiview_attention:
            if self.attn_multiview is None:
                raise RuntimeError("Multiview attention was not initialized.")
            multiview = rearrange(normalized, "(b n) l c -> b (n l) c", n=num_views)
            position_masks = options.get("position_attn_mask")
            position_mask = None
            if isinstance(position_masks, dict):
                position_mask = position_masks.get(multiview.shape[1])
            multiview_output = self.attn_multiview(
                multiview,
                encoder_hidden_states=multiview,
                attention_mask=position_mask,
            )
            multiview_output = rearrange(
                multiview_output, "b (n l) c -> (b n) l c", n=num_views
            )
            multiview_scale = 1.0 if self.is_turbo else options.get("mva_scale", 1.0)
            hidden_states = hidden_states + multiview_scale * multiview_output

        hidden_states = hidden_states + self.transformer.attn2(
            self.transformer.norm2(hidden_states),

View on GitHub (pinned to 0132848349)

Solutions

1. Replace transformer blocks with use_multiview_attention=True before running multiview inference
2. Set num_in_batch=1 in cross_attention_kwargs if you genuinely only need one view
3. Confirm every block that sees num_views>1 was covered by _replace_transformer_blocks

Example fix

# before
_replace_transformer_blocks(unet, use_multiview_attention=False)
out = block(x, cross_attention_kwargs={"num_in_batch": 4, ...})

# after
_replace_transformer_blocks(unet, use_multiview_attention=True)
out = block(x, cross_attention_kwargs={"num_in_batch": 4, ...})

num_views = int(opts.get('num_in_batch', 1))
if num_views > 1 and getattr(block, 'use_multiview_attention', False):
    assert block.attn_multiview is not None, 'attn_multiview not initialized'

def block_has_multiview_attn(block) -> bool:
    return getattr(block, 'use_multiview_attention', False) and block.attn_multiview is not None

try:
    out = block(x, cross_attention_kwargs=opts)
except RuntimeError as e:
    if 'Multiview attention was not initialized' in str(e) and int(opts.get('num_in_batch', 1)) == 1:
        out = block(x, cross_attention_kwargs=opts)
    else:
        raise

When it happens

Trigger: Calling the paint block forward with num_in_batch > 1 and use_multiview_attention=True on a block whose attn_multiview is None (built with multiview attention disabled or not replaced).

Common situations: Generating multiview images with a model whose blocks were replaced with use_multiview_attention=False; using a single-view checkpoint/pipeline path for a multiview request.

Related errors

• Reference attention was not initialized.
• Hunyuan3D Paint expects square latents and a matching view c
• encoder_hidden_states is required when encoder_key_value is
• Didn't get guidance strength for guidance distilled model.
• Hunyuan3D reference attention requires a shared cache.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5e49eccf0118d7d9. Report an issue: GitHub.