sgl-project/sglang · error · RuntimeError

H3 conditioning projection produced no output

Error message

H3 conditioning projection produced no output

What it means

RuntimeError raised when the conditioning projection forward produced no output: self.weight is None and self.layers is empty. Normally prevented by the constructor's error 1682, this guards against a module constructed in a degraded/invalid state.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/encoders/minimax_h3_qwen3vl.py:186

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        if int(hidden_states.shape[-1]) != self.input_dim:
            raise ValueError(
                f"H3 conditioning projection expects width {self.input_dim}, "
                f"got {int(hidden_states.shape[-1])}"
            )
        normalized = (hidden_states.float() - self.mean_in) / self.std_in
        projected = normalized @ self.weight if self.weight is not None else None
        if self.layers:
            residual = normalized.to(self.layers[0].weight.dtype)
            for index, layer in enumerate(self.layers):
                residual = layer(residual)
                if index + 1 < len(self.layers):
                    residual = F.gelu(residual)
            residual = residual.float()
            projected = residual if projected is None else projected + residual
        if projected is None:
            raise RuntimeError("H3 conditioning projection produced no output")
        output = projected * self.std_out + self.mean_out
        if self.sink_out is not None and int(output.shape[-2]) > 0:
            output[..., 0, :] = self.sink_out
        return output


class MiniMaxH3Qwen3VLEncoder(TextEncoder):
    """Qwen3-VL multimodal backbone producing MiniMax H3 conditioning.

    The component loader builds and loads this module under the encoder-folding
    TP group. A TP=1/SP=8 DiT deployment therefore shards the encoder over all
    eight otherwise-idle ranks during encoding.
    """

    # The inherited text-layer list covers Qwen's language stack; reference
    # modes also execute the embedded visual tower.
    layer_names = [
        *TextEncoder.layer_names,

View on GitHub (pinned to 0132848349)

Solutions

1. Reconstruct the projection from a valid checkpoint and fix the underlying 1682 condition
2. Don't bypass __init__ (e.g. via pickle/deepcopy tricks) when cloning the module

assert proj.weight is not None or len(proj.layers) > 0, "projection cannot produce output"

When it happens

Trigger: Calling forward() on a projection built without W or MLP layers (e.g. object created bypassing __init__ validation, or state mutated after construction).

Common situations: Rare defensive path; most commonly seen after manually constructing the module or after a failed/partial initialization swallowed earlier errors.

Related errors

• H3 conditioning projection {bias_name} has shape {tuple(bias
• H3 conditioning projection contains unsupported tensors: {so
• H3 conditioning projection has neither W nor an MLP
• H3 conditioning projection MLP outputs width {layer_input_di
• H3 conditioning projection W has shape {tuple(self.weight.sh

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/c12f839aa2d36658. Report an issue: GitHub.