sgl-project/sglang · critical · ValueError

H3 conditioning projection contains unsupported tensors

Error message

H3 conditioning projection contains unsupported tensors: {sorted(tensors)}

What it means

The conditioning projection constructor found leftover tensors in the checkpoint state dict that it could not map to any known weight/bias/MLP pattern. After consuming recognized W, MLP, and normalization tensors, any remaining keys trigger this error listing the offenders.

Solutions

1. Inspect the tensor names listed in the error and strip or rename the unsupported ones before construction
2. Re-export only the projection submodule keys using the naming pattern expected by the loader
3. Check for a version mismatch between the checkpoint exporter and this sglang version

Example fix

# before
proj = MiniMaxH3ConditioningProjection(sd, ...)
# after
supported = {k: v for k, v in sd.items() if not k.startswith("unexpected_prefix.")}
proj = MiniMaxH3ConditioningProjection(supported, ...)

ALLOWED = re.compile(r"^(W$|mlp\..*weight$|mlp\..*bias$|mean_in|std_in|mean_out|std_out|sink_out$)")
bad = [k for k in sd if not ALLOWED.match(k)]
assert not bad, f"unsupported tensors: {bad}"

When it happens

Trigger: MiniMaxH3ConditioningProjection constructed with a state dict containing keys that don't match any recognized pattern (extra norms, unexpected prefixes, optimizer state, or renamed keys).

Common situations: Checkpoint format changed between releases, user exported a full model instead of just the projection module, or extra keys like 'norm.weight' variants with unknown prefixes were included.

Related errors

• H3 conditioning projection
• H3 conditioning projection expects encoder width
• H3 conditioning projection expects width
• H3 conditioning projection has neither W nor an MLP
• H3 conditioning projection MLP outputs width

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d1b5c871db723c4a. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/models/encoders/minimax_h3_qwen3vl.py:147

        for layer_index in layer_indices:
            weight_name = f"mlp.{layer_index}.weight"
            bias_name = f"mlp.{layer_index}.bias"
            weight = tensors.pop(weight_name)
            bias = tensors.pop(bias_name, None)
            if weight.ndim != 2 or int(weight.shape[1]) != layer_input_dim:
                raise ValueError(
                    f"H3 conditioning projection {weight_name} cannot follow "
                    f"width {layer_input_dim}: got {tuple(weight.shape)}"
                )
            if bias is not None and tuple(bias.shape) != (int(weight.shape[0]),):
                raise ValueError(
                    f"H3 conditioning projection {bias_name} has shape "
                    f"{tuple(bias.shape)}, expected ({int(weight.shape[0])},)"
                )
            layers.append(_FrozenLinear(weight, bias))
            layer_input_dim = int(weight.shape[0])
        if tensors:
            raise ValueError(
                "H3 conditioning projection contains unsupported tensors: "
                f"{sorted(tensors)}"
            )
        if self.weight is None and not layers:
            raise ValueError("H3 conditioning projection has neither W nor an MLP")
        if layers and layer_input_dim != self.output_dim:
            raise ValueError(
                f"H3 conditioning projection MLP outputs width {layer_input_dim}, "
                f"expected {self.output_dim}"
            )
        if self.weight is not None and tuple(self.weight.shape) != (
            self.input_dim,
            self.output_dim,
        ):
            raise ValueError(
                "H3 conditioning projection W has shape "
                f"{tuple(self.weight.shape)}, expected "
                f"({self.input_dim}, {self.output_dim})"

View on GitHub (pinned to 0132848349)