sgl-project/sglang · critical · ValueError

Unsupported activation: {hidden_act}. Arcee model in SGLang

Error message

Unsupported activation: {hidden_act}. Arcee model in SGLang only supports 'relu2'.

What it means

The Arcee MLP implementation only supports the squared-ReLU ('relu2') activation; the config's hidden_act must be exactly 'relu2'.

Source

Thrown at python/sglang/srt/models/arcee.py:88

        super().__init__()
        # Arcee uses a single up-projection, not a merged gate/up projection.
        self.up_proj = ColumnParallelLinear(
            hidden_size,
            intermediate_size,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("up_proj", prefix),
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("down_proj", prefix),
            reduce_results=reduce_results,
        )
        if hidden_act != "relu2":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Arcee model in SGLang only supports 'relu2'."
            )
        # The activation function is relu(x)^2
        self.act_fn = get_act_fn("relu2")

    def forward(self, x, forward_batch=None):
        x, _ = self.up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(x)
        return x


class ArceeAttention(nn.Module):
    def __init__(
        self,
        config: LlamaConfig,
        hidden_size: int,

View on GitHub (pinned to 0132848349)

Solutions

1. Set "hidden_act": "relu2" in the model's config.json
2. Use an original Arcee checkpoint that ships relu2

Example fix

// before
"hidden_act": "gelu"
// after
"hidden_act": "relu2"

assert config.hidden_act == "relu2"

When it happens

Trigger: Loading an Arcee checkpoint whose config.json hidden_act is not "relu2" (e.g. "gelu" or "silu").

Common situations: Fine-tunes or converted checkpoints with a different activation; typo or default value written by a conversion script.

Related errors

• Unknown activation function: {act_fn}
• The hpc_ops MoE runner backend runs a plain SiLU-and-mul; it
• Unsupported activation: {self.activation}
• Unsupported activation: {hidden_act}. Only xIELU is supporte
• Unsupported activation: {hidden_act}. Only silu is supported

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/da9c9dc1f56c4cc2. Report an issue: GitHub.