sgl-project/sglang · error · ValueError

expert-pack requires --disable-shared-experts-fusion so the

Error message

expert-pack requires --disable-shared-experts-fusion so the shared expert remains on the dense GGUF path

What it means

The expert-pack path stores experts in a fused GGUF-like pack and requires shared experts to stay on the dense path. If shared-experts fusion is enabled (layer.num_fused_shared_experts != 0), the fused MoE layout conflicts with the pack, so create_weights raises.

Source

Thrown at python/sglang/srt/layers/quantization/expert_pack.py:99

        self.hidden_size: int | None = None
        self.intermediate_size: int | None = None
        self.activation = "silu"
        self.swiglu_limit: float | None = None
        self.situ_beta: float | None = None
        self.situ_linear_beta: float | None = None

    def create_weights(
        self,
        layer: torch.nn.Module,
        num_experts: int,
        hidden_size: int,
        intermediate_size_per_partition: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ) -> None:
        del extra_weight_attrs
        if layer.num_fused_shared_experts:
            raise ValueError(
                "expert-pack requires --disable-shared-experts-fusion so the "
                "shared expert remains on the dense GGUF path"
            )
        if layer.moe_ep_size != 1 or layer.moe_tp_size != 1:
            raise ValueError("expert-pack v1 supports only single-GPU TP=EP=1")
        if num_experts != self.store.header.num_experts:
            raise ValueError("FusedMoE expert count does not match expert-pack")
        if params_dtype not in (torch.bfloat16, torch.float16):
            raise ValueError("expert-pack kernel requires BF16 or FP16 activations")
        gate_shape = self.store.entries[(layer.layer_id, 0, 0)].shape
        down_shape = self.store.entries[(layer.layer_id, 0, 2)].shape
        if gate_shape != (hidden_size, intermediate_size_per_partition):
            raise ValueError(
                f"expert-pack gate shape {gate_shape} does not match "
                f"{(hidden_size, intermediate_size_per_partition)}"
            )
        if down_shape != (intermediate_size_per_partition, hidden_size):
            raise ValueError(

View on GitHub (pinned to 0132848349)

Solutions

  1. Launch with --disable-shared-experts-fusion
  2. Verify the layer then has num_fused_shared_experts == 0
  3. Check other expert-pack constraints (TP=EP=1) at the same time

Example fix

# before
python -m sglang.launch_server --model ds-expert-pack --tp 1
# after
python -m sglang.launch_server --model ds-expert-pack --tp 1 --disable-shared-experts-fusion
Defensive patterns

Strategy: validation

Validate before calling

assert "--disable-shared-experts-fusion" in server_args, "expert-pack requires --disable-shared-experts-fusion"

Prevention

When it happens

Trigger: Launching an expert-pack quantized MoE model without --disable-shared-experts-fusion, so SGLang fuses the shared expert into the MoE layer before expert-pack weight creation.

Common situations: Default server args enabling shared-experts fusion on models like DeepSeek-V3 expert-pack builds.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/4b7c3432601aa855. Report an issue: GitHub.