sgl-project/sglang · critical · RuntimeError

Checkpoint provides gate weight {name!r} but the model built

Error message

Checkpoint provides gate weight {name!r} but the model built no g_proj (gating is disabled in the config). Set gating to True, "per-head", or "per-element" to load this checkpoint.

What it means

During weight loading, Laguna found a checkpoint tensor whose name contains .g_proj. but for which the built model has no parameter (laguna.py:823). The model was constructed with gating disabled, yet the checkpoint contains gate weights, so the load would silently drop them; the code raises to prevent a numerically-wrong model from serving.

Source

Thrown at python/sglang/srt/models/laguna.py:823

                param.weight_loader(
                    param,
                    loaded_weight,
                    name,
                    shard_id=shard_id,
                    expert_id=expert_id,
                )
                if layer_id is not None:
                    loaded_expert_shards.add((layer_id, expert_id, shard_id))
                matched_expert = True
                break
            if matched_expert:
                continue

            if name.endswith(".bias") and name not in params_dict:
                continue
            if name not in params_dict:
                if ".g_proj." in name:
                    raise RuntimeError(
                        f"Checkpoint provides gate weight {name!r} but the model built no "
                        "g_proj (gating is disabled in the config). Set gating to True, "
                        '"per-head", or "per-element" to load this checkpoint.'
                    )
                logger.warning("Parameter %s not found in params_dict", name)
                continue
            param = params_dict[name]
            weight_loader = getattr(param, "weight_loader", default_weight_loader)
            weight_loader(param, loaded_weight)

        # If any routed-expert tensor was silently dropped (e.g. a future
        # checkpoint renaming `gate_proj`, or a ckpt-vs-mapping shape mismatch),
        # fail loud here instead of generating garbage.
        expected = {
            (layer_id, expert_id, shard_id)
            for layer_id in moe_layer_ids
            for expert_id in range(self.config.num_experts)
            for shard_id in ("w1", "w2", "w3")

View on GitHub (pinned to 0132848349)

Solutions

1. Set the model config's gating to True, "per-head", or "per-element" to match the checkpoint
2. Re-fetch the matching config.json that shipped with the weights
3. Alternatively strip g_proj tensors from the checkpoint (not recommended - changes model numerics)

Example fix

// before (config.json)
"gating": false
// after
"gating": true

ckpt_names = list_checkpoint_keys(path)
if any(".g_proj." in n for n in ckpt_names):
    assert cfg.gating in (True, "per-head", "per-element")

When it happens

Trigger: Loading a gated Laguna checkpoint into a runtime whose config disables gating (gating=False/None) while the safetensors contain g_proj weights.

Common situations: Mismatched config.json and weight files (config from a non-gated variant, weights from a gated one); hand-editing the gating config field.

Related errors

• Unsupported activation: {hidden_act}. Only silu is supported
• {len(missing)} routed-expert tensors were not loaded (sample
• expected a tensor with at least one dimension
• dimension {dim} size {dim_size} must be divisible by 2 * gro
• f"Invalid prefix for SparseVideoGen2AttentionImpl: {prefix}"

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/40b49d2cdad85df9. Report an issue: GitHub.