sgl-project/sglang · error · ValueError

not a complete raw Muse Glimmer HF checkpoint: {len(missing)

Error message

not a complete raw Muse Glimmer HF checkpoint: {len(missing)} missing keys {missing[:4]}{'...' if len(missing) > 4 else ''}, {len(unexpected)} unexpected keys {unexpected[:4]}{'...' if len(unexpected) > 4 else ''}{hint}

What it means

sanitize validated the weights against the expected raw Muse Glimmer HF checkpoint key set and found both missing and unexpected keys, so it can neither load a raw checkpoint nor treat it as packaged. When the gate projections are missing and there are no unexpected keys, a hint is appended saying the weights look already fused and the config must declare muse_glimmer_mlx_format.

Source

Thrown at python/sglang/srt/hardware_backend/mlx/models/muse_glimmer_mlx.py:655

        text_keys = {
            k
            for k in weights
            if not any(marker in k for marker in _VISION_KEY_MARKERS)
            and "rotary_emb" not in k
        }
        expected = self._expected_raw_keys()
        missing = sorted(expected - text_keys)
        unexpected = sorted(text_keys - expected)
        if missing or unexpected:
            hint = ""
            gate_missing = all("output_gate_proj" in k for k in missing) and missing
            if gate_missing and not unexpected:
                hint = (
                    " (weights look already fused: if this is a packaged "
                    'artifact, its config.json must carry "muse_glimmer_mlx_format": '
                    f"{MUSE_GLIMMER_MLX_FORMAT_VERSION})"
                )
            raise ValueError(
                "not a complete raw Muse Glimmer HF checkpoint: "
                f"{len(missing)} missing keys {missing[:4]}"
                f"{'...' if len(missing) > 4 else ''}, "
                f"{len(unexpected)} unexpected keys {unexpected[:4]}"
                f"{'...' if len(unexpected) > 4 else ''}{hint}"
            )

        H = self.args.num_attention_heads
        D = self.args.head_dim
        hidden = self.args.hidden_size

        embed_shape = tuple(weights["model.embed_tokens.weight"].shape)
        if embed_shape != (self.args.vocab_size, hidden):
            raise ValueError(
                f"embed_tokens.weight has shape {embed_shape} but config says "
                f"(vocab_size, hidden_size) = ({self.args.vocab_size}, {hidden})"
            )
        raw_q_shape = tuple(weights["model.layers.0.self_attn.q_proj.weight"].shape)

View on GitHub (pinned to 0132848349)

Solutions

  1. If weights are already fused, add \"muse_glimmer_mlx_format\": <MUSE_GLIMMER_MLX_FORMAT_VERSION> to config.json
  2. If weights should be raw, verify all shards are present and from the same checkpoint
  3. Re-download or regenerate the complete checkpoint

Example fix

// before: fused weights, config.json lacks the marker -> missing gate keys
// after: add to config.json
"muse_glimmer_mlx_format": 1  // = MUSE_GLIMMER_MLX_FORMAT_VERSION
Defensive patterns

Strategy: fallback

Validate before calling

expected = expected_raw_keys(cfg)
missing = expected - set(weight_keys); unexpected = set(weight_keys) - expected
if missing and not unexpected and all("output_gate_proj" in k for k in missing):
    cfg.setdefault("muse_glimmer_mlx_format", VERSION)  # looks packaged

Try / catch

try:
    w = sanitize(weights)
except ValueError as e:
    if "look already fused" in str(e):
        cfg["muse_glimmer_mlx_format"] = VERSION; w = sanitize(weights)

Prevention

When it happens

Trigger: Loading a weight dict whose keys don't match the raw checkpoint schema — e.g. output_gate_proj.* keys absent (already fused) plus extra fused keys, or loading a different model's weights entirely.

Common situations: Loading a packaged artifact whose config.json lost the muse_glimmer_mlx_format marker, mixing safetensors shards from different models, or pointing the loader at a partially converted checkpoint.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5688bb583f19f62a. Report an issue: GitHub.