sgl-project/sglang · critical · RuntimeError

Some weights are not initialized from checkpoints: {unloaded

Error message

Some weights are not initialized from checkpoints: {unloaded_params}

What it means

Raised by DeepseekOCRModel.load_weights when, after consuming every checkpoint weight file, some registered parameters were never loaded. This is a strict completeness check: any parameter name the model expects but the checkpoint did not provide aborts weight loading.

Source

Thrown at python/sglang/srt/models/deepseek_ocr.py:1887

                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip experts that are not assigned to this worker.
                if (
                    "mlp.experts." in name or "mlp.shared_experts." in name
                ) and name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        unloaded_params = params_dict.keys() - loaded_params
        if unloaded_params:
            raise RuntimeError(
                f"Some weights are not initialized from checkpoints: {unloaded_params}"
            )
        self.post_load_weights()

    def post_load_weights(self):
        if _is_cpu and _is_cpu_amx_available:
            from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading

            layer_ids = int(self.config.num_hidden_layers)
            first_k_dense_replace_id = (
                self.config.first_k_dense_replace
                if hasattr(self.config, "first_k_dense_replace")
                else -1
            )
            moe_layer_freq_id = (
                self.config.moe_layer_freq
                if hasattr(self.config, "moe_layer_freq")
                else 1

View on GitHub (pinned to 0132848349)

Solutions

1. Verify the checkpoint matches this model class exactly (same repo/revision the model file was written for)
2. Re-download the checkpoint and confirm shard count/sizes match the index JSON
3. Update sglang to a version where deepseek_ocr.py's parameter names match your checkpoint (or vice versa)
4. If the gap is only tied/scalar meta weights, confirm your fork's load_weights handles them and the checkpoint actually contains them

expected = set(model.named_parameters().keys())
provided = set(safetensors keys via safetensors.safe_open per shard)
missing = expected - provided
if missing: raise SystemExit(f"checkpoint missing: {missing}")

try:
    model.load_weights(weights_iter)
except RuntimeError as e:
    if "not initialized from checkpoints" in str(e):
        log.error(e); sys.exit(2)  # wrong checkpoint — do not serve partially initialized weights
    raise

When it happens

Trigger: Loading a checkpoint whose safetensors/bin files lack tensors matching one or more parameter names in params_dict — e.g. a base-model checkpoint loaded into an OCR-adapted architecture, a quantized/trimmed checkpoint, or a weight-file list that skips a shard.

Common situations: Checkpoint and model code version mismatch (model defines new layers the old checkpoint predates), loading DeepSeek base weights into DeepseekOCR, partial or corrupted checkpoint downloads, or wrong --model-path pointing at an incompatible repo.

Related errors

• Duplicate tensor names detected across safetensors files. Re
• Incomplete Diffusers H3 fused parameters: {incomplete}
• qkv weight has incompatible output dim for grouped checkpoin
• The original encoder only has {num_hidden_layers} layers, bu
• RIFE weight file not found: {flownet_path} Expected layout:

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/f200a9913e6ebda7. Report an issue: GitHub.