sgl-project/sglang · error · AttributeError

lm_head is not available in encoder-only mode

Error message

lm_head is not available in encoder-only mode

What it means

The lm_head property raises AttributeError when the Kimi-K3 model is in encoder-only mode, because there is no language model and therefore no output head. Code that reads lm_head for logits-related metadata, tied-weight detection, or sampling setup will fail on an encoder-only instance.

Source

Thrown at python/sglang/srt/models/kimi_k3.py:3319

    def precompile_kernels_after_loading(self) -> None:
        if self.config.language_only:
            return
        if self.vision_tower.precompile_fused_rope():
            logger.info("Precompiled dynamic-token fused K3 vision RoPE kernel")
        if self.vision_tower.precompile_attention_backend():
            logger.info("Precompiled Kimi-K3 vision FA4 kernel")

    def get_input_embeddings(self):
        if self.language_model is None:
            raise AttributeError(
                "get_input_embeddings() is not available in encoder-only mode"
            )
        return self.language_model.model.embed_tokens

    @property
    def lm_head(self):
        if self.language_model is None:
            raise AttributeError("lm_head is not available in encoder-only mode")
        return self.language_model.lm_head

    def set_dspark_layers_to_capture(self, layer_ids: list[int]) -> None:
        if self.language_model is None:
            raise AttributeError(
                "DSPARK layer capture is not available in encoder-only mode"
            )
        self.language_model.set_dspark_layers_to_capture(layer_ids)

    def preprocess_mm_for_encoder(
        self,
        mm_data,
        modality,
        config,
        *,
        image_processor=None,
        use_gpu_preprocessing=False,
    ):

View on GitHub (pinned to 0132848349)

Solutions

1. Guard access with model.language_model is not None before reading lm_head
2. Mark encoder-only instances and branch generic logit/weight logic away from them
3. Confirm the deployment actually intended encoder-only mode

Example fix

// before
head = model.lm_head

// after
head = model.lm_head if model.language_model is not None else None

assert model.language_model is not None, "encoder-only instance has no lm_head"

def has_lm_head(m) -> bool:
    return getattr(m, "language_model", None) is not None

try:
    head = model.lm_head
except AttributeError:
    head = None

When it happens

Trigger: Accessing model.lm_head on an encoder-only Kimi-K3 build (language_model is None), e.g. during weight loading, tied-weight checks, or logit setup in generic paths.

Common situations: EPD encoder deployment where only the vision tower is loaded; shared utilities assuming every model exposes lm_head.

Related errors

• get_input_embeddings() is not available in encoder-only mode
• DSPARK layer capture is not available in encoder-only mode
• module {__name__!r} has no attribute {name!r}
• attn_res: nvb must be in [1, {_MAX_BANK_ROWS}], got {nvb}
• Unsupported text encoder output: expected `hidden_states`.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/9fbd9707e8c5561c. Report an issue: GitHub.