sgl-project/sglang · error · ValueError

Invalid target(s): {invalid_targets}. Valid targets: {self.V

Error message

Invalid target(s): {invalid_targets}. Valid targets: {self.VALID_TARGETS}

What it means

set_lora validates each entry of targets against the pipeline class's VALID_TARGETS class attribute; unknown target module names are rejected before any weights are touched (deliberately before offload is disabled, to avoid OOM on constrained deployments).

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/lora/pipeline.py:966

        Supports both single LoRA (backward compatible) and multiple LoRA adapters.

        cache_merged re-homes merged weights to a file-backed store so they
        stop costing anonymous host memory; pass it only for the startup
        (static) adapter, where the merged combination is stable.
        """
        merge_mode = self._resolve_lora_merge_mode(merge_weights, merge_mode)

        # Normalize inputs to lists for multi-LoRA support
        lora_nicknames, lora_paths, strengths, targets, lora_alphas = (
            self._normalize_lora_params(
                lora_nickname, lora_path, strength, target, lora_alpha
            )
        )

        # Validate targets
        invalid_targets = [t for t in targets if t not in self.VALID_TARGETS]
        if invalid_targets:
            raise ValueError(
                f"Invalid target(s): {invalid_targets}. Valid targets: {self.VALID_TARGETS}"
            )

        # Checked before disabling offload, which materializes every layer: on a
        # memory-constrained deployment that would OOM instead of returning the
        # unsupported-LoRA error. Offloaded placeholders still carry the name.
        self._reject_lora_on_packed_weights()

        # Disable layerwise offload before convert_to_lora_layers to ensure weights are accessible
        # This is critical because convert_to_lora_layers needs to save cpu_weight from actual weights,
        # not from offloaded placeholder tensors
        if not self.lora_initialized:
            with self._temporarily_disable_offload(
                target="all", use_module_names_only=True
            ):
                self.convert_to_lora_layers()

        # Check adapter presence and load missing adapters

View on GitHub (pinned to 0132848349)

Solutions

1. Print pipeline.VALID_TARGETS and use one of those names
2. Check the pipeline subclass matching your model — targets differ per architecture
3. Fix typos like 'attention' vs 'attn'

Example fix

# before
pipeline.set_lora(..., targets=['q_proj'])
# after
pipeline.set_lora(..., targets=['attn'])

invalid = [t for t in targets if t not in pipeline.VALID_TARGETS]
assert not invalid, f'invalid targets {invalid}; valid: {pipeline.VALID_TARGETS}'

def valid_targets(pipeline, targets: list[str]) -> bool:
    return all(t in pipeline.VALID_TARGETS for t in targets)

When it happens

Trigger: Calling set_lora(targets=['attn']) or any module name not in e.g. {'attn', 'fc1', 'fc2', ...} for your pipeline class.

Common situations: Using target names from a different model family (LLM PEFT targets like 'q_proj' on a DiT pipeline); typo; new model version renamed layers without updating VALID_TARGETS.

Understand the failure class

Background: "Invalid value" and "allowed values are" config errors: what your library rejected and how to fix it — this error's family across 41 libraries.

Related errors

• Invalid LoRA merge mode: {merge_mode}. Valid modes: {LORA_ME
• lora_alpha must be a positive integer
• bad compress_ratio {compress_ratio}
• The requested FlashAttention forward configuration exceeds S
• flashinfer_sparse_mla supports only GLM DSA with FP8 KV cach

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d3505a154afecae6. Report an issue: GitHub.