unslothai/unsloth · error · ValueError

{workflow} is not supported for the '{state.family.name}' mo

Error message

{workflow} is not supported for the '{state.family.name}' model family.

What it means

Raised by `_workflow_pipe` when an image-conditioned workflow (img2img, inpaint, etc.) is requested but the loaded model family does not declare a pipeline class for it (`class_name` is falsy). Each model family object carries a mapping of workflow names to diffusers pipeline class names; a family without the mapping cannot run that workflow. The error is intentionally a ValueError so API layers surface it as a client error.

Source

Thrown at studio/backend/core/inference/diffusion.py:4796

        return plan_diffusion_memory(
            target = target,
            device_memory = device_memory,
            model_dense_mib = model_dense_mib,
            companion_dense_mib = companion_mib,
            text_encoder_dense_mib = text_encoder_mib,
            runtime_headroom_mib = runtime_headroom,
            requested_mode = memory_mode,
            explicit_offload = cpu_offload,
        )

    def _workflow_pipe(self, state: _LoadState, class_name: Optional[str], workflow: str) -> Any:
        """The diffusers pipeline for an image-conditioned ``workflow``, built once and
        cached. ``Pipeline.from_pipe`` re-wires the loaded text-to-image pipe's resident
        modules (transformer/VAE/text-encoder, incl. any compiled/quantised state) into
        the workflow pipeline class, so there is no extra VRAM and no reload. Raises a
        clear ValueError when the family does not support the workflow."""
        if not class_name:
            raise ValueError(
                f"{workflow} is not supported for the '{state.family.name}' model family."
            )
        cached = self._aux_pipes.get(class_name)
        if cached is not None:
            return cached
        import diffusers

        # torch_dtype=None is load-bearing: from_pipe otherwise recasts EVERY component to fp32, which hard-crashes the
        # dense-quant path (torchao subclasses cannot swap_tensors). None reuses resident modules at their loaded dtype.
        pipe = getattr(diffusers, class_name).from_pipe(state.pipe, torch_dtype = None)
        # Publish to the shared aux cache only if THIS load is still current: from_pipe runs without _lock, so an unload can null _state and caching would hand out stale modules.
        with self._lock:
            if self._state is state:
                self._aux_pipes[class_name] = pipe
        return pipe

    def _controlnet_pipe(self, state: _LoadState, resolved_cn: Any, cancel: threading.Event) -> Any:
        """Build (once, cached) the family's diffusers ControlNet pipeline around the requested

View on GitHub (pinned to 203007d190)

Solutions

  1. Switch the workflow to one the family supports (typically txt2img) or reload a family that supports the requested workflow.
  2. If you maintain the family registry, add the missing workflow->pipeline-class entry (e.g. controlnet_pipeline_class / img2img class) for that family.
  3. In the UI, gate workflow options on the loaded family's declared capabilities instead of showing all.

Example fix

# before
diffusion.generate(prompt="...", workflow="inpaint", mask_image=mask)  # family has no inpaint class
# after
diffusion.generate(prompt="...", workflow="txt2img")  # or load a family that supports inpaint
Defensive patterns

Strategy: type-guard

Validate before calling

fam = diffusion.loaded_family()
if workflow not in fam.workflows:  # capability registry from the load response
    workflow = "txt2img"

Type guard

def supports_workflow(family, workflow: str) -> bool:
    """Family declares a diffusers pipeline class for this workflow."""
    return bool(family.workflow_classes.get(workflow))

Try / catch

try:
    diffusion.generate(prompt=p, workflow=wf)
except ValueError as e:
    if "is not supported for the" in str(e):
        diffusion.generate(prompt=p, workflow="txt2img")
    else:
        raise

Prevention

When it happens

Trigger: Calling generate() with a workflow like 'img2img' or 'inpaint' for a family whose workflow-class registry has no entry, e.g. an edit-only or t2i-only family. `_workflow_pipe(state, class_name, workflow)` is invoked after the class lookup fails.

Common situations: Frontends exposing all workflow buttons regardless of loaded model (offering inpaint on a model that only supports txt2img); loading a base checkpoint when the user selected an img2img workflow; new/lesser-known families added without complete workflow class mappings.

Related errors

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/0ebb17751882e8b8. Report an issue: GitHub.