Skip to content

Derive the GMS manifest from modular block declarations #15351

Open

Description

@yiyixuxu

#15017 adds diffusers serve, which describes a model with a manifest: its tasks, inputs, parameters and outputs.

Right now modular pipelines go through the same path as standard pipelines: the blocks are converted into a __call__-like signature, and the manifest is guessed from argument names and type hints. That makes sense for standard pipelines, where the signature is all we have. But modular blocks are declarative, and we shouldn't need to guess for our own pipelines. I think the blocks can state most of this directly. A few ideas to start with:

1. InputParam.metadata
InputParam already has a metadata dict (Mellon uses it). It could carry what the manifest needs:

  • the public name (num_inference_steps → steps)
  • the kind: text, image, video, audio, mask, or a plain setting
  • whether it's user-facing at all (latents, timesteps are not)

2. Tasks from workflows
Tasks are guessed from whether there's an image / video input, so a model gets at most two. Blocks already declare their workflows and trigger inputs in _workflow_map. One task per workflow would also cover workflows triggered by other inputs, e.g. references in MiniMax-H3.

3. Require type_hint on user-facing inputs
A parameter's schema comes from its type hint (int → integer, list[float] → array of numbers). An input without one can't be described and is dropped today. We could require a type_hint on every InputParam that's exposed, and check it in the tests.

4. Reuse it in save_pretrained()
If the manifest comes from the blocks alone, the same function can run in save_pretrained() and write a gms.json next to modular_model_index.json. serve already picks up a gms.json from the model repo, so new models would ship with their manifest.

What the code could look like
A made-up sketch to show the idea, not tested. The "gms" key and its fields are placeholders.

# in a block
InputParam("num_inference_steps", type_hint=int, default=28, metadata={"gms": {"name": "steps"}})
InputParam("image", type_hint=PIL.Image.Image, metadata={"gms": {"type": "image"}})
InputParam("latents", type_hint=torch.Tensor)  # no metadata, not exposed
def derive_manifest(model_id, blocks):
    tasks = {}
    for workflow, triggers in blocks._workflow_map.items():
        # the blocks that run for this workflow, so only its own inputs are listed
        workflow_blocks = blocks.get_workflow(workflow)

        inputs, parameters = {}, {}
        for param in workflow_blocks.inputs:
            gms = (param.metadata or {}).get("gms")
            if gms is None:  # not user-facing
                continue
            name = gms.get("name", param.name)

            if gms.get("type") in ("text", "image", "video", "audio", "mask"):
                inputs[name] = {"type": gms["type"], "required": param.required or param.name in triggers}
            else:
                schema = type_hint_to_schema(param.type_hint)  # int -> {"type": "integer"}
                if param.default is not None:
                    schema["default"] = param.default
                parameters[name] = schema

        tasks[workflow] = {"inputs": inputs, "parameters": parameters}

    return {"id": model_id, "tasks": tasks}

No rush since the PR is still WIP, opening this to collect ideas.

Posted by Claude on behalf of @yiyixuxu. The text was drafted together and approved by yiyixuxu.

Activity

  1. yiyixuxu commented on Nov 19, 2026

    @yiyixuxu
    CollaboratorAuthor

    cc @DN6

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions