For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Python class

ModelManifest

ModelManifest

class max.pipelines.lib.ModelManifest(*args, metadata=None, **kwargs)

source

Bases: dict[str, MAXModelConfig]

Registry mapping semantic role strings to MAXModelConfig instances.

Each model is identified by a role string (e.g. "main", "draft", "vae", "unet"). Single-model pipelines use the "main" key by convention; multi-component pipelines (diffusion, speculative decoding) store models under their respective roles.

ModelManifest is a dict[str, MAXModelConfig] subclass, so standard dict operations ([], in, len, items, etc.) work directly.

For diffusion pipelines constructed from model_index.json, the metadata property exposes non-component entries (e.g. _class_name, _diffusers_version, is_distilled) as a plain dict.

Parameters:

• args (Any)
• metadata (dict[str, Any] | None)
• kwargs (Any)

clear()

clear()

source

Remove all model configs from the manifest.

Return type:

None

from_model_path()

classmethod from_model_path(model_path, revision=None, **kwargs)

source

Create a registry from a single model path.

Inspects model_path for a model_index.json before constructing any MAXModelConfig.

If the model is a diffusion pipeline (has a model_index.json), the registry is automatically expanded into per-component MAXModelConfig instances. Extra kwargs are forwarded to each component’s MAXModelConfig.

For single-model repos, a MAXModelConfig is constructed from model_path and any extra kwargs, then stored under the "main" key.

Parameters:

• model_path (str) – HuggingFace repo ID or local path to the model.
• revision (str | None) – Optional HuggingFace repo revision (branch, tag, or commit hash). Defaults to the HuggingFace Hub default.
• **kwargs (Any) – Additional keyword arguments forwarded to MAXModelConfig (only valid for single-model repos).

Returns:

A new ModelManifest. For transformers-style models this has a single "main" entry; for diffusion models it contains one entry per component.

Return type:

ModelManifest

loader()

loader()

source

Returns a WeightLoader over the role-prefixed union.

Public entry point for multi-component pipelines (diffusion, speculative decoding). Each role’s loader is exposed under its dotted role prefix, so a query like "transformer.blocks.0.attn.qkv_proj.weight" routes to the transformer config’s loader with "blocks.0.attn.qkv_proj.weight".

Single-model pipelines should call manifest["main"].loader() instead – the role prefix is not useful when there’s only one source and the Module tree’s parameter names don’t carry it.

Resolution is lazy: per-role loaders defer to the underlying Weights source, so weight bytes only page in when the Module’s adapter chain actually queries them.

Returns:

A WeightLoader resolving f"{role}.{name}" keys against the per-role source loaders.

Return type:

WeightLoader

log_model_info()

log_model_info()

source

Logs model configuration information for every model in the manifest.

Iterates over each role and delegates to MAXModelConfig.log_model_info() for per-model details.

Return type:

None

main_architecture_name

property main_architecture_name: str

source

Returns the main architecture class name.

For non-diffusion models (those with a "main" key), delegates to MAXModelConfig.architecture_name which returns architectures[0] from the HuggingFace config.

For diffusion pipelines (no "main" key), returns metadata["_class_name"] (e.g. "FluxPipeline").

Raises:

ValueError – If the architecture name cannot be determined.

metadata

property metadata: dict[str, Any]

source

Non-component entries from model_index.json.

For diffusion pipelines built via from_model_path, this contains every key/value pair from model_index.json that is not a component (e.g. _class_name, _diffusers_version, is_distilled). For non-diffusion manifests, returns an empty dict.

model_name

property model_name: str

source

Returns the served model name for metrics and API responses.

For single-model pipelines ("main" key), delegates to MAXModelConfig.model_name. For diffusion pipelines, returns model_path from the first component config.

pop()

pop(*args)

source

Remove and return a model config by key.

Parameters:

args (Any)

Return type:

Any

resolve()

resolve()

source

Validates and resolves every config in the manifest.

Delegates to MAXModelConfig.resolve() for each component.

Return type:

None

total_weights_size

property total_weights_size: int

source

Total weight size in bytes across all components.

Walks every MAXModelConfig in the manifest and sums weights_size(). Components with no weight files (e.g. schedulers) contribute zero.

Raises:

RuntimeError – If the manifest has not been resolved via resolve() first.

update()

update(*args, **kwargs)

source

Update the manifest with new model configs.

Parameters:

• args (Any)
• kwargs (Any)

Return type:

None

with_override()

with_override(role, config=None, **field_overrides)

source

Return a new manifest with the given role updated.

Three usage patterns:

1. Partial field update on an existing component:

   manifest.with_override("transformer",
       weight_path=[Path("w.safetensors")],
       quantization_encoding="float4_e2m1fnx2",
   )

2. Full replacement or addition of a component:

   manifest.with_override("draft",
       config=MAXModelConfig(model_path="org/draft"),
   )

3. Add/replace with additional field tweaks:

   manifest.with_override("draft",
       config=base_cfg,
       quantization_encoding="q4_0",
   )

Parameters:

• role (str) – The semantic role string identifying the component.
• config (MAXModelConfig | None) – A complete MAXModelConfig to use as the base. When None, the existing config for role is used (the role must already exist).
• **field_overrides (Any) – Individual field values to set on the config via model_copy(update=...).

Returns:

A new ModelManifest — the original is not modified.

Raises:

ValueError – If config is None and role does not exist, or if neither config nor field_overrides are provided.

Return type:

ModelManifest

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit