All Products
Search

This Product

    MaxCompute:Use GU resources in MaxFrame

    Document Center

    MaxCompute:Use GU resources in MaxFrame

    Last Updated:Mar 26, 2026

    MaxFrame runs GPU-based tasks on MaxCompute through GU (GPU Unit) Quota—enabling distributed multi-modal data processing and model inference at scale with Python.

    Prerequisites

    Before you begin, make sure you have:

    • A MaxCompute project with a GU Quota assigned

    • MaxFrame installed and configured

    For information on purchasing GU resources, see MaxCompute AI compute resources.

    Set the engine priority

    The Distributed Python Engine (DPE) is the compute engine that handles GPU workloads in MaxFrame. GUs are suitable for scenarios such as deep learning inference, image and audio processing, and offline inference for Large Language Models (LLMs).

    To route GPU tasks to the DPE, place "DPE" first in engine_order:

    from maxframe.config import options

options.dag.settings = {
    "engine_order": ["DPE", "MCSQL"]
}

    Configure GU resources

    Two methods are available. Choose based on your use case:

    Method 1: Session-level quota Method 2: UDF-level parameters
    Scope Applies automatically to MaxFrame AI Functions Applies to a specific UDF
    UDF support Requires manual override for regular UDFs Fully supported
    Use when Running MaxFrame AI Functions Running custom UDFs that need GPU resources

    Method 1: Set a session-level GU Quota

    options.session.gu_quota_name = "xxxxxx"  # GU Quota name
    The session-level GU Quota takes effect automatically only for MaxFrame AI Functions. For regular UDFs, set GU parameters at the UDF level (Method 2).

    Method 2: Set GU parameters at the UDF level

    @with_running_options(engine="dpe", gu=8, gu_quota="xxxxx")
def vad_call(batch, config_path: str):
    ...
    Parameter Description
    engine="dpe" Routes the UDF to the DPE for GPU execution
    gu=8 Allocates 8 GUs (8 cards) to each instance
    gu_quota="os_xxxx" Specifies the GU Quota group to use

    Next steps