cosmos-xenna

What it is and what it does

Cosmos-xenna is a distributed pipeline framework built on Ray that simplifies the development of AI inference workflows. You define a series of stages—each implementing setup() for one-time initialization and process_data() for batch processing—and connect them in a PipelineSpec. The framework handles resource allocation (CPUs, GPUs), autoscaling to balance throughput across stages, backpressure management to prevent memory overflow, and real-time monitoring. It supports three execution modes: streaming (recommended, all stages concurrent with dynamic worker balancing), batch (sequential, simpler but materializes intermediate data), and serving (online real-time input/output via queues). Most code is Python; autoscaling and artifact distribution logic is written in Rust for performance. The framework is designed for production workloads like video captioning, image processing, and other multi-model inference chains.

Use it for

• Build multi-stage video or image processing pipelines that download data, run VLM inference, and upload results at scale
• Orchestrate complex AI workflows combining multiple models (e.g., detection → captioning → embedding) with automatic GPU/CPU allocation
• Process large datasets through streaming pipelines with automatic worker balancing to maximize throughput without manual tuning
• Deploy online serving scenarios where inference requests arrive in real time and results are pushed to output queues
• Manage resource-constrained inference by defining per-stage batch sizes and resource requirements to prevent OOM

Worth the install?

AI-flagged interpretation of the facts on this page. Verify before relying on it.

Install

cosmos-xenna on PyPI

Before you install

Medium install friction due to 9 runtime dependencies including Ray, a distributed computing framework, plus compiled wheels for x86_64 and aarch64. Active maintenance as of August 2026, though the package description notes that Cosmos-Xenna is no longer under active development and users are encouraged to migrate to Cosmos 3.

Requires Python >= 3.12. Ray cluster must be running or accessible; cosmos-xenna orchestrates work across Ray actors.

License in practice

License treatment is unclear—no SPDX identifier or raw license text is available in the package metadata. Verify the actual license before use in proprietary or restricted-license projects.

Quickstart

pip install cosmos-xenna

import cosmos_xenna.pipelines.v1 as pipelines_v1

class MyStage(pipelines_v1.Stage):
    @property
    def stage_batch_size(self) -> int:
        return 10
    @property
    def required_resources(self) -> pipelines_v1.Resources:
        return pipelines_v1.Resources(gpus=0, cpus=1.0)
    def process_data(self, samples):
        return samples

pipeline_spec = pipelines_v1.PipelineSpec(
    input_data=samples,
    stages=[MyStage()]
)
pipelines_v1.run_pipeline(pipeline_spec)

Package facts

Evidence: cosmos_xenna-0.5.7-cp312-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl; cosmos_xenna-0.5.7-cp312-abi3-manylinux_2_28_aarch64.whl

Tags

More Distributed Computing packages

See also onnxruntime-openvino · xgboost-ray · nvidia-nvshmem-cu13 · agent-framework-azure-cosmos · azure-mgmt-cosmosdb · nvidia-nccl-cu13 · nvidia-nvshmem-cu12 · cosmpy · abstract-hugpy-dev