Fabric SDK

Official Python SDK for Fabric - Distributed AI Compute Network

Submit AI workloads to the Fabric network programmatically.

Installation

pip install fabric-compute-sdk

Quick Start

Option 1: Email/Password Authentication

from fabric_sdk import FabricClient

# Initialize client with email/password
client = FabricClient(
    api_url="https://api.fabric.carmel.so",
    email="your@email.com",
    password="your_password"
)

# Submit a job
job = client.submit_job(
    workload_type="llm_inference",
    params={
        "prompt": "Explain quantum computing in simple terms",
        "max_length": 200,
        "temperature": 0.7,
        "use_gpu": True
    },
    job_name="My LLM Inference Job"
)

print(f"Job submitted: {job['id']}")

# Wait for completion
result = client.wait_for_job(job['id'], timeout=300)
print(f"Job completed in {result['duration_seconds']}s")
print(f"Cost: ${result['actual_cost']}")

Option 2: API Key Authentication (Recommended for Google OAuth users)

from fabric_sdk import FabricClient

# Initialize client with API key
client = FabricClient(
    api_url="https://api.fabric.carmel.so",
    api_key="fb_live_..."  # Get this from dashboard Settings > API Keys
)

# Works exactly the same!
job = client.submit_job(
    workload_type="llm_inference",
    params={"prompt": "Explain quantum computing"},
    job_name="My Job"
)

Why use API keys?

• No password needed (great for Google/GitHub OAuth users)
• More secure for CI/CD pipelines
• Easy to rotate and revoke
• Each project can have its own key

Workflows - Zero-Config Job Submission

Instead of configuring dozens of parameters, use workflows - preset configurations for common use cases.

Bioinformatics Examples

# Sequence alignment - automatically sets all alignment parameters
job = client.submit_job(
    workload_type="ml_preprocess",
    params={
        "data_url": "https://your-bucket.s3.amazonaws.com/sequences.fasta",
        "workflow": "sequence_alignment"
    }
)

# Variant analysis - optimized for genomic variants
job = client.submit_job(
    workload_type="ml_preprocess",
    params={
        "data_url": "https://your-bucket.s3.amazonaws.com/variants.vcf",
        "workflow": "variant_analysis"
    }
)

# Protein property calculation
job = client.submit_job(
    workload_type="ml_preprocess",
    params={
        "data_url": "https://your-bucket.s3.amazonaws.com/proteins.fasta",
        "workflow": "protein_properties"
    }
)

ML Training Examples

# Quick classification - fast training with reasonable defaults
job = client.submit_job(
    workload_type="ml_train",
    params={
        "data_url": "https://your-bucket.s3.amazonaws.com/labeled_data.csv",
        "workflow": "quick_classification"
    }
)

# High accuracy - slower but more thorough hyperparameter search
job = client.submit_job(
    workload_type="ml_train",
    params={
        "data_url": "https://your-bucket.s3.amazonaws.com/labeled_data.csv",
        "workflow": "high_accuracy"
    }
)

# Survival analysis - for time-to-event data
job = client.submit_job(
    workload_type="ml_train",
    params={
        "data_url": "https://your-bucket.s3.amazonaws.com/survival_data.csv",
        "workflow": "survival_analysis"
    }
)

Image Processing Examples

# Cell segmentation - optimized for microscopy images
job = client.submit_job(
    workload_type="image_preprocess",
    params={
        "image_url": "https://your-bucket.s3.amazonaws.com/cells.tiff",
        "workflow": "cell_segmentation_prep"
    }
)

# ImageNet preparation - resize and normalize for deep learning
job = client.submit_job(
    workload_type="image_preprocess",
    params={
        "image_url": "https://your-bucket.s3.amazonaws.com/image.jpg",
        "workflow": "imagenet_prep"
    }
)

# Fluorescence quantification - for DAPI, GFP, etc.
job = client.submit_job(
    workload_type="image_preprocess",
    params={
        "image_url": "https://your-bucket.s3.amazonaws.com/fluorescence.tiff",
        "workflow": "fluorescence_quantification"
    }
)

Image Classification Examples

# Zero-shot classification - classify without training
job = client.submit_job(
    workload_type="image_label",
    params={
        "image_url": "https://your-bucket.s3.amazonaws.com/image.jpg",
        "workflow": "zero_shot_classification",
        "candidate_labels": ["cat", "dog", "bird", "fish"]
    }
)

# Object detection - find and locate objects
job = client.submit_job(
    workload_type="image_label",
    params={
        "image_url": "https://your-bucket.s3.amazonaws.com/image.jpg",
        "workflow": "object_detection"
    }
)

Override Workflow Defaults

Workflows set sensible defaults, but you can override any parameter:

job = client.submit_job(
    workload_type="ml_train",
    params={
        "data_url": "https://...",
        "workflow": "quick_classification",
        # Override specific settings
        "max_iterations": 200,      # More iterations
        "test_size": 0.3,           # Larger test set
        "use_gpu": True             # Force GPU
    }
)

Available Workflow Categories

ML Preprocessing (25+ workflows)

• Bioinformatics: sequence_alignment, variant_analysis, protein_properties, phylogenetic_prep
• Sports/Trajectory: trajectory_analysis, play_classification
• General: basic_cleaning, anomaly_detection_prep, clustering_prep

ML Labeling (12+ workflows)

• Clustering: quick_clustering, auto_clustering, hierarchical_clustering, density_clustering
• Segmentation: customer_segmentation, fine_segmentation, binary_threshold

ML Training (20+ workflows)

• Classification: quick_classification, balanced_classification, high_accuracy, interpretable
• Regression: quick_regression, ridge_regression, lasso_regression
• Neural Nets: neural_net_starter, neural_net_deep
• Specialized: survival_analysis, differential_expression, biomarker_discovery

ML Inference (10+ workflows)

• classification_inference, regression_inference, batch_inference, anomaly_scoring

Image Preprocessing (15+ workflows)

• Standard: imagenet_prep, object_detection_prep, high_resolution_prep
• Microscopy: fluorescence_quantification, cell_segmentation_prep, dapi_analysis

Image Labeling (13+ workflows)

• zero_shot_classification, object_detection, visual_clustering, feature_extraction

Image Training (12+ workflows)

• quick_train, balanced_train, resnet_classifier, vit_classifier, cell_image_classifier

Image Inference (12+ workflows)

• image_classification, imagenet_resnet50, cell_classification, batch_inference

Features

• Dual Authentication - Email/password OR API keys
• 100+ Workflows - Preset configurations for common use cases
• API Key Management - Create, list, and revoke keys programmatically
• Job Submission - Submit 30+ production workload types
• Batch Submission - Submit 1000s of jobs in seconds (100x faster!)
• Custom Workloads - Upload and run your own Python code
• Parallel Processing - Distribute work across multiple nodes
• Job Monitoring - Track progress and get results
• Credit Management - Check balance and purchase credits
• Node Discovery - List available compute nodes
• Auto-Retry - Built-in network resilience
• Type Hints - Full TypeScript-style typing support

Supported Workload Types (30+)

ML Pipeline (8)

• ml_preprocess, ml_label, ml_train, ml_inference
• image_preprocess, image_label, image_train, image_inference

Compute & Simulation (5)

• cpu_compute_benchmark, gpu_compute_benchmark
• eigenvalue_decomposition, financial_forecast_simulation, agent_simulation

Data Processing (5)

• data_cleaning, feature_extraction, csv_vectorization
• data_augmentation, outlier_detection

AI Inference (7)

• llm_inference, llm_inference_batch, image_classification
• embedding_generation, sentiment_analysis
• text_summarization, question_answering

Media Processing (5)

• video_transcode, audio_to_text, video_object_detection
• image_resize_batch, video_summarization

Web & HTTP (2)

• http_microtask, custom_python

Custom Workloads

Upload and run your own Python code:

# Upload custom workload
workload = client.upload_custom_workload(
    name="my-data-processor",
    description="Process proprietary data format",
    zip_path="./my_workload.zip"
)

# Run it
job = client.submit_job(
    workload_type="custom_python",
    params={
        "custom_workload_id": workload["id"],
        "data_url": "https://..."
    }
)

Parallel Processing

Distribute work across multiple nodes:

workload = client.upload_custom_workload(
    name="batch-processor",
    zip_path="./batch_workload.zip",
    parallel_mode=True,
    files_per_shard=100  # 100 files per node
)

# Submit with 10,000 input files - runs on ~100 nodes in parallel
job = client.submit_job(
    workload_type="custom_python",
    params={"custom_workload_id": workload["id"]}
)

Documentation

Complete documentation, guides, and examples are available in the Fabric Dashboard.

Key Resources

• Workflows Reference - All 100+ preset workflows
• Custom Workloads - Build and deploy your own code
• ML Pipeline Guide - End-to-end machine learning
• API Reference - Complete endpoint documentation
• Examples - Ready-to-run code samples

Support

Access the Fabric Dashboard for documentation, support, and account management.

License

Apache License 2.0

Copyright 2025 Carmel Labs, Inc.