Frequently Asked Questions¶

Common questions about FRAMEWORM.

General¶

What is FRAMEWORM?¶

FRAMEWORM is a production-ready deep learning framework focused on generative models (VAE, GAN, Diffusion). It provides: - Built-in models - Complete MLOps stack - Production deployment tools - Extensive monitoring and tracking

Who is FRAMEWORM for?¶

Researchers building generative models
ML engineers deploying models to production
Teams needing reproducible experiments
Anyone wanting batteries-included generative AI

How does it compare to PyTorch Lightning?¶

FRAMEWORM is specialized for generative models with built-in MLOps.
Lightning is general-purpose and requires more setup for production.

See detailed comparison.

Is FRAMEWORM production-ready?¶

Yes! Includes: - Health checks & graceful shutdown - Rate limiting & authentication - Model versioning & A/B testing - Kubernetes deployment - Monitoring with Prometheus

Installation¶

What Python versions are supported?¶

Python 3.8, 3.9, 3.10, 3.11

Do I need a GPU?¶

No, but recommended for training. All features work on CPU.

How do I install with dependencies?¶

pip install frameworm[all]  # Everything
pip install frameworm[wandb]  # Just W&B
pip install frameworm[deployment]  # Just deployment tools

Training¶

How do I resume training from checkpoint?¶

checkpoint = torch.load('checkpoint.pt')
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])

trainer = Trainer(model, optimizer)
trainer.train(train_loader, epochs=50, start_epoch=checkpoint['epoch'])

Can I use custom optimizers?¶

Yes, pass any PyTorch optimizer:

from torch_optimizer import Ranger

optimizer = Ranger(model.parameters())
trainer = Trainer(model, optimizer)

How do I track experiments?¶

from frameworm.experiment import Experiment

with Experiment(name='my-exp') as exp:
    trainer.set_experiment(exp)
    trainer.train(...)

Automatically tracks: - Config - Metrics - Git commit - System info

How do I save/load models?¶

# Save
torch.save(model.state_dict(), 'model.pt')

# Load
model = get_model('vae')(config)
model.load_state_dict(torch.load('model.pt'))

Distributed Training¶

How do I use multiple GPUs?¶

Single machine:

model = torch.nn.DataParallel(model)

Multiple machines:

model = torch.nn.parallel.DistributedDataParallel(model)

See distributed training guide.

Does mixed precision work with distributed?¶

Yes! FRAMEWORM supports both together:

from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler()
# Use in training loop

Deployment¶

How do I deploy to production?¶

# 1. Export
frameworm export model.pt --format onnx

# 2. Serve
frameworm serve model.pt --port 8000

# 3. Deploy to K8s
kubectl apply -f k8s/deployment.yaml

See production deployment guide.

Can I use Docker?¶

Yes, example Dockerfile:

FROM python:3.10-slim
RUN pip install frameworm
COPY model.pt /app/
CMD ["frameworm", "serve", "/app/model.pt"]

How do I monitor in production?¶

Built-in Prometheus metrics:

from frameworm.monitoring import MetricsExporter

exporter = MetricsExporter(port=9090)
exporter.start()
# Metrics at http://localhost:9090/metrics

Connect to Grafana for visualization.

Customization¶

How do I create a custom model?¶

import torch.nn as nn
from frameworm.core import register_model

@register_model('my_model')
class MyModel(nn.Module):
    def __init__(self, config):
        super().__init__()
        # Your architecture

    def forward(self, x):
        return x

    def compute_loss(self, x, y=None):
        # Must return dict with 'loss' key
        return {'loss': ...}

Can I add custom metrics?¶

from frameworm.metrics import register_metric

@register_metric('my_score')
class MyMetric:
    def __call__(self, real, fake):
        # Compute metric
        return score

How do I create plugins?¶

frameworm plugins create my-plugin
cd frameworm_plugins/my-plugin
# Edit __init__.py and plugin.yaml
frameworm plugins load my-plugin

Troubleshooting¶

Training is slow¶

Check data loading:

   loader = DataLoader(dataset, num_workers=8, pin_memory=True)

Enable mixed precision:
```
   trainer.enable_mixed_precision()
```

Use gradient accumulation:

   trainer.enable_gradient_accumulation(steps=4)

Out of memory errors¶

Reduce batch size

Enable gradient checkpointing:

   trainer.enable_gradient_checkpointing()

Use gradient accumulation to simulate larger batches

Validation loss not decreasing¶

Check learning rate (try 10x smaller/larger)
Add learning rate scheduler
Check data normalization
Visualize samples to ensure data is correct

Import errors¶

Make sure dependencies are installed:

pip install frameworm[all]

Or install specific integration:

pip install wandb  # For W&B
pip install mlflow  # For MLflow

Performance¶

How fast is FRAMEWORM?¶

Comparable to PyTorch Lightning (~5% faster due to optimized data pipeline).

See benchmarks.

Can I use FP16/BF16?¶

Yes:

from torch.cuda.amp import autocast

with autocast():
    output = model(x)

Or use trainer helper:

trainer.enable_mixed_precision()

How do I profile my code?¶

from frameworm.utils.profiler import TrainingProfiler

profiler = TrainingProfiler()
with profiler:
    trainer.train(...)

profiler.print_report()

Integration¶

Does FRAMEWORM work with W&B?¶

Yes:

from frameworm.integrations import WandBIntegration

trainer.add_callback(WandBIntegration(project='my-project'))

Can I use my existing Lightning code?¶

Partial compatibility. See migration guide.

Does it support TensorBoard?¶

Yes, via callbacks:

from torch.utils.tensorboard import SummaryWriter

writer = SummaryWriter()
# Use in custom callback

Contributing¶

How do I contribute?¶

Fork the repo
Create feature branch
Make changes + tests
Submit PR

See CONTRIBUTING.md.

I found a bug¶

Please open an issue with: - Python/PyTorch versions - Minimal reproduction code - Error message

Can I request features?¶

Yes! Create a feature request.

License & Commercial Use¶

What license is FRAMEWORM?¶

MIT License - free for commercial use.

Can I use FRAMEWORM in my company?¶

Yes! No restrictions.

Do I need to cite FRAMEWORM?¶

Not required, but appreciated:

@software{frameworm2026,
  title={FRAMEWORM: Production-Ready Generative AI Framework},
  author={Your Name},
  year={2026},
  url={https://github.com/Aakash0440/frameworm}
}

Still have questions?¶

Discord: https://discord.gg/frameworm
GitHub Discussions: https://github.com/Aakash0440/frameworm/discussions
Email: Aakashali0440@gmail.com