Creating a Stack

When I have defined my scenario requirements, I want to configure a complete infrastructure stack, so I can get accurate cost estimates and ensure my setup meets my performance needs.

The Challenge

You’ve defined a scenario with a 500ms P95 latency requirement and $5,000/month budget. Now you need to assemble a stack to meet those constraints. But which model hits that latency target? Which serving framework handles your throughput needs? Should you use H100s or is A100 sufficient?

For platform leads and research engineers, stack configuration is where abstract requirements meet concrete infrastructure. The problem isn’t lack of options—it’s too many options. Every choice (model, provider, framework, hardware) interacts with every other choice. Without a structured way to capture these decisions, configuration sprawl takes over.

How Lattice Helps

Stacks are Lattice’s answer to configuration sprawl. A stack captures your complete infrastructure configuration: which model you’re using, how it’s configured, what framework runs it, and what hardware powers it. Once defined, your stack becomes context for everything else—cost calculations, performance estimates, and agent recommendations all reference your stack configuration.

Step 1: Open the Stack Configuration Form

In the Studio panel (right side), find the Stacks section. Click the + button in the section header to create a new stack. The New Stack modal opens with a completeness indicator showing your progress.

The form is organized into collapsible sections:

• Stack Identity: Name and description
• Model: Provider, model selection, temperature, max tokens
• Hardware: Cloud provider, instance family, GPU type, region

Start by giving your stack a descriptive name like “My Production Stack” and optionally set it as your default stack for new conversations.

Step 2: Configure Model Settings

Select your model provider from the available options:

Provider	Use Case
Anthropic	Claude models for reasoning and analysis
OpenAI	GPT models for general tasks
Google AI	Gemini models with large context
AWS Bedrock	Enterprise managed access
Azure OpenAI	Enterprise Azure integration
Mistral	Open-weight models
Cohere	Enterprise embeddings and RAG
Meta	Llama models self-hosted

After selecting a provider (e.g., Anthropic), choose your model from the dropdown. The example shows Claude Haiku 4.5 selected.

Configure generation parameters:

• Temperature: 0.7 is a good default for balanced creativity
• Max Tokens: 4096 for typical responses, higher for long-form content

Step 3: Configure Framework and Hardware

Scroll down to configure your observability and infrastructure settings.

Observability Stack:

Option	Features
LangSmith	LangChain-native tracing, debugging
Helicone	Logging, caching, rate limiting
Weights & Biases	Experiment tracking, model monitoring
Console	Basic console logging

Enable Tracing and Metrics toggles based on your monitoring needs.

Hardware Configuration:

For self-hosted or cloud deployments, specify:

• Cloud Provider: AWS, GCP, Azure, Lambda Labs, RunPod, or On-Premise
• Instance Family: CPU General Purpose, GPU Optimized, etc.
• GPU Type: Select from H100, A100, L40S, RTX 4090, or No Accelerator
• Region: Deployment region (e.g., us-east-1) affects latency and compliance
• Use Spot Instances: Enable for 60-70% cost savings with interruption risk

Step 4: Save and Apply

Click Create Stack to save your configuration. The stack immediately appears in the Studio panel’s Stacks section and becomes available in the Context Bar dropdown for quick selection during chat.

Managing Stacks:

From the Studio panel, you can:

• Click a stack card to view full configuration details
• Set as default: The default stack is automatically selected for new conversations
• Edit, Rename, or Delete: Hover over stack cards to reveal action buttons

What You’ve Accomplished

You now have a complete infrastructure stack that:

• Captures your model, framework, and hardware choices in one place
• Provides context for accurate cost calculations
• Links to scenarios for compatibility checking
• Serves as the foundation for all Research Agent recommendations

What’s Next

With your stack configured, explore these related capabilities:

• Link to Scenario: Connect your stack to a scenario for compatibility validation
• Cost Estimation: Get accurate TCO projections based on your stack
• Stack Comparison: Create multiple stacks to compare configurations side-by-side

---

Stack Configuration is available in Lattice. Define your infrastructure configuration once and reference it everywhere.