Extending the Engine

Start here when adding model support, decoding behavior, or kernel implementations. Keep each extension focused: add the component required for the new behavior, then verify it with a targeted test or smoke run before expanding the scope.

If the goal is a new dLLM inference algorithm, read Research Engine first. It maps block-level algorithms to Diffulex’s Block Buffer backend and lists the files a code agent should cover.

Add New Models

See Add New Models for the model, sampler, config, CLI, and verification workflow.

Add New Decoding Strategies

See Add a Decoding Strategy for the detailed strategy component walkthrough.

• Add New Models
• Add a Decoding Strategy
• Add New Kernel
• Research Engine

Strategy work usually touches these components:

• a request class for per-request state;

• a scheduler for lifecycle policy;

• a KV cache manager for page allocation and append rules;

• attention metadata for kernel inputs;

• a model runner for tensor preparation and execution.

Use a strategy template when possible. Override only the methods whose behavior is different.

Add New Kernel

See Add New Kernel for kernel placement, reference testing, integration, and profiling guidance.

Verification Checklist

For model and strategy extensions:

• import the package that should trigger registration;

• inspect available registry keys;

• construct a Config;

• run a tiny offline generation;

• run a limited benchmark or server smoke test;

• add focused regression tests for the behavior you changed.