feat: Initial block diffusion strategy support

[Copilot]

Motivation

This PR introduces a flexible framework to support multiple inference paradigms for Diffusion Language Models (dLLMs). Specifically, it implements a Strategy Registration Mechanism, which decouples the decoding logic from the model architecture. This design allows researchers and developers to rapidly prototype, register, and experiment with novel inference algorithms without modifying the core system.

Implementation Status

Features included in this PR:

• Strategy Registration Mechanism: A registry system to manage different decoding strategies dynamically.
• Abstract Strategy Interface: Defined base classes for implementing custom diffusion policies.
• Initial Block Diffusion Strategy: Implementation of the standard block-wise diffusion logic using the new registry.
• Model Integration: Basic integration for Fast-dLLM-v2, SDAR, LLaDA-2.0-mini/flash with the new strategy framework.
• Configuration Support: Ability to select strategies via configuration/arguments.

Features to be implemented in future PRs:

• Advanced Batching support (leveraging SGLang's core scheduling with new strategies)
• CUDA Graph optimization for dynamic strategies
• Support for Prefix Cache with diffusion steps
• Support for overlap scheduling
• More advanced decoding policies (e.g., specific low-confidence policies)

Our From-Scratch Diffusion Language Models

• Fast-dLLM-v2: https://huggingface.co/collections/Efficient-Large-Model/fast-dllm
• SDAR: https://huggingface.co/collections/JetLM/sdar
• LLaDA2.0-flash: https://huggingface.co/inclusionAI/LLaDA2.0-flash
• LLaDA2.0-mini: https://huggingface.co/inclusionAI/LLaDA2.0-mini

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.