[Feat] Fast-dLLM-v2 Decoding Strategy Support

[drewjin]

Motivation

Description

This Pull Request introduces support for the Fast-dLLM-v2 (fdv2) native inference strategy. Building upon the existing implementation of Block Diffusion (bd), we are now integrating the fdv2 strategy featuring a specialized Intra-block Dual-Cache mechanism.

The implementation involves a hierarchical state management system, which increases the complexity of the orchestration logic (refer to the attached diagram):

• Sub-block States: ACTIVE, TO_CACHE, and IN_CACHE.
• Parent-block States: CACHING and DECODING.

Key Scheduling Logic:
The state management significantly alters the scheduling behavior compared to traditional methods:

• In the CACHING state, the scheduler allocates tokens based on the full Parent-block size.
• In the DECODING state, the scheduler switches to allocating tokens at the Sub-block granularity.

TODO List

• Implement fdv2 strategy engine.
• Implement fdv2 attention metadata.
• Adapt fdv2 attention kernels for the new strategy.

[github-actions]

👋 Hi! Thank you for contributing to the Diffulex project.

Please remember to run pre-commit run --all-files in the root directory of the project to ensure your changes are properly linted and formatted. This will help ensure your contribution passes the format check.

We appreciate you taking this step! Our team will review your contribution, and we look forward to your awesome work! 🚀

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}