sig-framework

sig-framework: Framework Adapter Layer

The primary user-facing SIG. Most users enter FlagOS through PyTorch/vLLM/SGLang — the adapter experience determines whether users stay.

Scope

Responsible Modules:

• PyTorch-Plugin-FL — PyTorch adaptation
• vllm-plugin-FL — vLLM adaptation
• sglang-plugin-FL — SGLang adaptation
• TransformerEngine-FL — TransformerEngine adaptation
• Megatron-LM-FL — Megatron-LM adaptation
• verl-FL — veRL adaptation

In Scope:

• Framework adapter development and maintenance
• Upstream framework version tracking and compatibility testing
• Migration guides and compatibility matrix maintenance
• Framework-level performance optimization
• User troubleshooting support

Out of Scope:

• Operator implementation → sig-operator
• Distributed training orchestration → sig-training
• Communication → sig-network

Members

Chairs

Name	GitHub	Affiliation	Status
(TBD)	—	—	Pending TSC confirmation

Tech Leads

Name	GitHub	Affiliation	Status
(TBD)	—	—	Pending TSC confirmation

Approvers

See OWNERS

Communication

• Meetings: TBD (biweekly)
• WeChat Group: To be created
• Meeting Notes: meetings/

Decision Making

• Routine PR/FEP approval: Any Approver in OWNERS may approve
• Major SIG internal decisions: 2/3 majority vote
• Unresolved disputes escalate to TSC

Key Deliverables

• Compatibility Matrix: Compatibility table for each FlagOS version × each framework version
• Migration Guide: Minimal-change migration steps from native frameworks to FlagOS
• Troubleshooting FAQ: Common issues and solutions
• Example Directory: Standalone examples for each framework adapter

Subprojects

Subproject	Responsible Framework	Owner
pytorch-plugin	PyTorch	(TBD)
vllm-plugin	vLLM	(TBD)
sglang-plugin	SGLang	(TBD)
te-plugin	TransformerEngine	(TBD)
megatron-plugin	Megatron-LM	(TBD)
verl-plugin	veRL	(TBD)