🍭[Roadmap] ms-swift3.9-3.12

Models

1. Integrate the latest models (P0)
2. Padding-free/packing multimodal support for more models
3. Hybrid data training support for more multimodal models
4. Agent templates support more models

Training

1. RAY supportP0
   a. DPO
   b. GRPO
   c. PPO
2. Support for embedding/reranker/sequence classification
   a. ✅padding_free/packing/seq_parallel
3. ✅Mixed use of ulysses & ring attention
4. ✅Embedding & reranker dataset format refactoring
5. ms-swift directory structure refactor; decoupling templates from model_type P0

Megatron-SWIFT

1. New model support P0
2. ✅blockwise FP8 P0
3. ✅MTP support P0
4. Multimodal support: ViT partial TP/CP performance optimization
5. ✅GRPO P0
6. RLHF support: ✅Reward Model；✅KTO；✅GKD；MPO
7. More training task types supported: ✅sequence classification; embedding; reranker
8. Deepspeed/fsdp2 integration
9. ✅Weight conversion optimization (mcore-bridge): multi-machine conversion; GRPO compatibility; remove explicit conversion step; support direct LoRA export P0
10. ✅Storage: save per epoch; save_total_limit; async saving support P0
11. ✅VPP and heterogeneous PP
12. Support for Megatron infer/deploy/eval modules

RL

1. GRPO
   a. ✅Weight sync optimization（LoRA/MoE）P0
   b. ✅Training memory optimization
   c. ✅CHORD
   d. ✅training-inference mismatch (TIS/MIS/FP32 lm_head ...)
   e. reward async/parallel execution
   f. fully async rollout pipeline
2. ✅RLOO P0
3. ✅Reinforce++ P0
4. ✅KTO padding_free & packing support
5. ✅liger_kernel memory optimization
6. ✅GKD support for vLLM / offload

Inference & Deployment

1. ✅Reranker inference & deployment support
2. SGLang support for multimodal models
3. vLLM/SGLang support for classification models / RM

Quantization

1. Quantization support for more model types: embedding/reranker/multi-classification, etc.
2. GPTQ/AWQ upgrades; ✅GPTQ-v2

ms-swift3.6-3.8 Roadmap: https://github.com/modelscope/ms-swift/issues/4561

The roadmap is organized into Models, Training, Megatron-SWIFT, RL, Inference & Deployment, and Quantization; begin by selecting one unchecked item and reading its corresponding subsystem. No file, test, or entry point is named, so define the affected code path and an acceptance test before starting; done is the selected roadmap item implemented and verified.