modelscope / modelscope/ms-swift

🍭[Roadmap] ms-swift3.6-3.8

Open
#4,561 48 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

discussion good first issue
Dominant language
Python
Stars
15.7k
Forks
1.7k
Avg merge
1d 16h
Merged PRs (30d)
136

Description

模型

  1. ✅最新模型接入 P0
  2. Omni pending
    a. 部署支持输出音频
    b. 支持talker的训练
  3. All-to-All优化 pending
  4. ✅embedding: 支持推理、部署
  5. ✅reranker训练支持
    a. 推理部署支持
  6. 序列分类: 多标签/回归支持量化

训练

  1. RAY支持 P0
  2. ✅长文本ring attention
  3. ✅AutoTP P0
  4. ✅channel loss支持packing/padding_free
  5. ✅多模态packing优化
  6. ✅new_special_tokens支持
  7. ✅多模态packing/padding_free支持更多模型
  8. ✅混合模态训练支持更多模态
  9. ✅flash-attention-3
  10. ✅DFT

Megatron-SWIFT

  1. 新模型支持
    a. ✅多模态: qwen2.5-VL/qwen2.5-Omni P0
    b. ✅DeepSeekV3
    c. Llama4
  2. ✅fp8
    a. blockwise fp8 P0
  3. ✅LoRA支持
    a. ✅MoE LoRA支持router训练
  4. ✅支持提前预处理数据集
  5. RLHF支持 P0
    a. GRPO
    b. KTO
    c. ✅DPO
  6. ✅bshd格式支持
  7. swanlab支持 P0
  8. ✅loss_scale支持
  9. 分类/Embedding模型支持
  10. Deepspeed集成
  11. ✅channel loss

RL

  1. GRPO
    a. ✅多轮AsyncEngine
    b. Agent MCP
    c. sglang
    d. 效率对比benchmark
    e. ✅多机rollout
    f. ✅GSPO
    g. ✅DeepEyes
  2. ✅MPO
  3. ✅GKD
  4. DPO
    a. ✅packing支持
    b. ✅LD-DPO
  5. RLOO P1
  6. Reinforce++ P06.多模态PPO
  7. KTO padding_free支持
  8. RM 支持 pointwise 训练 P0

推理与部署

  1. ✅sglang接入推理部署
    a. 多模态模型
  2. vLLM支持分类模型和RM

量化导出

  1. ✅fp8/bnb支持多模态模型
  2. ✅qlora支持merge-lora P1
  3. ✅fp8量化

ms-swift3.9-3.11 roadmap: https://github.com/modelscope/ms-swift/issues/5721

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

No source files, tests, or entry points are named. Read the pending items in this roadmap and the linked ms-swift 3.9–3.11 roadmap (issue 5721), then select one narrowly scoped task; done would require an agreed implementation and project validation for that task.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.