GitHub Issues 工作流

## Background / 背景
Why this needs to be done. 1-2 sentences.
为什么要做这件事，1-2 句话说清楚。

## Objective / 目标
What specific outcome is expected.
具体要达成什么效果。

## Tasks / 具体任务
- [ ] Sub-task 1 (may reference other issues: #12)
- [ ] Sub-task 2
- [ ] Sub-task 3

## References / 参考资料
- Related code path: `unirl/xxx/yyy.py`
- Related paper / link

## Acceptance Criteria / 验收标准
How do we know this is done? e.g., pass a specific test, reach a performance target.
怎样算做完了？比如：跑通某个测试、性能达到某个指标。

Step 1: Create a Tracking Issue (break down large tasks)
第一步：创建 Tracking Issue（大任务拆分）
        │
        ▼
Step 2: Create sub-issues for each task, listed in the Tracking Issue
第二步：每个子任务单独开 issue，在 Tracking Issue 中列出
        │
        ▼
Step 3: Assign sub-issues to team members via Assignees
第三步：通过 Assignees 把子 issue 指派给具体组员
        │
        ▼
Step 4: Assignee comments progress & problems under the issue
第四步：组员在 issue 下评论进度和遇到的问题
        │
        ▼
Step 5: Submit PR and link the issue (write "Closes #xx" in PR description)
第五步：完成后提交 PR，PR 描述里写 "Closes #xx" 关联 issue
        │
        ▼
Step 6: Code Review → Merge → Issue auto-closed
第六步：Code Review → 合并 → Issue 自动关闭

Title: [Tracking] UniRL Training Pipeline Optimization
标题：[Tracking] UniRL 训练流程优化

## Background / 背景
The current training pipeline has performance bottlenecks in gradient
accumulation and lacks mixed precision support.
当前训练流程在梯度累积环节存在性能瓶颈，且缺少混合精度支持。

## Sub-tasks / 子任务
- [ ] #101 Refactor gradient accumulation in backward_train_step → @MemberA
      重构 backward_train_step 中的梯度累积逻辑
- [ ] #102 Add mixed precision (bf16/fp16) support → @MemberB
      添加混合精度训练支持
- [ ] #103 Write training performance benchmark → @MemberC
      编写训练性能 benchmark
- [ ] #104 Add unit tests for training step → @MemberD
      补充训练步骤的单元测试

## Milestones / 里程碑
- Mar 15: Complete #101, #102 / 3月15日前完成 #101 #102
- Mar 31: Complete #103, #104 / 3月底前完成 #103 #104

## Acceptance Criteria / 验收标准
- All sub-issues closed with merged PRs / 所有子 issue 已关闭并合并 PR
- Benchmark shows ≥20% training speedup / Benchmark 显示训练速度提升 ≥20%
- Unit test coverage ≥80% for training module / 训练模块单测覆盖率 ≥80%

# Create an issue with assignee and labels
# 创建 issue 并指派和打标签
gh issue create \
  --title "[Task] Refactor backward_train_step gradient accumulation" \
  --body "## Objective
Optimize gradient accumulation logic.

## Tasks
- [ ] Analyze current implementation
- [ ] Refactor code
- [ ] Add tests" \
  --assignee teammate-username \
  --label "module: training,priority: high"

# List open issues assigned to someone
# 查看某人被分配的 issue
gh issue list --assignee teammate-username --state open

# View a specific issue
# 查看某个 issue 的详情
gh issue view 101

# Close an issue (usually done automatically via PR)
# 关闭 issue（通常通过 PR 自动关闭）
gh issue close 101

# Add a comment to an issue
# 给 issue 添加评论
gh issue comment 101 --body "Progress update: completed step 1 and 2."

# Create labels (run once to set up the repo)
# 创建标签（初始化仓库时运行一次）
gh label create "priority: high" --color "d93f0b" --description "Must be done ASAP"
gh label create "priority: medium" --color "fbca04" --description "Should be done this sprint"
gh label create "priority: low" --color "0e8a16" --description "Nice to have"
gh label create "module: training" --color "1d76db" --description "Training pipeline"
gh label create "module: inference" --color "5319e7" --description "Inference & rollout"
gh label create "module: algorithm" --color "006b75" --description "Algorithm related"
gh label create "needs-discussion" --color "d4c5f9" --description "Needs design discussion"