Add 2026 spring T3-1-1 skill submission for qhl18

[qhl18]

pytest output:

## Competition

- Track: T3-1-1 NineToothed .skill Innovation
- Participant: 钱泓林
- GitHub ID: qhl18
- Branch: 2026-spring-qhl18-t3-1-1
- Skill path: `skills/competition/ninetoothed-operator-skill/`

## Summary

- 指导 AI 智能体按固定流程完成 NineToothed 算子开发、测试、benchmark、调试和 PR 集成，避免凭零散经验写代码
- 覆盖 elementwise/broadcast、reduction/block、layout-sensitive、benchmark/debug 四类算子任务
- 核心工作流：需求理解 → 选择 DSL 表达 → 实现 → PyTorch reference 测试 → benchmark → generated source/AOT 检查 → 失败诊断 → PR 集成

## Scope and limitations

- Applicable tasks: elementwise/broadcast (add)、reduction/block (softmax)、layout-sensitive with non-contiguous/stride/offset (transpose_add)、benchmark/generated source/failing test 诊断
- Out of scope: 动态 shape、bfloat16/float64、修改 NineToothed 编译器核心、多 GPU/分布式

## Installation and usage

```bash
git clone https://github.com/InfiniTensor/ninetoothed.git
cd ninetoothed
pip install -e .
# 在 AI 编程工具中让 agent 读取 skills/competition/ninetoothed-operator-skill/SKILL.md
# 运行自测:
cd skills/competition/ninetoothed-operator-skill
pytest examples/task-01/test_add.py examples/task-02/test_softmax.py examples/task-03/test_transpose_add.py -v

Pytest output

Commands executed:

pytest skills/competition/ninetoothed-operator-skill/examples/task-01/test_add.py -v
pytest skills/competition/ninetoothed-operator-skill/examples/task-02/test_softmax.py -v
pytest skills/competition/ninetoothed-operator-skill/examples/task-03/test_transpose_add.py -v
ruff check skills/competition/ninetoothed-operator-skill/
ruff format --check skills/competition/ninetoothed-operator-skill/
python scripts/check_contributing_style.py

Observed results:

--- task-01/test_add.py ---
test_add_1d_same_shape[98432-dtype0-cuda] PASSED [ 20%]
test_add_2d_same_shape[512-512-dtype0-cuda] PASSED [ 40%]
test_add_2d_1d_broadcast[512-512-dtype0-cuda] PASSED [ 60%]
test_add_2d_row_broadcast[512-512-dtype0-cuda] PASSED [ 80%]
test_add_col_broadcast[512-512-dtype0-cuda] PASSED [100%]
============================== 5 passed in 9.55s ==============================

--- task-02/test_softmax.py ---
test_softmax_matches_pytorch[1823-781-dtype0-cuda] PASSED [ 20%]
test_softmax_matches_pytorch[37-129-dtype0-cuda] PASSED [ 40%]
test_softmax_matches_pytorch[4-1024-dtype0-cuda] PASSED [ 60%]
test_softmax_is_numerically_stable[dtype0-cuda] PASSED [ 80%]
test_softmax_rejects_non_2d[cuda] PASSED [100%]
============================== 5 passed in 1.94s ==============================

--- task-03/test_transpose_add.py ---
test_transpose_add_contiguous[257-129-dtype0-cuda] PASSED [ 20%]
test_transpose_add_contiguous[64-512-dtype0-cuda] PASSED [ 40%]
test_transpose_add_non_contiguous_input[dtype0-cuda] PASSED [ 60%]
test_transpose_add_empty_strided_input[dtype0-cuda] PASSED [ 80%]
test_transpose_add_rejects_bad_bias_shape[cuda] PASSED [100%]
============================== 5 passed in 10.73s =============================

Total: 15/15 passed

ruff check: All checks passed!
ruff format --check: 10 files already formatted
check_contributing_style.py: passed

说明：以上输出来自最终 commit a1902d5，全部测试在 GPU 上真实运行通过，无 skip、无失败。

Environment:

• OS: Windows 11
• GPU: NVIDIA GeForce RTX 5060 Laptop GPU (8GB)
• CUDA: 12.8
• Python: 3.12.13
• NineToothed commit/version: 0.26.0

Self-test coverage

• Elementwise/broadcast: add with broadcast，5/5 PASSED，覆盖 1D/2D same shape、2D+1D broadcast、row/column broadcast
• Reduction/block: softmax，5/5 PASSED，覆盖非 2 的幂长度 (781,129)、长 block (1024)、数值稳定性、非 2D 拒绝
• Layout-sensitive: transpose_add，5/5 PASSED，覆盖 contiguous、slicing non-contiguous、empty_strided 自定义 stride、错误 bias shape 异常
• Performance/diagnosis: add benchmark same/vector/row 三档 + row broadcast 未 expand 导致 arrangement shape 不一致的失败诊断与修复闭环

With-skill vs no-skill comparison

• Protocol: 同一 AI agent、同一 NineToothed 仓库、同类算子任务和相同时间预算下，对比安装 skill 前后的表现
• No-skill result: 测试常遗漏或写法不规范、常只测 contiguous、benchmark 常缺失、失败时大改无记录、合规披露易遗漏
• With-skill result: 固定 PyTorch reference + GPU 验证、强制 non-contiguous/stride/offset 测试、固定 warmup/sync/baseline/ratio benchmark 框架、分类排查 + 文档化诊断闭环、REFERENCE + HONOR_CODE 模板
• Conclusion and limitations: skill 显著提升测试覆盖率、性能意识和合规性；局限在于 softmax 仍为 naive 实现 (ratio=1187x)，性能优化空间较大

Signed disclosures

• HONOR_CODE.md: skills/competition/ninetoothed-operator-skill/HONOR_CODE.md (signed by 钱泓林, 2026-07-07)
• REFERENCE.md: skills/competition/ninetoothed-operator-skill/REFERENCE.md (官方资料、外部代码、AI 辅助范围均已披露)

Proposal and final report

• Proposal: 初赛阶段已提交，如有公开链接将补充
• Final report: skills/competition/ninetoothed-operator-skill/reports/钱泓林_九齿skill创新挑战_T3-1-1_赛题报告.pdf
• Evidence/self-test material: 真实 pytest/benchmark 输出已贴入上方 Pytest output；原始日志本地保留

Compliance

• No API keys or credentials
• No hidden answers or hard-coded evaluation task names
• No deleted, weakened, or bypassed tests
• No fabricated benchmark results
• References and AI assistance are disclosed

Known limitations

• 不覆盖动态 shape（需 shape guard，超出固定 tile 模式）
• 未完整验证 bfloat16/float64
• softmax (T2) 为 naive 实现，ratio=1187x，有较大优化空间
• 自测为单 GPU 环境，未覆盖多 GPU/分布式