Skip to content

[A1] 张境航 - 完成 A1 Basics - #95

Open
Elvira-hang920 wants to merge 4 commits into
OpenMOSS:mainfrom
Elvira-hang920:a1/Elvira-hang920
Open

[A1] 张境航 - 完成 A1 Basics#95
Elvira-hang920 wants to merge 4 commits into
OpenMOSS:mainfrom
Elvira-hang920:a1/Elvira-hang920

Conversation

@Elvira-hang920

Copy link
Copy Markdown
Contributor

PR Checklist

基本信息

  • 同学真名:张境航
  • 作业编号:A1

修改范围

  • [√ ] 本 PR 只包含我本人本次作业的文件。
  • [√ ] 我没有修改其他同学、students/_template、公共题面或仓库配置。
  • [√ ] PR 标题符合 [A编号] 姓名 - 简短说明

公开性与安全

  • [√ ] 我理解 GitHub 中所有内容均为公开资料。
  • [√ ] 公开内容不含内部主机名、IP、账号、路径、数据或未公开项目。
  • [√ ] GitHub 和飞书正文均不包含 Secret、Token、Cookie、密码或私钥。
  • [√ ] 我检查了 git diff --cached,没有提交大型数据、模型权重、缓存或完整日志。

双层提交

  • [√ ] README.md 提供可公开、已脱敏的作业报告。
  • [√ ] README.md 已填写飞书补充文档链接。

A1 额外检查

若本 PR 不是 A1,请将本节标记为不适用。

  • [√ ] 已提交 README.mdsubmission/cs336_basics/submission/tests/adapters.py
    submission/scripts/logs/
  • [√ ] 报告文件为 README.md(Markdown)。
  • [√ ] assignment1-basics 工作仓库位于 SummerQuest 的兄弟目录,且没有被提交到本 PR。
  • [√ ] 已在 ../assignment1-basics 运行官方测试,并使用同步脚本更新个人提交目录。

A0 额外检查

若本 PR 不是 A0,请将本节标记为不适用。

  • 已完成公开 GitHub profile 和组内飞书 profile。
  • 已在个人服务器实际运行 nvidia-smigpustat 并记录退出码。
  • A0 的组内验收材料已放入 README 链接的飞书补充文档。

给助教的说明

<说明未完成项、环境限制或希望重点审核的内容。不要粘贴密钥和内部地址。>

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d9985f217b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment on lines +240 to +241
q=self.rope(q,token_positions)
k=self.rope(k,token_positions)

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Add a head axis before applying RoPE

When run_multihead_self_attention_with_rope is called with per-example token_positions shaped (batch, sequence_length), q and k are already shaped (batch, num_heads, sequence_length, head_dim). Passing the positions through unchanged makes the RoPE caches shape (batch, sequence_length, head_dim/2), which PyTorch aligns against the heads dimension and either errors or broadcasts positions incorrectly unless batch == num_heads or batch == 1. Insert a singleton head dimension for batched positions before applying RoPE so every head uses the same per-example positions.

Useful? React with 👍 / 👎.

## 基本信息

- 作业题面版本:26.0.3
- 完成范围:<填写>

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Replace the A1 README placeholders

scripts/validate_repo.py rejects assignment READMEs containing <...> placeholders, and this README still has <填写> here plus several more template placeholders below. As submitted, repository validation fails for this A1 directory before grading, so these template fields need real values or an explicit .

Useful? React with 👍 / 👎.


## 飞书补充文档

- 链接:<粘贴飞书 Doc 或 Wiki 链接>

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Add the required Feishu supplement link

The repo validator requires each assignment README to contain a Feishu doc/wiki URL, but this line still contains the template placeholder instead of a URL. This causes scripts/validate_repo.py to report missing Feishu supplement URL for this submission, so add the actual org-visible Feishu supplement link.

Useful? React with 👍 / 👎.

@WillQvQ WillQvQ left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

审核结论:驳回,请修改后重新请求审核。

当前标题 A1/elvira hang920 不符合仓库要求的 [A编号] 姓名 - 简短说明 格式。请将标题修改为类似:

[A1] 张境航 - 完成 A1 Basics

标题修正后可继续在当前 PR 更新,无需另开 PR。

审核者:GPT-5.6 Sol Ultra

@Elvira-hang920 Elvira-hang920 changed the title A1/elvira hang920 [A1] 张境航 - 完成 A1 Basics Jul 21, 2026

@WillQvQ WillQvQ left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

复审结论:标题已修正,但当前提交仍需修改后再审核。

需要补齐的阻塞项:

  1. A1 必交目录缺失:当前没有 submission/scripts/logs/
  2. assignments/A1/README.md 仍基本是模板占位内容,没有实际实验结果。请补充完整日志/损失曲线、学习率扫描(含至少一次真实发散或明显不稳定)、batch size 从 1 到显存上限(包含 64 和 128)、四项消融(no RMSNorm、Post-Norm、NoPE、SiLU FFN)、OpenWebText 同架构同训练步数实验,以及生成结果(至少 256 个新 token,除非提前 EOS)。
  3. model.py 中 RoPE 对形状为 (batch, sequence_length)token_positions 只增加到了二维,未增加 head 维;对 (batch, num_heads, sequence_length, head_dim) 的 q/k 会广播错误或报错。请在应用 RoPE 前为批量位置增加 singleton head 维,并补充/运行相应测试。

飞书链接、Git 总结详细程度、模拟密钥文件细节,以及无 GPU 环境均不作为本次阻塞项。当前未发现真实凭据或超大文件,因此 PR 保持开放,修正后可继续在本 PR 请求复审。

审核者:GPT-5.6 Sol Ultra

@WillQvQ WillQvQ left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

复审结论:RoPE 的批量位置广播问题已经修复,但当前提交仍需修改。

剩余阻塞项:

  1. A1 必交目录 submission/scripts/ 仍缺失,尚未提交训练、数据编码与生成脚本。
  2. 题面要求的日志目录是作业根目录下的 logs/。当前只新增了位置不符且内容为空的 submission/logs/README.md,不能作为实际实验日志。
  3. assignments/A1/README.md 明确说明训练 pipeline、OpenWebText 流程和实际训练实验尚未完成;请至少补齐可复现的训练流程与真实实验记录,并按题面覆盖学习率扫描、batch size 探测、四项消融、OpenWebText 对照及生成结果。若硬件条件有限,可以如实说明限制并采用可运行的缩减规模,但不能完全缺少训练与实验材料。

本次未发现真实凭据、超大文件、数据集、模型权重或其他禁止提交文件,因此 PR 保持开放;飞书链接、Git 总结详细程度、模拟密钥文件细节、机器/GPU 信息均不作为阻塞项。

审核者:GPT-5.6 Sol Ultra

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants