[DO NOT MERGE][TEST ONLY] Create AGENTS.md

[Stonepia]

This PR is for test only

[laifenxiawucha]

@laifenxiawucha 1

[laifenxiawucha]

@laifenxiawucha review

[laifenxiawucha]

@laifenxiawucha review

[laifenxiawucha]

@laifenxiawucha 1

[laifenxiawucha]

@laifenxiawucha review

[laifenxiawucha]

@laifenxiawucha review

[laifenxiawucha]

Requested in this mention.

I see a ping from @laifenxiawucha but the message "@laifenxiawucha 1" doesn't clearly indicate what you'd like me to help with.

Could you please clarify what you need? For example:

• A code review of this AGENTS.md PR?
• Feedback on the AGENTS.md content or structure?
• Something else specific?

[laifenxiawucha]

@laifenxiawucha review

[laifenxiawucha]

reply with skills：
Requested in this mention.

Status: Request Changes

High (Must-Fix)

Missing dispatch target will cause build/runtime failure

• Problem: Function log_softmax_sparse_xpu is removed but remains registered in dispatch configuration
• Evidence: log_softmax_sparse_xpu deleted from src/ATen/native/sparse/xpu/SparseSoftmax.cpp:23-26 while yaml/native/native_functions.yaml:line неизвестен still contains SparseXPU: log_softmax_sparse_xpu
• Why it matters: Dispatch lookup will fail at runtime when log_softmax is called on sparse XPU tensors, causing undefined linker error or missing function crash
• Suggestion: Restore log_softmax_sparse_xpu function or remove/update the yaml entry to match

OR-incorrect error message after dimension check change

• Problem: Error message text does not match the actual dimension being checked
• Evidence: src/ATen/native/transformers/Attention.cpp:147-150 checks qkv_weight.dim() == 4 but error message reads "expected 2-D \qkv_weight`"`
• Why it matters: Developers and users will be misleading, failing quickly or inconsistently, handling wrong shapes leads to incorrect buffer interprets
• Suggestion: Change error message from "expected 2-D\qkv_weight`"to"expected 4-D`qkv_weight`"`

Inconsistent backward/forward status

• Problem: Backward function log_softmax_backward_sparse_xpu remains while its forward counterpart is removed
• Evidence: Forward removal at src/ATen/native/sparse/xpu/SparseSoftmax.cpp:23-26, backward present at src/ATen/native/sparse/xpu/SparseSoftmax.cpp:37-40
• Why it matters: Calling backward without a forward path is nonsensical; this breaks expected lifecycle for log_softmax and may leave dead code
• Suggestion: Remove log_softmax_backward_sparse_xpu as well or restore the forward to maintain the API contract

Medium

Safety guard removed without justification

• Problem: NestedTensor-with-mask restriction check removed, comments misaligned
• Evidence: Guard TORCH_CHECK(!mask || !query.is_nested(), "NestedTensor with mask is not supported yet") removed from src/ATen/native/transformers/Attention.cpp:128-130 with no analysis
• Why it matters: Removes documented safety envelope; could enable unsupported behavior that yields incorrect results; inconsistent with CUDA/backend assumptions
• Suggestion: Restore the check and prove that the unsupported case is truly handled or add test + comment explaining why it’s safe

Untested logic changes

• Problem: Logic changes (dimension check and guard removal) lack corresponding tests
• Evidence: No new tests added for native_multi_head_attention_xpu dimension changes; no tests for removed guard covering when mask + nested query appear
• Why it matters: CI will not catch regressions; behavior changes may affect multi-head attention correctness on XPU
• Suggestion: Add a test that verifies 4-D qkv_weight shape validation; add or update test exercising mask + nested query if indeed supported now

Low / Suggestion

Clarify function renaming

• Problem: Kernel name changed from softmax_sparse_xpu_kernel to softmax_forward_sparse_xpu_kernel without explanation
• Evidence: src/ATen/native/sparse/xpu/SparseSoftmax.cpp:18 changed to call softmax_forward_sparse_xpu_kernel instead of softmax_sparse_xpu_kernel
• Why it matters: Unclear whether this is a deliberate refactoring or a local inconsistency; hampers readability
• Suggestion: Add inline comment or PR description line explaining rename and noting that backward kernels remained unchanged

Residual Risk

• Dispatch reachability for sparse softmax/log_softmax not verified for all paths; may affect features beyond the changed site
• Dimension behavior impact on Windows/accelerate pipelines not verified; missing blast radius analysis
• No explanation of why 4-D qkv_weight is now expected (e.g., whether this adds a new head/grouped-attention mode or just aligns with CUDA semantics)

NOTES

• This PR is tagged [DO NOT MERGE][TEST ONLY], suggesting it may be a test harness rather than a ready change. Regardless, the comments above reflect concrete correctness issues in the current diffs and should be addressed before any merge.

[laifenxiawucha]

@laifenxiawucha 你有什么skills，怎么才能用这些skills

[laifenxiawucha]

Requested in this mention.

你好！目前我支持以下 skills，它们会根据你的 @mention 内容自动触发：

可用 skills 及适用场景：

• general-pr-review - 通用代码审查：适用于 PyTorch/C++/Python PR 或单个提交的代码 review，关注正确性、回归风险、潜在 bug 和测试覆盖
• xpu-pr-review - XPU/SYCL 专用代码审查：适用于 Intel GPU 相关改动，检查 CUDA 对等性、dispatch 注册、fallback 正确性、XPU 特定测试覆盖等
• xpu-issue-triage - Issue 分类处理：帮助你判断 issue 是否可执行、缺少什么信息、如何分类问题及下一步建议
• xpu-docstring - 文档编写/审查：帮助撰写或改进 XPU API 的 docstring 和简短说明
• xpu-at-dispatch-v2 - Dispatch 迁移：处理 AT_DISPATCH_V2 迁移、宏清理或验证 XPU dispatch 转换是否正确
• xpu-add-uint-support - 无符号整数支持：添加或审查 uint16/uint32/uint64 支持，检查缺失的 unsigned dispatch 覆盖

如何使用：

在 PR、issue 或 commit 评论中 @我，直接用自然语言提出你的需求即可，例如：

• "@laifenxiawucha 请帮我 review 这个 PR 的改动的正确性"
• "@laifenxiawucha 这个 issue 需要补充什么信息？"
• "@laifenxiawucha 帮我检查 torch-xpu-ops 中的 AT_DISPATCH 迁移是否完整"

系统会根据你的问题自动选择最合适的 skill 来处理。如果问题跨越多个主题（比如既要 review 又要改 docstring），我会组合使用多个 skills（然后输出独立的检查项）。

[laifenxiawucha]

### no skills reply：
Thanks for the review request. Looking at this PR:

Overall Assessment

Despite being marked as "DO NOT MERGE [TEST ONLY]", this PR includes non-trivial code changes that introduce regressions and should not be in a test-only branch.

Issues Found

src/ATen/native/sparse/xpu/SparseSoftmax.cpp

• The log_softmax_sparse_xpu function has been completely removed. There's no replacement visible in this diff, which deletes functionality.
• The kernel call changed from softmax_sparse_xpu_kernel to softmax_forward_sparse_xpu_kernel - this API change needs to be verified against the kernel implementation.
• This suggests an incomplete refactor that breaks the log_softmax API.

src/ATen/native/transformers/Attention.cpp

• Removed the TORCH_CHECK that validated NestedTensor with mask is not supported. This removes a safety guard rather than adding the missing feature.
• Changed qkv_weight.dim() validation from 2D to 4D. The comment "This aligns with CUDA" suggests this might be correct, but this appears to change the API contract without seeing callers updated.
• These changes together seem like they're attempting to add NestedTensor support with 4D qkv_weight, but the implementation appears incomplete.

Recommendation

Either:

1. Remove all code changes and keep only the AGENTS.md file (if this is truly a test PR), or
2. Complete the refactor properly: restore log_softmax_sparse_xpu, update all callers of softmax_sparse_xpu, add proper NestedTensor mask support instead of just removing the check, and verify the 4D qkv_weight change propagates correctly through the stack.

The current state is not mergeable as-is.