xpu-release-notes

name	xpu-release-notes
description	Generate PyTorch release notes worksheet for XPU by searching git log between release branches. Trigger on "xpu-release-notes", "generate XPU release notes".

Generate Release Notes from Git History

Generate and complete the release notes worksheet for xpu.

Usage

/xpu-release-notes <version> <torch-release-notes-clone-path>

Where <version> is the PyTorch release version (e.g., 2.11.0) and <torch-release-notes-clone-path> is the path to the cloned https://github.com/meta-pytorch/torch-release-notes repository.

Instructions

Step 0: Validate inputs

1. Check that the version and torch-release-notes-clone-path arguments were provided. If missing, ask the user for the missing arguments.
2. Check that the torch-release-notes-clone-path exists. If it doesn't exist, tell the user to clone it.
3. Confirm the version directory exists (e.g., torch-release-notes-clone-path/2.12.0/). If not, tell the user to update to latest main branch of torch-release-notes.
4. Confirm the worksheet file exists at <torch-release-notes-clone-path>/<version>/todo/result_xpu.md. If it's already in done/, tell the user it's already completed and ask if they want to re-process it.

Step 1: Collect all the commits between the previous pyTorch release and the current release

Fetch the previous and current release branches from the PyTorch repo:

git fetch origin release/<prev_version_minor>:release/<prev_version_minor> release/<current_version_minor>:release/<current_version_minor>

For example, for version 2.12.0:

• Previous branch: release/2.11
• Current branch: release/2.12

Step 2: Find relevant commits

Search git log between the two release branches for commits related to xpu, and save the results to a .txt for review. Use the following command:

git log --oneline origin/release/<prev>..origin/release/<current> | grep -Ei "xpu|sycl|oneapi|intel"

Step 3: Read /result_xpu.md to understand task

• Read <torch-release-notes-clone-path>/<version>/todo/result_xpu.md worksheet file with the categorized commits. Follow the instructions in the worksheet to ensure proper categorization, clear descriptions, and formatting.

• Reference previous version's completed notes to understand more about the categories and style.

Step 4: Finish the /result_xpu.md worksheet

• Remove all the items categorized.
• for each commit, not only read the title, but also the PR descriptions to understand the context and impact of
• put the commit in the existed categories.

Format each entry as:

- Clear description of the change ([#NNNNN](https://github.com/pytorch/pytorch/pull/NNNNN))

Grouping: Merge related PRs into a single entry when they implement the same logical change:

- Improve Inductor UT coverage for XPU ([#174053](https://github.com/pytorch/pytorch/pull/174053), [#174054](https://github.com/pytorch/pytorch/pull/174054), ...)

Style rules:

• Start with a verb or "Support/Enable/Add/Fix"
• Use backticks for code identifiers: torch.compile, conv2d, addmm
• Be specific about what changed and on which device/backend
• Don't start entries with [tag] prefixes — integrate the context into the description

Step 5: Write the worksheet file

Write the complete worksheet to <version>/done/result_xpu.md (or <version>/todo/result_xpu.md if not yet finalized).

Step 6: Review with user

Present the categorized notes for user review. Common feedback patterns:

• Items in wrong categories
• Items that should be merged into one entry
• Items that should be dropped
• Items missing that the user knows about

Iterate until the user is satisfied, then ensure the file is in done/.

Example: XPU for 2.12.0

For XPU, the search terms were xpu, sycl, oneapi, Intel GPU. The process found ~80 commits, which were filtered down to ~25 entries across categories. Key decisions:

• Pin updates and CI infra → dropped entirely
• C++20 enforcement → devs (affects source builders)
• Header includes, warning fixes → not user facing
• Test coverage expansion → one grouped entry in improvements
• torch.accelerator.Graph support → new feature (new API surface)