add-uint-support
Add unsigned integer (uint) type support to PyTorch operators by updating AT_DISPATCH macros.…
Update the third_party/kineto submodule in PyTorch to the latest commit from this kineto repo and commit the change. Use when updating the kineto submodule hash for a release.
$ npx -y skills add pytorch/pytorch --skill kineto-release --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/kineto-releaseContext preview
The summary Claude sees to decide when to auto-load this skill.
Update the third_party/kineto submodule in PyTorch to the latest commit from this kineto repo and commit the change. Use when updating the kineto submodule hash for a release.
name: kineto-release description: Update the third_party/kineto submodule in PyTorch to the latest commit from this kineto repo and commit the change. Use when updating the kineto submodule hash for a release. disable-model-invocation: true argument-hint: "[commit-hash]" allowed-tools: Bash(git:*), Read, AskUserQuestion
Update the `third_party/kineto` submodule in a local PyTorch repo to point to a commit from this Kineto repo.
If a commit hash is provided via `$ARGUMENTS`, use that as the target hash. Otherwise, use the tip of main from this repo.
Before starting, you need the path to the local PyTorch repo. If you do not already know it from prior conversation context, ask the user:
> "What is the path to your local PyTorch repo?"
The PyTorch repo is required — do not proceed without it. Store the path as `$PYTORCH` for the steps below.
1. **(PyTorch) Get the current submodule hash:**
cd $PYTORCH && git submodule status third_party/kineto | awk '{print $1}' | sed 's/^[-+]//'2. **(Kineto) Get the target hash.** If `$ARGUMENTS` is provided, use it. Otherwise, get the tip of main from this repo:
git log -1 --format='%H' main
Do NOT fetch or pull from remote -- you do not have permissions. The repo is already up to date.
3. **(PyTorch) Create a branch** named `release_YYYY_MM_DD_<short_new_hash>` (using today's date and the short target commit hash) and switch to it.
4. **(Kineto) List included commits** between old and new hash:
git log --format='- %s %h' <old_hash>..<new_hash> | sed 's/(#/(pytorch\/kineto#/g'
5. **(PyTorch) Update the submodule** to the new hash. The submodule's local object store likely won't have the target commit, so fetch from the local Kineto repo first:
cd $PYTORCH/third_party/kineto && git fetch <path_to_kineto_repo> main && git checkout <new_hash>
6. **(PyTorch) Commit** with a message in this format:
Update third_party/kineto submodule to <short_new_hash> Includes the following commits: - <message1> <hash1> - <message2> <hash2> - ...
If the current hash already matches the target, inform the user that kineto is already up to date and stop.
Tensors and Dynamic neural networks in Python with strong GPU acceleration
Add unsigned integer (uint) type support to PyTorch operators by updating AT_DISPATCH macros.…
Debug AOTInductor (AOTI) errors and crashes. Use when encountering AOTI segfaults, device…
Convert PyTorch AT_DISPATCH macros to AT_DISPATCH_V2 format in ATen C++ code. Use when…
Query PyTorch CI, GitHub Actions, HUD, Grafana, and infrastructure metrics. Use when users…
Choose 32-bit vs 64-bit index math in PyTorch CUDA kernels. Use when fixing large-tensor…
Sub-triages issues in the oncall:distributed queue by assigning distributed module labels,…