Add pytorch-cuda-fix-xpu-alignment agent skill by laifenxiawucha · Pull Request #3466 · intel/torch-xpu-ops

laifenxiawucha · 2026-04-24T07:41:46Z

Summary

Add a new agent skill (pytorch-cuda-fix-xpu-alignment) that discovers CUDA and other backend bug fixes in pytorch/pytorch and validates whether the same bugs exist on XPU.

Input / Output

Input: A time window (e.g., "last 1 day") and target repository (pytorch/pytorch).

Output: A local triage_scan_<date>.md file containing:

Candidate list with URLs, titles, and dates
Filter decisions (reject/pass) with reasons for each candidate
Adapted XPU reproducer scripts for passed candidates
Local XPU nightly validation results (confirmed / not-reproduced / unverified)
Filter and validation statistics summary

Workflow example

User: "Scan pytorch/pytorch for CUDA bug fixes in the last 1 day and check if they affect XPU."

The skill then:

Collect — Search pytorch/pytorch via GitHub MCP for recent backend bug-fix signals (issues, PRs, commits). Save a lightweight candidate list to triage_scan_<date>.md.
Process one at a time — For each candidate:
- Filter: Read details, reject if infra-only / build-CI-only / docs-typo-only; otherwise pass
- Reproduce: Extract or adapt the upstream regression test for torch.xpu
- Validate: Run on local XPU nightly, record confirmed / not-reproduced / unverified
Summarize — Append filter statistics and validation statistics to the scan file.

What is included

File	Purpose
`.github/skills/pytorch-cuda-fix-xpu-alignment/SKILL.md`	Skill definition and end-to-end workflow
`.github/skills/pytorch-cuda-fix-xpu-alignment/references/local-xpu-validation-reference.md`	Environment setup, CUDA→XPU device mapping, run commands, bug confirmation criteria
`.github/copilot-instructions.md`	Registered in the skills index (1 row added)

Adds a new agentskills.io-compatible skill package that mines pytorch/pytorch for backend-divergence bug fixes and generates minimal Python reproducers for XPU nightly validation. Skill layout: SKILL.md - instructions, ranking rubric, guardrails scripts/ - find_xpu_python, run_collect_env, run_with_xpu_python, update_torch_xpu_nightly references/ - GitHub MCP query patterns, local validation guide

Copilot

Pull request overview

Skill used: xpu-ops-pr-review.

Adds a new agentskills.io-compatible skill package (pytorch-cuda-xpu-triage) intended to help an agent mine pytorch/pytorch CUDA-fix signals and generate minimal XPU reproducers for local nightly validation.

Changes:

Introduces a new skill definition (SKILL.md) plus workflow references for GitHub MCP querying and local XPU validation.
Adds helper bash scripts to (a) locate an XPU-capable Python, (b) run scripts / collect_env with it, and (c) install/upgrade XPU nightly wheels.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
.github/skills/pytorch-cuda-xpu-triage/SKILL.md	Defines the triage/mining workflow, ranking rubric, output format, and guardrails.
.github/skills/pytorch-cuda-xpu-triage/references/github-mcp-reference.md	Documents GitHub MCP setup and query patterns for mining upstream fixes.
.github/skills/pytorch-cuda-xpu-triage/references/local-xpu-validation-reference.md	Provides a checklist and suggested boilerplate for validating repros on XPU nightly.
.github/skills/pytorch-cuda-xpu-triage/scripts/find_xpu_python.sh	Auto-detects an interpreter by probing candidates for torch.xpu support.
.github/skills/pytorch-cuda-xpu-triage/scripts/run_with_xpu_python.sh	Runs an arbitrary Python script with the detected interpreter.
.github/skills/pytorch-cuda-xpu-triage/scripts/run_collect_env.sh	Runs `torch.utils.collect_env` with the detected interpreter.
.github/skills/pytorch-cuda-xpu-triage/scripts/update_torch_xpu_nightly.sh	Upgrades/install torch/vision/audio from the XPU nightly wheel index with optional locking and dry-run.

Copilot

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds a new GitHub Copilot skill to mine recent CUDA/backend fixes in pytorch/pytorch, qualify XPU-relevant candidates, and generate local XPU repro/validation steps.

Changes:

Introduces pytorch-cuda-xpu-triage skill definition and end-to-end workflow guidance.
Adds reference docs for GitHub MCP search patterns and local XPU validation commands/criteria.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
.github/skills/pytorch-cuda-xpu-triage/SKILL.md	Defines the skill, search/qualification rubric, and expected output/handoff format.
.github/skills/pytorch-cuda-xpu-triage/references/github-mcp-reference.md	Documents read-only GitHub search/query patterns and a narrowing heuristic.
.github/skills/pytorch-cuda-xpu-triage/references/local-xpu-validation-reference.md	Documents local XPU validation assumptions, commands, and bug confirmation criteria.

Stonepia

Overall LGTM, we could quick enable for the init stage

Stonepia · 2026-04-24T08:56:20Z

Next step should land this skills. Currently, this skill only working with local OpenCode/Codex env

Copilot

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds a new “pytorch-cuda-xpu-triage” skill to guide discovery of upstream CUDA/backend fixes in pytorch/pytorch and produce minimal local XPU repro/validation steps.

Changes:

Adds a skill definition (SKILL.md) describing search strategy, qualification gates, and output format.
Adds GitHub MCP query/tooling reference guidance.
Adds local XPU validation and evidence collection reference guidance.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
.github/skills/pytorch-cuda-xpu-triage/SKILL.md	Defines the triage skill’s workflow, candidate rubric, and required guardrails.
.github/skills/pytorch-cuda-xpu-triage/references/github-mcp-reference.md	Documents GitHub search/tool patterns to discover and narrow candidates.
.github/skills/pytorch-cuda-xpu-triage/references/local-xpu-validation-reference.md	Provides local XPU run/validation checklists and environment/evidence collection steps.

github-actions · 2026-04-24T15:44:44Z

Performance outliers, please check!

🔴 [-1, 80%), should be regression

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	mnasnet1_0	1.030685	0.755972
torchbench_bfloat16_training	dcgan	0.799755	0.765201

🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	mobilenet_v3_large	0.967016	0.842025
torchbench_bfloat16_training	densenet121	0.823402	0.842823
torchbench_bfloat16_training	resnext50_32x4d	1.047162	0.860684
huggingface_bfloat16_training	AllenaiLongformerBase	0.894286	0.948284

github-actions · 2026-04-29T03:18:55Z

Performance outliers, please check!

🔴 [-1, 80%), should be regression

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
timm_models_bfloat16_training	deit_base_distilled_patch16_224	0.742053	0.751471
timm_models_bfloat16_training	vit_base_patch16_siglip_256	0.733104	0.758308
timm_models_bfloat16_training	dm_nfnet_f0	0.650604	0.774531
timm_models_bfloat16_training	visformer_small	0.742822	0.802383
timm_models_bfloat16_training	nfnet_l0	0.743742	0.819781
timm_models_bfloat16_training	mobilenetv3_large_100	0.780790	0.820245
timm_models_bfloat16_training	beit_base_patch16_224	0.782675	0.854589
timm_models_bfloat16_training	adv_inception_v3	0.796822	0.861245
timm_models_bfloat16_training	mobilevit_s	0.722264	0.956258

🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
timm_models_bfloat16_training	convnextv2_nano.fcmae_ft_in22k_in1k	0.814999	0.817716
timm_models_bfloat16_training	repvgg_a2	0.843529	0.847274
timm_models_bfloat16_training	inception_v3	0.804447	0.848304
timm_models_bfloat16_training	deit_tiny_patch16_224.fb_in1k	0.874167	0.850921
timm_models_bfloat16_training	ghostnet_100	0.849341	0.857362
timm_models_bfloat16_training	swin_base_patch4_window7_224	0.883423	0.863834
timm_models_bfloat16_training	mobilenetv2_100	0.834205	0.877436
timm_models_bfloat16_training	tf_efficientnet_b0	0.811884	0.935734

…riteria

…iteria

…entries

…vice mapping

…iles

Stonepia · 2026-04-29T09:12:13Z

Overall seems ok for now. I have the concern about the pipeline:

Please specify in the PR description on what you are going to deliver (for example, what is the input of this skill and what is the output). We need an example to indicate what is the workflow of your skills.
The naming convention is not good. You are not actually doing the cuda-xpu-triage, you are just aligning the cuda fix for xpu on this skill, then give the reproducer. So this is not the issue triage, but an issue discovery task

…row only)

…ub-mcp-reference - Rename skill folder from pytorch-cuda-xpu-triage to pytorch-cuda-fix-xpu-alignment (this is fix discovery + alignment, not issue triage) - Delete github-mcp-reference.md; inline the example query into SKILL.md Step 1 - Update copilot-instructions.md skill index entry - Update SKILL.md frontmatter and description

After validating a bug on XPU, determine whether the fix belongs in pytorch/pytorch (upstream XPU kernel) or intel/torch-xpu-ops (local kernel). Add routing statistics to the summary.

Remove dispatch-coverage.md and triage-patterns.md — the domain knowledge they contained (coverage signals, triage patterns, key interpretations) is either already in SKILL.md's workflow steps or something the agent should derive from code. Rewrite batch-scan-workflow.md: strip JSON schemas, shell command examples, and Markdown templates. Keep only workflow steps and constraints. Inline the 'where to look' file paths from dispatch-coverage.md into SKILL.md Step 2. Before: 4 files, 342 lines After: 2 files, 112 lines (comparable to PR #3466's 109 lines)

Copilot AI review requested due to automatic review settings April 24, 2026 07:41

laifenxiawucha mentioned this pull request Apr 24, 2026

Add xpu-ops-issue-creation agent skill #3467

Open

Copilot started reviewing on behalf of laifenxiawucha April 24, 2026 07:42 View session

Copilot AI reviewed Apr 24, 2026

View reviewed changes

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/scripts/find_xpu_python.sh Outdated

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/scripts/run_collect_env.sh Outdated

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/scripts/run_with_xpu_python.sh Outdated

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/scripts/update_torch_xpu_nightly.sh Outdated

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

laifenxiawucha added 2 commits April 24, 2026 01:25

refactor: simplify triage skill for GitHub Copilot

a14181a

fix: remove stray line from triage skill reference

33acdc3

Copilot AI review requested due to automatic review settings April 24, 2026 08:39

refactor: drop redundant triage skill section

fbd7507

Copilot AI reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/references/local-xpu-validation-reference.md Outdated

Comment thread .github/skills/pytorch-cuda-xpu-triage/references/local-xpu-validation-reference.md Outdated

docs: clarify triage skill local run commands

d91ddc5

Stonepia reviewed Apr 24, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/references/local-xpu-validation-reference.md Outdated

Stonepia approved these changes Apr 24, 2026

View reviewed changes

Stonepia requested review from ZhaoqiongZ and chuanqi129 April 24, 2026 08:56

docs: update collect_env instructions for triage skill

9bfb710

Copilot AI review requested due to automatic review settings April 24, 2026 09:03

Copilot started reviewing on behalf of laifenxiawucha April 24, 2026 09:04 View session

Copilot AI reviewed Apr 24, 2026

View reviewed changes

Copilot started reviewing on behalf of laifenxiawucha April 24, 2026 09:28 View session

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/references/local-xpu-validation-reference.md Outdated

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-fix-xpu-alignment/references/local-xpu-validation-reference.md

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/SKILL.md Outdated

laifenxiawucha added 12 commits April 29, 2026 00:51

docs: address review — simplify filter, fix duplicates, broaden bug c…

3a5a963

…riteria

docs: add device mapping, fallback detection, doc-reject, fix perf cr…

6bff154

…iteria

docs: align wording with upstream-first approach, add device mapping …

cad78c9

…entries

docs: simplify search examples, add GitHub MCP tooling, generalize de…

2dbb2e8

…vice mapping

docs: require saving scan results locally for auditability

2876838

docs: persist filter decisions and validation results to local scan file

6fbe1c2

docs: require saving reproducer scripts to local files

f258aa0

docs: embed reproducer scripts in scan file instead of separate .py f…

671e8e3

…iles

docs: include full run output in scan file

8ab166e

docs: add filtering statistics summary to scan file

b0159cf

docs: add validation statistics summary to scan file

741780e

docs: restructure workflow to process candidates one at a time

d46237b

laifenxiawucha requested a review from Stonepia April 29, 2026 09:04

Stonepia reviewed Apr 29, 2026

View reviewed changes

Comment thread .github/skills/pytorch-cuda-xpu-triage/references/github-mcp-reference.md Outdated

laifenxiawucha added 2 commits April 29, 2026 02:19

merge: resolve conflict in copilot-instructions.md (add triage skill …

e75dc43

…row only)

laifenxiawucha changed the title ~~Add pytorch-cuda-xpu-triage agent skill~~ Add pytorch-cuda-fix-xpu-alignment agent skill Apr 29, 2026

docs: restore original descriptive wording in skill description

7107a65

laifenxiawucha requested a review from Stonepia April 29, 2026 09:39

feat: add routing step to determine target repo for confirmed bugs

6cf55ae

After validating a bug on XPU, determine whether the fix belongs in pytorch/pytorch (upstream XPU kernel) or intel/torch-xpu-ops (local kernel). Add routing statistics to the summary.

Conversation

laifenxiawucha commented Apr 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Input / Output

Workflow example

What is included

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Stonepia left a comment

Choose a reason for hiding this comment

Uh oh!

Stonepia commented Apr 24, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Apr 24, 2026

Performance outliers, please check!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Apr 29, 2026

Performance outliers, please check!

Uh oh!

Uh oh!

Stonepia commented Apr 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

laifenxiawucha commented Apr 24, 2026 •

edited

Loading