pull.yml

# SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
# All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause

# A workflow to send CI-related helpful information to PRs
name: pull
on:
  pull_request:

concurrency:
  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.run_id}}
  cancel-in-progress: true

run-name: CI status hello ${{ github.event.pull_request.number }} - ${{ github.event.pull_request.head.sha }}
jobs:
  status_hello:
    name: send CI hello status
    runs-on: ubuntu-latest
    permissions:
      statuses: write
    steps:
      - name: Set CI hello status
        run: |
          curl \
          -X POST \
          -H "Accept: application/vnd.github+json" \
          -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
          https://api.github.com/repos/${{ github.repository }}/statuses/${{ github.event.pull_request.head.sha }} \
          -d "{\"state\":\"success\",\"target_url\":\"https://github.com/NVIDIA/Fuser/wiki/Bot-Commands\",\"description\":\"Authorized users: comment !build or !test to trigger CI pipelines. See wiki.\",\"context\":\"CI notes\"}"

  pr-agent-tools:
    name: PR Agent tools
    runs-on: ubuntu-latest
    permissions:
      pull-requests: write
      issues: write
      packages: read
    container:
      image: ghcr.io/nvidia/fuser:ci-llm-workflow
      credentials:
        username: ${{ github.actor }}
        password: ${{ secrets.GITHUB_TOKEN }}
      env:
        GITHUB__USER_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        CONFIG__PUBLISH_OUTPUT: true

        OPENAI__KEY: ${{ secrets.LLM_OPENAI__KEY }}
        OPENAI__API_BASE: ${{ secrets.LLM_OPENAI__API_BASE }}
        CONFIG__MODEL: ${{ secrets.LLM_CONFIG__MODEL }}
        CONFIG__CUSTOM_MODEL_MAX_TOKENS: 131072

        CONFIG__MAX_MODEL_TOKENS: 65536
        CONFIG__PUBLISH_OUTPUT_PROGRESS: false 

        PR_REVIEWER__REQUIRE_SCORE_REVIEW: false
        PR_REVIEWER__REQUIRE_TESTS_REVIEW: true
        PR_REVIEWER__REQUIRE_ESTIMATE_EFFORT_TO_REVIEW: true
        PR_REVIEWER__REQUIRE_CAN_BE_SPLIT_REVIEW: false
        PR_REVIEWER__REQUIRE_SECURITY_REVIEW: false
        PR_REVIEWER__REQUIRE_TICKET_ANALYSIS_REVIEW: false

        PR_REVIEWER__ENABLE_REVIEW_LABELS_EFFORT: false
        PR_REVIEWER__ENABLE_REVIEW_LABELS_SECURITY: false

        PR_REVIEWER__PERSISTENT_COMMENT: true
        PR_REVIEWER__FINAL_UPDATE_MESSAGE: false

        PR_REVIEWER__EXTRA_INSTRUCTIONS: |
          To review a pull request (PR) for the Nvfuser project effectively, follow this structured approach:

          Overall Principles Check:
          
          Ensure the PR provides actual data and focuses on significant performance aspects.
          Verify that a clear performance goal is set and that feedback was sought early.

          Preliminary Evaluation:
          
          Scope and Context: Confirm the PR clearly states its purpose and the significance of the problem it addresses.
          SOL Analysis: Check if the PR uses a roofline model or existing implementations (e.g., CUTLASS) as a target for expected performance.
          Current Status and Data: Ensure the PR includes current performance metrics and reasons for any gaps.

          Approach Assessment:
          
          Technical Details: Review the technical aspects of the approach and expected performance gains.
          Limitations and Comparisons: Evaluate if multiple approaches were considered and if trade-offs are discussed.

          Results Analysis:
          
          Correctness and Tests: Confirm all tests pass and new tests were added if necessary.
          Performance Data: Check for thorough performance evaluation and clear data presentation.
          Gap Analysis: Assess the explanation of performance gaps and their importance.
          Regressions: Evaluate the impact of any regressions and whether benefits outweigh drawbacks, supported by quantitative data.

          Conclusion and Documentation:
          
          Summarize the PR's effectiveness and potential next steps.
          Ensure all guidelines are met and documentation is comprehensive.

          Insight from Failed Approaches:
          
          Consider insights from discarded methods, even if they weren't the final solution.

    steps:
      - name: PR Agent review
        run: python /app/pr_agent/cli.py --pr_url ${{ github.event.pull_request.html_url }} review