vllm Compare Requirements Skill Skill

vllm Compare Requirements Skill

Purpose

Compare vllm requirements files and Dockerfiles between versions to identify dependency changes, providing intelligent analysis for AIPCC package onboarding workflows.

For accelerator builds (ROCm, CUDA, TPU, XPU), Dockerfile comparison is critical because they specify exact commits/branches for dependencies built from source (PyTorch, Triton, Flash Attention, etc.) - information not available in requirements files.

Note: Each variant compares only the files that actually exist in the vllm repository. For example, CUDA has cuda.txt but no cuda-build.txt, while ROCm has both rocm.txt and rocm-build.txt.

Skill Type

Executable Script - Runs a Python script to fetch and compare requirements files from the vllm GitHub repository.

When to Use

Use this skill when you need to:

Compare vllm runtime AND build requirements between versions before upgrading
Identify new packages that need onboarding to the AIPCC wheels builder
Understand dependency changes for specific hardware variants (ROCm, CUDA, etc.)
Analyze Dockerfile ARG changes (build commits/branches for ROCm/CUDA)
Debug build failures related to dependency mismatches

Note: Variant comparisons include BOTH runtime and build dependencies to ensure complete coverage.

Usage

Basic Syntax

./scripts/compare_reqs.py <version1> <version2> <variant|file> [--pretty]

Arguments

version1: First version to compare (e.g., v0.13.0, v0.14.0rc1)
version2: Second version to compare
variant|file: Either:
- Variant name: rocm, cuda, cpu, tpu, xpu (auto-includes runtime + build requirements + Dockerfiles)
- Specific file: rocm-build.txt, common.txt, docker/Dockerfile.rocm, etc.
--pretty: Show clean categorized output (default)
--no-pretty: Show simple diff output

Examples

# Compare ROCm runtime + build requirements + Dockerfiles
./scripts/compare_reqs.py v0.13.0 v0.14.0 rocm

# Compare CUDA runtime + build requirements + Dockerfiles
./scripts/compare_reqs.py v0.13.0 v0.14.0 cuda

# Compare specific file only
./scripts/compare_reqs.py v0.13.0 v0.14.0 common.txt
./scripts/compare_reqs.py v0.13.0 v0.14.0 rocm-build.txt
./scripts/compare_reqs.py v0.13.0 v0.14.0 docker/Dockerfile.rocm_base

# All variants (based on what files actually exist in vllm repo)
./scripts/compare_reqs.py v0.13.0 v0.14.0 rocm   # common.txt + rocm.txt + rocm-build.txt + Dockerfiles
./scripts/compare_reqs.py v0.13.0 v0.14.0 cuda   # common.txt + cuda.txt + Dockerfile
./scripts/compare_reqs.py v0.13.0 v0.14.0 cpu    # common.txt + cpu.txt + cpu-build.txt + Dockerfile.cpu
./scripts/compare_reqs.py v0.13.0 v0.14.0 tpu    # common.txt + tpu.txt + Dockerfile.tpu
./scripts/compare_reqs.py v0.13.0 v0.14.0 xpu    # common.txt + xpu.txt + Dockerfile.xpu (Intel GPU)

Output Format

The script provides clean, categorized output with a summary table followed by detailed changes:

=== Comparing rocm variant (build + Dockerfiles): v0.13.0 -> v0.14.0rc1 ===

📊 Change Summary Table:

File                 Package                             Old Version               New Version               Type      
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
common.txt           protobuf                            -                         >= 6.30.0                 Added
common.txt           grpcio                              -                         >=1.76.0                  Added
rocm-build.txt       torch                               ==2.9.0                   ==2.9.1                   Changed
rocm-build.txt       triton                              ==3.5.0                   ==3.5.1                   Changed
docker/Dockerfil...  PYTORCH_BRANCH=1c57644d                                                                 Changed
docker/Dockerfil...  MORI_BRANCH=2d02c6a9                -                                                   Added

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📄 common.txt
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 Changed:
  xgrammar == 0.1.27 → xgrammar == 0.1.29
  mistral_common[image] >= 1.8.5 → mistral_common[image] >= 1.8.8

➕ Added:
  protobuf >= 6.30.0 # Required by LlamaTokenizer, gRPC.
  grpcio>=1.76.0

➖ Removed:
  scipy # Required for phi-4-multimodal-instruct

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📄 rocm-build.txt
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 Changed:
  torch==2.9.0 → torch==2.9.1
  triton==3.5.0 → triton==3.5.1

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🐳 docker/Dockerfile.rocm_base
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 Changed:
  BASE_IMAGE=rocm/dev-ubuntu-22.04:7.1-complete → BASE_IMAGE=rocm/dev-ubuntu-22.04:7.0-complete
  PYTORCH_BRANCH=1c57644d → PYTORCH_BRANCH=89075173
  AITER_BRANCH=59bd8ff2 → AITER_BRANCH=6af8b687

➕ Added:
  MORI_BRANCH=2d02c6a9
  MORI_REPO=https://github.com/ROCm/mori.git
  RIXL_BRANCH=50d63d94
  RIXL_REPO=https://github.com/vcave/RIXL.git

AI Analysis Requirements

After running the script and displaying the output, you must provide intelligent impact analysis:

Required Analysis Sections

Summary: Brief overview of changes
Impact Level: Low/Medium/High based on:
- Number of new packages
- Breaking changes (major version bumps)
- Removed dependencies
- Dockerfile ARG changes (especially BASE_IMAGE, PYTORCH_BRANCH, etc.)
AIPCC Wheels Builder Impact:
- ✅ No Action Required: Stable dependencies
- ⚠️ Action Required: New packages to onboard, version updates
- 🚨 Breaking Changes: Major issues, base image changes
Package Details: For each new package:
- Purpose (from requirement comments)
- Complexity assessment (pure Python vs compiled)
- PyPI link
- Any ambiguities (multiple packages with same name)
Dockerfile Analysis (for ROCm/CUDA):
- Base image changes (ROCm/CUDA version changes)
- Source build commit changes (PyTorch, Triton, etc.)
- New dependencies added (MORI, RIXL, UCX, etc.)
Next Steps: Concrete action items
Context: Infer purpose of the update

Example Analysis

## Summary of Changes: vllm v0.13.0 → v0.14.0rc1

### Impact Level: Medium

### AIPCC Wheels Builder Impact:

#### ✅ No Action Required:
- ROCm version remains at 6.4
- PyTorch ecosystem stable (minor patch updates only)

#### ⚠️ Action Required:
**New packages to onboard:**
1. **`grpcio>=1.76.0`** (NEW)
   - Purpose: gRPC support
   - Complexity: Compiled extension (C++)
   - PyPI: https://pypi.org/project/grpcio/
   - Action: Onboard grpcio 1.76.0+

**Version updates:**
- torch: 2.9.0 → 2.9.1 (patch - safe)
- triton: 3.5.0 → 3.5.1 (patch - safe)

**Removed dependencies:**
- scipy: Was required for phi-4-multimodal-instruct
  - Risk: Low (model-specific, may not be widely used)

#### Next Steps:
1. Onboard grpcio>=1.76.0 to wheels builder
2. Verify torch 2.9.1 and triton 3.5.1 availability
3. Test vllm 0.14.0rc1 build with new dependencies
4. Validate phi-4 model functionality without scipy

**Context:** This release adds gRPC support and updates to PyTorch 2.9.1.

Integration with AIPCC Workflows

Package Onboarding

When onboarding a new vllm version:

Run comparison to identify new dependencies
Analyze complexity of each new package
Prioritize onboarding based on impact
Generate action plan

Debugging Build Failures

When builds fail:

Compare requirements between working and failing versions
Correlate changes with error messages
Identify missing or mismatched dependencies
Recommend fixes

Release Planning

When planning releases:

Compare requirements across multiple versions
Estimate onboarding effort
Identify blocking dependencies
Generate timeline

Technical Details

Language: Python 3
Dependencies: Standard library only (urllib)
Network: Fetches files from GitHub (requires internet)
Colors: Uses ANSI color codes for terminal output
Exit codes: 0 for success, 1 for failure

Error Handling

The script gracefully handles:

Version not found (404 errors)
File not found in specific version
Network errors
Invalid input

Agent Skills: vllm Compare Requirements Skill

Install this agent skill to your local

Skill Files