用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/opendatahub-io/ai-helpers --skill vllm-compare-reqs命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | vllm-compare-reqs |
| description | Use this skill to compare vllm requirements files between versions |
| allowed-tools | Bash |
| user-invocable | true |
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: Starting from v0.20, vllm moved build requirements into a build/ subdirectory (e.g., build/rocm.txt instead of rocm-build.txt). The script automatically handles both layouts, falling back to the old paths for older versions. This makes cross-version comparisons (e.g., v0.19 vs v0.20) work seamlessly.
Executable Script - Runs a Python script to fetch and compare requirements files from the vllm GitHub repository.
Use this skill when you need to:
Note: Variant comparisons include BOTH runtime and build dependencies to ensure complete coverage.
./scripts/compare_reqs.py <version1> <version2> <variant|file> [--pretty]
v0.13.0, v0.14.0rc1)rocm, cuda, cpu, tpu, xpu (auto-includes runtime + build requirements + Dockerfiles)build/rocm.txt, common.txt, docker/Dockerfile.rocm, etc. (old-style paths like rocm-build.txt also accepted)# Compare ROCm runtime + build requirements + Dockerfiles
./scripts/compare_reqs.py v0.19.0 v0.20.2 rocm
# Compare CUDA runtime + build requirements + Dockerfiles
./scripts/compare_reqs.py v0.19.0 v0.20.2 cuda
# Compare specific file only
./scripts/compare_reqs.py v0.19.0 v0.20.2 common.txt
./scripts/compare_reqs.py v0.19.0 v0.20.2 build/rocm.txt
./scripts/compare_reqs.py v0.19.0 v0.20.2 docker/Dockerfile.rocm_base
# Cross-version comparison (old layout vs new layout - handled automatically)
./scripts/compare_reqs.py v0.18.0 v0.20.2 rocm
# All variants (auto-detects correct file paths per version)
./scripts/compare_reqs.py v0.19.0 v0.20.2 rocm # common.txt + rocm.txt + build/rocm.txt + Dockerfiles
./scripts/compare_reqs.py v0.19.0 v0.20.2 cuda # common.txt + cuda.txt + build/cuda.txt + Dockerfile
./scripts/compare_reqs.py v0.19.0 v0.20.2 cpu # common.txt + cpu.txt + build/cpu.txt + Dockerfile.cpu
./scripts/compare_reqs.py v0.19.0 v0.20.2 tpu # common.txt + tpu.txt + Dockerfile.tpu
./scripts/compare_reqs.py v0.19.0 v0.20.2 xpu # common.txt + xpu.txt + Dockerfile.xpu (Intel GPU)
The script provides clean, categorized output with a summary table followed by detailed changes:
=== Comparing rocm variant (runtime + build + Dockerfiles): v0.19.0 -> v0.20.2 ===
📊 Change Summary Table:
File Package Old Version New Version Type
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
common.txt protobuf - >= 6.30.0 Added
common.txt grpcio - >=1.76.0 Added
build/rocm.txt torch ==2.9.0 ==2.9.1 Changed
build/rocm.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
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📄 build/rocm.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
After running the script and displaying the output, you must provide intelligent impact 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.
When onboarding a new vllm version:
When builds fail:
When planning releases:
The script detects and reports errors for:
rocm-build.txt vs new build/rocm.txt)