feat(vllm-tensorizer): Bump vLLM to v0.20.2 on CUDA 13.2 / Ubuntu 24.04

[JustinPerlman]

Summary

• Bump vLLM to v0.20.2
• Upgrade base images from CUDA 12.9.1 / Ubuntu 22.04 to CUDA 13.2.1 / Ubuntu 24.04 (both builder and final)

Ubuntu 24.04 compatibility fixes

• Remove python3-pip from apt in builder-base and add rm -f /usr/lib/python3.*/EXTERNALLY-MANAGED before pip bootstrap — on Ubuntu 24.04, apt-installed pip has no RECORD file and blocks pip self-upgrade
• Purge python3-jwt in the final base stage before pip installs — same root cause: Debian-managed PyJWT has no RECORD file and blocks vLLM's dependency resolution
• Fix cuda-python version spec from ~=${CUDA_VERSION} to ~=${CUDA_VERSION%.*} — patch-level CUDA versions (e.g. 13.2.1) don't match available cuda-python releases; strip to major.minor
• Install wheel package in lmcache-builder and restore it to builder-base pip install

Relevant information: vllm-project/vllm@6c964bd

[github-actions]

@JustinPerlman Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/25751418629
Image: ghcr.io/coreweave/ml-containers/vllm-tensorizer:jperlman-vllm0.20.2-cc65ad3-v0.20.2

[abatilo]

This seems reasonable to me but I'd feel better if @Eta0 could take a peek

[JustinPerlman]

This seems reasonable to me but I'd feel better if @Eta0 could take a peek

Fair enough lol

[alexeldeib]

Pure 13.2, no matrix with 12.9? 🫣I would really like having both options…if it’s a giant pain on vllm side it’s fine, but I think you then need to validate this actually works on b40/rtxp6000 with latest supported/installed drivers cw ships

[alexeldeib]

I am still not aware of a cuda + driver combo that has decent support and works as expected, but haven’t followed too closely lately