feat: Refactor vLLM build to use upstream container image

[mc-nv]

Description

This PR refactors the vLLM backend build process to use the official NVIDIA vLLM container image as a base instead of building vLLM from scratch. This change simplifies the build process by removing complex installation logic and leveraging pre-built, optimized vLLM images.

The key motivation is to:

• Reduce build complexity and maintenance overhead
• Leverage tested and optimized upstream vLLM containers
• Improve build reliability by removing custom vLLM installation code
• Allow flexibility to use custom inference images when needed

Changes

• Removed vLLM installation logic: Deleted ~45 lines of complex vLLM installation code including architecture-specific dependencies (MKL for x86_64, NVPL for aarch64), pytorch-triton custom builds, and private index URL handling
• Added inference image support: Introduced new INFERENCE_IMAGE parameter to allow specifying custom base images for the final container
• Default vLLM image fallback: When vLLM backend is enabled and no custom inference image is provided, automatically uses nvcr.io/nvidia/vllm:{version}-py3 as the base image
• Simplified BASE_IMAGE logic: BASE_IMAGE now intelligently switches to INFERENCE_IMAGE when building vLLM backends
• Extended --image flag: Added support for --image inference, to specify custom inference base images

Affected Files

• build.py - Main build orchestration script
  • Lines 1204-1224: New inference image selection logic
  • Lines 1472-1509: Removed vLLM installation code (42 lines deleted)
  • Lines 1682-1688: Added INFERENCE_IMAGE to dockerfile argument map
  • Line 2926: Extended --image flag validation to support "inference"

[whoisj]

Approved w/ 1 suggestion.