[gaow0007]

Summary

Upgrade the PPU (torch2100 + vLLM) Docker image from PPU SDK 2.1.1 to SDK 2.2.0, and rework how PPU-custom packages are installed and pinned.

• Base image: switch to the SDK 2.2.0 base (sdk2.2.0-pytorch2.10.0-ubuntu24.04-cuda13.0-py312-20261008).
• constraints_ppu.txt: bump every pinned component to its +vX.X.X.ppu2.2.0 build and expand coverage to the full PPU stack (torch, triton, xformers, flash-attn/3, flashinfer, transformer-engine, apex, deep-ep, deep-gemm, megatron-core, sglang-router, cupy, etc.).
• Dockerfile:
  • Replace flash-attn-3 2.8.2 with the SDK 3.0.0 build.
  • Swap cupy to the numpy-1-compatible SDK build and satisfy vLLM's cupy-cuda12x dependency with a local metapackage alias.
  • Install vLLM (PPU build) with a controlled dependency pass: last numpy-1-compatible opencv (4.11), then the remaining metadata deps.
  • Keep numpy on 1.26.4 and pin numba/ray/sglang-router/deepspeed to versions consistent with the SDK 2.2.0 stack.
  • Simplify the FLA install (backup PPU kda shim → install fla-org → restore shim); torchvision now comes pinned via constraints instead of a separate index-cleared install.
  • Extend the final import smoke test to cover flash-attn-3, numpy, cupy, scipy, sklearn.