[Feature]: can we trim rocm/pytorch image?

[tantara]

Suggestion Description

we've seen dramatic increase in docker image size from rocm 7.2 to 7.14. here is my investigation

   Image                        Compressed pull size    Dominant layer
  ━━━━━━━━━━━━━━━━━━━━━━━━━━━  ━━━━━━━━━━━━━━━━━━━━━━  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
   ROCm 7.2.4 / PyTorch 2.10     10.39 GB / 9.68 GiB    ROCm apt packages: 7.91 GB; PyTorch wheels: 2.08 GB
  ───────────────────────────  ──────────────────────  ─────────────────────────────────────────────────────
   ROCm 7.14 / PyTorch 2.10     19.14 GB / 17.82 GiB    Wheel install: 18.64 GB, 97.4%
  ───────────────────────────  ──────────────────────  ─────────────────────────────────────────────────────
   ROCm 10.0 / PyTorch 2.12     20.19 GB / 18.80 GiB    Wheel install: 19.69 GB, 97.5%

  These exact registry values agree with the rounded layer sizes shown by Docker Hub: 7.2.4, 7.14, and 10.0.

  The responsible layers are:

  ROCm 7.2.4
  sha256:4442b2c3...   7.912 GB   apt-get install rocm ...
  sha256:dd9255a7...   2.078 GB   pip install torch/vision/audio/triton/apex

  ROCm 7.14
  sha256:17935fbf...  18.636 GB   install_pytorch_wheels.py --device-targets device-all

  ROCm 10.0
  sha256:bf9ca4f9...  19.688 GB   install_pytorch_wheels.py --device-targets device-all

Operating System

No response

GPU

No response

ROCm Component

No response