PyGranso 2.0 ⚡

[rydeveraumn]

PyGranso 2.0 Prep 🚀 — Memory Leak Fix, CUDA OSQP Backend, Config Flag, & E2E Tests

Summary 📌

This PR prepares PyGranso 2.0 by:

• 🧼 Fixing a memory leak that limited model size
• ⚡ Adding an optional CUDA-accelerated OSQP backend
• 🔧 Introducing a config parameter to select the QP solver backend
• ✅ Validating across example notebooks and large-scale problems (e.g., topology optimization)

Motivation 💡

For 2.0 we need better scalability and GPU efficiency. This patch removes a long-standing memory retention issue and enables a CUDA OSQP pathway for QP subproblems throughout the algorithm.

What’s Changed 🛠️

• Memory leak fix 🧠🩹

  • Eliminated autograd graph / tensor retention in solver loops while creating historical snap shots
  • Removed in-place ops on saved tensors along differentiable paths
  • Verified steady memory usage in long runs

• CUDA OSQP backend 🧩⚙️

  • New integration layer for OSQP built with CUDA
  • Graceful fallback to CPU OSQP when CUDA build isn’t available

• Configuration 🧭

  • Added qp_backend with values: "osqp" (default) or "osqp-cuda"

    • Optional bool alias use_cuda_osqp: bool = False

  • Example:

    # choose one style depending on your API
    solver = PyGranso(..., qp_backend="osqp-cuda")
    # or
    solver = PyGranso(..., use_cuda_osqp=True)

• Testing & examples 🧪📓

  • Ran all example notebooks end-to-end
  • Exercised large problems (topology optimization) to verify stability, correctness, and memory behavior
  • Added/updated regression tests where applicable

Usage Notes 📝

• Optional CUDA OSQP: install the wheel matching your CUDA/Python/OS, then set qp_backend="osqp-cuda". If not present, PyGranso auto-falls back to CPU OSQP.
• No action needed for CPU-only users 👍

Backward Compatibility 🔄

• No breaking changes—default remains "osqp".

Documentation 📚

• Updated README/config docs for qp_backend / use_cuda_osqp and install notes for the CUDA OSQP wheel
• Added troubleshooting tips (e.g., selecting the correct wheel)

QA Checklist ✅

• Stable memory over long training runs 🧮
• Example notebooks pass 🟢
• Large-scale topology optimization completes with expected results 🏗️
• CPU OSQP path unchanged 🖥️
• CUDA OSQP path enabled + clean fallback 🧷
• Docs updated 🖊️

[rydeveraumn]

PyGranso 2.0 Prep 🚀 — Memory Leak Fix, CUDA OSQP Backend, Config Flag, & E2E Tests

Summary 📌

This PR prepares PyGranso 2.0 by:

• 🧼 Fixing a memory leak that limited model size
• ⚡ Adding an optional CUDA-accelerated OSQP backend
• 🔧 Introducing a config parameter to select the QP solver backend
• ✅ Validating across example notebooks and large-scale problems (e.g., topology optimization)

Motivation 💡

For 2.0 we need better scalability and GPU efficiency. This patch removes a long-standing memory retention issue and enables a CUDA OSQP pathway for QP subproblems throughout the algorithm.

What’s Changed 🛠️

• Memory leak fix 🧠🩹

  • Eliminated autograd graph / tensor retention in solver loops while creating historical snap shots
  • Removed in-place ops on saved tensors along differentiable paths
  • Verified steady memory usage in long runs

• CUDA OSQP backend 🧩⚙️

  • New integration layer for OSQP built with CUDA
  • Graceful fallback to CPU OSQP when CUDA build isn’t available

• Configuration 🧭

  • Added opts.cuda_osqp_enabled with values: True or False`

• Testing & examples 🧪📓

  • Ran all example notebooks end-to-end
  • Exercised large problems (topology optimization) to verify stability, correctness, and memory behavior
  • Added / updated regression tests where applicable

Usage Notes 📝

• Optional CUDA OSQP: install the wheel matching your CUDA/Python/OS, then set opts.cuda_osqp_enabled=True. If not present, PyGranso auto-falls back to CPU OSQP.
• No action needed for CPU-only users 👍

Backward Compatibility 🔄

• No breaking changes—default remains "osqp".

Documentation 📚

• Updated README/config docs for qp_backend / use_cuda_osqp and install notes for the CUDA OSQP wheel
• Added troubleshooting tips (e.g., selecting the correct wheel)

QA Checklist ✅

• Stable memory over long training runs 🧮
• Example notebooks pass 🟢
• Large-scale topology optimization completes with expected results 🏗️
• CPU OSQP path unchanged 🖥️
• CUDA OSQP path enabled + clean fallback 🧷
• Docs updated 🖊️