[codex] Sync Torch-OSQP release-readiness package

.github/workflows/torch-osqp-core.yml

@@ -0,0 +1,54 @@
+name: Torch OSQP core
+
+on:
+  push:
+  pull_request:
+  workflow_dispatch:
+
+jobs:
+  deterministic-core:
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            python: "3.10"
+            torch: "torch==2.8.0"
+          - os: ubuntu-latest
+            python: "3.12"
+            torch: "torch"
+          - os: ubuntu-latest
+            python: "3.13"
+            torch: "torch"
+          - os: windows-latest
+            python: "3.10"
+            torch: "torch==2.8.0"
+          - os: windows-latest
+            python: "3.12"
+            torch: "torch"
+          - os: macos-latest
+            python: "3.10"
+            torch: "torch==2.8.0"
+          - os: macos-latest
+            python: "3.12"
+            torch: "torch"
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 120
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+      - name: Install CPU test environment
+        shell: bash
+        run: |
+          python -m pip install --upgrade pip
+          if [[ "$RUNNER_OS" == "macOS" ]]; then
+            python -m pip install "${{ matrix.torch }}"
+          else
+            python -m pip install "${{ matrix.torch }}" --index-url https://download.pytorch.org/whl/cpu
+          fi
+          python -m pip install numpy scipy osqp pytest gurobipy
+          python -m pip install -e . --no-deps
+      - name: Run deterministic and end-to-end gates
+        run: python -m pytest tests -q

.github/workflows/torch-osqp-cuda-promotion.yml

@@ -0,0 +1,57 @@
+name: Torch OSQP CUDA promotion gate
+
+on:
+  workflow_dispatch:
+
+jobs:
+  cuda-real-hardware:
+    runs-on: [self-hosted, cuda]
+    timeout-minutes: 120
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Verify preinstalled CUDA PyTorch
+        run: >-
+          python -c "import torch; assert torch.cuda.is_available();
+          print(torch.__version__, torch.cuda.get_device_name())"
+      - name: Install non-Torch test dependencies
+        run: >-
+          python -m pip install numpy scipy osqp pytest gurobipy
+      - name: Install PyGRANSO without replacing CUDA PyTorch
+        run: python -m pip install -e . --no-deps
+      - name: CUDA float64 correctness and stress
+        run: >-
+          python torch_osqp_stability.py
+          --device cuda
+          --dtype float64
+          --seeds 100
+          --stress-seeds 100
+          --time-limit-seconds 7200
+          --output artifacts/cuda-float64
+      - name: CUDA qualified float32 correctness
+        run: >-
+          python torch_osqp_stability.py
+          --device cuda
+          --dtype float32
+          --seeds 100
+          --stress-seeds 100
+          --time-limit-seconds 7200
+          --output artifacts/cuda-float32
+      - name: End-to-end five-times promotion ceiling
+        run: >-
+          python bench_pygranso_osqp_workloads.py
+          --workloads B1 B2 B3
+          --repeats 5
+          --warmups 2
+          --maxit 20
+          --maximum-slowdown 5
+          --enforce-gate
+          --export-csv artifacts/cuda-performance.csv
+          --export-md artifacts/cuda-performance.md
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: torch-osqp-cuda-promotion
+          path: artifacts

.github/workflows/torch-osqp-nightly.yml

@@ -0,0 +1,46 @@
+name: Torch OSQP nightly stability
+
+on:
+  schedule:
+    - cron: "17 5 * * *"
+  workflow_dispatch:
+
+jobs:
+  cpu-stability:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest, macos-latest]
+        dtype: [float32, float64]
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 120
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Install current stable CPU environment
+        shell: bash
+        run: |
+          python -m pip install --upgrade pip
+          if [[ "$RUNNER_OS" == "macOS" ]]; then
+            python -m pip install torch
+          else
+            python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          fi
+          python -m pip install numpy scipy osqp pytest gurobipy
+          python -m pip install -e . --no-deps
+      - name: Generate fixed-seed evidence
+        run: >-
+          python torch_osqp_stability.py
+          --device cpu
+          --dtype ${{ matrix.dtype }}
+          --seeds 100
+          --stress-seeds 100
+          --time-limit-seconds 7200
+          --output artifacts/${{ matrix.os }}-${{ matrix.dtype }}
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: torch-osqp-${{ matrix.os }}-${{ matrix.dtype }}
+          path: artifacts/${{ matrix.os }}-${{ matrix.dtype }}

.gitignore

@@ -79,4 +79,9 @@ examples/TMLR/user_grad_constr_dl_orthogonal_constraint.py.lprof
 examples/TMLR/constr_dl_orthogonal_constraint.py.lprof
 examples/TODO
 
-*neural-structural-optimization*
+*neural-structural-optimization*
+
+# Generated Torch-OSQP evidence and PDF render scratch. Release artifacts are
+# force-added deliberately after their source commit is cleanly identified.
+tmp/
+output/

README.md

@@ -56,10 +56,54 @@ By default, `pip` or `uv` may install a CPU-only build of PyTorch. For **GPU (CU
 
 Set `opts.torch_device = torch.device("cuda")` when calling PyGRANSO to use the GPU.
 
+### OSQP backend options
+
+PyGRANSO uses OSQP for its internal quadprog-compatible QP subproblems. The
+Torch route is a correctness-first dense reference implementation with a
+replaceable linear-solver boundary. It is not a sparse large-scale solver.
+
+- `opts.osqp_algebra = "auto"` follows `opts.torch_device`: CPU uses builtin
+  OSQP; a validated accelerator uses Torch inside the supported KKT and memory
+  envelope. Unsupported or unsuccessful Torch solves visibly fall back to
+  builtin OSQP and retain structured fallback diagnostics.
+- `opts.osqp_algebra = "builtin"` forces CPU OSQP.
+- `opts.osqp_algebra = "torch"` explicitly requests the dense Torch reference
+  route on `opts.torch_device`. Above the validated `n + m <= 2400` KKT limit,
+  it warns and attempts the solve rather than silently changing backend.
+- `opts.osqp_settings` overrides common settings shared by builtin and Torch.
+  Defaults are dtype-aware (`1e-8` for float64 and `1e-5` for float32) and
+  enable 10-pass Ruiz scaling, deterministic adaptive rho, polishing, and
+  structurally compatible warm starts.
+
+Float64 is authoritative through estimated KKT conditioning around `1e8`.
+Float32 is a qualified route with a conservative conditioning envelope around
+`1e2`; harder float32 cases are retained as stress evidence rather than
+claimed support.
+
+Archived sparse-CG and CUDA Graph settings are recognized for one migration
+release but raise an actionable deprecation error. The research implementation
+is preserved on `archive/sparse-cg-cuda-graph`.
+
+Torch support is promoted backend-by-backend. CPU and CUDA require their own
+release gates. ROCm remains unclaimed until real-hardware CI is available. MPS
+is float32-only and remains unclaimed until reusable LU is validated on Apple
+hardware; float64 MPS auto requests return the builtin CPU float64 solution.
+
+The current local NVIDIA run passed the fixed-seed correctness buckets but
+failed the end-to-end promotion gate at 12.48x, 21.33x, and 43.46x builtin CPU
+runtime on representative B1/B2/B3 workloads, so CUDA remains unpromoted and
+`auto` falls back visibly to builtin OSQP.
+
+PyGRANSO does not differentiate through the OSQP QP solve; autograd is used to
+compute the objective and constraint gradients before QP construction.
+
 ### Verify installation
 
 - **CPU:** `python test_cpu.py`
 - **CUDA:** `python test_cuda.py`
+- **Torch OSQP core:** `python -m pytest tests -q`
+- **Stability evidence:** `python torch_osqp_stability.py --seeds 100`
+- **Dense reference benchmark:** `python bench_osqp_dense_reference.py`
 
 Then check the [example folder](./examples) or the [example section](https://ncvx.org/examples) on the documentation website to get started.
 
@@ -163,4 +207,3 @@ Thanks to other contributors and bug reporters:
 - [Ying Cui](https://sites.google.com/site/optyingcui/home): Advised the adversarial robustness problems.
 
 - [Chen Jiang](https://github.com/shoopshoop): Tested perceptual attack example (ex6). Tested PyGRANSO on Win10. Debugged updatePenaltyParameter function.
-