Update SOCP dependencies for cuOpt 26.06

Author: jgoldberg-nvidia

README.md

@@ -95,7 +95,11 @@ source $HOME/.local/bin/env  # (sh, bash, zsh)
 # source $HOME/.local/bin/env.fish  # (fish)
 
 # Install with CUDA-specific dependencies
-uv sync --extra cuda13 # this container image has cuda13
+uv sync --extra cuda13 # full CUDA 13 stack currently tracks cuOpt/cuML 26.04
+# On CUDA 12 hosts, use the full cuOpt/cuML 26.06 stack:
+# uv sync --extra cuda12
+# For direct SOCP testing on CUDA 13 with cuOpt 26.06:
+# uv sync --extra cuda13-socp
 
 # Optional: Install development tools
 uv sync --extra cuda13 --extra dev
@@ -107,13 +111,13 @@ uv run python -m ipykernel install --user --name=portfolio-opt --display-name "P
 uv run jupyter lab --no-browser --NotebookApp.token=''
 ```
 
-**Note:** If you use a different container image than the suggested one above, during uv sync, please use the `--extra cuda12` or `--extra cuda13` flag to install the GPU packages (cuOpt, cuML) matching your container's CUDA version. The `uv sync` command automatically creates a virtual environment and installs all dependencies from `uv.lock`.
+**Note:** If you use a different container image than the suggested one above, during uv sync, use `--extra cuda12` for the full cuOpt/cuML 26.06 CUDA 12 stack or `--extra cuda13` for the current full CUDA 13 cuOpt/cuML stack. As of the cuOpt 26.06 release, `cuml-cu13` 26.06 is not published, so CUDA 13 SOCP testing with cuOpt 26.06 uses `--extra cuda13-socp`; that extra is cuOpt-only and is intended for direct SOCP preview/Mean-Variance variance-cap solves, not GPU KDE/CVaR rebalancing. The `uv sync` command automatically creates a virtual environment and installs all dependencies from `uv.lock`.
 
 **Tip:** To check your CUDA version, run `nvidia-smi` and look for "CUDA Version" in the output.
 
 **Important Notes:**
 - If you encounter "No space left on device" errors, set `UV_CACHE_DIR` to an alternate cache location: `export UV_CACHE_DIR=/path/to/cache/directory`
-- The `cuda12` and `cuda13` extras are mutually exclusive - install only one based on your system's CUDA version
+- The `cuda12`, `cuda13`, and `cuda13-socp` extras are mutually exclusive - install only one based on your system's CUDA version and workflow
 - If you plan to run the Streamlit demo from this container, include `-p 8501:8501` when starting Docker. Docker port mappings cannot be added to an already-running container; restart the container with the port published if it was omitted.
 
 #### Using the Jupyter Kernel

demo/README_streamlit.md

@@ -45,7 +45,7 @@ Open `http://localhost:8501` for local runs. On a remote GPU instance, forward o
 - Python 3.11+
 - Streamlit, Plotly, and Squarify from `demo/requirements.txt`
 - CVXPY and cuFOLIO core dependencies from `pyproject.toml`
-- Optional: NVIDIA GPU + CUDA with the matching `cuda12` or `cuda13` extra for cuOpt acceleration
+- Optional: NVIDIA GPU + CUDA with `cuda12` for the full cuOpt/cuML 26.06 stack, `cuda13` for the current full CUDA 13 stack, or `cuda13-socp` for CUDA 13 SOCP-only cuOpt 26.06 preview
 
 ## Troubleshooting
 
@@ -57,4 +57,4 @@ Open `http://localhost:8501` for local runs. On a remote GPU instance, forward o
 uv run python -c 'from cufolio.utils import download_data; download_data("data/stock_data", datasets=["sp500"])'
 ```
 
-**GPU Solver Unavailable**: The UI can still boot without a GPU, but the full GTC comparison and SOCP preview need cuOpt. Install the matching CUDA extra and run on an NVIDIA GPU instance.
+**GPU Solver Unavailable**: The UI can still boot without a GPU, but the full GTC comparison and SOCP preview need cuOpt. Use `--extra cuda12` for full cuOpt/cuML 26.06 validation, `--extra cuda13` for the current full CUDA 13 stack, or `--extra cuda13-socp` for CUDA 13 SOCP-only preview with cuOpt 26.06.

pyproject.toml

@@ -7,7 +7,7 @@ requires-python = ">=3.11"
 dependencies = [
     "numpy>=2.0.0",
     "pandas>=2.0",
-    "cvxpy>=1.7.0",
+    "cvxpy>=1.9.1",
     "scikit-learn>=1.5",
     "seaborn>=0.13",
     "yfinance>=0.2.0",
@@ -21,12 +21,15 @@ dev = [
     "pre-commit==4.3.0",
 ]
 cuda12 = [
-    "cuml-cu12==26.8.*",
-    "cuopt-cu12==26.8.*",
+    "cuml-cu12==26.6.*",
+    "cuopt-cu12==26.6.*",
 ]
 cuda13 = [
-    "cuml-cu13==26.8.*",
-    "cuopt-cu13==26.8.*",
+    "cuml-cu13==26.4.*",
+    "cuopt-cu13==26.4.*",
+]
+cuda13-socp = [
+    "cuopt-cu13==26.6.*",
 ]
 
 [tool.ruff]
@@ -66,18 +69,21 @@ conflicts = [
         { extra = "cuda12" },
         { extra = "cuda13" },
     ],
+    [
+        { extra = "cuda12" },
+        { extra = "cuda13-socp" },
+    ],
+    [
+        { extra = "cuda13" },
+        { extra = "cuda13-socp" },
+    ],
 ]
 
 [[tool.uv.index]]
 name = "nvidia"
 url = "https://pypi.nvidia.com"
 
 [tool.uv.sources]
-# Build cvxpy from the upstream master branch instead of the PyPI wheel so
-# we can pull in QP-via-cuOpt fixes ahead of the next release. uv.lock pins
-# the resolved commit SHA for reproducibility; bump with
-# `uv lock --upgrade-package cvxpy`.
-cvxpy = { git = "https://github.com/cvxpy/cvxpy.git", branch = "master" }
 cuml-cu12 = { index = "nvidia" }
 cuopt-cu12 = { index = "nvidia" }
 cuml-cu13 = { index = "nvidia" }

skills/cufolio/SKILL.md

@@ -44,8 +44,8 @@ Do not use it for generic finance summaries, price forecasting, neural-network t
 ## Prerequisites
 
 - Python environment with the installed `cufolio` package.
-- NVIDIA GPU runtime with cuOpt and cuML installed. Mean-Variance SOCP workflows require a cuOpt build with QCQP/SOCP support, such as the 26.08 line or newer.
-- CUDA extra matching the host, such as `uv sync --extra cuda12` or `uv sync --extra cuda13`.
+- NVIDIA GPU runtime with cuOpt and cuML installed. Mean-Variance SOCP workflows require a cuOpt build with QCQP/SOCP support, such as the 26.06 line or newer.
+- CUDA extra matching the host and workflow: `uv sync --extra cuda12` for full cuOpt/cuML 26.06 on CUDA 12, `uv sync --extra cuda13` for the current full CUDA 13 stack, or `uv sync --extra cuda13-socp` for CUDA 13 SOCP-only validation with cuOpt 26.06.
 - `cvxpy` exposing `cp.CUOPT`.
 - Network access on first run if the default price CSV must be downloaded.
 
@@ -243,7 +243,8 @@ Use the package APIs instead of reimplementing portfolio math or simulation loop
 ## Limitations
 
 - Requires an NVIDIA GPU with cuOpt and cuML; CPU solvers are intentionally disallowed.
-- Mean-Variance SOCP variance caps require cuOpt QCQP/SOCP support. Use the 26.08 line or newer when installing CUDA extras.
+- Mean-Variance SOCP variance caps require cuOpt QCQP/SOCP support. Use the 26.06 line or newer when installing CUDA extras.
+- `cuda13-socp` intentionally installs cuOpt without cuML because `cuml-cu13` 26.06 is not published yet; use it for direct SOCP/QCQP validation, not GPU KDE CVaR workflows.
 - Cardinality plus SOCP is treated as unsupported unless cuFOLIO exposes explicit mixed-integer conic support.
 - CPU-only eval containers can still validate routing, data handling, and reporting behavior, but they cannot produce a valid cuOpt solve. In that case, report the missing GPU/cuOpt runtime explicitly.
 - Default price data is a historical snapshot and may omit current constituents.
@@ -253,8 +254,8 @@ Use the package APIs instead of reimplementing portfolio math or simulation loop
 
 - Missing default CSV or `FileNotFoundError`: explain that cuFOLIO will fetch public market data with `download_data("data/stock_data", datasets=["sp500"])`; run it only after user confirmation.
 - `SolverError` or missing `cp.CUOPT`: install the CUDA extra matching the host and verify with `python -c "import cvxpy as cp; print(hasattr(cp, 'CUOPT'), cp.installed_solvers())"`.
-- `ImportError` for `cuml` or GPU KDE failures: confirm cuML is present with `python -c "import cuml"` and keep `KDESettings(device="GPU")`.
-- SOCP setup fails before solving: verify the installed `cuopt` package is on the 26.08 line or newer and that `MeanVarianceParameters.var_limit` is positive.
+- `ImportError` for `cuml` or GPU KDE failures: confirm cuML is present with `python -c "import cuml"` and keep `KDESettings(device="GPU")`. If using `cuda13-socp`, this is expected for CVaR/KDE; switch to `cuda12` or `cuda13` for cuML workflows.
+- SOCP setup fails before solving: verify the installed `cuopt` package is on the 26.06 line or newer and that `MeanVarianceParameters.var_limit` is positive.
 - Ordinary optimization returns all cash: set `c_max=0.0` in `CvarParameters`.
 - Solver reports infeasible or no solution: check for contradictory bounds, too few tickers for the requested caps/cardinality, or a date filter that leaves too little data; report the smallest constraint change that would make the request feasible.
 - Requested tickers are absent from the default CSV: report them and proceed with the remaining requested tickers.