[CI] --no-deps for installation with uv (#1493)

Author: vmoens

.github/unittest/linux/scripts/install.sh

@@ -44,7 +44,11 @@ else
 fi
 
 printf "* Installing tensordict\n"
-pip install -e .
+# Install runtime deps explicitly (except torch/torchvision which are handled above),
+# then install tensordict without resolving dependencies to avoid any solver changing
+# the PyTorch build (stable vs nightly).
+python -m pip install -U packaging pyvers importlib_metadata
+python -m pip install -e . --no-deps
 
 # smoke test
 python -c "import functorch"

.github/workflows/benchmarks.yml

@@ -79,7 +79,10 @@ jobs:
           uv pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu126
           uv pip install "pybind11[global]" "setuptools" "wheel" "ninja"
           uv pip install pytest pytest-benchmark
-          uv pip install -e .
+          # Do not resolve runtime dependencies here: we want to keep the PyTorch build (nightly)
+          # that was explicitly installed above, and avoid any solver replacing it.
+          uv pip install -e . --no-deps
+          uv run --active python -c "import torch; print(torch.__version__); assert 'dev' in torch.__version__, 'Expected a PyTorch nightly build'"
           ${{ matrix.device == 'CPU' && 'export CUDA_VISIBLE_DEVICES=' || '' }}
       - name: check GPU presence
         if: matrix.device == 'GPU'
@@ -162,8 +165,6 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           gh-pages-branch: gh-pages
           auto-push: true
-          # Use regular comments instead of PR reviews to avoid permission issues
-          comment-style: 'github'
 
       - name: Store GPU benchmark results
         uses: benchmark-action/github-action-benchmark@v1
@@ -179,5 +180,3 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           gh-pages-branch: gh-pages
           auto-push: true
-          # Use regular comments instead of PR reviews to avoid permission issues
-          comment-style: 'github'

.github/workflows/benchmarks_pr.yml

@@ -68,7 +68,7 @@ jobs:
           uv pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu126
           uv pip install "pybind11[global]" "setuptools" "wheel" "ninja"
           uv pip install pytest pytest-benchmark
-          uv pip install -e .
+          uv pip install -e . --no-deps
           ${{ matrix.device == 'CPU' && 'export CUDA_VISIBLE_DEVICES=' || '' }}
       - name: check GPU presence
         if: matrix.device == 'GPU'

.github/workflows/docs.yml

@@ -57,16 +57,21 @@ jobs:
         # Install PyTorch (CPU version for docs)
         python -m pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cpu -U --quiet --root-user-action=ignore
 
+        # Install tensordict runtime deps explicitly (except torch/torchvision which are handled above),
+        # then install tensordict without resolving dependencies to avoid any solver changing the
+        # PyTorch build (stable vs nightly).
+        python -m pip install -U packaging pyvers importlib_metadata --quiet --root-user-action=ignore
+
         # Set up versioning environment (conditional on release branch/tag)
         export GITHUB_REF_TYPE=${{ github.ref_type }}
         export GITHUB_REF_NAME=${{ github.ref_name }}
         source .github/scripts/version_script.sh
 
         # Install tensordict -- install as develop if not on a release branch
         if [[ ${{ github.ref_type }} == branch && ${{ github.ref_name }} == release/* ]] || [[ ${{ github.ref_type }} == tag ]]; then
-          python -m pip install .
+          python -m pip install . --no-deps
         else
-          python -m pip install -e .
+          python -m pip install -e . --no-deps
         fi
 
         # Install docs requirements

README.md

@@ -143,6 +143,31 @@ To enjoy the latest features, one can use
 pip install tensordict-nightly
 ```
 
+**With uv + PyTorch nightlies**:
+
+If you're using a **PyTorch nightly** (e.g. installed from the PyTorch nightly wheel index), then for **editable**
+installs you should install tensordict with **`--no-deps`**.
+
+This avoids uv re-resolving `torch` from its configured indexes (by default: PyPI), which can otherwise replace an
+existing nightly with the latest stable `torch`.
+
+This is an **uv-specific** behavior; plain `pip install -e .` typically won’t replace an already-installed PyTorch
+nightly.
+
+To keep a nightly `torch` with an editable install:
+
+- install without resolving deps (**recommended**):
+
+```bash
+uv pip install -e . --no-deps
+```
+
+- or (less recommended) explicitly point uv at the PyTorch nightly wheel index (CPU shown; use the appropriate backend directory like `cu126/`):
+
+```bash
+uv pip install -e . --prerelease=allow -f "https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html"
+```
+
 **With Conda**:
 
 Install `tensordict` from `conda-forge` channel.

docs/source/index.rst

@@ -56,6 +56,12 @@ or via a `git clone` if you're willing to contribute to the library:
   $ cd tensordict
   $ pip install -e .
 
+.. note::
+
+  If you're using **uv** with a **PyTorch nightly** installed from the PyTorch nightly wheel index,
+  prefer installing tensordict editable with ``--no-deps`` (or configure uv to use the nightly wheel index),
+  otherwise uv may re-resolve and replace the existing nightly ``torch`` with a stable build from PyPI.
+
 Tutorials
 =========