ci: Test Python 3.15 with torch nightly

.github/actions/install-deps/action.yml

@@ -24,6 +24,13 @@ runs:
       shell: bash
       run: uv venv
 
+    # Installed before the other dependencies, so that it also works on Python versions for which
+    # no stable torch wheel exists yet.
+    - name: Install torch nightly
+      if: ${{ inputs.torch_channel == 'nightly' }}
+      shell: bash
+      run: uv pip install --pre torch torchvision -P torch -P torchvision --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+
     - name: Install dependencies (options=[${{ inputs.options }}], groups=[${{ inputs.groups }}])
       shell: bash
       run: |
@@ -50,9 +57,8 @@ runs:
         echo $cmd
         eval $cmd
 
-    - name: Upgrade to torch nightly
+    - name: Print torch nightly version
       if: ${{ inputs.torch_channel == 'nightly' }}
       shell: bash
       run: |
-        uv pip install --pre torch torchvision -P torch -P torchvision --extra-index-url https://download.pytorch.org/whl/nightly/cpu
         uv run python -c "import torch, torchvision; print('torch:', torch.__version__, '| torchvision:', torchvision.__version__)"

.github/workflows/checks.yml

@@ -45,6 +45,11 @@ jobs:
             extra_groups: 'lower_bounds'
           # Latest: torch nightly build.
           - torch_channel: 'nightly'
+          # Upcoming Python version, not yet supported by stable torch. cvxpy is excluded because one
+          # of its dependencies (sparsediffpy) has no wheel yet and can't be built without BLAS.
+          - python-version: '3.15'
+            torch_channel: 'nightly'
+            options: 'quadprog_projector fairgrad'
 
     steps:
       - name: Checkout repository

.gitignore

@@ -6,8 +6,8 @@ worktrees/
 CLAUDE.md
 .claude/
 
-# Profiling results
-tests/profiling/results/
+# Profiler results
+tests/profiler/results/
 
 # Trajectories results
 tests/trajectories/results/

tests/paths.py

@@ -3,5 +3,5 @@
 TORCHJD_DIR = Path(__file__).parent.parent
 TESTS_DIR = Path(__file__).parent
 PLOTS_RESULTS_DIR = TESTS_DIR / "plots" / "results"
-PROFILING_RESULTS_DIR = TESTS_DIR / "profiling" / "results"
+PROFILER_RESULTS_DIR = TESTS_DIR / "profiler" / "results"
 TRAJECTORIES_RESULTS_DIR = TESTS_DIR / "trajectories" / "results"

tests/profiling/plot_memory_timeline.py → tests/profiler/plot_memory_timeline.py

@@ -13,7 +13,7 @@
 matplotlib.use("Agg")  # Use non-GUI backend to avoid tkinter dependency
 import matplotlib.pyplot as plt
 import numpy as np
-from paths import PROFILING_RESULTS_DIR
+from paths import PROFILER_RESULTS_DIR
 
 
 @dataclass
@@ -63,7 +63,7 @@ def plot_memory_timelines(experiment: str, folders: list[str]) -> None:
     cpu_timelines = []
     cuda_timelines = []
     for folder in folders:
-        path = PROFILING_RESULTS_DIR / folder / f"{experiment}.json"
+        path = PROFILER_RESULTS_DIR / folder / f"{experiment}.json"
         cpu_timeline, cuda_timeline = extract_memory_timelines(path)
         cpu_timelines.append(cpu_timeline)
         cuda_timelines.append(cuda_timeline)
@@ -104,7 +104,7 @@ def plot_memory_timelines(experiment: str, folders: list[str]) -> None:
 
     fig.tight_layout()
 
-    output_dir = Path(PROFILING_RESULTS_DIR / "memory_timelines")
+    output_dir = Path(PROFILER_RESULTS_DIR / "memory_timelines")
     output_dir.mkdir(parents=True, exist_ok=True)
     output_path = output_dir / f"{experiment}.png"
     print(f"\nSaving plot to: {output_path}")

tests/profiling/run_profiler.py → tests/profiler/run_profiler.py

@@ -23,7 +23,7 @@
 )
 from utils.tensors import make_inputs_and_targets
 
-from tests.paths import PROFILING_RESULTS_DIR
+from tests.paths import PROFILER_RESULTS_DIR
 from torchjd.aggregation import UPGrad, UPGradWeighting
 from torchjd.autogram import Engine
 
@@ -98,7 +98,7 @@ def _save_and_print_trace(
     batch_size: int,
 ) -> None:
     filename = f"{factory}-bs{batch_size}-{DEVICE.type}.json"
-    output_dir = PROFILING_RESULTS_DIR / method_name
+    output_dir = PROFILER_RESULTS_DIR / method_name
     output_dir.mkdir(parents=True, exist_ok=True)
     trace_path = output_dir / filename