agilerl 2.7.0.dev1__tar.gz → 2.7.0.dev3__tar.gz

{agilerl-2.7.0.dev1 → agilerl-2.7.0.dev3}/.github/workflows/linux-tests.yml

@@ -73,22 +73,20 @@ jobs:
             - name: Reset coverage data
               run: rm -f .coverage .coverage.*
 
-            - name: Run non-LLM tests (parallel)
-              run: uv run pytest -m "not llm" --exitfirst --cov=agilerl --cov-report= --durations=0 --durations-min=1.0
-
-            - name: Run LLM tests (sequential, exclusive GPU)
-              run: uv run pytest -n0 -m "llm" --exitfirst --cov=agilerl --cov-append --cov-report= --durations=0 --durations-min=1.0
-
-            - name: Combine coverage data
-              run: |
-                  if ls .coverage.* >/dev/null 2>&1; then
-                      uv run coverage combine
-                  else
-                      echo "No parallel coverage shards found; skipping combine."
-                  fi
-
-            - name: Generate coverage report
-              run: uv run coverage xml -o coverage.xml
+            # Single phase: `-n auto --dist loadgroup` from pyproject gives 8
+            # workers on the gha-runner-scale-set node. `vllm`- and `gpu`-marked
+            # tests share the 4 `gputest0..gputest3` xdist groups defined in
+            # `tests/conftest.py`, which caps GPU-touching concurrency at 4
+            # workers regardless of -n; the remaining ~4 workers fan out across
+            # CPU-only tests. See the `pytest_collection_modifyitems` docstring
+            # for the GPU-memory and port-race rationale behind the 4-group cap.
+            #
+            # Single invocation = single coverage run; pytest-cov auto-combines
+            # the per-worker `.coverage.*` shards before writing `coverage.xml`,
+            # so we don't need a manual `coverage combine` step (which was
+            # tripping over corrupted shards under the old two-phase setup).
+            - name: Run tests
+              run: uv run pytest --exitfirst --cov=agilerl --cov-report=xml --durations=0 --durations-min=1.0
 
             - name: Upload coverage reports to Codecov
               uses: codecov/codecov-action@v3

{agilerl-2.7.0.dev1 → agilerl-2.7.0.dev3}/.github/workflows/macos-tests.yml

@@ -42,7 +42,6 @@ jobs:
                   python-version: ${{ matrix.python-version }}
 
             - name: Install dependencies
-              # swig is needed to build box2d-py from source (no pre-built wheels for py3.10+).
               run: |
                   uv sync --locked --all-groups --extra all
 

{agilerl-2.7.0.dev1 → agilerl-2.7.0.dev3}/.gitignore

@@ -149,3 +149,7 @@ wandb/
 
 # Cursor
 .cursor
+
+# AI assistant instructions
+CLAUDE.md
+AGENTS.md

{agilerl-2.7.0.dev1 → agilerl-2.7.0.dev3}/.pre-commit-config.yaml

@@ -1,6 +1,9 @@
 ---
 # See https://pre-commit.com for more information
 # See https://pre-commit.com/hooks.html for more hooks
+
+# Vendored binary test fixtures; opaque to every hook here.
+exclude: ^tests/assets/tiny_llm/
 ci:
     autoupdate_branch: nightly
     skip: [codeql-python]
@@ -32,7 +35,7 @@ repos:
                 - --skip=*.css,*.js,*.map,*.scss,*.svg
                 - --ignore-words-list=magent,pres,roate
     - repo: https://github.com/astral-sh/ruff-pre-commit
-      rev: v0.15.11
+      rev: v0.15.12
       hooks:
           - id: ruff
             name: Ruff Linter
@@ -47,7 +50,7 @@ repos:
 
     - repo: https://github.com/astral-sh/uv-pre-commit
       # uv version.
-      rev: 0.11.7
+      rev: 0.11.8
       hooks:
           - id: uv-lock
 

{agilerl-2.7.0.dev1 → agilerl-2.7.0.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agilerl
-Version: 2.7.0.dev1
+Version: 2.7.0.dev3
 Summary: AgileRL is a deep reinforcement learning library focused on improving RL development through RLOps.
 Author-email: Nick Ustaran-Anderegg <dev@agilerl.com>
 License-Expression: Apache-2.0

{agilerl-2.7.0.dev1 → agilerl-2.7.0.dev3}/agilerl/algorithms/cispo.py

@@ -2,7 +2,6 @@
 
 from __future__ import annotations
 
-from functools import partial
 from typing import Any
 
 from agilerl.algorithms.grpo import GRPO, _signatures_without_loss_type
@@ -14,11 +13,9 @@ class CISPO(GRPO):
     Paper: https://arxiv.org/abs/2506.13585
     """
 
-    _init_with_cispo = partial(GRPO.__init__, loss_type="cispo")
-
     def __init__(self, *args: Any, **kwargs: Any) -> None:
         """Initialize a CISPO agent with fixed ``loss_type``."""
-        self._init_with_cispo(self, *args, **kwargs)
+        super().__init__(*args, loss_type="cispo", **kwargs)
 
 
 _CISPO_CLASS_SIG, _CISPO_INIT_SIG = _signatures_without_loss_type()