[Feature] Tool services

.github/unittest/llm/scripts_llm/environment.yml

@@ -22,3 +22,4 @@ dependencies:
     - transformers
     - datasets
     - vllm
+    - mcp

.github/unittest/llm/scripts_llm/install.sh

@@ -61,3 +61,17 @@ python -m pip install -e . --no-build-isolation
 
 # smoke test
 python -c "import torchrl"
+
+# Install MCP dependencies for tool execution tests
+printf "* Installing MCP dependencies (uvx, Deno)\n"
+
+# Install uvx (universal package runner)
+pip install uvx
+
+# Install Deno (required by mcp-run-python)
+curl -fsSL https://deno.land/install.sh | sh
+export PATH="$HOME/.deno/bin:$PATH"
+
+# Verify installations
+uvx --version || echo "Warning: uvx not installed"
+deno --version || echo "Warning: Deno not installed"

docs/source/reference/index.rst

@@ -10,6 +10,7 @@ API Reference
     llms
     modules
     objectives
+    services
     trainers
     utils
     config

docs/source/reference/llms.rst

@@ -930,7 +930,9 @@ Tools are usually implemented as transforms, and appended to a base environment
 such as :class:`~torchrl.envs.llm.ChatEnv`.
 
 An example of a tool transform is the :class:`~torchrl.envs.llm.transforms.PythonInterpreter` transform, which is used
-to execute Python code in the context of the LLM.
+to execute Python code in the context of the LLM. The PythonInterpreter can optionally use a shared 
+:class:`~torchrl.envs.llm.transforms.PythonExecutorService` for efficient resource usage across multiple environments.
+See :ref:`ref_services` for more details on the service registry system.
 
     >>> from torchrl.envs.llm.transforms import PythonInterpreter
     >>> from torchrl.envs.llm import ChatEnv
@@ -1141,6 +1143,7 @@ By following these design principles, reward transforms can be effectively integ
     KLRewardTransform
     MCPToolTransform
     PolicyVersion
+    PythonExecutorService
     PythonInterpreter
     RayDataLoadingPrimer
     RetrieveKL