---

name: llm-vla-and-services description: "Use TorchRL LLM post-training, VLA data/actions, services, rendering, and weight-update surfaces safely." disable-model-invocation: true metadata: disco-role: operating license: MIT

TorchRL LLM, VLA, Services, and Rendering

Use this sub-skill when a task mentions TorchRL LLM post-training, RLHF, GRPO, SFT, chat environments, LLM inference wrappers, tool execution services, Vision-Language-Action data, action tokenizers, service registry/lifecycle, rendering, video outputs, or LLM/VLA weight synchronization.

Route by task

• For ChatEnv, History, PromptData, reward data, LLMCollector, TransformersWrapper, vLLMWrapper, SGLangWrapper, GRPO, SFT, distillation, policy version tracking, or vLLM/SGLang weight sync, read LLM and RLHF workflows.
• For validate_vla_tensordict, VLA canonical TensorDict keys, RobotDatasetMetadata, action chunking, UniformActionTokenizer, VocabTailActionTokenizer, ActionScaling, TinyVLA, LeRobot, or OpenX, read VLA data and actions and run check_vla_schema.py for a CPU-safe schema smoke test.
• For Service, get_services, owner/client lifecycles, direct/process/Ray placement, PythonExecutorService, PythonInterpreter, rlrender, render checkpoint loading, or video output, read services and rendering and run smoke_services.py for import/lifecycle checks.
• For optional extras, model downloads, GPU serving memory, tokenizer/template mismatch, policy-version drift, Ray cleanup, VLA shape/key failures, render codecs, or display issues, read troubleshooting.

Boundaries

This sub-skill owns TorchRL's LLM/VLA/service/render-specific surfaces and the safe checks bundled here. Route generic collector topology, replay buffers, and weight updaters that are not LLM-specific to the collectors-and-replay area. Route generic actor/critic/key wiring to modules-and-policies, and generic loss module or trainer workflows to objectives-and-training.

Optional GPU serving backends, downloaded model checkpoints, Ray clusters, robot datasets, and codec/display stacks are reference-only here unless the user explicitly provisions those dependencies and hardware.