torch-agent-training
Route and compose public single-agent PyTorch agents, models, memories, resources, trainers, and experiment controls in skrl 2.1.0.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- torch-agent-training
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Route and compose public single-agent PyTorch agents, models, memories, resources, trainers, and experiment controls in skrl 2.1.0.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/algorithm-and-component-selection.md9.5 KB
- 参考文档references/troubleshooting.md10.0 KB
- 参考文档references/workflow-and-api.md11.4 KB
- 脚本scripts/torch_ppo_components.py5.0 KB
03 SKILL.md 正文
Torch agent training
Use this route for a single-agent PyTorch skrl 2.1.0 workflow: choose an
algorithm, build its model-role dictionary, allocate rollout/replay memory,
configure preprocessors/noises/schedulers, select a trainer, and make a safe
train/evaluate/resume plan. Start with the bounded component check in
scripts/torch_ppo_components.py, then use
the references for the exact role keys and construction details.
Route quickly
- Establish the observation, optional state, and action spaces through the
environment route; this skill does not choose or implement Gymnasium
wrappers. See
../../references/framework-selection.mdand../../references/troubleshooting.mdin the root skill before committing to a backend or trainer. - Choose an algorithm and its required model roles in
references/algorithm-and-component-selection.md. Match the policy distribution to the action space and use a deterministic value/critic model where the table requires it. - Compose the models,
RandomMemory(when required), configuration, and agent usingreferences/workflow-and-api.md. Keepmemory_sizeand the agent's rollout/update settings consistent. - Use
SequentialTrainerfor the normal loop,StepTrainerwhen the caller owns each iteration, andParallelTraineronly after checking its process, pickling, shared-memory, and hardware constraints. - Disable writes during a smoke/evaluation probe, load a checkpoint into an
architecture-equivalent agent, and only then enable training or logging.
Check workflow-specific failure modes in
references/troubleshooting.md.
Boundaries
- Environment wrappers and space conversion: route to the
environment-integrationsibling. - IPPO/MAPPO, multi-agent trainers, and Runner/YAML: route to
../../sub-skills/multi-agent-and-runner/SKILL.md. - JAX and Warp: not covered here.
- Framework-wide backend selection and package-wide install/import diagnosis belong to the root route; this sub-skill owns Torch-specific composition.
Completion checklist
Before handing a workflow to a researcher, verify: torch and skrl import;
all model dictionary keys are exact; each model's input/output role is
compatible with the space; the memory device and environment count agree; the
trainer config has finite, intentional timesteps; logging/checkpoint paths
are explicitly chosen; and resume/evaluation uses an architecture-compatible
agent. A CPU smoke establishes import and CPU behavior only—it does not
prove CUDA or another accelerator.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).