multi-agent-and-runner
Route skrl 2.1.0 IPPO/MAPPO multi-agent workflows, framework Runner configuration, model and memory dictionaries, trainer scopes, and distributed/checkpoint boundaries.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- multi-agent-and-runner
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Route skrl 2.1.0 IPPO/MAPPO multi-agent workflows, framework Runner configuration, model and memory dictionaries, trainer scopes, and distributed/checkpoint boundaries.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/multi-agent-workflows.md11.1 KB
- 参考文档references/runner-configuration.md10.9 KB
- 参考文档references/troubleshooting.md10.4 KB
03 SKILL.md 正文
Multi-agent and Runner
Use this route when the request names IPPO, MAPPO, PettingZoo or another multi-agent environment, per-agent model/memory dictionaries, a centralized state/value function, Runner/YAML configuration, trainer scopes, or multi-process/distributed execution. It owns the multi-agent contract and declarative Runner machinery; it does not replace environment wrapping or the single-agent framework guides.
Route quickly
- Establish a framework-wrapped multi-agent environment first. Confirm
possible_agents, currentagents,observation_spaces,state_spaces,action_spaces,device, andnum_envs. For PettingZoo or Isaac Lab, start with../environment-integration/SKILL.md. - Choose the learning semantics in
references/multi-agent-workflows.md: IPPO is decentralized training/execution with an independent value model; MAPPO is centralized training/decentralized execution through a state-based value model. Both expose the same per-agent constructor shape. - Build nested dictionaries keyed by every
possible_agentsentry:models[agent_id]["policy"],models[agent_id]["value"], and, for training,memories[agent_id]. A single-agent model dictionary is not a multi-agent dictionary. Validate role keys, spaces, and model initialization before constructing the algorithm. - For declarative setup, use
references/runner-configuration.md. A Runner needs a wrapped environment and top-levelmodels,memory,agent, andtrainerconfiguration. Treat YAML as trusted input: Runner evaluates selected component strings and expression fields. - Select a trainer deliberately. A single IPPO/MAPPO object normally uses
SequentialTrainer,StepTrainer, or direct loop ownership.ParallelTraineris a Torch multiprocessing facility, not the same thing as a multi-agent environment. Scope validation and simultaneous-agent limitations are in the workflow reference. - Before resuming, check framework-specific checkpoint format, experiment
directory, rank ownership of logging, and model-role compatibility. Use
references/troubleshooting.mdfor shape, key, migration, YAML, scope, and distributed failures.
Framework boundaries
- Torch: IPPO and MAPPO are available; Torch additionally supports a
policy/value shared model through
shared_modeland automatic mixed precision.ParallelTraineris Torch-only. - JAX: IPPO and MAPPO are available with separate policy/value models.
Runner does not support shared policy/value models and removes the
unsupported mixed-precision setting. Call
init_state_dictfor generated models before use. - Warp: the Warp Runner and sequential trainer cover single-agent Warp
agents. This release has no Warp IPPO/MAPPO mapping; do not route a Warp
multi-agent request to this path as if it were supported. See
../warp-agent-training/SKILL.mdfor the single-agent boundary. - Framework-specific model construction belongs to
../torch-agent-training/SKILL.md,../jax-agent-training/SKILL.md, or../warp-agent-training/SKILL.md. Link there for mixin or device details rather than duplicating those guides.
Completion checklist
A safe handoff has: a wrapper-derived, immutable possible_agents list; all
model and memory dictionaries covering it; policy/action and value/state space
compatibility; an explicit choice between IPPO and MAPPO; a finite trainer
timestep plan; validated scopes if multiple agents share a vectorized
environment; trusted Runner/YAML component names if used; explicit logging and
checkpoint paths; and a framework/distributed boundary statement. Do not run
full multi-agent training or an unreviewed user YAML merely to diagnose its
shape.
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).