rollout-and-tools
Use when modifying verl rollout engines, multi-turn agent loops, function tools, tool schemas, rollout traces, or generation-server flows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- rollout-and-tools
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use when modifying verl rollout engines, multi-turn agent loops, function tools, tool schemas, rollout traces, or generation-server flows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/rollout-workflows.md4.4 KB
- 参考文档references/tooling-api.md4.2 KB
- 参考文档references/troubleshooting.md3.5 KB
- 脚本scripts/validate_function_tool.py3.7 KB
03 SKILL.md 正文
Rollout And Tools
Use this sub-skill when the task touches verl rollout behavior rather than trainer optimizer policy: rollout backend config, async generation servers, multi-turn agent loops, tool calling, function-tool schema authoring, tokenization sanity checks, trace output, or rollout visualization.
Route The Task
- For rollout backend choices and config fields, read Rollout workflows.
- For
@function_tool,BaseTool, schema loading, and tool dispatch, read Tooling API. - For tokenization warnings, tool parser issues, trace volume, and viewer pitfalls, read Troubleshooting.
- To preflight a Python file used by
actor_rollout_ref.rollout.multi_turn.function_tool_path, run validate_function_tool.py.
Safe Defaults
- Prefer
actor_rollout_ref.rollout.mode=async; sync rollout mode is removed. - Use
actor_rollout_ref.rollout.multi_turn.enable=Trueplusactor_rollout_ref.rollout.agent.default_agent_loop=tool_agentfor built-in tool-calling loops. - Use function tools for stateless
fn(**parameters)calls; use nativeBaseTooltools for lifecycle, per-trajectory state, sandbox resources, ortools_kwargs. - Keep
tokenization_sanity_check_mode=strictuntil the model chat template has been validated; useignore_strippableonly for whitespace-only template drift. - Treat rollout viewer as an interactive diagnostic over generated JSONL rollout data, not as a required runtime dependency.
Verification Targets
- CPU-safe candidates: function-tool registration/return tests,
ToolAgentLoop._call_toolerror-handling tests, Qwen3 parser safety tests, and vLLM CLI argument parsing tests. - When changing multi-turn or tool parsing behavior, include a case where a malformed tool call returns a readable
ToolResponseinstead of crashing the loop.
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).