react-inference
Configure, validate, and troubleshoot Tongyi DeepResearch ReAct inference workflows, data, tools, and rollout outputs.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- react-inference
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Configure, validate, and troubleshoot Tongyi DeepResearch ReAct inference workflows, data, tools, and rollout outputs.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/data-formats.md4.8 KB
- 参考文档references/tool-contracts.md7.3 KB
- 参考文档references/workflows.md9.1 KB
- 脚本scripts/build_react_env.py15.9 KB
- 脚本scripts/validate_deepresearch_dataset.py9.8 KB
03 SKILL.md 正文
React Inference
Use this sub-skill when a user needs to set up, inspect, or troubleshoot Tongyi DeepResearch ReAct inference: .env values, JSON/JSONL input data, file references, local vLLM serving, OpenRouter or other OpenAI-compatible adaptation, rollout arguments, ReAct loop behavior, and tool contracts.
Do not use this sub-skill for benchmark judging or metric aggregation; route that to benchmark-evaluation. Do not use it to choose among WebAgent-family projects; route that to webagent-family. Do not run full local inference unless the user has explicitly provided model weights, GPUs, service credentials, and runtime approval.
Operating Procedure
- Classify the route.
- Local vLLM route: the repository launcher expects a local model path, eight visible CUDA devices, ports
6001through6008, and a.envfile. - OpenRouter/OpenAI-compatible route: future agents must adapt the ReAct model call in the working copy to the provider base URL, API key, model id, and reasoning-content concatenation convention; do not run the eight-port launcher for this route.
- Local vLLM route: the repository launcher expects a local model path, eight visible CUDA devices, ports
- Preflight configuration. Use
scripts/build_react_env.pyto print a safe template, generate an.env, or validate one. TreatMODEL_PATH,DATASET, andOUTPUT_PATHas strict fields. Treat Serper, Jina, summary API, Dashscope, and SandboxFusion credentials as required only when the corresponding tool is enabled or likely to be called. - Preflight input data. Use
scripts/validate_deepresearch_dataset.pyon JSON/JSONL files. If questions contain uploaded-file markers, supply the matchingfile_corpusdirectory and verify that all referenced files are present. - Explain execution shape before launching. The launcher loads
.env, validatesMODEL_PATH, starts eight vLLM servers, waits for/v1/modelsreadiness, then runsrun_multi_react.pywith rollout, split, worker, model, and sampling arguments. - Reason from the ReAct transcript. Model messages alternate assistant output with
<tool_call>...</tool_call>and user<tool_response>...</tool_response>. A valid final answer must be enclosed in<answer>...</answer>. - Troubleshoot by layer. Check configuration, dataset shape, model serving, external service credentials, tool-call JSON, file-corpus placement, token/time limits, and missing
<answer>termination in that order.
Bundled References
references/workflows.md— local vLLM and OpenAI-compatible routes,.envvariables, launcher behavior,run_multi_react.pyarguments, output layout, and ReAct loop.references/data-formats.md— JSON/JSONL input records, file references, file-corpus placement, and rollout output fields.references/tool-contracts.md—search,visit,google_scholar,PythonInterpreter, andparse_filecontracts and credentials.references/troubleshooting.md— symptom-driven checks for common setup, runtime, and ReAct termination failures.
Bundled Scripts
scripts/build_react_env.py --help— stdlib helper to generate or validate a DeepResearch-style.envwithout secrets or network calls.scripts/validate_deepresearch_dataset.py --help— stdlib helper to validate question/answer JSON or JSONL records and uploaded-file references.
Evidence Basis
This sub-skill distills the repository README, FAQ, environment example, requirements list, ReAct launcher, rollout runner, agent loop, prompt/tool sources, and the provided example JSONL/file-corpus fixtures. It is self-contained; future agents should not need to reopen those source files for routine inference setup questions.
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).