moe
Routes Hivemind remote expert and mixture-of-experts workflows, including server setup, custom expert registration, and client-side expert routing.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- moe
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Routes Hivemind remote expert and mixture-of-experts workflows, including server setup, custom expert registration, and client-side expert routing.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/api-reference.md4.2 KB
- 参考文档references/troubleshooting.md3.8 KB
- 参考文档references/workflows.md3.3 KB
03 SKILL.md 正文
Mixture of Experts
Use this route when the task is about hivemind.moe.Server, ModuleBackend, RemoteExpert, RemoteMixtureOfExperts, RemoteSwitchMixtureOfExperts, custom expert registration, or the hivemind-server command.
Include
- Hosting experts with
hivemind.moe.server.Serverorbackground_server. - Remote expert clients with
RemoteExpertand the MoE routing layers. - Declaring and discovering experts with
declare_expertsandget_experts. - Custom expert classes via
register_expert_classandcustom_module_path. - Checkpointing and backend scheduling behavior in
ModuleBackend. - The
hivemind-serverconsole command.
Exclude
- DHT bootstrap and network reachability debugging; route that to
dht. - Optimizer wrapping, tensor averaging, and ALBERT training; route those to
collaborative-training. - Benchmarks and throughput experiments.
Start here
- Read
references/api-reference.mdfor the verified server/client signatures. - Read
references/workflows.mdfor the standard server, client, and custom-expert flows. - Read
references/troubleshooting.mdwhen the server cannot publish experts, clients cannot find experts, or a custom module fails to load. - Run
../../scripts/check_install.pyfrom the generated skill tree to confirm the installed package and CLI entry points.
What to remember
hivemind-serveris the package command for serving experts.background_server(...)is the fastest way to spin up a temporary expert host in tests or small demos.client_modematters when a client can dial out but should not accept inbound connections.custom_module_pathlets the server load extra expert definitions from a Python file.Server.create(...)accepts either explicit expert UIDs or generated patterns.- On CUDA-capable hosts, the server may default to GPU execution unless you request
--device cpu.
Good follow-up questions
- "How do I host experts and connect to them remotely?"
- "How do I register a custom expert class?"
- "How do I fetch experts from the DHT?"
- "Why does
hivemind-serversay an expert is missing?" - "How do I checkpoint expert state?"
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).