expert-cache
Debug and extend mixtral-offloading's MoE expert cache, storage wrapper, and sparse routing internals.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- expert-cache
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Debug and extend mixtral-offloading's MoE expert cache, storage wrapper, and sparse routing internals.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/expert-cache-api.md2.9 KB
- 参考文档references/storage-and-routing.md2.3 KB
- 参考文档references/troubleshooting.md2.2 KB
- 脚本scripts/inspect_cache_plan.py2.4 KB
03 SKILL.md 正文
Expert Cache
Use this sub-skill when the task is about expert offloading internals rather
than simply running the demo: ExpertCache, MixtralExpertWrapper,
SparseMoeWrapper, expert UIDs, LRU eviction groups, pinned offload storage,
or cache-capacity errors.
Read this when
- The user sees
Cache is full,No evictable experts,already loading experts, duplicate expert UID assertions, or eviction-group assertions. - The task is to change offloading policy, add prefetching, debug storage movement, or inspect sparse MoE routing.
- The user wants to understand how
offload_per_layerbecomesmain_sizeandoffload_sizeinsideExpertCache.
Route map
- Read references/expert-cache-api.md for
ExpertInfo,EvictionGroupInfo,ExpertCache, and method invariants. - Read references/storage-and-routing.md
for
MixtralExpertWrapperstorage replacement andSparseMoeWrapper.forwardrouting flow. - Use scripts/inspect_cache_plan.py to compute cache capacities and example expert UIDs without loading model weights.
- Read references/troubleshooting.md for capacity, reentrancy, storage, pinned-memory, and routing-shape failures.
What this sub-skill owns
- Expert-cache capacity and eviction invariants.
- Expert UID layout and eviction-group behavior.
- Storage wrapping and state-dict storage hooks.
- Sparse MoE top-k routing through cached experts.
What to route elsewhere
- Use ../inference-workflow/SKILL.md for full model construction, generation, and offload config selection.
- Use ../quantization-kernels/SKILL.md for HQQ quantized MLP layers, packing, and Triton kernel failures.
Safe verification
Most cache internals require real MixtralExpertWrapper objects backed by
CUDA UntypedStorage. For safe planning, use the bundled script and source/API
reasoning. Do not construct the full cache with dummy objects unless their
storage type, size, and device match the expected wrapper invariants.
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).