측정: Qwen3.8 expert 라우팅 국소성 (층별 LRU-64/128 적중률)
Maintainers usually reply within 1 day
Nobody has claimed this yet.
Assessment
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Newbie friendliness
- 48/100
- Issue type
- Feature
- Clarity
- Clearly specified
- Activity status
- Active
- Tech stack
- rust
- Domain
- machine-learning
Research direction
Start by checking llama.cpp’s eval callback and the built model’s tensor names to confirm whether ffn_moe_topk exposes the per-token expert IDs. Run CPU offload on separate prose and code samples, each with thousands of tokens, and calculate per-layer LRU-64, LRU-128, and static top-32 hit rates. Report the results with the text sources; the LRU-64 result determines whether the residency assumption holds.
Written by the indexing model from the issue text.
Description
Qwen3.8(MoE)의 토큰별 top-k expert id를 실제 텍스트에서 모아, 층마다 expert 재사용이 얼마나 되는지 재 주세요. 적응형 residency(자주 쓰는 expert를 카드에 남기는 방식) 설계가 이 값을 전제로 합니다.
- 지금 가진 값: 층별 LRU-64 적중 약 67 %, LRU-128 약 81 %, 정적 top-32 적중 약 10 %. llama.cpp #27861에서 가져온 재생값이라 부풀었을 수 있습니다.
- 할 일: llama.cpp의 eval callback으로 MoE 층의 top-k id(텐서 이름은 빌드에서 확인해 주세요.
ffn_moe_topk로 보입니다)를 덤프합니다. 학습에 안 쓰였을 법한 산문과 코드 각각 수천 토큰을 넣습니다. - 보고: 층별 LRU-64, LRU-128 적중률과 정적 top-32 적중률. 산문과 코드를 따로, 사용한 텍스트 출처와 함께.
- 하드웨어: 라우팅은 하드웨어와 무관하니 CPU offload로 충분합니다. GPU가 없어도 됩니다.
- 판정: LRU-64가 60 % 이상이면 적응형 residency의 입력으로 그대로 씁니다. 40 % 미만이면 그 전제를 버리고 설계를 다시 봅니다.
작업을 맡으실 분은 댓글로 알려 주세요. AI로 편하게 쓰셔도 됩니다(CONTRIBUTING.ko.md).
- Dominant language
- Rust
- Stars
- 21
- Forks
- 4
- Avg merge
- 2h 42m
- Merged PRs (30d)
- 692
Getting set up
- No Dockerfile or Docker Compose file
- Has a pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from midagedev/bloomery
-
good first issue hardware help wanted no-box serve
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
Maintainers usually reply within 1 day
-
good first issue hardware help wanted no-box
Difficulty 2/5 1-3 hours Newbie friendliness 73/100
Maintainers usually reply within 1 day
-
model new-model
Difficulty 5/5 Over a week Newbie friendliness 20/100
Maintainers usually reply within 1 day
-
gate-gpu-qwen4exp-e2e q38dense red since b825b640c (BF16 residual streams)May be free again A pull request for this issue was closed without being merged. Open
Difficulty 4/5 1-2 days Newbie friendliness 48/100
midagedev/bloomery#935 · 2 comments ·
Maintainers usually reply within 1 day
-
gpu-evidence hardware help wanted needs-gpu
Difficulty 2/5 1-3 hours Newbie friendliness 58/100
midagedev/bloomery#926 · 2 comments ·
Maintainers usually reply within 1 day
All issues in midagedev/bloomery
Similar issues
-
good first issue help wanted r-docs
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
servo/rust-smallvec#824 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 83/100
Maintainers usually reply within 1 day
-
priority middle
Difficulty 1/5 Under an hour Newbie friendliness 62/100
KATO-Hiro/AtCoderClans#12940 ·
Maintainers usually reply within 1 day
-
bug
Difficulty 2/5 1-3 hours Newbie friendliness 83/100
Maintainers usually reply within 1 day
-
o8 next debug build aborts on Glass to Dark appearance switch (nil identifier messaged before check)Openbug
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
hurttlocker/o8#5112 ·