ML-systems researcher serving 35B MoEs from SSD at 20 tok/s inside 3 GiB — pushing inference past the memory wall
1
papers here
1
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Insufficient data to score publication output. This axis would measure total volume and consistency of research production — papers per year across venues, first/corresponding-author share, and coverage beyond the single arXiv paper Evolve has reviewed so far.
Impact—
Not yet scoreable: citation counts and downstream influence are absent from the current snapshot. This axis would measure citations, adoption of Edge0's one-token-ahead routing-prediction approach, and uptake of SSD-based MoE serving by other groups.
Focus100/100
Scored 100 because all reviewed output sits in a single field: ML systems / efficient large-model inference (field distribution 1/1). The one catalog paper targets MoE serving under extreme memory limits, a sharply defined problem. With only one reviewed paper, this perfect concentration is an early-stage signal rather than an established specialization.
Collaboration32/100
Scored 32 from a distinct co-author count of 4 on the reviewed paper: Yiming Wang, Runyuan Cai, Hanze Liu, and Xiaodong Zeng. The first-author paper reflects a small, close-knit team. In absolute terms this is a limited footprint, but with one paper in the catalog the score says more about sample size than about collaboration style.
Activity—
No score yet: the snapshot lacks publication-cadence data. This axis would measure research pace over time — yearly paper counts, submission recency, sustained productivity — none of which a single 2026 arXiv paper can establish.
Reproducibility100/100
Scored 100 from reproMean 100 over 1 reproducibility-assessed paper; the reviewed work is rated high for reproducibility readiness. Evolve's review describes it as unusually honest systems work, consistent with openly stated limitations such as missing external-offloading baselines. As with focus, the perfect score rests on a single-paper base.
Research summary
Yu Lin is an early-stage machine-learning-systems researcher whose reviewed work centers on efficient inference for large sparse models — specifically serving mixture-of-experts architectures under severe memory constraints. Evolve's catalog currently holds one paper, first-authored: 'The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction' (arXiv, 2026), co-authored with Yiming Wang, Runyuan Cai, Hanze Liu, and Xiaodong Zeng. Its system, Edge0, runs a 35B-parameter MoE from an SSD at 20 tokens per second inside 3 GiB of memory. The recipe combines routing prediction trained one token ahead of demand, an unmerged recovery LoRA distilled to absorb mispredictions, and INT4 quantization. Evolve's verdict calls it 'solid, unusually honest systems work' while flagging that headline gains lack external-offloading baselines; reproducibility readiness is rated high, yielding a repro mean of 100 on a single paper. Bibliometric signals mirror this narrow footprint: all reviewed output sits in one field (focus 100), collaboration spans four distinct co-authors (score 32), and output, impact, and activity axes are unscored for insufficient data. Because 'Yu Lin' is a common name, this profile is resolved strictly from public bibliographic data and the reviewed paper.
Collaboration pattern
All reviewed work is a first-author collaboration with four co-authors — Yiming Wang, Runyuan Cai, Hanze Liu, and Xiaodong Zeng — giving a distinct co-author count of 4 and a collaboration score of 32 on Evolve's 0–100 scale. The single reviewed paper points to a small, tightly-knit team typical of early-career systems groups; broader collaboration patterns cannot be assessed until more papers enter the catalog.
Research keywords
Highlights
Serving a 35B MoE from SSD inside 3 GiB
One-token-ahead routing prediction with a recovery LoRA
Rated honest, highly reproducible systems work
Timeline
- 2026Publishes first-author arXiv paper 'The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction' with co-authors Yiming Wang, Runyuan Cai, Hanze Liu, and Xiaodong Zeng.
- 2026The paper's Edge0 system serves a 35B MoE from SSD at 20 tok/s inside 3 GiB, using one-token-ahead routing prediction plus an unmerged recovery LoRA.
- 2026Evolve's review rates the work solid and unusually honest, with high reproducibility readiness (repro mean 100 over 1 paper), while flagging missing external-offloading baselines.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- 'Yu Lin' is a common name; this profile is name-resolved from public bibliographic data and Evolve's reviewed papers only — no institution, country, or ORCID is on record.
- Evidence base is one reviewed arXiv paper (2026); the focus (100) and reproducibility (100) scores rest on that single-paper sample.
- Output, impact, and activity dimensions are unscored because citation counts, publication cadence, and volume data are absent from the current snapshot.
- Coverage is limited to Evolve-reviewed work from arXiv; publication venues, career history, and non-reviewed papers are not represented.