Conferences: How to Consume Without Attending¶

You will not attend these conferences. You have a Zoho job and a family. You will consume the papers and talks asynchronously. That is enough to stay current with the frontier.

The value is not the networking. The value is the filter — the top systems venues have already rejected 90% of the mediocre work. What survives peer review at MLSys/OSDI/SOSP is very likely worth reading.

The venue map¶

Venue	Focus	Cadence	Frontier for inference?
MLSys	ML systems (kernels, serving, training)	Annual, ~May	YES — primary
OSDI	Operating systems	Biennial (even years), ~July	YES — vLLM, DistServe, Sarathi-Serve here
SOSP	Operating systems	Biennial (odd years), ~Oct	YES — sibling to OSDI
ASPLOS	Arch + PL + OS	Annual, ~Apr	Some (GPU arch, memory)
ISCA	Computer architecture	Annual, ~June	Some (hardware papers)
HPCA	High-perf comp arch	Annual, ~Feb/Mar	Some (accelerators)
PPoPP	Parallel programming	Annual, ~Mar	Some (CUDA, kernels)
SC / Supercomputing	HPC	Annual, ~Nov	Some (large-scale training)
NeurIPS	ML — Efficient Natural Language Processing workshop	Annual, ~Dec	Workshops matter more than main track
ICML	ML — ES-FoMo (Efficient Systems for FMs) workshop	Annual, ~July	Workshops matter more than main track
ICLR	ML — quant/distillation papers	Annual, ~May	Some
USENIX ATC	Systems, wider scope than OSDI	Annual, ~July	Some
FAST	File & storage systems	Annual, ~Feb	Mooncake (Best Paper 2025) was here

Top-3 for inference engineering: MLSys, OSDI, SOSP. If you only follow three, follow those.

The workshops that matter more than the main track¶

For an inference engineer, the workshops are often more valuable than the main NeurIPS/ICML tracks. Main-track ML papers are 80% modeling. The efficiency workshops are 100% relevant.

ES-FoMo (Efficient Systems for Foundation Models) — ICML co-located. Started 2023. Home for FlashAttention-style work, quant, serving. Search: es-fomo <year> → the workshop site has the accepted papers list.
Efficient NLP — NeurIPS co-located workshop.
MLArchSys — ISCA co-located.
HAET (Hardware Aware Efficient Training) — ICML/NeurIPS.
CoLM — Conference on Language Modeling. New (2024+). Emerging venue for LM-specific systems work.

Rule: when a workshop’s accepted-papers list drops, skim titles in 15 min. Flag 1-2 for full reading.

How to consume each conference (the 3-hour protocol)¶

For each edition you care about, spend ~3 hours total:

Titles pass (30 min) — Open the accepted-papers page. Read every title. Star anything with keywords: serving, inference, kernel, attention, quant, KV cache, MoE, speculative, pipeline, disaggregat*, roofline, GPU, H100/B200, parallelism.
Abstracts pass (60 min) — Read abstracts for the ~10-20 starred titles. Cut to ~5 papers.
Video pass (60 min) — Watch author talks on YouTube for the 5. Most systems venues post talks 2-4 months after the conference.
Deep-read pass (30 min or defer) — Add the top 1-2 papers to your reading queue. Deep-read on a weekend.

Total: one lazy Saturday morning per conference edition. Do this ~5x/year and you are current.

Where the accepted-papers lists live¶

Bookmark these pages once and revisit each year:

MLSys — https://mlsys.org/ — click Accepted Papers.
OSDI — https://www.usenix.org/conference/osdi (biennial: OSDI 20/22/24/26).
SOSP — https://sosp.org/ (biennial: SOSP 21/23/25/27).
ASPLOS — https://asplos-conference.org/.
ISCA — https://iscaconf.org/.
HPCA — https://hpca-conf.org/.
PPoPP — https://ppopp.sigplan.org/.
SC — https://sc.supercomputing.org/.
NeurIPS — https://neurips.cc/, then the workshop page for ENLSP / ES-FoMo.
ICML — https://icml.cc/, then the workshop page for ES-FoMo.
USENIX ATC / FAST — https://www.usenix.org/conferences.

How to find the videos: USENIX venues (OSDI/ATC/FAST/SOSP) put talks on their site + YouTube channel @USENIXAssociation. MLSys uploads to @MLSysConf. NeurIPS/ICML videos go on NeurIPS/ICML sites (paywalled for a year, then free on YouTube).

Landmark inference papers by venue (historical anchors)¶

Use this to calibrate what “good” looks like at each venue.

MLSys:

FlashAttention (Dao et al., 2022 poster + NeurIPS 22) — MLSys 23 best paper adjacent work.
AWQ (Lin et al., MLSys 2024) — best paper.
FlashInfer (Ye et al., MLSys 2025) — best paper.

OSDI:

Orca (Yu et al., OSDI ‘22) — continuous batching origin.
Sarathi-Serve (OSDI ‘24) — chunked prefill.
DistServe (OSDI ‘24) — prefill/decode disaggregation.

SOSP:

vLLM / PagedAttention (Kwon et al., SOSP ‘23) — the paper you’re standing on.

FAST:

Mooncake (2025 best paper) — KV-cache-centric disaggregated architecture.

NeurIPS:

SGLang / RadixAttention (NeurIPS ‘24) — programmable serving.
The whole speculative-decoding lineage (Leviathan, Medusa, EAGLE).

ICLR:

GPTQ, AWQ (also MLSys), SmoothQuant precursors.

Notice the pattern: the systems venues (OSDI/SOSP/MLSys/FAST) are where the infrastructure breakthroughs land. The ML venues (NeurIPS/ICML/ICLR) are where the algorithm breakthroughs land. As an inference engineer you need both, but the systems venues are your primary.

Annual rhythm (mark your calendar once)¶

Month	What lands	Action
Feb	FAST accepted list, HPCA papers	Skim FAST for storage/KV-cache papers
Mar	HPCA proceedings, PPoPP videos	Watch 1-2 talks
Apr	ASPLOS, ICLR camera-ready	Skim ICLR quant papers
May	MLSys — the big one	3-hour deep dive on accepted list
June	ISCA	Skim GPU arch papers
July	OSDI (even years) / USENIX ATC, ICML + ES-FoMo workshop	3-hour OSDI dive. Read ES-FoMo.
Aug	MLSys videos on YouTube	Watch the 5 starred talks
Sept	Camera-ready season for NeurIPS	—
Oct	SOSP (odd years), MLSys re-review	3-hour SOSP dive if odd year
Nov	SC (Supercomputing), OSDI/SOSP videos	Watch remaining talks
Dec	NeurIPS + ENLSP workshop	3-hour NeurIPS dive

Total time investment: ~15 hours/year for full frontier awareness. That is one work-day. Non-negotiable.

What to do at each conference (if you do attend one, someday)¶

If Zoho ever sends you or you decide to spend PTO+savings:

MLSys is the one. ~500 attendees, hallway-track culture is welcoming, systems people are approachable, poster sessions have actual conversations.
OSDI/SOSP — bigger, more academic, harder to break in as an industry person. But: legendary talks, papers you will actually read for the rest of your career.
NeurIPS/ICML — do not attend for systems work. Too big (~15k people), too model-focused. Only go if you have a workshop paper.
NVIDIA GTC — free (or cheap), San Jose in March, corporate but excellent tech talks. Not peer-reviewed but you’ll learn a lot about H200/B200/Blackwell realities.

Rule of thumb: don’t attend a conference until you have a paper, a poster, or a talk. Before that, the video archive is 90% of the value at 0% of the cost.

The counterfactual: don’t over-index on conferences¶

Two failure modes:

Conference-tourism — treating “read all of MLSys 2026” as a checkbox. You will not remember 90% of it. Read 5 papers deeply, not 50 shallowly.
Frontier-chasing — reading every new paper the day it drops. Most conference papers don’t matter in 2 years. Wait for citations. If a paper has >200 citations after a year, it’s real. Under that, defer.

The papers in 09_papers/ are the already-filtered canon. Your conference consumption is to add 3-5 papers per year to that canon, not to replace it.

Cross-references¶

Blog canon: 10_communities/04_blog_canon.md — many blog posts are conference-paper distillations.
YouTube: 10_communities/06_youtube_channels.md — MLSys/USENIX YouTube channels are the video sources.
Paper canon: 09_papers/ — the historical anchors above are already in there.

Verified. Actionable. This is not the section that wins you the job — it’s the section that keeps you from becoming obsolete after you get it.