Dim Sum & Demo at Ghirardelli Square with Traction Layer AI @ SGLang Summit
Supported SGLang: LTS branches, backported fixes, a hardened build and engineers to call.
A small seated dinner for teams that need an SGLang build they can stay on.
After day one of the SGLang Summit, join Traction Layer AI for dim sum at Ghirardelli Square. Dinner is at Palette Tea House, a Michelin Guide-listed dim sum house a short walk from Fort Mason, with a short talk and a live demo. Twenty seats, one room, and most of the evening left for conversation.
The question for the table: how do you stay on one SGLang version for a year and still get the security fixes?
Upstream ships a new release roughly every two weeks, and fixes land only on the newest one. The version you qualified in March gets no security patches in September. Your change board meets quarterly. Your security team wants a build it can verify for itself. If that is your situation, this dinner is for you.
On the menu
Talk: "Upstream SGLang, on a branch you can stand on." One long-term-support branch per quarter, twelve months of backported security and stability fixes, an end-of-life date published at cut, and CUDA 12.x and 13 lanes. Shankaran Sitarama, Co-founder & CEO, Traction Layer AI.
Live demo: a build you can verify. A signed image checked with cosign before it runs, and the hardened profile walked through against its published checklist: pickle off, admin routes keyed, ZeroMQ on localhost, non-root.
Also on the table: Traction Tuner. Scheduler settings adjusted against a live workload, with verify-or-revert on every change.
Open table. What you need backported, how long a branch should live, air-gapped installs, and what support has to look like before your team signs off.
The API does not change. It is the SGLang engine you already run: pinned, patched, hardened, signed and supported.
Schedule
6:00 PM: Doors, tea and dim sum
6:45 PM: Talk
7:05 PM: Live demo
7:30 PM: Dinner and open table
9:00 PM: Close
Who should come
Platform teams running SGLang in production who cannot upgrade every two weeks
Teams that qualify a version once and have to run it for a year, often under a quarterly change board
Security and compliance leads who have to approve an open inference engine
Operators of regulated, sovereign or air-gapped sites
Anyone still on CUDA 12 who needs a maintained lane
Requesting a seat
Complimentary. Seats are limited, so register
Confirmations go out on a rolling basis. If your plans change, release your seat so someone else can take it.
Vegetarian options are available. Tell us about dietary needs when you register.
Getting there
Palette Tea House, 900 North Point St, Suite B201A, Ghirardelli Square, San Francisco. A short walk east along the waterfront from Fort Mason Center.
About Traction Layer AI
Traction Layer AI builds Traction Runtime for SGLang: an independent downstream distribution of the upstream engine with quarterly long-term-support branches, backported security fixes, a hardened profile, signed builds and support from the engineers who build the branches. The API stays OpenAI-compatible. It runs where your data lives: your cloud account, your data center, or a site with no network at all.
NVIDIA Inception member. AWS, Microsoft Azure and Red Hat partner.
tractionlayer.ai · hello@tractionlayer.ai
SGLang is an open-source project released under the Apache License 2.0 and hosted by LMSYS. This is an independent side event. Traction Layer AI is not affiliated with, sponsored by or endorsed by the SGLang project.