Skip to content

Sglang fast recovery - #385

Open
liusy58 wants to merge 19 commits into
lm-sys:mainfrom
QiuMike:sglang_fast_recovery
Open

Sglang fast recovery#385
liusy58 wants to merge 19 commits into
lm-sys:mainfrom
QiuMike:sglang_fast_recovery

Conversation

@liusy58

@liusy58 liusy58 commented Aug 8, 2026

Copy link
Copy Markdown
Contributor

No description provided.

@alexnails alexnails left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

first pass :)

Comment thread blog/2026-07-28-sglang-fast-recovery.md Outdated
## Public Roadmap

The Weight Cache Daemon is just the **first step** — there is still a lot to build, and we are excited about the road ahead. Phase 1 today covers TP + PP, single- and multi-node launch, per-GPU zero-copy CUDA IPC, and unquantized plus block-wise FP8. Beyond that, many high-impact directions remain open:

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Another direction is other hardwares having this feature supported assuming they have similar functionality (AMD, Intel both have similar IPC mechanism)

Comment thread blog/2026-07-28-sglang-fast-recovery.md Outdated

## TL;DR

Nowadays, SOTA models are getting much bigger and reloading the model service after a crash is very expensive. Therefore, we introduce the **Weight Cache Daemon**, a persistent GPU process that holds post-quantized model weights in GPU memory and serves them to new SGLang engine instances via CUDA IPC zero-copy mapping. This reduces weight loading from minutes to sub-second times.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

State of the Art (SOTA)

Comment thread blog/2026-07-28-sglang-fast-recovery.md Outdated

## TL;DR

Nowadays, SOTA models are getting much bigger and reloading the model service after a crash is very expensive. Therefore, we introduce the **Weight Cache Daemon**, a persistent GPU process that holds post-quantized model weights in GPU memory and serves them to new SGLang engine instances via CUDA IPC zero-copy mapping. This reduces weight loading from minutes to sub-second times.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"Therefore, we are introducing"

Comment thread blog/2026-07-28-sglang-fast-recovery.md Outdated

## TL;DR

Nowadays, SOTA models are getting much bigger and reloading the model service after a crash is very expensive. Therefore, we introduce the **Weight Cache Daemon**, a persistent GPU process that holds post-quantized model weights in GPU memory and serves them to new SGLang engine instances via CUDA IPC zero-copy mapping. This reduces weight loading from minutes to sub-second times.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

weight loading from minutes to seconds.

liusy58 added 2 commits August 9, 2026 10:22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants