Repository navigation
vllm: serve qwen3.6-35b-a3b on demand via llama-swap - #1006
Merged
Merged
Conversation
The always-on vLLM service pinned most of the A6000 even when idle, which blocks other GPU jobs such as structure prediction. llama-swap starts the container on the first request and stops it after ten minutes, so VRAM is only held while an agent is working. vLLM runs from the upstream image pinned by digest: nixpkgs vLLM trails new model support, and every CUDA torch/vLLM bump is a multi-hour rebuild for a leaf service nothing else depends on. Qwen3.6-35B-A3B keeps a KV cache in only 10 of 40 layers, so FP8 weights plus fp8 KV still fit the full 262144 context on 48GB with room for parallel sub-agent requests. The API listens on Naru only, which spans our own hosts and already encrypts and authenticates peers, so nginx, synced certificates and Tailnet DNS are unnecessary. A sops-provided API key still guards against a compromised peer, since agents execute arbitrary code.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The always-on vLLM service pinned most of the A6000 even when idle, which blocks other GPU jobs such as structure prediction. llama-swap starts the container on the first request and stops it after ten minutes, so VRAM is only held while an agent is working.
vLLM runs from the upstream image pinned by digest: nixpkgs vLLM trails new model support, and every CUDA torch/vLLM bump is a multi-hour rebuild for a leaf service nothing else depends on. Qwen3.6-35B-A3B keeps a KV cache in only 10 of 40 layers, so FP8 weights plus fp8 KV still fit the full 262144 context on 48GB with room for parallel sub-agent requests.
The API listens on Naru only, which spans our own hosts and already encrypts and authenticates peers, so nginx, synced certificates and Tailnet DNS are unnecessary. A sops-provided API key still guards against a compromised peer, since agents execute arbitrary code.