25launches

The first launch of tamserve.

tamserve page

tamserve

Launching today

Multi-camera object tracking with about 4x less GPU memory

Hunted by IlyasHunter
Visit github.com (opens in a new tab)

tamserve is a Python serving engine that runs Meta's SAM 2.1 and EdgeTAM tracking models on live video from many cameras. One copy of the model serves every camera, batching their frames the way vLLM batches chat requests, and memory stays flat however long a stream runs. Point at an object, or name it and SAM 3 finds it. With EdgeTAM on one RTX 4090, 8 cameras took 2.2 GB of GPU memory instead of 8.6 GB with the original code, at the same accuracy. Needs an NVIDIA GPU. Apache-2.0.

Meet the team

Hunted by Ilyas. The makers are not on 25launches yet.

H (Hunter)
  • tamserve screenshot 1
  • tamserve screenshot 2
  • tamserve screenshot 3
  • tamserve screenshot 4

Discussion 1

Ask questions, give feedback. Comments don't count in the ranking.

Sign in to join the discussion.

  • IlyasHunter Pinned

    Putting SAM 2 on several live cameras usually means one copy of the model per camera, and GPU memory runs out fast. tamserve shares one copy across all of them, the way LLM servers batch requests, without changing what the model outputs. The 4x figure is the author's own benchmark (8 cameras, EdgeTAM, one RTX 4090), and the repo already has 261 stars.