ELSEIF
Your brief EB
525 stories from 214 feeds 1270 clusters Refreshed 12 minutes ago next pull 21:07

AI Signal 239 2 feeds carried it

vLLM TT Plugin adds Tenstorrent accelerator support with mesh-compiled execution

The vLLM TT Plugin integrates Tenstorrent accelerators into vLLM via the out-of-tree platform plugin mechanism, automatically discovering hardware when TT-Metal is importable and preserving the existing OpenAI-compatible API surface.

WHY IT MATTERS

This gives engineers a non-GPU path for LLM serving where parallelism is compiled into a single mesh program rather than configured as runtime ranks. The plugin demonstrates that vLLM's plugin interfaces are general enough to express hardware architectures that differ fundamentally from GPUs without modifying vLLM core.

Written by elseif from the cluster below · every claim links back to a source

The three things worth knowing

01

The plugin registers Tenstorrent hardware automatically when TT-Metal is importable, requiring no changes to client code or the OpenAI-compatible API.

02

Tenstorrent's mesh architecture compiles parallelism into a single traced program for the whole mesh, eliminating tensor-parallel and pipeline-parallel rank configuration.

03

Supported models include Llama 3.1/3.2/3.3, Qwen 2.5/3 variants, Mistral, Gemma 3/4, DeepSeek V3, and GPT-OSS, with multimodal coverage for several families.

THE CLUSTER

Same story, 2 feeds.

ORDERED BY FIRST SEEN
vLLM Blog Serving LLMs on Tenstorrent Hardware: Inside the vLLM TT Plugin Open ↗
vllm.ai via Lobsters Serving LLMs on Tenstorrent Hardware: Inside the vLLM TT Plugin Open ↗