返回资讯中心
外部精选
软件工程
#开发者实时流

PyTorch-native TPU back end for vLLM

| Documentation | Blog | User Forum | Developer Slack (#sig-tpu) |

Hacker News 最新zniturah2026/10/19 · 27 天前9 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

| Documentation | Blog | User Forum | Developer Slack (#sig-tpu) |

vLLM TorchTPU is the PyTorch-native TPU backend for vLLM. It is an out-of-tree vLLM platform plugin, packaged as vllm-torchtpu, that runs vLLM's standard PyTorch model definitions on Google Cloud TPUs through TorchTPU (torch_tpu, via torch.compile(backend="tpu")) and a library of custom JAX/Pallas kernels for attention, Mixture-of-Experts, and quantized matmul.

The goal is simple: the same vLLM you already use — same CLI, same OpenAI-compatible server, same model definitions — running at TPU speed with no model code changes.

Looking for the JAX backend? The original TPU backend for vLLM, tpu-inference, has been rebranded as vLLM JAX. It remains available for serving pure JAX model implementations on TPU. vllm-torchtpu is the recommended path for serving PyTorch model definitions on TPU.

TPU generation Status

TPU v7x (Ironwood) Recommended — validated in CI on every pull request and nightly

TPU v6e / v5e / v5p Experimental — not covered by CI

See Cloud TPU Setup for project setup, quota, and provisioning commands.

Install

vllm-torchtpu requires Python 3.12 on a Cloud TPU VM. Install the plugin and a matching vLLM with a single command:

pip install
--extra-index-url https://download.pytorch.org/whl/cpu
vllm-torchtpu

For uv, editable source installs, cross-repo development against a local vllm checkout, and Docker images, see the full Installation guide and Docker guide.

Start an OpenAI-compatible server exactly as you would on any other vLLM platform:

vllm serve meta-llama/Llama-3.2-1B-Instruct
--tensor-parallel-size 1
--max-model-len 4096

Then send a request:

A stage may instead span a prefill context-parallel group (--prefill-context-parallel-size), which splits each chunk's attention over the stage's cores; MoE models then need --enable-expert-parallel. This needs a torch_tpu that compiles a stage's kernels and collectives for the stage's devices only (the 2026-09-28 build of google-pytorch/torch_tpu#4387); on a torch_tpu that builds world-size meshes only, the platform refuses such stages. A two-core stage must not be the two cores of one chip: set VLLM_TPU_DEVICE_ORDER to the device each local rank binds, e.g. 0,2,1,3,4,6,5,7 for four stages of two cores on one host.

curl http://localhost:8000/v1/chat/completions
-H "Content-Type: application/json"
-d '{ "model": "meta-llama/Llama-3.2-1B-Instruct", "messages": [{"role": "user", "content": "Explain how TPUs accelerate matrix multiplication."}], "max_tokens": 128 }'

Tip

The first startup compiles TPU graphs for the configured shapes and can take several minutes. Set VLLM_XLA_CACHE_PATH to persist compiled artifacts across restarts — see Compilation & Warmup Cache.

The Quickstart covers offline batch inference with vllm.LLM, multi-device serving with TP/EP, and multimodal examples.

vllm-torchtpu can serve most PyTorch models supported by vLLM, with first-class TPU kernel, quantization, and parallelism support for Qwen3 / Qwen3.5, DeepSeek-V3 / R1 / V4, Kimi-K2.5 / K3, GLM-5, Llama 3 / 4, GPT-OSS, Gemma 3 / 4, Mixtral, Qwen-VL and Gemma VL families, and diffusion models via vllm-omni.

TPU v7x (Ironwood) serving recipes

TPU v6e (Trillium) serving recipes

Kubernetes deployment

Multi-host serving · Parallelism & scaling · Disaggregated serving (P/D)

Quantization · Speculative decoding

🤝 Contribute

We welcome contributions of all kinds — bug reports, model and feature requests, documentation fixes, and code.

Developer references: Testing & CI · Profiling · Benchmarking · Coding guidance · Optimization best practices

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →