vllm-project/tpu-inference: Moving to vllm-project/vllm-jax. TPU inference for vLLM, with unified JAX and PyTorch support.

vLLM JAX

[!IMPORTANT]

tpu-inference is now vLLM JAX (vllm-jax) and is focused on serving JAX models on TPU with vLLM.

> This repository is moving to vllm-project/vllm-jax. Until the move is complete, development continues in vllm-project/tpu-inference; open issues and pull requests there.

> PyTorch-native TPU serving has moved to its own project, vllm-torchtpu, which ships its own releases and is where all future PyTorch-native TPU support in vLLM will live. If you are serving PyTorch models on TPU, head there.

🤝 Contribute to the Project
_{Looking to help? Click a badge below to find issues that need your attention.}

Latest News

Announcing Gemma 4 on vLLM Byte for byte, the most capable open models - available on TPUs on Day 0!

Previous News 🔥

Pytorch Conference Learn how Spotify uses vLLM with both GPUs and TPUs to drive down costs and improve user experience.
Ray Summit, November 3-5 in San Francisco!
JAX DevLab on November 18th in Sunnyvale!
[2025/10] vLLM TPU: A New Unified Backend Supporting PyTorch and JAX on TPU

About

vLLM JAX is powered by vllm-jax (formerly tpu-inference), an expressive and powerful new hardware plugin unifying JAX and PyTorch under a single lowering path within the vLLM project. The new backend now provides a framework for developers to:

Push the limits of TPU hardware performance in open source.
Provide more flexibility to JAX and PyTorch users by running PyTorch model definitions performantly on TPU without any additional code changes, while also extending native support to JAX.
Retain vLLM standardization: keep the same user experience, telemetry, and interface.

Recommended models and features

Although vLLM JAX’s new unified backend makes out-of-the-box high performance serving possible with any model supported in vLLM, the reality is that we're still in the process of implementing a few core components.

For this reason, we’ve provided Recommended Models and Recommended Features pages detailing the models and features that are validated through unit, integration, and performance testing.

Get started

Get started with vLLM on TPUs by following the quickstart guide.

Visit our documentation to learn more.

Compatible TPU Generations

Recommended: v7x, v5e, v6e
Experimental: v3, v4, v5p

Recipes

TPU Support Matrix Dashboard

Below is the live status of our supported models, features, and kernels. Click on any category to expand the detailed support table. It is automatically updated from our detailed Support Matrices.

Last Updated: 2026-10-22 06:52 AM UTC

🚦 Status Legend

- ✅ Passing: Tested and works as expected. Ready for use.

- ❌ Failing: Known to be broken or not functional. Help is wanted to fix this!

- 🧪 Experimental: Works, but unoptimized or pending community validation.

- 📝 Planned: Not yet implemented, but on the official roadmap.

- ⛔️ Unplanned: There is no benefit to adding this.

- ❓ Untested: The functionality exists but has not been recently or thoroughly verified.

>

📐 View Matrix Aggregation Rules (v6e/v7x & C+P)

> - 🛠️ Correctness + Performance (C + P)

- ❌ Failing: If either check fails.

- ✅ Passing: If BOTH checks pass successfully.

- ❓ Untested: If any check is untested (and neither fails).

> - 🌐 Hardware Rollups (v6e + v7x)

- ❌ Failing: If the feature fails on either v6e or v7x.

- ✅ Passing: If the feature passes on BOTH v6e and v7x.

- ❓ Untested: If either generation is untested (and neither fails).

Release Support Matrices

Click to expand support matrices

Stable support status for official releases and production deployments.

✅ Tested Models

| Model | Type | Unit Test | Correctness Test | Performance Test | | --- | --- | --- | --- | --- | | google/gemma-4-26B-A4B-it | Multimodal | ✅ | ✅ | ✅ | | google/gemma-4-31B-it | Multimodal | ✅ | ✅ | ✅ | | google/gemma-4-E2B-it | Multimodal | ✅ | ✅ | ✅ | | google/gemma-4-E4B-it | Multimodal | ✅ | ✅ | ✅ | | google/gemma-3-27b-it | Text | ✅ | ✅ | ✅ | | meta-llama/Llama-3.1-8B-Instruct | Text | ✅ | ✅ | ✅ | | meta-llama/Llama-3.3-70B-Instruct | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-30B-A3B | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-32B | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-4B | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-Coder-480B-A35B-Instruct | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3.5-397B-A17B | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-Embedding-8B | Embedding | ✅ | ✅ | ❓ | | Qwen/Qwen2.5-VL-7B-Instruct | Multimodal | ✅ | ✅ | ❓ | | deepseek-ai/DeepSeek-R1 | Text | ✅ | ✅ | ❓ | | moonshotai/Kimi-K2.6 | Text | ✅ | ✅ | ❓ | | openai/gpt-oss-120b | Text | ✅ | ✅ | ❓ | | deepseek-ai/DeepSeek-OCR | Multimodal | ❓ | ❓ | ❓ | | meta-llama/Llama-4-Maverick-17B-128E-Instruct/multimodal | Multimodal | ❓ | ❓ | ❓ | | Qwen/Qwen3-Omni-30B-A3B-Instruct | Multimodal | ❓ | ❓ | ❓ | | Qwen/Qwen3-VL-8B-Instruct | Multimodal | ❓ | ❓ | ❓ | | Qwen/Qwen3.5-9B | Multimodal | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-Math-V2 | Text | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-V3.1 | Text | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-V3.2 | Text | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-V3.2-Speciale | Text | ❓ | ❓ | ❓ | | MiniMaxAI/MiniMax-M2.5 | Text | ❓ | ❓ | ❓ | | mistralai/Mistral-Small-4-119B-2603/text-only | Text | ❓ | ❓ | ❓ | | moonshotai/Kimi-K2-Thinking | Text | ❓ | ❓ | ❓ | | openai/gpt-oss-20b | Text | ❓ | ❓ | ❓ | | Qwen/Qwen3.8-2.4T-A95B | Text | ❓ | ❓ | ❓ | | zai-org/GLM-5 | Text | ❓ | ❓ | ❓ |

🚀 Advanced Capabilities

Core Features

Feature Flax Torchax Default

async scheduler ✅ ✅ ✅

Chunked Prefill ✅ ✅ ✅

DCN-based P/D disaggregation ✅ ✅ ✅

KV Cache Offload ✅ ✅ ✅

LoRA_Torch ✅ ✅ ✅

Multimodal Inputs ✅ ✅ ✅

Out-of-tree model support ✅ ✅ ✅

Prefix Caching ✅ ✅ ✅

Single Program Multi Data ✅ ✅ ✅

Speculative Decoding: Eagle3 ✅ ✅ ✅

Speculative Decoding: Ngram ✅ ✅ ✅

Speculative Decoding: DFlash ✅ ❓ ✅

hybrid kv cache ❓ ❓ ❓

multi-host ❓ ❓ ❓

runai_model_streamer_loader ❓ ❓ ❓

sampling_params ❓ ❓ ❓

Step Pooling (Embedding) ❓ ❓ ❓

structured_decoding ❓ ❓ ❓

Parallelism Techniques

Feature Flax Torchax

Single-host Multi-host Single-host Multi-host

PP ✅ ✅ ✅ ✅

DP ✅ ❓ ✅ ❓

EP ✅ ❓ ✅ ❓

TP ✅ ❓ ✅ ❓

CP ❓ ❓ ❓ ❓

SP (vote to prioritize) ❓ ❓ ❓ ❓

Quantization Methods

Checkpoint dtype Method Supported
Hardware Acceleration Flax Torchax

FP4 W4A16 mxfp4 v7 ❓ ❓

FP8 W8A16 compressed-tensor v7 ❓ ❓

FP8 W8A8 compressed-tensor v7 ❓ ❓

INT4 W4A16 awq v5, v6 ❓ ❓

INT8 W8A8 compressed-tensor v5, v6 ❓ ❓

NVFP4 W4A16 modelopt_fp4 v7 ❓ ❓

Note:

- This table only tests checkpoint loading compatibility.

🔬 Microbenchmark Kernel Support

Category Test W16A16 W8A8 W8A16 W4A4 W4A8 W4A16

Moe Fused MoE ❓ ❓ ❓ ❓ ❓ ❓

gmm ❓ ❓ ❓ ❓ ❓ ❓

Dense All‑gather matmul ❓ ❓ ❓ ❓ ❓ ❓

Attention Generic Ragged Paged
Attention V3* ❓ ❓ ❓ ❓ ❓ ❓

MLA ❓ ❓ ❓ ❓ ❓ ❓

Ragged Paged
Attention V3 Head_Dim
64* ❓ ❓ ❓ ❓ ❓ ❓

Note:

- For attention kernels, W[x]A[y] denotes KV cache as W, A as compute, and x, y as bit precision.

Feature	Flax	Torchax	Default
async scheduler	✅	✅	✅
Chunked Prefill	✅	✅	✅
DCN-based P/D disaggregation	✅	✅	✅
KV Cache Offload	✅	✅	✅
LoRA_Torch	✅	✅	✅
Multimodal Inputs	✅	✅	✅
Out-of-tree model support	✅	✅	✅
Prefix Caching	✅	✅	✅
Single Program Multi Data	✅	✅	✅
Speculative Decoding: Eagle3	✅	✅	✅
Speculative Decoding: Ngram	✅	✅	✅
Speculative Decoding: DFlash	✅	❓	✅
hybrid kv cache	❓	❓	❓
multi-host	❓	❓	❓
runai_model_streamer_loader	❓	❓	❓
sampling_params	❓	❓	❓
Step Pooling (Embedding)	❓	❓	❓
structured_decoding	❓	❓	❓

Feature	Flax	Torchax
PP	✅	✅	✅	✅
DP	✅	❓	✅	❓
EP	✅	❓	✅	❓
TP	✅	❓	✅	❓
CP	❓	❓	❓	❓
SP (vote to prioritize)	❓	❓	❓	❓

Checkpoint dtype	Method	Supported Hardware Acceleration	Flax	Torchax
FP4 W4A16	mxfp4	v7	❓	❓
FP8 W8A16	compressed-tensor	v7	❓	❓
FP8 W8A8	compressed-tensor	v7	❓	❓
INT4 W4A16	awq	v5, v6	❓	❓
INT8 W8A8	compressed-tensor	v5, v6	❓	❓
NVFP4 W4A16	modelopt_fp4	v7	❓	❓

Category	Test	W16A16	W8A8	W8A16	W4A4	W4A8	W4A16
Moe	Fused MoE	❓	❓	❓	❓	❓	❓
gmm	❓	❓	❓	❓	❓	❓
Dense	All‑gather matmul	❓	❓	❓	❓	❓	❓
Attention	Generic Ragged Paged Attention V3*	❓	❓	❓	❓	❓	❓
MLA	❓	❓	❓	❓	❓	❓
Ragged Paged Attention V3 Head_Dim 64*	❓	❓	❓	❓	❓	❓

Nightly Support Matrices

Click to expand support matrices

Support status for the latest nightly/main branch developments.

✅ Tested Models

| Model | Type | Unit Test | Correctness Test | Performance Test | | --- | --- | --- | --- | --- | | google/gemma-4-26B-A4B-it | Multimodal | ✅ | ✅ | ✅ | | google/gemma-4-31B-it | Multimodal | ✅ | ✅ | ✅ | | google/gemma-3-27b-it | Text | ✅ | ✅ | ✅ | | meta-llama/Llama-3.1-8B-Instruct | Text | ✅ | ✅ | ✅ | | meta-llama/Llama-3.3-70B-Instruct | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-30B-A3B | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-4B | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3-Coder-480B-A35B-Instruct | Text | ✅ | ✅ | ✅ | | Qwen/Qwen3.5-397B-A17B | Text | ✅ | ✅ | ✅ | | Qwen/Qwen2.5-VL-7B-Instruct | Multimodal | ✅ | ✅ | ❌ | | Qwen/Qwen3-Embedding-8B | Embedding | ✅ | ✅ | ❓ | | deepseek-ai/DeepSeek-R1 | Text | ✅ | ✅ | ❓ | | moonshotai/Kimi-K2.6 | Text | ✅ | ✅ | ❓ | | openai/gpt-oss-120b | Text | ✅ | ✅ | ❓ | | google/gemma-4-E2B-it | Multimodal | ✅ | ❌ | ❓ | | google/gemma-4-E4B-it | Multimodal | ✅ | ❌ | ❓ | | Qwen/Qwen3-32B | Text | ❌ | ❓ | ❓ | | deepseek-ai/DeepSeek-OCR | Multimodal | ❓ | ❓ | ❓ | | Qwen/Qwen3-Omni-30B-A3B-Instruct | Multimodal | ❓ | ❓ | ❓ | | Qwen/Qwen3-VL-8B-Instruct | Multimodal | ❓ | ❓ | ❓ | | Qwen/Qwen3.5-9B | Multimodal | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-Math-V2 | Text | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-V3.1 | Text | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-V3.2 | Text | ❓ | ❓ | ❓ | | deepseek-ai/DeepSeek-V3.2-Speciale | Text | ❓ | ❓ | ❓ | | MiniMaxAI/MiniMax-M2.5 | Text | ❓ | ❓ | ❓ | | moonshotai/Kimi-K2-Thinking | Text | ❓ | ❓ | ❓ | | openai/gpt-oss-20b | Text | ❓ | ❓ | ❓ | | zai-org/GLM-5 | Text | ❓ | ❓ | ❓ |

🚀 Advanced Capabilities

Core Features

Feature Flax Torchax Default

async scheduler ✅ ✅ ✅

Chunked Prefill ✅ ✅ ✅

DCN-based P/D disaggregation ✅ ✅ ✅

KV Cache Offload ✅ ✅ ✅

LoRA_Torch ✅ ✅ ✅

Multimodal Inputs ✅ ✅ ✅

Out-of-tree model support ✅ ✅ ✅

Prefix Caching ✅ ✅ ✅

Single Program Multi Data ✅ ✅ ✅

Speculative Decoding: Eagle3 ✅ ✅ ✅

Speculative Decoding: DFlash ✅ ✅ ✅

Speculative Decoding: Ngram ✅ ✅ ✅

hybrid kv cache ❓ ❓ ❓

multi-host ❓ ❓ ❓

runai_model_streamer_loader ❓ ❓ ❓

sampling_params ❓ ❓ ❓

Step Pooling (Embedding) ❓ ❓ ❓

structured_decoding ❓ ❓ ❓

Parallelism Techniques

Feature Flax Torchax

Single-host Multi-host Single-host Multi-host

PP ✅ ✅ ✅ ✅

DP ✅ ❓ ✅ ❓

EP ✅ ❓ ✅ ❓

TP ✅ ❓ ❌ ❓

CP ❓ ❓ ❓ ❓

SP (vote to prioritize) ❓ ❓ ❓ ❓

Quantization Methods

Checkpoint dtype Method Supported
Hardware Acceleration Flax Torchax

FP4 W4A16 mxfp4 v7 ❓ ❓

FP8 W8A16 compressed-tensor v7 ❓ ❓

FP8 W8A8 compressed-tensor v7 ❓ ❓

INT4 W4A16 awq v5, v6 ❓ ❓

INT8 W8A8 compressed-tensor v5, v6 ❓ ❓

NVFP4 W4A16 modelopt_fp4 v7 ❓ ❓

Note:

- This table only tests checkpoint loading compatibility.

🔬 Microbenchmark Kernel Support

Category Test W16A16 W8A8 W8A16 W4A4 W4A8 W4A16

Moe Fused MoE ❓ ❓ ❓ ❓ ❓ ❓

gmm ❓ ❓ ❓ ❓ ❓ ❓

Dense All‑gather matmul ❓ ❓ ❓ ❓ ❓ ❓

Attention Generic Ragged Paged
Attention V3* ❓ ❓ ❓ ❓ ❓ ❓

MLA ❓ ❓ ❓ ❓ ❓ ❓

Ragged Paged
Attention V3 Head_Dim
64* ❓ ❓ ❓ ❓ ❓ ❓

Note:

- For attention kernels, W[x]A[y] denotes KV cache as W, A as compute, and x, y as bit precision.

Feature	Flax	Torchax	Default
async scheduler	✅	✅	✅
Chunked Prefill	✅	✅	✅
DCN-based P/D disaggregation	✅	✅	✅
KV Cache Offload	✅	✅	✅
LoRA_Torch	✅	✅	✅
Multimodal Inputs	✅	✅	✅
Out-of-tree model support	✅	✅	✅
Prefix Caching	✅	✅	✅
Single Program Multi Data	✅	✅	✅
Speculative Decoding: Eagle3	✅	✅	✅
Speculative Decoding: DFlash	✅	✅	✅
Speculative Decoding: Ngram	✅	✅	✅
hybrid kv cache	❓	❓	❓
multi-host	❓	❓	❓
runai_model_streamer_loader	❓	❓	❓
sampling_params	❓	❓	❓
Step Pooling (Embedding)	❓	❓	❓
structured_decoding	❓	❓	❓

Feature	Flax	Torchax
PP	✅	✅	✅	✅
DP	✅	❓	✅	❓
EP	✅	❓	✅	❓
TP	✅	❓	❌	❓
CP	❓	❓	❓	❓
SP (vote to prioritize)	❓	❓	❓	❓

Checkpoint dtype	Method	Supported Hardware Acceleration	Flax	Torchax
FP4 W4A16	mxfp4	v7	❓	❓
FP8 W8A16	compressed-tensor	v7	❓	❓
FP8 W8A8	compressed-tensor	v7	❓	❓
INT4 W4A16	awq	v5, v6	❓	❓
INT8 W8A8	compressed-tensor	v5, v6	❓	❓
NVFP4 W4A16	modelopt_fp4	v7	❓	❓

Category	Test	W16A16	W8A8	W8A16	W4A4	W4A8	W4A16
Moe	Fused MoE	❓	❓	❓	❓	❓	❓
gmm	❓	❓	❓	❓	❓	❓
Dense	All‑gather matmul	❓	❓	❓	❓	❓	❓
Attention	Generic Ragged Paged Attention V3*	❓	❓	❓	❓	❓	❓
MLA	❓	❓	❓	❓	❓	❓
Ragged Paged Attention V3 Head_Dim 64*	❓	❓	❓	❓	❓	❓

🤝 Contribute

We're thrilled you're interested in contributing to the vLLM JAX project! Your help is essential for making our tools better for everyone. There are many ways to get involved, even if you're not ready to write code.

Ways to Contribute:

🐞 Submit Bugs & Suggest Features: See an issue or have an idea? Open a new issue to let us know.
👀 Provide Feedback on Pull Requests: Lend your expertise by reviewing open pull requests and helping us improve the quality of our codebase.
📚 Improve Our Documentation: Help us make our guides clearer. Fix a typo, clarify a confusing section, or write a new recipe.

If you're ready to contribute code, our Contributing Guide is the best place to start. It covers everything you need to know, including:

Tips for finding an issue to work on (we recommend starting with our good-first issues!.

🌟 Contributors Wall

A huge thank you to everyone who has helped build and improve vllm-project/vllm-jax (formerly vllm-project/tpu-inference)!

🌟 Contribution Type Legend & Ranking

| Emoji | Contribution | Meaning |

| :---: | :--- | :--- |

| 💻 | Code | Submitted merged pull requests or code changes. |

| 🐛 | Issues | Opened valid issues or bug reports. |

| 👀 | Reviews | Reviewed pull requests and provided feedback. |

🏆 Ranking: Contributors are sorted from highest to lowest based on their total effort score (Total Commits + Unique Issues Opened + PRs Reviewed). If there is a tie, contributors are displayed alphabetically.

_{xiangxu-google}
💻

_jrplatin
🐛 👀 💻

_{buildkite-bot}
💻

_kyuyeunk
🐛 👀 💻

_py4
💻

_fenghuizhang
💻

You finished the article!

Want to go deeper or explore more stories?

Full Article on GitHub Trending More Tech News

Reading Mode

Adjust typography and theme to your preference.

Typography

Text Scale

View Original

vllm-project/tpu-inference: Moving to vllm-project/vllm-jax. TPU inference for vLLM, with unified JAX and PyTorch support.

Latest News

About

Recommended models and features

Get started

Recipes

TPU Support Matrix Dashboard

Release Support Matrices

Nightly Support Matrices

🤝 Contribute

🌟 Contributors Wall

Reading Mode

Related Stories

I didn't know I was lacking in Vitamin D until I collapsed at work

Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 Astra

ICE wants its mask back, and hired a company to ask

Northflank Breached

Medicare Will Pay $50k for Your Heart Attack. It Should Pay $100 to Prevent It

Related Stories

Feature	Flax		Torchax
Feature	Single-host	Multi-host	Single-host	Multi-host
PP	✅	✅	✅	✅
DP	✅	❓	✅	❓
EP	✅	❓	✅	❓
TP	✅	❓	✅	❓
CP	❓	❓	❓	❓
SP (vote to prioritize)	❓	❓	❓	❓