[Aligned with vllm 0.23.0 now][RFC]: vLLM-Omni 2026 Q2 Roadmap

[hsliuustc0106]

Motivation.

This issue outlines the development roadmap for vllm-omni in Q2 2026. We are collecting your feedbacks and finalize it later.

Proposed Change.

1) Model Support

• Omni Models [RFC]: Omni-Modality Q2 Roadmap #2207
• TTS Models [RFC]: TTS Development Roadmap - Q2 2026 #2115
• Diffusion Models [RFC]: vLLM-Omni Diffusion Module — Q2 2026 Roadmap #2226 including world models and vla models

2) Feature Support

• Quantization [RFC]: Continuous Quantization Support  #1854
• Prefix Cache [RFC]: Enable Prefix Caching with Hidden-State I/O (Multi-round / Service Scenarios) #1184
• KV offload [Core] Support KV Cache CPU Offloading #1330
• Model Runner V2 [RFC]: Migrate to Model Runner V2 #1770
• Streaming audio/video input/output [Feature] add session based audio streaming input  #2208
• RL integration with veRL [RFC] Multi-modal Generation RL 2026Q2 Roadmap verl-project/verl#5755
• Diffusion Optimizations [RFC]: Continuous Diffusion Model Acceleration Support #1217
• Diffusion continuous batching: [Feat] Support step-boundary abort in diffusion #1769
• Diffusers backend [RFC]: Diffusers Backend Integration for Extended Model Coverage #2403

3) Large Scale Deployment (Disaggregated Serving)

• EPDG disaggregation: x(E)y(P/D)zG [RFC][calling for help from the community]: Large-Scale Multi-Stage Serving Architecture for vLLM-Omni #2336
• targeting models: Bagel/HY-Image3.0/Qwen-Omni/Ming-Omni

4) Hardware Support

• ROCm (AMD): [RFC]: vLLM-Omni ROCm 2026 Q2 Roadmap #2413
• Intel XPU: [RFC]: Omni Model Runner Q2 Roadmap #2574
• Ascend NPU: [RFC]: vLLM-Omni NPU 2026 Q2 Roadmap #2223
• MUSA: [Roadmap][Feature] Support Moore Threads (MUSA) GPUs #2347

5) CI/CD & Usability

• Quality Gates: L1-L5 level tests enhancement&refactor [RFC]: Restructure vLLM-Omni Test Layout, Fixture Scope, and Support Modules #2299 [RFC]: Framework for Tiny Model Testing #4197 [RFC]: Conditional testing strategy for CI #4254
• Benchmarking performance: vllm-omni-kanban
• Developer Usability: vllm-omni-skills
• User guidance: vllm/recipes

Feedback Period.

No response

CC List.

@ywang96 @Gaohan123 @youkaichao @Isotr0py @ZJY0516 @david6666666 @SamitHuang @linyueqian @wtomin @tzhouam @gcanlin @tjtanaa @xuechendi

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[SamitHuang]

VAE stage disaggregation #2089

[tzhouam]

For Q2 we are going to have the following features included:

1. Migrate to Model Runner V2 [RFC]: Migrate to Model Runner V2 #1770
2. Output Processor Refactor for clear architecture and supporting multiple multimodal outputs. [RFC]: Decouple Model-Specific Logic from Model Runner #1595
3. Additional information interface aligned. [RFC]: Rebase Additional Information into Model Intermediate Buffer #1351
4. Prefix Caching enabled [RFC]: Enable Prefix Caching with Hidden-State I/O (Multi-round / Service Scenarios) #1184
5. Support World Models [RFC]: World Model Support #1987

[hrshovon]

grpc support?