vllm
vllm copied to clipboard
vllm-project
[Roadmap] vLLM Roadmap Q4 2025
This page is accessible via roadmap.vllm.ai
This is a living document! For each item here, we intend to link the RFC as well as discussion Slack channel in the vLLM Slack
In the Q3 2025, we fully removed V0 code path and made vLLM excel in large scale serving with mature wide E and prefill disaggregation. In this quarter, our goal is to continue to drive down the CPU overhead, enhance vLLM on frontier clusters, and strengthen our RL integrations.
We list help wanted item as 🙋in areas that the committer group is seeking more dedicated contributions.
Engine Core
- [ ] Async Scheduling on by default (#sig-default-async-sched)
- [ ] Optimize Input Preparation (“Persistent Batch V2”)
- [ ] Multimodal Processing Simplification (#sig-multi-modality)
- [ ] 🙋Speculative Decoding Enhancements (Suffix Decoding, CUDA Graph/torch.compile support) (#feat-spec-decode)
Large Scale Serving
- [ ] Elastic Experts (#feat-elastic-ep)
- [ ] Transfer KV Cache through CPU (#feat-v1-cpu-offloading)
- [ ] Long context serving
- [ ] 🙋GB200 optimizations (#sig-nvidia)
- [ ] 🙋AMD support (#sig-amd)
- [ ] 🙋EPLB algorithm testing and enhancements in production
Reinforcement Learning (#sig-post-training)
- [ ] Full determinism and batch invariance
- [ ] Add more testing cases for popular integrations
- [ ] Custom checkpoint loader, custom model format
- [ ] Simple data parallel router for scale out
- [ ] 🙋Enhance weight loading speed for syncing and resharding
- [ ] 🙋Study a way to enable multi-turn long horizon scheduling to avoid preemption
Performance and UX Enhancement
- [ ] Continue to drive down startup time (#feat-startup-ux)
- [ ] Refactor tool use parsing to leverage grammar structural tag (#feat-tool-calling)
- [ ] Refactor CI (#sig-ci, #ci-sprint)
- [ ] Turn on torch compile fusion by default, no extra flags on default case (#sig-torch-compile)
- [ ] Prefix caching for Hybrid models (https://github.com/vllm-project/vllm/issues/26201)
- [ ] 🙋Model Bash: profile and optimize newer model architectures on different hardwares (NVIDIA Hopper, Blackwell, AMD MI3xx) (#sig-model-bash)
- DeepSeek V3.2
- Qwen3MoE, Qwen3 VL, Qwen3 Next
- gpt-oss
If any of the items you wanted is not on the roadmap, your suggestion and contribution is strongly welcomed! Please feel free to comment in this thread, open feature request, or create an RFC.
Historical Roadmap: #20336, #15735, #11862, #9006, #5805, #3861, #2681, #244
An important feature from V0 which wasn't yet implemented in V1 is concurrent partial prefills. Please consider prioritizing it 🙏 An issue that tracks it: https://github.com/vllm-project/vllm/issues/21674 Thanks!
Should we consider supporting E/P/D disaggregation for large-scale multimodal model serving? It's a beneficial feature for large-batch or encode-compute-heavy MLLM deployment scenarios. https://github.com/vllm-project/vllm/pull/25233
I hope we could land speculative decoding with draft models this quarter. Posting here to raise awareness :)
https://github.com/vllm-project/vllm/pull/24322
What's the plan for the proxy/router component used in disaggregated deployments?
"elastic expert" is a specialist with deep knowledge of the Elastic Stack (Elasticsearch, Kibana, Logstash, Beats) who can design, deploy, manage, and troubleshoot solutions using these technologies.
Can we please get Flash Attention 3 support for RTX6000 Blackwell GPUs? Really want to try native support for FP4 computation.
