Title: [Question] DeepSeek-V4 Support Plan

[xulunfan]

Summary

Hi RTP-LLM team, thank you for the great work on this inference engine!

I noticed that the latest release supports DeepSeek V1/V2/V3/R1, but DeepSeek-V4 is not yet mentioned in the supported models list. Meanwhile, other frameworks like vLLM (v0.20+) and SGLang have already added Day-0/early support for DeepSeek-V4.

Questions

1. Is DeepSeek-V4 support currently on the RTP-LLM roadmap?
2. If yes, is there an estimated timeline for when it might be available?
3. Are there any specific technical challenges with V4 that the team is working on? (e.g., MoE changes, new kernel requirements, etc.)

Context

RTP-LLM already has excellent optimizations for DeepSeek-V3 (DeepEP, DeepGEMM, FlashMLA, EPLB, TP/DP/EP hybrid deployment), achieving impressive throughput numbers (Prefill 42.6K TPS, Decode 14.7K TPS on H800). Given the architectural continuity from V3 to V4, it would be great to know when we can expect V4 support.

Thanks!

[Vinkle-hzt]

Hi, thanks for your interest and for the detailed context!

DeepSeek-V4 support is already under active development. We currently have it supported on the development branch feat/dsv4_on_dev, together with a number of performance optimizations.

We are still doing further validation, benchmarking, and polishing before merging it into main. Once we are fully ready, we will merge the branch and update the supported models documentation accordingly.

At the moment we do not have a precise ETA to share, but you can follow the feat/dsv4_on_dev branch for progress.

Thanks again for following RTP-LLM!