vLLM Ascend Roadmap Q1 2025

[Yikun]

This is a living document!

Note that: vLLM Ascend 0.7.3 (match vLLM v0.7.3) is main release for 2025 Q1, see more in link.

Supported models track: #260

Hardware Plugin

• [RFC]: Hardware pluggable vllm#11162

Basic support

Initial vLLM Ascend support will start to support with basic hardware compatibility support.

• (P0) Chunked Prefill
• (P1) Automatic Prefix Caching (Improve performance)
• (P1) Speculative decoding
• (P1) Guided Decoding: [Feature]: Add support for Guided Decoding #177
• (P1) Multi step scheduler: support multistep decoding main #222
• LoRA
• Prompt adapter

Feature support

• (P0) V1 engine support: [RFC] V1 engine support #9
• (P1) EP support Expert Parallelism (EP) Support for DeepSeek Models vllm#12583
• (P1) Custom op [RFC]: Add support for custom ops #156
• (P1) MTP v0.7.3 Add MTP support for deepseek #236
• disaggregated prefill [Feature][Disaggregated] Support XpYd disaggregated prefill with MooncakeStore vllm#12957
• Scheduler plugin [do not merge] V1 scheduler interface vllm#12544
• RLHF Post train support - verl: Add Ascend NPU support for verl volcengine/verl#338
• RLHF Post train suppor - OpenRLHF: [WIP] support Ascend NPU backend OpenRLHF/OpenRLHF#605

Model support

• (P0) DeepSeek V3 / DeepSeek R1: [New Model]: DeepSeek V3 / R1 #72
• (P0) Llama3
• (P0) Qwen2.5
• (P0) Qwen2-VL: [New Model]: Qwen2-VL #246
• (P0) Qwen2.5-VL: [New Model]: Qwen2.5-VL #75
• (P1) BAAI/bge-m3: [New Model]: BAAI/bge-m3 #235
• MiniCPM
• GLM4
• InternLM
• llava
• GLM4v
• InternVL

Performance

• add vllm-ascend perf website like vLLM does https://perf.vllm.ai/
• focus on llama3, qwen2.5, qwen2-vl, deepseek v3/R1, improve the performance

Quality

• Full UT coverage
• Model e2e test
• Multi card/node e2e test

Docs

• README
• vllm-ascend website: https://vllm-ascend.readthedocs.org/
• Quick start / Installation / Turtorial
• User guide: supported feature / models
• Developer guide: Contributing / Versioning policy

CI and Developer Productivity

• vllm-ascend Docker image: [CI] Add container image build ci #64
• Ascend CI for main / dev branch: [Core] Init vllm-ascend #3

[wuhuikx]

chunked prefill will also be supported in Q1 for functionality.

[Yikun]

chunked prefill will also be supported in Q1 for functionality.

I updated Chunked Prefill as P0 priority.

[piekey1994]

Looking forward to lora ^^

[linfan]

When will GPTQ / AWQ / W8A8 support available ?

[Yikun]

Move forward: #448