[RFC]: Doc enhancement

[wangxiyuan]

Motivation.

This RFC is used for trace the community work of official doc improvement. Currently, there are 3 important section need update.

1. Tutorials
2. User doc
3. Developer doc

I'll list all the work items below. Everyone is welcome to take the task.

Proposed Change.

Tutorials

Problem:

1. Moe model guide is missing. For example Qwen3 Moe
2. No detial parallel case example
3. QwQ is not very popular.
4. V1 Engine should be used by default

Propose Change:
After improvment. The content be more clear(required : 313T+64GB):

• Single NPU(Qwen3-8B) aclgraph mode + eager mode @leo-pony Doc Enhancement: Single NPU(Qwen3-8B) aclgraph mode + eager mode #1374
• Single NPU(Qwen2.5-VL-7B）eager mode @shen-shanshan [Doc] Add Qwen2.5-VL eager mode doc #1394
• Single NPU(Qwen2.5-audio）eager mode @shen-shanshan [Doc] Add qwen2-audio eager mode tutorial #1371
• Single NPU(Qwen3 8B embedding）eager mode @wangxiyuan
• Multi NPU 2 card (Qwen3 MOE-30B) aclgraph mode + TP2 @leo-pony
• Multi NPU 4 card (Qwen3 32B) aclgraph mode + TP2 + DP2 + W8A8（optional） @22dimensions
• Multi Node 2 node (DeepSeek V3 0528 W8A8) TP4+DP4+EP Graph mode @Potabk
• Multi Node 4 node (DeepSeek R1) TP16+DP2 Graph mode @MengqingCao
• Multi Node 4 node (DeepSeek V3 0528 ) 1P1D Graph mode @wangxiyuan

User Guide

user guide should contain the usage for vLLM Ascend

Problem:

1. A lot of usage is missing.

Propose Change:

• Feature Support Matrix (Need refresh)

  • index
  • Graph mode Guide (Need refresh) @wangxiyuan [Doc] Fix doc typo #1424
  • Quantization Guide Add user guide for quantization #1206 @22dimensions
  • Disaggregated prefill Guide @zhangxinyuehfad
  • EP Guide (for example, how to use MC2、duel batch、EPLB...) @MengqingCao
  • Sleep Mode Guide @Potabk [Doc] Add sleep mode doc #1295
  • Guided Decoding Guide @shen-shanshan
  • Spec Decode Guide @mengwei805
  • Lora Guide @paulyu12 [DOC] add LoRA user guide #1265
  • VLLM supported feature(func call ...) @wangxiyuan

• Model Support Matrix (Need refresh) @zhangxinyuehfad

• Environment Vars (Need refresh) @shen-shanshan
  I think we should change the shown way. Just like Additoinal config, to make it more clear by hand.

• Additional Config (Need refresh) @wangxiyuan [Doc] Fix doc typo #1424

• Release note

Developer Guide

Problem:

1. There is no feature or code guide for developers at all.

Propose Change:

• How to contribute (Need refresh) @Yikun [1/N][CI] Move linting system to pre-commits hooks #1256
• Version Policy @Yikun No need to refresh
• CI system @Yikun
• Test Guide @MengqingCao [Doc] Update FAQ and add test guidance #1360
• Design Documents
  • Patch @wangxiyuan [Doc] Add patch doc #1414
  • Code Architecture @MengqingCao
  • Ops and Custom ops @Yikun
    - pta ops
    - custom ops
    - fused moe
  • Modeling @shen-shanshan
    what model is added/patched, why
    how to add a new model [docs] Update guidance on how to implement and register new models #1126
  • Attention @wangxiyuan
    - default attention
    - mla
  • Communicator @leo-pony
  • Disaggregated Prefill @ganyi1996ppo
  • Graph mode @zzzzwwjj
  • Quantization @22dimensions
• Evaluation (Need refresh) @zhangxinyuehfad
• release guide @wangxiyuan

[wangxiyuan]

@Yikun please help to update the content.

[tt545571022]

Multi Step ???

[Potabk]

#1295 sleep mode doc, need review

[wangxiyuan]

@tt545571022 Multi Step is the feature only for V0 Engine. So it's not listed here. But if any contributor would like to write one, we're happy to add it.

[shen-shanshan]

Need Review:

• [Doc] Add qwen2-audio eager mode tutorial #1371
• [Doc] Add performance tuning doc to main #1392
• [Doc] Add Qwen2.5-VL eager mode doc #1394
• [Doc] Add guidance on how to implement and register new models #1426
• [Doc] Add Structured Output guide #1499
• [Doc] Refactor env vars module and change the way of doc generation #1504
• [Doc] Add model costomization doc #1737

[leo-pony]

Ready for review:

• Single NPU(Qwen3-8B) aclgraph mode + eager mode
• Multi-npu qwen3-MoE-32B Tutorials

[tt545571022]

Is DBO included in the graph docs? If not, I hope that the implementation of supplementary dbo can be introduced in this document.

[Potabk]

Multi Node 2 node (DeepSeek V3 0528 W8A8) TP8+DP2 Graph mode runs filed, can only run eager-mode + vllm-ascend scheduler

[ChenTaoyu-SJTU]

Need Review:

• [Doc] Add ModelRunner_prepare_inputs doc #1493