docs: Add multimodal documentation vllm, sglang, and trtllm backends

[indrajit96]

Overview:

Add comprehensive multimodal guides for vLLM, Sglang and TRT-LLM backends documenting architectures, deployment modes, input formats, and known limitations.

Details:

• New: docs/backends/vllm/multimodal_vllm_guide.md - Complete vLLM multimodal reference
  -New: docs/backends/trtllm/multimodal_trtllm_guide.md - Complete TRT-LLM multimodal reference
  -New: dynamo/docs/backends/sglang/multimodal_sglang_guide.md - Complete SGlang multimodal reference

[rmccorm4]

Can you update https://github.com/ai-dynamo/dynamo/blob/main/docs/multimodal/multimodal_intro.md at the bottom with links to each of the backend specific docs as a central location?

[rmccorm4]

re: https://github.com/ai-dynamo/dynamo/actions/runs/19580969130/job/56078496698?pr=4510

#15 5.473 checking consistency... /workspace/dynamo/docs/backends/sglang/multimodal_sglang_guide.md: WARNING: document isn't included in any toctree [toc.not_included]
#15 5.474 /workspace/dynamo/docs/backends/trtllm/multimodal_trtllm_guide.md: WARNING: document isn't included in any toctree [toc.not_included]
#15 5.474 /workspace/dynamo/docs/backends/vllm/multimodal_vllm_guide.md: WARNING: document isn't included in any toctree [toc.not_included]

Needs these files added to docs/hidden_toctree.rst

[rmccorm4]

@krishung5 can you help review the docs here? Main point is to clearly document what is supported in each backend today with relation to multimodality, and highlight at least 1 key example or model for each.

[krishung5]

What does the KV sharing here mean? Does it just mean PD disagg?

[krishung5]

Thanks for putting up this doc, great work! I think one minor comment is that, we have multimodal doc for all three frameworks, so maybe we can link them in the guide here somehow? i.e. vllm, trtllm here and here, and sglang

[krishung5]

I think it could be a bit confusing on --modality multimodal as users might not be familiar with the launch scripts. I understand we want to keep it short here in the bullet points, but maybe we can do something like this?

SIMPLE AGGREGATED (agg.sh):
  Client → Frontend (Rust) → Worker [image load, encode, P+D] → Response
  • 2 components • worker flag  `--modality multimodal` • Easiest setup

[krishung5]

Maybe add actual link here?

[krishung5]

Qq -

Yes/No (KV cache - UCX or NIXL)

Does this mean
Yes(KV cache transfer using NIXL)
No(KV cache transfer using UCX)

[krishung5]

Do we want to highlight the rust processor?

Suggested change

	Client → Frontend (Rust) → Worker [image load, encode, P+D] → Response
	Client → Frontend (Rust processor) → Worker [image load, encode, P+D] → Response

[krishung5]

Could we clarify here it's the EPD not the simple aggregated one? Do we want to add a section for the simple aggregated workflow? I think one confusion that I got frequently from people is EPD vs simple/traditional. If we could align on the wording that would be very helpful.

[krishung5]

same here, can add the actual link to the script.