GitLab

Needs more testing but good enough for now. I get the same results with this as with `vllm serve`.

Co-authored-by: hongkuanz <hongkuanz@nvidia.com>
Co-authored-by: Piotr Tarasiewicz <ptarasiewicz@nvidia.com>
Co-authored-by: Piotr Tarasiewicz Nvidia <ptarasiewicznv@Piotrs-MacBook-Pro.local>
Co-authored-by: Neelay Shah <neelays@ipp2-0493.ipp2u1.colossus.nvidia.com>
Co-authored-by: Neelay Shah <neelays@ipp1-1941.ipp1a1.colossus.nvidia.com>
Co-authored-by: ishandhanani <ishandhanani@gmail.com>
Co-authored-by: Neelay Shah <neelays@4u8g-gen-0078.ipp3a2.colossus.nvidia.com>
Co-authored-by: ptarasiewiczNV <104908264+ptarasiewiczNV@users.noreply.github.com>

`cargo build --locked` won't let you use "1.85.0" if you only have "stable" installed, even if those are the same thing right now.

Engine, `tio` support and docs.

Proof of concept / experimental.

Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

triton-distributed-llm component and support in tio

Docs in README

Co-authored-by: Graham King <grahamk@nvidia.com>

Co-authored-by: Ryan Olson <rolson@nvidia.com>

This means we don't need to explain the parts to the users until they are ready. We use what they provide and default the rest.

Allows all of this and more:
- `tio out=tdr://test`
- `tio out=tdr://llama_8b_pool`
- `tio in=tdr://corp_ai_research_group/model_next-20250226`
- `tio out=tdr://AIRE.NIM.migrate.mistralrs.1802`

Python, API, etc all untouched.

- Setup venv

```
uv venv
source .venv/bin/activate
uv pip install pip
uv pip install sgl-kernel --force-reinstall --no-deps
uv pip install "sglang[all]==0.4.2" --find-links https://flashinfer.ai/whl/cu124/torch2.4/flashinfer/
```

- Build: `cargo build --release --features sglang`

- Run single node (make sure you're in the venv): `./tio out=sglang ~/llm_models/my_model`

- Run Deepseek multi-gpu / multi-node:

Node 1:
```
tio in=http out=sglang --model-path ~/llm_models/DeepSeek-R1-Distill-Llama-70B/ --tensor-parallel-size 8 --num-nodes 2 --node-rank 0 --dist-init-addr 10.217.98.122:9876
```

Node 2:
```
tio in=none out=sglang --model-path ~/llm_models/DeepSeek-R1-Distill-Llama-70B/ --tensor-parallel-size 8 --num-nodes 2 --node-rank 1 --dist-init-addr 10.217.98.122:9876
```

Co-authored-by: aflowers <aflowers@nvidia.com>

Signed-off-by: Meenakshi Sharma <163925564+nvda-mesharma@users.noreply.github.com>
Signed-off-by: Neelay Shah <neelays@nvidia.com>
Co-authored-by: Neelay Shah <neelays@nvidia.com>
Co-authored-by: Ryan Olson <ryanolson@users.noreply.github.com>
Co-authored-by: Meenakshi Sharma <163925564+nvda-mesharma@users.noreply.github.com>
Co-authored-by: Biswa Panda <biswapanda@users.noreply.github.com>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

Add backend type `EngineConfig::StaticCore` that wraps the engine in a preprocessor (prompt templating and tokenization).

Add example engine `echo_core` (`out=echo_core`) which takes and returns tokens. A nice side effect is that it echos the full prompt template with system prompt, whereas `echo_full` echos only user prompt.

![image](https://github.com/user-attachments/assets/27ec0a7b-a27d-4e69-96ea-1ffa0822ea90)

Signed-off-by: Ryan McCormick <rmccormick@nvidia.com>

Signed-off-by: Neelay Shah <neelays@nvidia.com>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>