GitLab

In a distributed system we don't know if the remote workers need pre-processing done ingress-side or not. Previously Client required us to decide this before discovering the remote endpoints, which was fine because pre-processing was worker-side.

As part of moving pre-processing back to ingress-side we need to split this into two steps:
- Client discovers the endpoints, and (later PR) will fetch their Model Deployment Card.
- PushRouter will use the Model Deployment Card to decide if they need pre-processing or not, which affects the types of the generic parameters.

Part of #743

Signed-off-by: Olga Andreeva <124622579+oandreeva-nv@users.noreply.github.com>

This will allow an ingress-side pre-processor to see it without needing a model checkout.

Currently pre-processing is done in the worker, which has access to the model deployment card ("MDC") files (`config.json`, `tokenizer.json` and `tokenizer_config.json`) locally. We want to move the pre-processor to the ingress side to support KV routing. That requires ingress side (i.e the HTTP server), on a different machine than the worker to be able to see those three files.

To support that this PR makes the worker upload the contents of those files to the NATS object store, and publishes the MDC with those NATS urls to the key-value store. 

The key-value store has an interface so any store (nats, etcd, redis, etc) can be supported. Implementations for memory and NATS are provided.

Fetching the MDC from the store, doing pre-processing ingress side, and publishing a card backed by a GGUF, are all for a later commit.

Part of #743 

It's different enough that I made a new engine vllm0_8 and renamed the previous engine to vllm0_7.

`dynamo-run out=vllm` now expects 0.8. This matches the container change in #690.

For older use `dynamo-run out=vllm0_7`.

As a first step towards KV routing:
- introduce a `--router-mode` in dynamo-run that only does random and round-robin right now. Not that interesting yet.
- Make the vllm engine publish the KV events received from our patched vllm.

Now we "just" need to connect the two. Easy right?

Also upgrade the cargo resolver to v3, the default.

New clippy lints:
- `next_back()` instead of `last()` for a double-ended iterator. That avoids walking the whole list.
- ` repeat_n` instead of `repeat.take`. That avoids cloning.
- Doc indenting

Moved all of `lib/llm/src/engines` to their own crates as e.g. `lib/engines/mistralrs`. This will allow publishing of the `dynamo-llm` crate as it won't have any github dependencies.

The only engines in dynamo-llm will be the demo `echo` ones.
Co-authored-by: Graham King <grahamk@nvidia.com>

Put the arguments in a JSON file:
```
{
    "dtype": "half",
    "trust_remote_code": true
}
```

Pass it like this:
```
dynamo-run out=sglang ~/llm_models/Llama-3.2-3B-Instruct --extra-engine-args sglang_extra.json
```

Requested here https://github.com/ai-dynamo/dynamo/issues/290 (`dtype`) and here https://github.com/ai-dynamo/dynamo/issues/360 (`trust_remote_code`).

This lets us do:
```
dynamo-run out=llamacpp <gguf_file>
```

Previously a `--model-config <hf-repo>` was also required, to configure our tokenizer.

Co-authored-by: Olga Andreeva <oandreeva@oandreeva-mlt.client.nvidia.com>

```
dynamo-run in=batch:prompts.jsonl out=mistralrs ~/llm_models/Llama-3.2-3B-Instruct/
```

The file has genai format, one entry per line:
```
{"text": "the prompt"}
{"text": ..etc
```

The prompt is evaluated and the output written to `output.jsonl` in the
same folder as the input.

At the end of the run various statistics are printed:
> Ran 5 files in 8s 679ms. Tokens in: 40 (5/s). Tokens out: 346 (43/s)

This is also helpful for pushing load into the system and stressing the
various components. Not intended for performance measurement, it's a
batch inference tool.

- Any engine can take the name of a Hugging Face repository. It will be downloaded before calling the engine.

- The default engine (previously always mistralrs) depends on what is compiled in.

- Text can be piped in and will result in a single run of the model.

All of those together mean if you build with `--features vllm` you can do this and it will download the model and run it with vllm, answer your question, and exit:
```
echo "What is the capital of Costa Rica?"  | dynamo-run Qwen/Qwen2.5-3B-Instruct
```
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

Command line arguments are passed to the python engine like this:
```
dynamo-run out=pystr:my_python_engine.py -- -n 42 --custom-arg Orange --yes
```

The python engine receives the arguments in `sys.argv`. The argument list will include some standard ones as well as anything after the `--`.

This input:
```
dynamo-run out=pystr:my_engine.py /opt/models/Llama-3.2-3B-Instruct/ --model-name llama_3.2 --tensor-parallel-size 4 -- -n 1
```

is read like this:
```
async def generate(request):
    .. as before ..

if __name__ == "__main__":
    print(f"MAIN: {sys.argv}")
```

and produces this output:
```
MAIN: ['my_engine.py', '--model-path', '/opt/models/Llama-3.2-3B-Instruct/', '--model-name', 'llama3.2', '--http-port', '8080', '--tensor-parallel-size', '4', '--base-gpu-id', '0', '--num-nodes', '1', '--node-rank', '0', '-n', '1']
```

This allows quick iteration on the engine setup. Note how the `-n` `1` is included. Flags `--leader-addr` and `--model-config` will also be added if provided to `dynamo-run`.

- Latest from repo, many improvements
- Support most of the OpenAI request features (temperature, top_p, etc)
- Download models from Hugging Face if necessary

For the `echo` and `pystr` engines we previously required the user to pass `--model-name <x>` so we would have a name for the model. If the input is HTTP we do need this to match on the users' JSON request.

If the input is Text we don't need a name. So if the input is Text and we don't already have a name for the model, give it one.

Co-authored-by: Biswa Panda <biswa.panda@gmail.com>

There are two etcd keys:
- The service
- The model

The second one is the interesting one for us. Previously we confused the two.

Instead of using `out=pystr:<my.py>` we can now do this:
```
dynemo-run out=pytok:/home/graham/my_python_engine.py --model-path <hf-repo-checkout>
```

That engine will receive and respond with tokens. Here's an example engine file:
```
import asyncio

async def generate(request):
    yield {"token_ids":[791]}
    await asyncio.sleep(0.1)
    yield {"token_ids":[6864]}
    await asyncio.sleep(0.1)
    yield {"token_ids":[315]}
    await asyncio.sleep(0.1)
    yield {"token_ids":[9822]}
    await asyncio.sleep(0.1)
    yield {"token_ids":[374]}
    await asyncio.sleep(0.1)
    yield {"token_ids":[12366]}
    await asyncio.sleep(0.1)
    yield {"token_ids":[13]}
```

Also reduce duplication by making the bindings engine use the llm lib engine.

1. Create `my_engine.py`

```
import asyncio

async def generate(request):
    yield {"id":"1","choices":[{"index":0,"delta":{"content":"The","role":"assistant"}}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
    await asyncio.sleep(0.1)
    yield {"id":"1","choices":[{"index":0,"delta":{"content":" capital","role":"assistant"}}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
    await asyncio.sleep(0.1)
    yield {"id":"1","choices":[{"index":0,"delta":{"content":" of","role":"assistant"}}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
    await asyncio.sleep(0.1)
    yield {"id":"1","choices":[{"index":0,"delta":{"content":" France","role":"assistant"}}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
    await asyncio.sleep(0.1)
    yield {"id":"1","choices":[{"index":0,"delta":{"content":" is","role":"assistant"}}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
    await asyncio.sleep(0.1)
    yield {"id":"1","choices":[{"index":0,"delta":{"content":" Paris","role":"assistant"}}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
    await asyncio.sleep(0.1)
    yield {"id":"1","choices":[{"index":0,"delta":{"content":".","role":"assistant"}}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
    await asyncio.sleep(0.1)
    yield {"id":"1","choices":[{"index":0,"delta":{"content":"","role":"assistant"},"finish_reason":"stop"}],"created":1841762283,"model":"Llama-3.2-1B-Instruct","system_fingerprint":"local","object":"chat.completion.chunk"}
```

2. Build

```
cargo build --release --feature python
```

3. Run

```
dynemo-run out=pystr:my_engine.py --name test
```

And here's a distributed system, with your engine:

- Node 1: `dynemo-run in=http out=dyn://test`
- Node 2: `dynemo-run in=dyn://test out=pystr:my_engine.py`

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

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

Engine, `tio` support and docs.

Proof of concept / experimental.

triton-distributed-llm component and support in tio

Docs in README

- 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
```

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>

Add support in tio for distributed components and discovery.

Node 1:
```
tio in=http out=tdr://ns/backend/mistralrs
```

Node 2:
```
tio in=tdr://ns/backend/mistralrs out=mistralrs ~/llm_models/Llama-3.2-3B-Instruct
```

This will use etcd to auto-discover the model and NATS to talk to it. You can run multiple workers on the same endpoint and it will pick one at random each time.

The `ns/backend/mistralrs` are purely symbolic, pick anything as long as it has three parts, and it matches the other node.

Signed-off-by: Ryan Olson <ryanolson@users.noreply.github.com>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

You can now run an HF repo directly:
```
tio ~/llm_models/Llama-3.2-1B-Instruct/
```

or a GGUF
```
tio ~/llm_models/Llama-3.2-1B-Instruct-Q4_K_M.gguf
```

Also cleanup kv_router so I can merge.

This allows us to run a real model.

Build:
```
cargo build --release --features mistralrs,cuda
```

Run:
```
./target/release/tio in=text out=mistralrs --model-path Llama-3.2-1B-Instruct-Q4_K_M.gguf
```

Why [mistral.rs](https://github.com/EricLBuehler/mistral.rs)?

- It has no dependencies. You don't need a container or a virtual env to get started.
- It supports CUDA, Metal (MacOS) and CPU-only. Everyone can join the AI revolution.
- It starts fast and serves fast (with CUDA). That makes it fun to experiment with.
- It runs many models, not just Mistral, that's just it's name.