GitLab

. New mistralrs and llamacpp version
. mistralrs: Handle Gemma 3 and Llama 4 as vision models
. Update the dynamo-run docs to use Qwen 3
. Our pre-processor now supports Llama 4's newer multi-modal `config.json`
. Upgrade minijinja to handle Qwen 3's prompt template

For Llama 4 we'll need to limit the max seq len. vllm says:
> To serve at least one request with the models's max seq len (10485760), (240.00 GiB KV cache is needed,...

I was able to run Llama 4 with llamacpp and a quantized GGUF, with Dynamo doing the pre-processing.

Signed-off-by: Graham King <graham@gkgk.org>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

vllm and sglang are now the sub-process engines from #954

Also updated docs on doing vllm and sglang multi-gpu (tensor parallel) and multi-node (pipeline parallel).

New vllm and sglang engines that run in a sub-process. Will hopefully replace the existing embedded python engines.
    
Why?
    
  - Pure Python, does not require knowing Rust to work on it. Much simpler to maintain.
  - No embedded Python interpreter which avoids linking libpython and avoids the MacOS virtualenv issues.
  - Should have better performance as it's "native" vllm / sglang.
  - Works with any version of vllm (including v1!) and sglang. Less upgrade struggle.

Adding this to a Python script makes it register on the network so that `dynamo-run` can discover it and send it requests:
```
from dynamo.llm import register_llm

MODEL = "Qwen/Qwen2.5-0.5B-Instruct"
await register_llm(endpoint, MODEL, 3)
```

Full vllm example, with pre-processing in dynamo:
- `dynamo-run in=text out=dyn://dynamo.backend.generate`
- `cd lib/bindings/python/examples/hello_world`
- `python server_vllm.py`

This builds on top of the work to move pre-processor to ingress side. It means we can decouple Rust and Python using NATS as the bus.

The `register_llm` call does this:

- Download the model from HF if necessary
- Load the model deployment card from the HF folder or extract from GGUF
- Push the tokenizer config etc into NATS object store so ingress can access it from a different machine
- Publish the model deployment card to ETCD

Part of https://github.com/ai-dynamo/dynamo/issues/743

Allow `hf://` prefix on command line. 

Closes GitHub issue: https://github.com/ai-dynamo/dynamo/issues/829

Adds support for specifying default request parameters through a json template file that can be applied across all inference requests. This enables consistent parameter settings while still allowing per-request overrides.

Changes:
- Add --request-template CLI flag to specify template file path
- Integrate template support in HTTP, batch and text input modes
- Template values can be overridden by individual request parameters
- Example template.json:
```
{
    "model": "Qwen2.5-3B-Instruct",
    "temperature": 0.7,
    "max_completion_tokens": 4096
}
```

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 

#### Overview:

This PR adds a command-line verbosity flag (-v, -vv) to dynamo-run to control log levels.
- Added new verbosity flag to Flags struct:
  - -v: Sets log level to debug
  - -vv: Sets log level to trace
  - No flag (default): Keeps log level at info

#### Details:
- closes GitHub issue: https://github.com/ai-dynamo/dynamo/issues/567

"echo_core" is an engine that echoes the post-processed request back to you so you can see the template. Good for testing. It needed an extra flag set to work correctly.

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`.

Signed-off-by: yexiaochuan <tap91624@gmail.com>

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

Adds `@dynamo_worker(static = True)` to create a static worker which has a predictable name and hence does not require discovery or `etcd` to be running. There can only be a single static worker per namespace / component / endpoint trio.

This contrasts with the default dynamic `dynamo_worker` endpoints we have now, which get a unique random name (based on namespace/component/endpoint), and are discovered by ingress components using etcd.

Also change the hello_world example to use `dynamo_worker(static = True)` so that it is exercised and demonstrated somewhere.

For NIM.

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.

That ensures it gets removed when the process stops.

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

Under load it sometimes drops a request. The request gets added to the batch (sequence) and immediately gets a FinishReason Stop. Not sure why. It doesn't happen with the default scheduler (non-paged attention), so switch to that for now.

Previously several parts of the stack ensured max tokens (for this single request) was set.

Now only text input sets it (to 8k). Everything else leaves as is, potentially blank. The engines themselves have very small defaults, 16 for vllm and 128 for sglang.

Also fix dynamo-run CUDA startup message to only print if we're using an engine that would benefit from it (mistralrs, llamacpp).

```
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.

Engines mistralrs, sglang and vllm included by default. Can be disabled like this: `cargo build --no-default-features --features <add-back-what-you-want>`.

Added `--feature vulkan` option, for llamacpp.

Build time message if CUDA or Metal would help and are missing. That's the best we can do:
> warning: dynamo-run@0.1.0: CUDA not enabled, re-run with `--features cuda`

Runtime message if CUDA, Metal or Vulkan are enabled:
> 2025-03-14T21:59:26.501937Z  INFO dynamo_run: CUDA on

Runtime message if they are missing:
> 2025-03-14T22:02:37.439404Z  INFO dynamo_run: CPU mode. Rebuild with `--features cuda|metal|vulkan` for better performance

Defaut engine message includes available engines:
> 2025-03-14T21:59:26.503612Z  INFO dynamo_run: Using default engine: mistralrs. Use out=<engine> to specify one of echo_core, echo_full, mistralrs, llamacpp, sglang, vllm, pystr, pytok

The really important outcome is that this should now "just work":
```
cargo install dynamo-run
dynamo-run Qwen/Qwen2.5-3B-Instruct
```

Sadly you still need `--features cuda|metal` for performance, I couldn't automate that.

- Mac doesn't have `pipe2` syscall so use plain `pipe`.
- rtnetlink isn't a dependency on mac so don't use the type

Previously we tokenized and counted tokens to stop when max tokens was reached. Now we let the mistral.rs engine do it which saves the extra tokenization step.

Also dynamo-run prints which engines are compiled in in help message, and some minor lint fixes.

"netlink" doesn't exist on Mac. We print the primary network interface to help multi-node setup, which is also unlikely on Mac.

- 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>