dynamo_run.md 24.3 KB
Newer Older
1
# Running Dynamo (`dynamo run`)
2

3
This guide explains the `dynamo run` command.
4
5
6
7

`dynamo-run` is a CLI tool for exploring the Dynamo components. It's also an example of how to use components from Rust. If you use the Python wheel, it's available as `dynamo run` .

It supports these engines: mistralrs, llamacpp, sglang, vllm, and tensorrt-llm. `mistralrs` is the default.
8
9
10

Usage:
```
11
dynamo-run in=[http|text|dyn://<path>|batch:<folder>] out=echo_core|echo_full|mistralrs|llamacpp|sglang|vllm|dyn [--http-port 8080] [--model-path <path>] [--model-name <served-model-name>] [--model-config <hf-repo>] [--tensor-parallel-size=1] [--context-length=N] [--num-nodes=1] [--node-rank=0] [--leader-addr=127.0.0.1:9876] [--base-gpu-id=0] [--extra-engine-args=args.json] [--router-mode random|round-robin|kv] [--kv-overlap-score-weight=2.0] [--kv-gpu-cache-usage-weight=1.0] [--kv-waiting-requests-weight=1.0] [--verbosity (-v|-vv)]
12
13
```

14
Example: `dynamo run Qwen/Qwen3-0.6B`
15

16
Set the environment variable `DYN_LOG` to adjust the logging level; for example, `export DYN_LOG=debug`. It has the same syntax as `RUST_LOG`.
17

18
19
20
21
22
23
24
To adjust verbosity, use `-v` to enable debug logging or `-vv` to enable full trace logging. For example:

```bash
dynamo-run in=http out=mistralrs -v  # enables debug logging
dynamo-run in=text out=llamacpp -vv  # enables full trace logging
```

25
## Quickstart with pip and vllm
26

27
If you used `pip` to install `dynamo`, you have the `dynamo-run` binary pre-installed with the `vllm` engine. You must be in a virtual environment with vllm installed to use this engine. To compile from source, see [Full usage details](#full-usage-details) below.
28

29
30
The vllm and sglang engines require [etcd](https://etcd.io/) and [nats](https://nats.io/) with jetstream (`nats-server -js`). Mistralrs and llamacpp do not.

31
### Use model from Hugging Face
32

33
To automatically downloads Qwen3 4B from Hugging Face (16 GiB download) and starts it in interactive text mode:
34
```
35
dynamo run out=vllm Qwen/Qwen3-4B
36
37
```

38
The general format for HF download follows this pattern:
39
40
```
dynamo run out=<engine> <HUGGING_FACE_ORGANIZATION/MODEL_NAME>
41
42
```

43
For gated models (such as meta-llama/Llama-3.2-3B-Instruct), you must set an `HF_TOKEN` environment variable.
44

45
The parameter can be the ID of a HuggingFace repository (which will be downloaded), a GPT-Generated Unified Format (GGUF) file, or a folder containing safetensors, config.json, or similar (perhaps a locally checked out HuggingFace repository).
46

47
### Run a model from local file
48

49
50
51
52
53
To run a model from local file:
- Download the model from Hugging Face
- Run the model from local file

See the following sections for details.
54

55
56
57
58
59
#### Download model from Hugging Face
One of the models available from HUgging Face should be high quality and fast on almost any machine: https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF
For example, try https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/blob/main/Llama-3.2-3B-Instruct-Q4_K_M.gguf

To download model file:
60
```
61
curl -L -o Llama-3.2-3B-Instruct-Q4_K_M.gguf "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/Llama-3.2-3B-Instruct-Q4_K_M.gguf?download=true"
62
```
63
#### Run model from local file
64
65
66
To run the model:

*Text interface*
67
```
68
dynamo run Llama-3.2-3B-Instruct-Q4_K_M.gguf # or path to a Hugging Face repo checkout instead of the GGUF file
69
```
70

71
*HTTP interface*
72
```
73
dynamo run in=http out=mistralrs Llama-3.2-3B-Instruct-Q4_K_M.gguf
74
```
75
You can also list models or send a request:
76

77
*List the models*
78
```
79
curl localhost:8080/v1/models
80
```
81

82
*Send a request*
83
```
84
curl -d '{"model": "Llama-3.2-3B-Instruct-Q4_K_M", "max_completion_tokens": 2049, "messages":[{"role":"user", "content": "What is the capital of South Africa?" }]}' -H 'Content-Type: application/json' http://localhost:8080/v1/chat/completions
85
```
86

87
88
89
### Distributed System

You can run the ingress side (HTTP server and pre-processing) on one machine, for example a CPU node, and the worker on a different machine (a GPU node).
90

91
You will need [etcd](https://etcd.io/) and [nats](https://nats.io) with jetstream installed and accessible from both nodes.
92

93
**Node 1:** OpenAI compliant HTTP server, optional pre-processing, worker discovery:
94

95
```
96
dynamo-run in=http out=dyn
97
```
98

99
**Node 2:** Vllm engine. Receives and returns requests over the network:
100

101
```
102
dynamo-run in=dyn://llama3B.backend.generate out=vllm ~/llms/Llama-3.2-3B-Instruct
103
```
104

105
106
This uses etcd to auto-discover the model and NATS to talk to it. You can
run multiple instances on the same endpoint; it picks one based on the
107
`--router-mode` (round-robin by default if left unspecified).
108

109
Run `dynamo-run --help` for more options.
110

111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
### Network names

The `in=dyn://` URLs have the format `dyn://namespace.component.endpoint`. For quickstart just use any string `dyn://test`, `dynamo-run` will default any missing parts for you. The pieces matter for a larger system.

* *Namespace*: A pipeline. Usually a model. e.g "llama_8b". Just a name.
* *Component*: A load balanced service needed to run that pipeline. "backend", "prefill", "decode", "preprocessor", "draft", etc. This typically has some configuration (which model to use, for example).
* *Endpoint*: Like a URL. "generate", "load_metrics".
* *Instance*: A process. Unique. Dynamo assigns each one a unique instance_id. The thing that is running is always an instance. Namespace/component/endpoint can refer to multiple instances.

If you run two models, that is two pipelines. An exception would be if doing speculative decoding. The draft model is part of the pipeline of a bigger model.

If you run two instances of the same model ("data parallel") they are the same namespace+component+endpoint but different instances. The router will spread traffic over all the instances of a namespace+component+endpoint. If you have four prefill workers in a pipeline, they all have the same namespace+component+endpoint and are automatically assigned unique instance_ids.

Example 1: Data parallel load balanced, one model one pipeline two instances.
```
Node 1: dynamo-run in=dyn://qwen3-32b.backend.generate out=sglang /data/Qwen3-32B --tensor-parallel-size 2 --base-gpu-id 0
Node 2: dynamo-run in=dyn://qwen3-32b.backend.generate out=sglang /data/Qwen3-32B --tensor-parallel-size 2 --base-gpu-id 2
```

Example 2: Two models, two pipelines.
```
Node 1: dynamo-run in=dyn://qwen3-32b.backend.generate out=vllm /data/Qwen3-32B
Node 2: dynamo-run in=dyn://llama3-1-8b.backend.generate out=vllm /data/Llama-3.1-8B-Instruct/
```

Example 3: Different endpoints.

The KV metrics publisher in VLLM adds a `load_metrics` endpoint to the current component. If the `llama3-1-8b.backend` component above is using patched vllm it will also expose `llama3-1-8b.backend.load_metrics`.

140
Example 4: Multiple component in a pipeline.
141
142
143
144
145

In the P/D disaggregated setup you would have `deepseek-distill-llama8b.prefill.generate` (possibly multiple instance of this) and `deepseek-distill-llama8b.decode.generate`.

For output it is always only `out=dyn`. This tells Dynamo to auto-discover the instances, group them by model, and load balance appropriately (depending on `--router-mode` flag). The old syntax of `dyn://...` is still accepted for backwards compatibility.

146
147
148
149
### KV-aware routing

**Setup**

150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
Currently, only patched vllm supports KV-aware routing.

To set up KV-aware routing on patched vllm:

1. Ensure that `etcd` and `nats` (see [Quickstart with pip and vllm](#quickstart-with-pip-and-vllm)) are running and accessible from all nodes.
1. Create a virtualenv: `uv venv kvtest` and source its `activate`.
1. Use `pip` to **either**:
   1. Install Dynamo's vllm branch:
      ```
      uv pip install ai-dynamo-vllm
      ```
       **or**
   1. Install upstream vllm 0.8.4:
      ```
      uv pip install vllm==0.8.4
      ```
      And then patch it:
      ```
      cd kvtest/lib/python3.12/site-packages
      patch -p1 < $REPO_ROOT/container/deps/vllm/vllm_v0.8.4-dynamo-kv-disagg-patch.patch
      ```
1. Build the C bindings:
   ```
   cd $REPO_ROOT/lib/bindings/c
   cargo build`.
   ```
1. Put the library you just built on library path:
   ```
   export LD_LIBRARY_PATH=$REPO_ROOT/target/debug/
   ```
If you patched locally (instead of installing `ai-dynamo-vllm`), edit vllm's `platforms/__init__.py` to undo a patch change:
181
182
183
184
185
186
187
```
    #vllm_version = version("ai_dynamo_vllm")
    vllm_version = version("vllm")
```

**Start the workers**

188
The workers are started normally:
189
190
191
192
193
194
195
196

```
dynamo-run in=dyn://dynamo.endpoint.generate out=vllm /data/llms/Qwen/Qwen3-4B
```

**Start the ingress node**

```
197
dynamo-run in=http out=dyn --router-mode kv
198
199
```

200
The only difference from the distributed system above is `--router-mode kv`. The patched vllm announces when a KV block is created or removed. The Dynamo router run finds the worker with the best match for those KV blocks and directs the traffic to that node.
201

202
For performance testing, compare a typical workload with `--router-mode random|round-robin` to see if it can benefit from KV-aware routing.
203

204
## Full usage details
205

206
`dynamo run` executes `dynamo-run`. `dynamo-run` is also an example of what can be built in Rust with the `dynamo-llm` and `dynamo-runtime` crates. The following guide shows how to build from source with all the features.
207

208
### Getting Started
209

210
211
212
#### Setup

##### Step 1: Install libraries
213
**Ubuntu:**
214
```
215
sudo apt install -y build-essential libhwloc-dev libudev-dev pkg-config libssl-dev libclang-dev protobuf-compiler python3-dev cmake
216
217
```

218
**macOS:**
219
220
221
222
223
224
225
- [Homebrew](https://brew.sh/)
```
# if brew is not installed on your system, install it
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
```
- [Xcode](https://developer.apple.com/xcode/)

226
227
```
brew install cmake protobuf
228

229
## Check that Metal is accessible
230
231
xcrun -sdk macosx metal
```
232
If Metal is accessible, you should see an error like `metal: error: no input files`, which confirms it is installed correctly.
233

234
##### Step 2: Install Rust
235
```
236
237
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source $HOME/.cargo/env
238
```
239

240
##### Step 3: Build
241

242
- Linux with GPU and CUDA (tested on Ubuntu):
243
```
244
cargo build --features cuda
245
```
246

247
- macOS with Metal:
248
```
249
cargo build --features metal
250
251
```

252
- CPU only:
253
```
254
cargo build
255
256
```

257
258
259
260
261
262
263
Optionally you can run `cargo build` from any location with arguments:

```
--target-dir /path/to/target_directory` # specify target_directory with write privileges
--manifest-path /path/to/project/Cargo.toml` # if cargo build is run outside of `launch/` directory
```

264
The binary is called `dynamo-run` in `target/debug`
265
```
266
cd target/debug
267
268
```

269
270
Build with `--release` for a smaller binary and better performance, but longer build times. The binary will be in `target/release`.

271
272
273
274
275
276
277


#### Defaults
The input defaults to `in=text`. The output defaults to `out=mistralrs` engine, unless it is disabled with `--no-default-features` in which case vllm is used.
### Running Inference with Pre-built Engines

#### mistralrs
278
279
280
281

[mistral.rs](https://github.com/EricLBuehler/mistral.rs) is a pure Rust engine that is fast to run, fast to load, supports GGUF as well as safetensors, and runs well on CPU as well as GPU. For those reasons it is the default engine.

```
282
dynamo-run Qwen/Qwen3-4B
283
284
285
286
287
```

is equivalent to

```
288
dynamo-run in=text out=mistralrs Qwen/Qwen3-4B
289
290
```

291
292
If you have multiple GPUs, mistral.rs does automatic tensor parallelism. You do not need to pass any extra flags to dynamo-run to enable it.

293
#### llamacpp
294

295
[llama.cpp](https://github.com/ggml-org/llama.cpp) is built for CPU by default. For an optimized build pass the appropriate feature flag (highly recommended):
296

297
```
298
cargo build --features cuda|metal|vulkan -p dynamo-run
299
300
```

301
302
303
304
305
306
For GNU OpenMP support add the `openmp` feature. On Ubuntu this requires `libgomp1` (part of `build-essential`) at build and runtime.

```
cargo build --features cuda,openmp -p dynamo-run
```

307
```
308
309
dynamo-run out=llamacpp ~/llms/gemma-3-1b-it-q4_0.gguf
dynamo-run out=llamacpp ~/llms/Qwen3-0.6B-Q8_0.gguf # From https://huggingface.co/ggml-org
310
```
311

312
Note that in some cases we are unable to extract the tokenizer from the GGUF, and so a Hugging Face checkout of a matching model must also be passed. Dynamo uses the weights from the GGUF and the pre-processor (`tokenizer.json`, etc) from the `--model-config`:
313
```
314
dynamo-run out=llamacpp ~/llms/Llama-4-Scout-17B-16E-Instruct-UD-IQ1_S.gguf --context-length 32768 --model-config ~/llms/Llama-4-Scout-17B-16E-Instruct
315
316
317
```

If you have multiple GPUs, llama.cpp does automatic tensor parallelism. You do not need to pass any extra flags to dynamo-run to enable it.
318

319
#### sglang
320

321
322
The [SGLang](https://docs.sglang.ai/index.html) engine requires [etcd](https://etcd.io/) and [nats](https://nats.io/) with jetstream (`nats-server -js`) to be running.

323
324
1. Setup the python virtual env:

325
326
327
328
329
330
331
```
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/
```
332

333
334
2. Run

335
Any example above using `out=sglang` can work, but our sglang backend is also multi-gpu.
336
337

```
338
339
cd target/debug
./dynamo-run in=http out=sglang --model-path ~/llms/DeepSeek-R1-Distill-Llama-70B/ --tensor-parallel-size 8
340
341
```

342
To pass extra arguments to the sglang engine see [Extra engine arguments](#extra-engine-arguments).
343

344
**Multi-GPU**
345

346
347
348
349
350
Pass `--tensor-parallel-size <NUM-GPUS>` to `dynamo-run`.

```
dynamo-run out=sglang ~/llms/Llama-4-Scout-17B-16E-Instruct/ --tensor-parallel-size 8
```
351

352
To specify the GPU to start from, pass `--base-gpu-id <num>`; for example, on a shared eight GPU machine where GPUs 0–3 are already in use:
353
```
354
dynamo-run out=sglang <model> --tensor-parallel-size 4 --base-gpu-id 4
355
356
```

357
**Multinode:**
358

359
Dynamo only manages the leader node (node rank 0). The follower nodes are started in the [normal sglang way](https://docs.sglang.ai/references/deepseek.html#running-examples-on-multi-node).
360

361
Leader node:
362
```
363
364
365
366
367
368
dynamo-run out=sglang /data/models/DeepSeek-R1-Distill-Llama-70B/ --tensor-parallel-size 16 --node-rank 0 --num-nodes 2 --leader-addr 10.217.98.122:5000
```

All follower nodes. Increment `node-rank` each time:
```
python3 -m sglang.launch_server --model-path /data/models/DeepSeek-R1-Distill-Llama-70B --tp 16 --dist-init-addr 10.217.98.122:5000 --nnodes 2 --node-rank 1 --trust-remote-code
369
```
370
371
372
373

- Parameters `--leader-addr` and `--dist-init-addr` must match and be the IP address of the leader node. All followers must be able to connect. SGLang is using [PyTorch Distributed](https://docs.pytorch.org/tutorials/beginner/dist_overview.html) for networking.
- Parameters `--tensor-parallel-size` and `--tp` must match and be the total number of GPUs across the cluster.
- `--node-rank` must be unique consecutive integers starting at 1. The leader, managed by Dynamo, is 0.
Graham King's avatar
Graham King committed
374

375
#### vllm
Graham King's avatar
Graham King committed
376

377
378
379
Using the [vllm](https://github.com/vllm-project/vllm) Python library. Slow startup, fast inference. Supports both safetensors from HF and GGUF files, but is very slow for GGUF - prefer llamacpp.

The vllm engine requires requires [etcd](https://etcd.io/) and [nats](https://nats.io/) with jetstream (`nats-server -js`) to be running.
Graham King's avatar
Graham King committed
380
381
382

We use [uv](https://docs.astral.sh/uv/) but any virtualenv manager should work.

383
1. Setup:
Graham King's avatar
Graham King committed
384
385
386
387
```
uv venv
source .venv/bin/activate
uv pip install pip
388
uv pip install vllm==0.8.4 setuptools
Graham King's avatar
Graham King committed
389
390
```

391
392
393
```{note}
If you're on Ubuntu 22.04 or earlier, you must add `--python=python3.10` to your `uv venv` command.
```
Graham King's avatar
Graham King committed
394

395
2. Build:
Graham King's avatar
Graham King committed
396
```
397
cargo build
398
cd target/debug
Graham King's avatar
Graham King committed
399
400
```

401
402
403
404
3. Run
Inside that virtualenv:

**HF repo:**
Graham King's avatar
Graham King committed
405
```
406
./dynamo-run in=http out=vllm ~/llms/Llama-3.2-3B-Instruct/
Graham King's avatar
Graham King committed
407
408

```
409

410
To pass extra arguments to the vllm engine see [Extra engine arguments](#extra-engine-arguments).
411

412
413
vllm attempts to allocate enough KV cache for the full context length at startup. If that does not fit in your available memory pass `--context-length <value>`.

414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
If you see an error similar to the following:
```text
2025-06-28T00:32:32.507Z  WARN dynamo_run::subprocess: Traceback (most recent call last):
2025-06-28T00:32:32.507Z  WARN dynamo_run::subprocess:   File "/tmp/.tmpYeq5qA", line 29, in <module>
2025-06-28T00:32:32.507Z  WARN dynamo_run::subprocess:     from dynamo.llm import ModelType, WorkerMetricsPublisher, register_llm
2025-06-28T00:32:32.507Z  WARN dynamo_run::subprocess: ModuleNotFoundError: No module named 'dynamo'
```
Then run
```
uv pip install maturin
pip install patchelf
cd lib/bindings/python
maturin develop
```
this builds the Python->Rust bindings into that missing dynamo module. Rerun dynamo-run, the problem should be resolved.

430
**Multi-GPU**
431

432
Pass `--tensor-parallel-size <NUM-GPUS>` to `dynamo-run`.
433

434
To specify which GPUs to use set environment variable `CUDA_VISIBLE_DEVICES`.
435

436
**Multinode:**
437

438
vllm uses [ray](https://docs.vllm.ai/en/latest/serving/distributed_serving.html#running-vllm-on-multiple-nodes) for pipeline parallel inference. Dynamo does not change or manage that.
439

440
441
442
443
Here is an example on two 8x nodes:
- Leader node: `ray start --head --port=6379`
- Each follower node: `ray start --address='<HEAD_NODE_IP>:6379`
- Leader node: `dynamo-run out=vllm ~/llms/DeepSeek-R1-Distill-Llama-70B/ --tensor-parallel-size 16`
444

445
The `--tensor-parallel-size` parameter is the total number of GPUs in the cluster. This is often constrained by a model dimension such as being a divisor of the number of attention heads.
446

447
Startup can be slow so you may want to `export DYN_LOG=debug` to see progress.
448

449
Shutdown: `ray stop`
450

451
#### trtllm
452

453
454
455
456
457
Using [TensorRT-LLM's LLM API](https://nvidia.github.io/TensorRT-LLM/llm-api/), a high-level Python API.

You can use `--extra-engine-args` to pass extra arguments to LLM API engine.

The trtllm engine requires requires [etcd](https://etcd.io/) and [nats](https://nats.io/) with jetstream (`nats-server -js`) to be running.
458
459
460

##### Step 1: Build the environment

461
See instructions [here](https://github.com/ai-dynamo/dynamo/blob/main/examples/tensorrt_llm/README.md#build-docker) to build the dynamo container with TensorRT-LLM.
462
463
464

##### Step 2: Run the environment

465
See instructions [here](https://github.com/ai-dynamo/dynamo/blob/main/examples/tensorrt_llm/README.md#run-container) to run the built environment.
466
467
468
469
470

##### Step 3: Execute `dynamo run` command

Execute the following to load the TensorRT-LLM model specified in the configuration.
```
471
dynamo-run in=http out=trtllm TinyLlama/TinyLlama-1.1B-Chat-v1.0
472
473
```

474
#### Echo Engines
475
476
477

Dynamo includes two echo engines for testing and debugging purposes:

478
##### echo_core
479

480
The `echo_core` engine accepts pre-processed requests and echoes the tokens back as the response. This is useful for testing pre-processing functionality as the response includes the full prompt template.
481
482
483
484
485

```
dynamo-run in=http out=echo_core --model-path <hf-repo-checkout>
```

486
487
488
489
490
491
492
Note that to use it with `in=http` you need to tell the post processor to ignore stop tokens from the template by adding `nvext.ignore_eos` like this:
```
curl -N -d '{"nvext": {"ignore_eos": true}, "stream": true, "model": "Qwen2.5-3B-Instruct", "max_completion_tokens": 4096, "messages":[{"role":"user", "content": "Tell me a story" }]}' ...
```

The default `in=text` sets that for you.

493
##### echo_full
494
495
496
497

The `echo_full` engine accepts un-processed requests and echoes the prompt back as the response.

```
498
dynamo-run in=http out=echo_full --model-name my_model
499
500
```

501
##### Configuration
502
503
504
505
506
507
508
509
510

Both echo engines use a configurable delay between tokens to simulate generation speed. You can adjust this using the `DYN_TOKEN_ECHO_DELAY_MS` environment variable:

```
# Set token echo delay to 1ms (1000 tokens per second)
DYN_TOKEN_ECHO_DELAY_MS=1 dynamo-run in=http out=echo_full
```

The default delay is 10ms, which produces approximately 100 tokens per second.
511

512
#### Batch mode
513

514
`dynamo-run` can take a jsonl file full of prompts and evaluate them all:
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532

```
dynamo-run in=batch:prompts.jsonl out=llamacpp <model>
```

The input file should look like this:
```
{"text": "What is the capital of France?"}
{"text": "What is the capital of Spain?"}
```

Each one is passed as a prompt to the model. The output is written back to the same folder in `output.jsonl`. At the end of the run some statistics are printed.
The output looks like this:
```
{"text":"What is the capital of France?","response":"The capital of France is Paris.","tokens_in":7,"tokens_out":7,"elapsed_ms":1566}
{"text":"What is the capital of Spain?","response":".The capital of Spain is Madrid.","tokens_in":7,"tokens_out":7,"elapsed_ms":855}
```

533
534
535
536
537
538
539
540
541
542
543
544
545
### Extra engine arguments
The vllm and sglang backends support passing any argument the engine accepts.
Put the arguments in a JSON file:
```
{
    "dtype": "half",
    "trust_remote_code": true
}
```
Pass it like this:
```
dynamo-run out=sglang ~/llms/Llama-3.2-3B-Instruct --extra-engine-args sglang_extra.json
```
546
547
548
549
550
551
552
553
554
555
556
557
558
559

The tensorrtllm backend also support passing any argument the engine accepts. However, in this case config should be a yaml file.

```
backend: pytorch
kv_cache_config:
  event_buffer_max_size: 1024
```

Pass it like this:
```
dynamo-run in=http out=trtllm TinyLlama/TinyLlama-1.1B-Chat-v1.0 --extra-engine-args trtllm_extra.yaml
```

560
### Writing your own engine in Python
561
562
563
564
565
566
567
568
569
570
571
572

The [dynamo](https://pypi.org/project/ai-dynamo/) Python library allows you to build your own engine and attach it to Dynamo.

The Python file must do three things:
1. Decorate a function to get the runtime
2. Register on the network
3. Attach a request handler

```
from dynamo.llm import ModelType, register_llm
from dynamo.runtime import DistributedRuntime, dynamo_worker

573
574
575
576
   # 1. Decorate a function to get the runtime
   #
   @dynamo_worker(static=False)
   async def worker(runtime: DistributedRuntime):
577
578
579
580
581

    # 2. Register ourselves on the network
    #
    component = runtime.namespace("namespace").component("component")
    await component.create_service()
582
    model_path = "Qwen/Qwen3-0.6B" # or "/data/models/Qwen3-0.6B"
583
584
    model_type = ModelType.Backend
    endpoint = component.endpoint("endpoint")
585
586
    # Optional last param to register_llm is model_name. If not present derives it from model_path
    await register_llm(model_type, endpoint, model_path)
587
588
589
590
591
592

    # Initialize your engine here
    # engine = ...

    # 3. Attach request handler
    #
593
    await endpoint.serve_endpoint(RequestHandler(engine).generate)
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611

class RequestHandler:

    def __init__(self, engine):
        ...

    async def generate(self, request):
        # Call the engine
        # yield result dict
        ...

if __name__ == "__main__":
    uvloop.install()
    asyncio.run(worker())
```


The `model_path` can be:
612
- A HuggingFace repo ID, optionally prefixed with `hf://`. It is downloaded and cached locally.
613
614
615
616
617
618
619
620
- The path to a checkout of a HuggingFace repo - any folder containing safetensor files as well as `config.json`, `tokenizer.json` and `tokenizer_config.json`.
- The path to a GGUF file, if your engine supports that.

The `model_type` can be:
- ModelType.Backend. Dynamo handles pre-processing. Your `generate` method receives a `request` dict containing a `token_ids` array of int. It must return a dict also containing a `token_ids` array and an optional `finish_reason` string.
- ModelType.Chat. Your `generate` method receives a `request` and must return a response dict of type [OpenAI Chat Completion](https://platform.openai.com/docs/api-reference/chat). Your engine handles pre-processing.
- ModelType.Completion. Your `generate` method receives a `request` and must return a response dict of the older [Completions](https://platform.openai.com/docs/api-reference/completions). Your engine handles pre-processing.

621
622
623
624
625
`register_llm` can also take the following kwargs:
- `model_name`: The name to call the model. Your incoming HTTP requests model name must match this. Defaults to the hugging face repo name, the folder name, or the GGUF file name.
- `context_length`: Max model length in tokens. Defaults to the model's set max. Only set this if you need to reduce KV cache allocation to fit into VRAM.
- `kv_cache_block_size`: Size of a KV block for the engine, in tokens. Defaults to 16.

626
Here are some example engines:
627
628
629
630
631
632
633
634

- Backend:
    * [vllm](https://github.com/ai-dynamo/dynamo/blob/main/lib/bindings/python/examples/hello_world/server_vllm.py)
    * [sglang](https://github.com/ai-dynamo/dynamo/blob/main/lib/bindings/python/examples/hello_world/server_sglang.py)
- Chat:
    * [sglang](https://github.com/ai-dynamo/dynamo/blob/main/lib/bindings/python/examples/hello_world/server_sglang_tok.py)

More fully-featured Backend engines (used by `dynamo-run`):
635
636
637
- [vllm](https://github.com/ai-dynamo/dynamo/blob/main/launch/dynamo-run/src/subprocess/vllm_inc.py)
- [sglang](https://github.com/ai-dynamo/dynamo/blob/main/launch/dynamo-run/src/subprocess/sglang_inc.py)