Deploying vLLM on Kubernetes is a scalable and efficient way to serve machine learning models. This guide walks you through deploying vLLM using the [vLLM production stack](https://github.com/vllm-project/production-stack). Born out of a Berkeley-UChicago collaboration, [vLLM production stack](https://github.com/vllm-project/production-stack) is an officially released, production-optimized codebase under the [vLLM project](https://github.com/vllm-project), designed for LLM deployment with:
***Upstream vLLM compatibility** – It wraps around upstream vLLM without modifying its code.
***Ease of use** – Simplified deployment via Helm charts and observability through Grafana dashboards.
***High performance** – Optimized for LLM workloads with features like multi-model support, model-aware and prefix-aware routing, fast vLLM bootstrapping, and KV cache offloading with [LMCache](https://github.com/LMCache/LMCache), among others.
If you are new to Kubernetes, don't worry: in the vLLM production stack [repo](https://github.com/vllm-project/production-stack), we provide a step-by-step [guide](https://github.com/vllm-project/production-stack/blob/main/tutorials/00-install-kubernetes-env.md) and a [short video](https://www.youtube.com/watch?v=EsTJbQtzj0g) to set up everything and get started in **4 minutes**!
## Pre-requisite
Ensure that you have a running Kubernetes environment with GPU (you can follow [this tutorial](https://github.com/vllm-project/production-stack/blob/main/tutorials/00-install-kubernetes-env.md) to install a Kubernetes environment on a bare-medal GPU machine).
## Deployment using vLLM production stack
The standard vLLM production stack install uses a Helm chart. You can run this [bash script](https://github.com/vllm-project/production-stack/blob/main/tutorials/install-helm.sh) to install Helm on your GPU server.
To install the vLLM production stack, run the following commands on your desktop:
And then you can send out a query to the OpenAI-compatible API to check the available models:
```bash
curl -o- http://localhost:30080/models
```
Expected output:
```json
{
"object":"list",
"data":[
{
"id":"facebook/opt-125m",
"object":"model",
"created":1737428424,
"owned_by":"vllm",
"root":null
}
]
}
```
To send an actual chatting request, you can issue a curl request to the OpenAI `/completion` endpoint:
```bash
curl -X POST http://localhost:30080/completions \
-H"Content-Type: application/json"\
-d'{
"model": "facebook/opt-125m",
"prompt": "Once upon a time,",
"max_tokens": 10
}'
```
Expected output:
```json
{
"id":"completion-id",
"object":"text_completion",
"created":1737428424,
"model":"facebook/opt-125m",
"choices":[
{
"text":" there was a brave knight who...",
"index":0,
"finish_reason":"length"
}
]
}
```
### Uninstall
To remove the deployment, run:
```bash
sudo helm uninstall vllm
```
------
### (Advanced) Configuring vLLM production stack
The core vLLM production stack configuration is managed with YAML. Here is the example configuration used in the installation above:
```yaml
servingEngineSpec:
runtimeClassName:""
modelSpec:
-name:"opt125m"
repository:"vllm/vllm-openai"
tag:"latest"
modelURL:"facebook/opt-125m"
replicaCount:1
requestCPU:6
requestMemory:"16Gi"
requestGPU:1
pvcStorage:"10Gi"
```
In this YAML configuration:
***`modelSpec`** includes:
*`name`: A nickname that you prefer to call the model.
*`repository`: Docker repository of vLLM.
*`tag`: Docker image tag.
*`modelURL`: The LLM model that you want to use.
***`replicaCount`**: Number of replicas.
***`requestCPU` and `requestMemory`**: Specifies the CPU and memory resource requests for the pod.
***`requestGPU`**: Specifies the number of GPUs required.
***`pvcStorage`**: Allocates persistent storage for the model.
**NOTE:** If you intend to set up two pods, please refer to this [YAML file](https://github.com/vllm-project/production-stack/blob/main/tutorials/assets/values-01-2pods-minimal-example.yaml).
**NOTE:** vLLM production stack offers many more features (*e.g.* CPU offloading and a wide range of routing algorithms). Please check out these [examples and tutorials](https://github.com/vllm-project/production-stack/tree/main/tutorials) and our [repo](https://github.com/vllm-project/production-stack) for more details!
Using Kubernetes to deploy vLLM is a scalable and efficient way to serve machine learning models. This guide will walk you through the process of deploying vLLM with Kubernetes, including the necessary prerequisites, steps for deployment, and testing.
Deploying vLLM on Kubernetes is a scalable and efficient way to serve machine learning models. This guide walks you through deploying vLLM using native Kubernetes.
## Prerequisites
*[Deployment with CPUs](#deployment-with-cpus)
*[Deployment with GPUs](#deployment-with-gpus)
Before you begin, ensure that you have the following:
Alternatively, you can deploy vLLM to Kubernetes using any of the following:
- If you have your HuggingFace models cached somewhere else, update `hf_cache_dir` below.
- If you have your HuggingFace models cached somewhere else, update `hf_cache_dir` below.
- If you don't have an existing HuggingFace cache you will want to start `vllm0` and wait for the model to complete downloading and the server to be ready. This will ensure that `vllm1` can leverage the model you just downloaded and it won't have to be downloaded again.
- If you don't have an existing HuggingFace cache you will want to start `vllm0` and wait for the model to complete downloading and the server to be ready. This will ensure that `vllm1` can leverage the model you just downloaded and it won't have to be downloaded again.
- The below example assumes GPU backend used. If you are using CPU backend, remove `--gpus all`, add `VLLM_CPU_KVCACHE_SPACE` and `VLLM_CPU_OMP_THREADS_BIND` environment variables to the docker run command.
- The below example assumes GPU backend used. If you are using CPU backend, remove `--gpus device=ID`, add `VLLM_CPU_KVCACHE_SPACE` and `VLLM_CPU_OMP_THREADS_BIND` environment variables to the docker run command.
- Adjust the model name that you want to use in your vLLM servers if you don't want to use `Llama-2-7b-chat-hf`.
- Adjust the model name that you want to use in your vLLM servers if you don't want to use `Llama-2-7b-chat-hf`.
```console
```console
mkdir -p ~/.cache/huggingface/hub/
mkdir -p ~/.cache/huggingface/hub/
hf_cache_dir=~/.cache/huggingface/
hf_cache_dir=~/.cache/huggingface/
docker run -itd --ipc host --privileged --network vllm_nginx --gpus all --shm-size=10.24gb -v $hf_cache_dir:/root/.cache/huggingface/ -p 8081:8000 --name vllm0 vllm --model meta-llama/Llama-2-7b-chat-hf
@@ -14,7 +14,7 @@ Let's say we want to serve the popular QWen model by running `vllm serve Qwen/Qw
...
@@ -14,7 +14,7 @@ Let's say we want to serve the popular QWen model by running `vllm serve Qwen/Qw
2. After confirming the existence of the model, vLLM loads its config file and converts it into a dictionary. See this [code snippet](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L185-L186) for the implementation.
2. After confirming the existence of the model, vLLM loads its config file and converts it into a dictionary. See this [code snippet](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L185-L186) for the implementation.
3. Next, vLLM [inspects](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L189) the `model_type` field in the config dictionary to [generate](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#190-L216) the config object to use. There are some `model_type` values that vLLM directly supports; see [here](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L48) for the list. If the `model_type` is not in the list, vLLM will use [AutoConfig.from_pretrained](https://huggingface.co/docs/transformers/en/model_doc/auto#transformers.AutoConfig.from_pretrained) to load the config class, with `model`, `--revision`, and `--trust_remote_code` as the arguments. Please note that:
3. Next, vLLM [inspects](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L189) the `model_type` field in the config dictionary to [generate](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L190-L216) the config object to use. There are some `model_type` values that vLLM directly supports; see [here](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L48) for the list. If the `model_type` is not in the list, vLLM will use [AutoConfig.from_pretrained](https://huggingface.co/docs/transformers/en/model_doc/auto#transformers.AutoConfig.from_pretrained) to load the config class, with `model`, `--revision`, and `--trust_remote_code` as the arguments. Please note that:
- HuggingFace also has its own logic to determine the config class to use. It will again use the `model_type` field to search for the class name in the transformers library; see [here](https://github.com/huggingface/transformers/tree/main/src/transformers/models) for the list of supported models. If the `model_type` is not found, HuggingFace will use the `auto_map` field from the config JSON file to determine the class name. Specifically, it is the `AutoConfig` field under `auto_map`. See [DeepSeek](https://huggingface.co/deepseek-ai/DeepSeek-V2.5/blob/main/config.json) for an example.
- HuggingFace also has its own logic to determine the config class to use. It will again use the `model_type` field to search for the class name in the transformers library; see [here](https://github.com/huggingface/transformers/tree/main/src/transformers/models) for the list of supported models. If the `model_type` is not found, HuggingFace will use the `auto_map` field from the config JSON file to determine the class name. Specifically, it is the `AutoConfig` field under `auto_map`. See [DeepSeek](https://huggingface.co/deepseek-ai/DeepSeek-V2.5/blob/main/config.json) for an example.
- The `AutoConfig` field under `auto_map` points to a module path in the model's repository. To create the config class, HuggingFace will import the module and use the `from_pretrained` method to load the config class. This can generally cause arbitrary code execution, so it is only executed when `--trust_remote_code` is enabled.
- The `AutoConfig` field under `auto_map` points to a module path in the model's repository. To create the config class, HuggingFace will import the module and use the `from_pretrained` method to load the config class. This can generally cause arbitrary code execution, so it is only executed when `--trust_remote_code` is enabled.
@@ -6,11 +6,16 @@ To enable various optimizations in vLLM such as [chunked prefill](#chunked-prefi
...
@@ -6,11 +6,16 @@ To enable various optimizations in vLLM such as [chunked prefill](#chunked-prefi
Here are the main features of {class}`~vllm.multimodal.processing.BaseMultiModalProcessor`:
Here are the main features of {class}`~vllm.multimodal.processing.BaseMultiModalProcessor`:
## Prompt Replacement Detection
## Prompt Update Detection
One of the main responsibilies of HF processor is to replace input placeholder tokens (e.g. `<image>` for a single image) with feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size). The information about which tokens have been replaced is key to finding the correspondence between placeholder feature tokens and multi-modal inputs.
One of the main responsibilies of HF processor is to update the prompt with placeholder tokens. For example:
In vLLM, this information is specified using {class}`~vllm.multimodal.processing.PromptReplacement` in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`. Given this specification, we can automatically detect whether HF has replaced the input placeholder tokens by checking whether the feature placeholder tokens exist in the prompt.
- Insert feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size) at the start of the string.
- Replace existing input placeholder tokens (e.g. `<image>` for a single image) with feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size).
The information about which tokens have been updated is key to finding the correspondence between placeholder feature tokens and multi-modal inputs.
In vLLM, this information is specified using {class}`~vllm.multimodal.processing.PromptUpdate` in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates`. We can automatically detect whether HF has updated the prompt by checking the existence of the updated tokens.
## Tokenized Prompt Inputs
## Tokenized Prompt Inputs
...
@@ -22,7 +27,7 @@ Consider that HF processors follow these main steps:
...
@@ -22,7 +27,7 @@ Consider that HF processors follow these main steps:
1. Tokenize the text
1. Tokenize the text
2. Process multi-modal inputs
2. Process multi-modal inputs
3. Perform prompt replacement
3. Perform prompt updates
And we require that:
And we require that:
...
@@ -44,16 +49,16 @@ Moreover, since the tokenized text has not passed through the HF processor, we h
...
@@ -44,16 +49,16 @@ Moreover, since the tokenized text has not passed through the HF processor, we h
We work around the first issue by requiring each model to define how to generate dummy text based on the number of multi-modal inputs, via {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`. This lets us generate dummy text corresponding to the multi-modal inputs and input them together to obtain the processed multi-modal data.
We work around the first issue by requiring each model to define how to generate dummy text based on the number of multi-modal inputs, via {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`. This lets us generate dummy text corresponding to the multi-modal inputs and input them together to obtain the processed multi-modal data.
(mm-automatic-prompt-replacement)=
(mm-automatic-prompt-updating)=
### Automatic prompt replacement
### Automatic prompt updating
We address the second issue by implementing model-agnostic code in
We address the second issue by implementing model-agnostic code in
{meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_prompt_replacements` to automatically replace input placeholder tokens with feature placeholder tokens based on the specification outputted by {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`.
{meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_prompt_updates` to automatically update the prompt with feature placeholder tokens based on the specification outputted by {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates`.
### Summary
### Summary
With the help of dummy text and automatic prompt replacement, our multi-modal processor can finally accept both text and token prompts with multi-modal data. The detailed logic is shown in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_main`.
With the help of dummy text and automatic prompt updating, our multi-modal processor can finally accept both text and token prompts with multi-modal data. The detailed logic is shown in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_main`.
## Processor Output Caching
## Processor Output Caching
...
@@ -61,4 +66,4 @@ Some HF processors, such as the one for Qwen2-VL, are [very slow](gh-issue:9238)
...
@@ -61,4 +66,4 @@ Some HF processors, such as the one for Qwen2-VL, are [very slow](gh-issue:9238)
When new data is passed in, we first check which items are in the cache, and which ones are missing. The missing items are passed into the HF processor in a single batch and cached, before being merged with the existing items in the cache.
When new data is passed in, we first check which items are in the cache, and which ones are missing. The missing items are passed into the HF processor in a single batch and cached, before being merged with the existing items in the cache.
Since we only process the missing multi-modal data items, the number of input placeholder tokens no longer corresponds to the number of the multi-modal inputs, so they can't be passed alongside the text prompt to HF processor. Therefore, we process the text and multi-modal inputs separately, using [dummy text](#mm-dummy-text) to avoid HF errors. Since this skips HF's prompt replacement code, we apply [automatic prompt replacement](#mm-automatic-prompt-replacement) afterwards to keep the output tokens and multi-modal data consistent with each other.
Since we only process the missing multi-modal data items, the number of input placeholder tokens no longer corresponds to the number of the multi-modal inputs, so they can't be passed alongside the text prompt to HF processor. Therefore, we process the text and multi-modal inputs separately, using [dummy text](#mm-dummy-text) to avoid HF errors. Since this skips HF's prompt updating code, we apply [automatic prompt updating](#mm-automatic-prompt-updating) afterwards to keep the output tokens and multi-modal data consistent with each other.
Ensure the v1 LLM Engine exposes a superset of the metrics available in v0.
## Objectives
- Achieve parity of metrics between v0 and v1.
- The priority use case is accessing these metrics via Prometheus as this is what we expect to be used in production environments.
- Logging support - i.e. printing metrics to the info log - is provided for more ad-hoc testing, debugging, development, and exploratory use cases.
## Background
Metrics in vLLM can be categorized as follows:
1. Server-level metrics: these are global metrics that track the state and performance of the LLM engine. These are typically exposed as Gauges or Counters in Prometheus.
2. Request-level metrics: these are metrics that track the characteristics - e.g. size and timing - of individual requests. These are typically exposed as Histograms in Prometheus, and are often the SLO that an SRE monitoring vLLM will be tracking.
The mental model is that the "Server-level Metrics" explain why the "Request-level Metrics" are what they are.
### v0 Metrics
In v0, the following metrics are exposed via a Prometheus-compatible `/metrics` endpoint using the `vllm:` prefix:
These are documented under [Inferencing and Serving -> Production Metrics](project:../../serving/metrics.md).
### Grafana Dashboard
vLLM also provides [a reference example](https://docs.vllm.ai/en/latest/getting_started/examples/prometheus_grafana.html) for how to collect and store these metrics using Prometheus and visualize them using a Grafana dashboard.
The subset of metrics exposed in the Grafana dashboard gives us an indication of which metrics are especially important:
-`vllm:e2e_request_latency_seconds_bucket` - End to end request latency measured in seconds
-`vllm:request_success_total` - Number of finished requests by their finish reason: either an EOS token was generated or the max sequence length was reached
-`vllm:request_queue_time_seconds` - Queue Time
-`vllm:request_prefill_time_seconds` - Requests Prefill Time
-`vllm:request_decode_time_seconds` - Requests Decode Time
-`vllm:request_max_num_generation_tokens` - Max Generation Token in Sequence Group
See [the PR which added this Dashboard](gh-pr:2316) for interesting and useful background on the choices made here.
### Prometheus Client Library
Prometheus support was initially added [using the aioprometheus library](gh-pr:1890), but a switch was made quickly to [prometheus_client](gh-pr:2730). The rationale is discussed in both linked PRs.
### Multi-process Mode
In v0, metrics are collected in the engine core process and we use multi-process mode to make them available in the API server process. See <gh-pr:7279>.
### Built in Python/Process Metrics
The following metrics are supported by default by `prometheus_client`, but the are not exposed with multiprocess mode is used:
-`python_gc_objects_collected_total`
-`python_gc_objects_uncollectable_total`
-`python_gc_collections_total`
-`python_info`
-`process_virtual_memory_bytes`
-`process_resident_memory_bytes`
-`process_start_time_seconds`
-`process_cpu_seconds_total`
-`process_open_fds`
-`process_max_fds`
This is relevant because if we move away from multiprocess mode in v1,
we get these back. However, it's questionable how relevant these are
if they don't aggregate these stats for all processes that make up a
vLLM instance.
### v0 PRs and Issues
For background, these are some of the relevant PRs which added the v0 metrics:
-<gh-pr:1890>
-<gh-pr:2316>
-<gh-pr:2730>
-<gh-pr:4464>
-<gh-pr:7279>
Also note the ["Even Better Observability"](gh-issue:3616) feature where e.g. [a detailed roadmap was laid out](gh-issue:3616#issuecomment-2030858781).
## v1 Design
### v1 PRs
For background, here are the relevant v1 PRs relating to the v1
metrics issue <gh-issue:10582>:
-<gh-pr:11962>
-<gh-pr:11973>
-<gh-pr:10907>
-<gh-pr:12416>
-<gh-pr:12478>
-<gh-pr:12516>
-<gh-pr:12530>
-<gh-pr:12561>
-<gh-pr:12579>
-<gh-pr:12592>
-<gh-pr:12644>
### Metrics Collection
In v1, we wish to move computation and overhead out of the engine core
process to minimize the time between each forward pass.
The overall idea of V1 EngineCore design is:
- EngineCore is the inner loop. Performance is most critical here
- AsyncLLM is the outer loop. This is overlapped with GPU execution
(ideally), so this is where any "overheads" should be if
possible. So AsyncLLM.output_handler_loop is the ideal place for the
metrics bookkeeping if possible.
We will achieve this by collecting metrics in the frontend API server,
and base these metrics on information we can glean from the
`EngineCoreOutputs` returned by the engine core process to the
frontend.
### Interval Calculations
Many of our metrics are the time interval between various events in
the processing of a request. It is best practice to use timestamps
based on "monotonic time" (`time.monotonic()`) rather than "wall-clock
time" (`time.time()`) to calculate intervals as the former is
unaffected by system clock changes (e.g. from NTP).
It's also important to note that monotonic clocks differ between
processes - each process has its own reference. point. So it is
meaningless to compare monotonic timestamps from different processes.
Therefore, in order to calculate an interval, we must compare two
monotonic timestamps from the same process.
### Scheduler Stats
The engine core process will collect some key statistics from the
scheduler - e.g. the number of requests that were scheduled or waiting
after the last scheduler pass - and include those statistics in
`EngineCoreOutputs`.
### Engine Core Events
The engine core will also record the timestamp of certain per-request
events so that the frontend can calculate the interval between these
events.
The events are:
-`QUEUED` - when the request was received by the engine core and
added to the scheduler queue.
-`SCHEDULED` - when the request was first scheduled for execution.
-`PREEMPTED` - the request has been put back in the waiting queue
in order to make room for other requests to complete. It will be
re-scheduled in future and re-start its prefill phase.
-`NEW_TOKENS` - when the output included in `EngineCoreOutput` was
generated. Since this is common to all requests in a given
iteration, we use a single timestamp on `EngineCoreOutputs` to
record this event.
And the calculated intervals are:
- Queue interval - between `QUEUED` and most recent `SCHEDULED`.
- Prefill interval - between most recent `SCHEDULED` and the subsequent
first `NEW_TOKENS`.
- Decode interval - between first (after the most recent `SCHEDULED`) and
last `NEW_TOKENS`.
- Inference interval - between most recent `SCHEDULED` and last `NEW_TOKENS`.
- Inter-token interval - between successive `NEW_TOKENS`.
@@ -183,7 +183,7 @@ When a request is finished, we free all its blocks if no other requests are usin
...
@@ -183,7 +183,7 @@ When a request is finished, we free all its blocks if no other requests are usin
When the head block (least recently used block) of the free queue is cached, we have to evict the block to prevent it from being used by other requests. Specifically, eviction involves the following steps:
When the head block (least recently used block) of the free queue is cached, we have to evict the block to prevent it from being used by other requests. Specifically, eviction involves the following steps:
1. Pop the block from the head of the free queue. This is the LRU black to be evicted.
1. Pop the block from the head of the free queue. This is the LRU block to be evicted.
2. Remove the block ID from the Cache Block.
2. Remove the block ID from the Cache Block.
3. Remove the block hash.
3. Remove the block hash.
...
@@ -191,7 +191,7 @@ When the head block (least recently used block) of the free queue is cached, we
...
@@ -191,7 +191,7 @@ When the head block (least recently used block) of the free queue is cached, we
In this example, we assume the block size is 4 (each block can cache 4 tokens), and we have 10 blocks in the KV-cache manager in total.
In this example, we assume the block size is 4 (each block can cache 4 tokens), and we have 10 blocks in the KV-cache manager in total.
**Time 1: The cache is empty and a new request comes in.** We allocate 4 blocks. 3 of them are already full and cached. The fourth block is partially full with 2 of 4 tokens.
**Time 1: The cache is empty and a new request comes in.** We allocate 4 blocks. 3 of them are already full and cached. The fourth block is partially full with 3 of 4 tokens.
@@ -203,7 +203,7 @@ In this example, we assume the block size is 4 (each block can cache 4 tokens),
...
@@ -203,7 +203,7 @@ In this example, we assume the block size is 4 (each block can cache 4 tokens),
:alt: Example Time 3
:alt: Example Time 3
:::
:::
**Time 4: Request 1 comes in with the 14 prompt tokens, where the first 11 tokens are the same as request 0.** We can see that only 2 blocks (11 tokens) hit the cache, because the 3rd block only matches 3 of 4 tokens.
**Time 4: Request 1 comes in with the 14 prompt tokens, where the first 10 tokens are the same as request 0.** We can see that only the first 2 blocks (8 tokens) hit the cache, because the 3rd block only matches 2 of 4 tokens.
@@ -221,7 +221,7 @@ In this example, we assume the block size is 4 (each block can cache 4 tokens),
...
@@ -221,7 +221,7 @@ In this example, we assume the block size is 4 (each block can cache 4 tokens),
:alt: Example Time 6
:alt: Example Time 6
:::
:::
**Time 7: Request 2 comes in with the 33 prompt tokens, where the first 16 tokens are the same as request 0\.** Note that even the block order in the free queue was `7 - 8 - 9 - 4 - 3 - 2 - 6 - 5 - 1 - 0`, the cache hit blocks (i.e., 0, 1, 2) are touched and removed from the queue before allocation, so the free queue becomes `7 - 8 - 9 - 4 - 3 - 6 - 5`. As a result, the allocated blocks are 0 (cached), 1 (cached), 2 (cached), 7, 8, 9, 4, 3 (evicted).
**Time 7: Request 2 comes in with the 29 prompt tokens, where the first 12 tokens are the same as request 0\.** Note that even the block order in the free queue was `7 - 8 - 9 - 4 - 3 - 2 - 6 - 5 - 1 - 0`, the cache hit blocks (i.e., 0, 1, 2) are touched and removed from the queue before allocation, so the free queue becomes `7 - 8 - 9 - 4 - 3 - 6 - 5`. As a result, the allocated blocks are 0 (cached), 1 (cached), 2 (cached), 7, 8, 9, 4, 3 (evicted).
In vLLM's V1 architecture, `torch.compile` is enabled by default and is a critical part of the framework. This document gives a simple walk-through example to show how to understand the `torch.compile` usage.
Throughout the example, we will run a common Llama model using v1, and turn on debug level logging to show all the details. The command to be used is `VLLM_USE_V1=1 VLLM_LOGGING_LEVEL=DEBUG vllm serve meta-llama/Llama-3.2-1B`.
## Compilation Cache
In the very verbose logs, we can see:
```
INFO 03-07 03:06:55 [backends.py:409] Using cache directory: ~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0 for vLLM's torch.compile
```
vLLM will take all the available factors into consideration, and decide a directory to store all the compilation artifact. This means, you can directly copy the whole `~/.cache/vllm/torch_compile_cache` directory in your deployment scenario to save a great amount of compilation time, and hence accelerating the starting time of the vLLM instance.
The factors considered include:
- All the related configs (see the `compute_hash` functions in the [config.py](gh-file:vllm/config.py))
- PyTorch configs (see the `compute_hash` functions in the [compiler_interface.py](gh-file:vllm/compilation/compiler_interface.py))
- The model's forward function and the relevant functions called by the forward function (see below)
With all these factors taken into consideration, usually we can guarantee that the cache is safe to use, and will not cause any unexpected behavior. Therefore, the cache is enabled by default. If you want to debug the compilation process, or if you suspect the cache is causing some issues, you can disable it by setting the environment variable `VLLM_DISABLE_COMPILE_CACHE=1`.
A unique aspect of vLLM's `torch.compile` integration, is that we guarantee all the compilation finishes before we serve any requests. No requests will trigger new compilations. Otherwise, the engine would be blocked on that request, and the response time will have unexpected spikes.
## Python Code Compilation
In the very verbose logs, we can see:
```
DEBUG 03-07 03:06:52 [decorators.py:203] Start compiling function <code object forward at 0x7f08acf40c90, file "xxx/vllm/model_executor/models/llama.py", line 339>
DEBUG 03-07 03:06:54 [backends.py:370] Traced files (to be considered for compilation cache):
DEBUG 03-07 03:07:07 [backends.py:462] Computation graph saved to ~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/computation_graph.py
DEBUG 03-07 03:07:07 [wrapper.py:105] Dynamo transformed code saved to ~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/transformed_code.py
```
This is about the Python code compilation, i.e. graph capture by Dynamo. It tries to trace the function with code `xxx/vllm/model_executor/models/llama.py:339`, which is the `forward` function of the model we compile. During the forward pass, there are also other functions called and inlined by Dynamo, as shown by the logs, including some PyTorch functions from `xxx/torch/nn/modules/module.py` (used by PyTorch `nn.Module`, because module attribute access will trigger a function call), some communication / attention / activation functions from vLLM. All the traced files will be considered when we decide the cache directory to use. This way, any code change in the above files will trigger compilation cache miss, and therefore recompilation.
The result of the Dynamo compilation, is a new function stored in `~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/transformed_code.py`. Usually, this function unpacks tensors from the module, and then pass it to the traced computation graph. The computation graph is stored in `~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/computation_graph.py`.
## Computation Graph Processing
The computation graph has shape annotations for every tensor. The inputs are input ids, position ids, weights and buffers from the model, and the outputs are the final hidden states. Note that lm head projection and sampling operations are not considered in the graph.
Most of the inputs to the computation graph has static shape, since they are model weights and buffers, and will not change during the lifetime of the model. Only the input ids and position ids have symbolic shapes, i.e. the shape can change from batch to batch. However, they will share the same symbolic shapes. That is to say, the only changing size to the computation graph, is the batch size (number of tokens processed in the current forward pass).
The attention operation is complicated, and it needs to interact with kv caches, with complicated shapes. Fortunately, the output of the attention operation just share the same shape as the input query of the attention operation. Therefore, we wrap the whole attention operation into a PyTorch custom op `torch.ops.vllm.unified_attention_with_output`, so that Dynamo will not try to inspect any of the internal operations. This way, although attention operation is complicated, we can still capture the model's computation graph as a full-graph, from Dynamo's perspective.
The computation graph is further split into pieces, by the `splitting_ops` (usually this is the attention operation). Therefore, in the `~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/computation_graph.py` file, we can see lots of submodules, each submodule is a piece of graph after splitting:
- Attention operation itself is a submodule.
- The part of computation graph, from one attention operation to the next attention operation, is a submodule.
Every submodule can be identified by its index, and will be processed individually.
## Computation Graph Compilation
In the very verbose logs, we can also see:
```
DEBUG 03-07 03:52:37 [backends.py:134] store the 0-th graph for shape None from inductor via handle ('fpegyiq3v3wzjzphd45wkflpabggdbjpylgr7tta4hj6uplstsiw', '~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/inductor_cache/iw/ciwzrk3ittdqatuzwonnajywvno3llvjcs2vfdldzwzozn3zi3iy.py')
DEBUG 03-07 03:52:39 [backends.py:134] store the 1-th graph for shape None from inductor via handle ('f7fmlodmf3h3by5iiu2c4zarwoxbg4eytwr3ujdd2jphl4pospfd', '~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/inductor_cache/ly/clyfzxldfsj7ehaluis2mca2omqka4r7mgcedlf6xfjh645nw6k2.py')
...
DEBUG 03-07 03:52:45 [backends.py:134] store the 15-th graph for shape None from inductor via handle ('f7fmlodmf3h3by5iiu2c4zarwoxbg4eytwr3ujdd2jphl4pospfd', '~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/inductor_cache/ly/clyfzxldfsj7ehaluis2mca2omqka4r7mgcedlf6xfjh645nw6k2.py')
DEBUG 03-07 03:52:45 [backends.py:134] store the 16-th graph for shape None from inductor via handle ('fvj3ccoi7m34f3dnr4itmu55mmun44l5xymwhrjlwisylsk7q6jy', '~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/inductor_cache/tf/ctfftkglj7b4lcttq5cymx6cew372uoauupqn6ldsvpiucavqcjc.py')
```
This means the first piece of computation graph (with shape `None` for symbolic shape) is compiled by Inductor (with a key `fpegyiq3v3wzjzphd45wkflpabggdbjpylgr7tta4hj6uplstsiw`). The compiled kernel is stored in `~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/inductor_cache/iw/ciwzrk3ittdqatuzwonnajywvno3llvjcs2vfdldzwzozn3zi3iy.py`. You can open the file to see what is the code Inductor finally runs.
One more detail: you can see that the 1-th graph and the 15-th graph have the same key, while the 0-th graph and the 16-th graph are different. This is expected, since we split the graph by the attention op, we get 3 unique subgraphs:
- the first layer before attention
- every middle layer, from one attention operation to the next attention operation
- the final layer after attention
If we already have the cache directory (e.g. run the same code for the second time), we will see the following logs:
```
DEBUG 03-07 04:00:45 [backends.py:86] Directly load the 0-th graph for shape None from inductor via handle ('fpegyiq3v3wzjzphd45wkflpabggdbjpylgr7tta4hj6uplstsiw', '~/.cache/vllm/torch_compile_cache/1517964802/rank_0_0/inductor_cache/iw/ciwzrk3ittdqatuzwonnajywvno3llvjcs2vfdldzwzozn3zi3iy.py')
```
This time, Inductor compilation is completely bypassed, and we will load from disk to read the compilation artifact we get from the last time.
The above example just uses Inductor to compile for a general shape (i.e. symbolic shape). We can also use Inductor to compile for some of the specific shapes, for example:
Then it will also compile a specific kernel just for batch size `1, 2, 4, 8`. At this time, all of the shapes in the computation graph are static and known, and we will turn on auto-tuning to tune for max performance. This can be slow when you run it for the first time, but the next time you run it, we can directly bypass the tuning and run the tuned kernel.
When all the shapes are known, `torch.compile` can compare different configs, and often find some better configs to run the kernel. For example, we can see the following log:
SingleProcess AUTOTUNE benchmarking takes 2.0428 seconds and 7.5727 seconds precompiling
```
It means, for a matrix multiplication with shape `8x2048x3072`, `torch.compile` tries triton template with various configs, and it is much faster than the default code (which dispatches to cublas library).
Unfortunately, because auto-tuning takes quite a long time (from seconds to minutes, depending on the model size and the batch size), even though it can be cached for later use, for the sake of user-friendliness, we turn it off by default. If you want to have max performance, it is recommended to try it, by compiling specific shapes.
## Cudagraph Capture
vLLM's V1 architecture uses piecewise cudagraph. The full computation graph is split as mentioned above, and we only capture the cudagraph for the piece of graph between attention operations (including the first graph before any attention operation, and the last graph after all the attention operation). This is based on a common observation: computation between attentions are usually token-wise and easy to deal with for cudagraph; while the attention operation is non-trival to be cudagraph compatible. Thus, by running the attention operation in eager mode while the rest operations in cudagraph, we keep the flexibility of the attention operation.
The piecewise cudagraph also has fine-grained memory management. The purpose is to only exclude the attention kernel from cudagraph, while keeping all the rest modules and the memory allocation operations in the cudagraph. This is why the attention operation in V1 has the output tensor as the input of the attention.
The cudagraphs are captured and managed by the compiler backend, and replayed when the batch size has corresponding cudagraph captured. The caller of the model (model runner) only needs to make sure it manages the input buffers correctly. All of the intermediate buffers are managed automatically by the compiler backend.
By default, vLLM will try to determine a set of sizes to capture cudagraph. You can also override it using the config `cudagraph_capture_sizes`:
@@ -110,7 +110,7 @@ In addition to serving LoRA adapters at server startup, the vLLM server now supp
...
@@ -110,7 +110,7 @@ In addition to serving LoRA adapters at server startup, the vLLM server now supp
LoRA adapters at runtime through dedicated API endpoints. This feature can be particularly useful when the flexibility
LoRA adapters at runtime through dedicated API endpoints. This feature can be particularly useful when the flexibility
to change models on-the-fly is needed.
to change models on-the-fly is needed.
Note: Enabling this feature in production environments is risky as user may participate model adapter management.
Note: Enabling this feature in production environments is risky as users may participate in model adapter management.
To enable dynamic LoRA loading and unloading, ensure that the environment variable `VLLM_ALLOW_RUNTIME_LORA_UPDATING`
To enable dynamic LoRA loading and unloading, ensure that the environment variable `VLLM_ALLOW_RUNTIME_LORA_UPDATING`
is set to `True`. When this option is enabled, the API server will log a warning to indicate that dynamic loading is active.
is set to `True`. When this option is enabled, the API server will log a warning to indicate that dynamic loading is active.
...
@@ -170,7 +170,7 @@ Now, you can specify a base_model_name alongside the name and path using JSON fo
...
@@ -170,7 +170,7 @@ Now, you can specify a base_model_name alongside the name and path using JSON fo
To provide the backward compatibility support, you can still use the old key-value format (name=path), but the `base_model_name` will remain unspecified in that case.
To provide the backward compatibility support, you can still use the old key-value format (name=path), but the `base_model_name` will remain unspecified in that case.
## Lora model lineage in model card
## LoRA model lineage in model card
The new format of `--lora-modules` is mainly to support the display of parent model information in the model card. Here's an explanation of how your current response supports this:
The new format of `--lora-modules` is mainly to support the display of parent model information in the model card. Here's an explanation of how your current response supports this:
To create a new 4-bit quantized model, you can leverage [AutoAWQ](https://github.com/casper-hansen/AutoAWQ).
To create a new 4-bit quantized model, you can leverage [AutoAWQ](https://github.com/casper-hansen/AutoAWQ).
Quantizing reduces the model's precision from FP16 to INT4 which effectively reduces the file size by ~70%.
Quantization reduces the model's precision from BF16/FP16 to INT4 which effectively reduces the total model memory footprint.
The main benefits are lower latency and memory usage.
The main benefits are lower latency and memory usage.
You can quantize your own models by installing AutoAWQ or picking one of the [400+ models on Huggingface](https://huggingface.co/models?sort=trending&search=awq).
You can quantize your own models by installing AutoAWQ or picking one of the [6500+ models on Huggingface](https://huggingface.co/models?sort=trending&search=awq).
```console
```console
pip install autoawq
pip install autoawq
```
```
After installing AutoAWQ, you are ready to quantize a model. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`:
After installing AutoAWQ, you are ready to quantize a model. Please refer to the `AutoAWQ documentation <https://casper-hansen.github.io/AutoAWQ/examples/#basic-quantization>`_ for further details. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`:
We recommend using the tokenizer from base model instead of GGUF model. Because the tokenizer conversion from GGUF is time-consuming and unstable, especially for some models with large vocab size.
We recommend using the tokenizer from base model instead of GGUF model. Because the tokenizer conversion from GGUF is time-consuming and unstable, especially for some models with large vocab size.
:::
:::
GGUF assumes that huggingface can convert the metadata to a config file. In case huggingface doesn't support your model you can manually create a config and pass it as hf-confing-path
```console
#If you model is not supported by huggingface you can manually provide a huggingface compatible config path
To create a new 4-bit or 8-bit GPTQ quantized model, you can leverage [GPTQModel](https://github.com/ModelCloud/GPTQModel) from ModelCloud.AI.
Quantization reduces the model's precision from BF16/FP16 (16-bits) to INT4 (4-bits) or INT8 (8-bits) which significantly reduces the
total model memory footprint while at-the-same-time increasing inference performance.
Compatible GPTQModel quantized models can leverage the `Marlin` and `Machete` vLLM custom kernels to maximize batching
transactions-per-second `tps` and token-latency performance for both Ampere (A100+) and Hopper (H100+) Nvidia GPUs.
These two kernels are highly optimized by vLLM and NeuralMagic (now part of Redhat) to allow world-class inference performance of quantized GPTQ
models.
GPTQModel is one of the few quantization toolkits in the world that allows `Dynamic` per-module quantization where different layers and/or modules within a llm model can be further optimized with custom quantization parameters. `Dynamic` quantization
is fully integrated into vLLM and backed up by support from the ModelCloud.AI team. Please refer to [GPTQModel readme](https://github.com/ModelCloud/GPTQModel?tab=readme-ov-file#dynamic-quantization-per-module-quantizeconfig-override)
for more details on this and other advanced features.
You can quantize your own models by installing [GPTQModel](https://github.com/ModelCloud/GPTQModel) or picking one of the [5000+ models on Huggingface](https://huggingface.co/models?sort=trending&search=gptq).
```console
pip install -U gptqmodel --no-build-isolation -v
```
After installing GPTQModel, you are ready to quantize a model. Please refer to the [GPTQModel readme](https://github.com/ModelCloud/GPTQModel/?tab=readme-ov-file#quantization) for further details.
Here is an example of how to quantize `meta-llama/Llama-3.2-1B-Instruct`:
# increase `batch_size` to match gpu/vram specs to speed up quantization
model.quantize(calibration_dataset,batch_size=2)
model.save(quant_path)
```
To run an GPTQModel quantized model with vLLM, you can use [DeepSeek-R1-Distill-Qwen-7B-gptqmodel-4bit-vortex-v2](https://huggingface.co/ModelCloud/DeepSeek-R1-Distill-Qwen-7B-gptqmodel-4bit-vortex-v2) with the following command:
@@ -10,7 +10,10 @@ Reasoning models return a additional `reasoning_content` field in their outputs,
...
@@ -10,7 +10,10 @@ Reasoning models return a additional `reasoning_content` field in their outputs,
vLLM currently supports the following reasoning models:
vLLM currently supports the following reasoning models:
-[DeepSeek R1 series](https://huggingface.co/collections/deepseek-ai/deepseek-r1-678e1e131c0169c0bc89728d)(`deepseek_r1`, which looks for `<think> ... </think>`)
| Model Series | Parser Name | Structured Output Support | Tool Calling |
@@ -76,7 +79,142 @@ Streaming chat completions are also supported for reasoning models. The `reasoni
...
@@ -76,7 +79,142 @@ Streaming chat completions are also supported for reasoning models. The `reasoni
}
}
```
```
Please note that it is not compatible with the OpenAI Python client library. You can use the `requests` library to make streaming requests.
OpenAI Python client library does not officially support `reasoning_content` attribute for streaming output. But the client support extra attributes in the response. You can use `hasattr` to check if the `reasoning_content` attribute is present in the response. For example:
```python
fromopenaiimportOpenAI
# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key="EMPTY"
openai_api_base="http://localhost:8000/v1"
client=OpenAI(
api_key=openai_api_key,
base_url=openai_api_base,
)
models=client.models.list()
model=models.data[0].id
messages=[{"role":"user","content":"9.11 and 9.8, which is greater?"}]
Remember to check whether the `reasoning_content` exists in the response before accessing it. You could checkout the [example](https://github.com/vllm-project/vllm/blob/main/examples/online_serving/openai_chat_completion_with_reasoning_streaming.py).
## Structured output
The reasoning content is also available in the structured output. The structured output engine like `xgrammar` will use the reasoning content to generate structured output.
```python
fromopenaiimportOpenAI
frompydanticimportBaseModel
# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key="EMPTY"
openai_api_base="http://localhost:8000/v1"
client=OpenAI(
api_key=openai_api_key,
base_url=openai_api_base,
)
models=client.models.list()
model=models.data[0].id
classPeople(BaseModel):
name:str
age:int
json_schema=People.model_json_schema()
prompt=("Generate a JSON with the name and age of one random person.")
The reasoning content is also available when both tool calling and the reasoning parser are enabled. Additionally, tool calling only parses functions from the `content` field, not from the `reasoning_content`.
Extract reasoning content from a complete model-generated string.
Extract reasoning content from a complete model-generated string.
...
@@ -132,20 +270,42 @@ class ExampleParser(ReasoningParser):
...
@@ -132,20 +270,42 @@ class ExampleParser(ReasoningParser):
The request object that was used to generate the model_output.
The request object that was used to generate the model_output.
Returns:
Returns:
Tuple[Optional[str], Optional[str]]
tuple[Optional[str], Optional[str]]
A tuple containing the reasoning content and the content.
A tuple containing the reasoning content and the content.
"""
"""
```
```
After defining the reasoning parser, you can use it by specifying the `--reasoning-parser` flag when making a request to the chat completion endpoint.
Additionally, to enable structured output, you'll need to create a new `Reasoner` similar to the one in `vllm/model_executor/guided_decoding/reasoner/deepseek_reasoner.py`.
The structured output engine like `xgrammar` will use `end_token_id` to check if the reasoning content is present in the model output and skip the structured output if it is the case.
Finally, you can enable reasoning for the model by using the `--enable-reasoning` and `--reasoning-parser` flags.
```bash
```bash
vllm serve <model_tag> \
vllm serve <model_tag> \
--enable-reasoning--reasoning-parser example
--enable-reasoning--reasoning-parser example
```
```
## Limitations
- The reasoning content is only available for online serving's chat completion endpoint (`/v1/chat/completions`).
- It is not compatible with the [`structured_outputs`](#structured_outputs) and [`tool_calling`](#tool_calling) features.
- The reasoning content is not available for all models. Check the model's documentation to see if it supports reasoning.
Note: Please use `--speculative_config` to set all configurations related to speculative decoding. The previous method of specifying the model through `--speculative_model` and adding related parameters (e.g., `--num_speculative_tokens`) separately will be deprecated in the next release.
@@ -162,7 +172,7 @@ A variety of speculative models of this type are available on HF hub:
...
@@ -162,7 +172,7 @@ A variety of speculative models of this type are available on HF hub:
## Speculating using EAGLE based draft models
## Speculating using EAGLE based draft models
The following code configures vLLM to use speculative decoding where proposals are generated by
The following code configures vLLM to use speculative decoding where proposals are generated by
an [EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency)](https://arxiv.org/pdf/2401.15077) based draft model.
an [EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency)](https://arxiv.org/pdf/2401.15077) based draft model. A more detailed example for offline mode, including how to extract request level acceptance rate, can be found [here](<gh-file:examples/offline_inference/eagle.py>).
@@ -194,11 +206,10 @@ A few important things to consider when using the EAGLE based draft models:
...
@@ -194,11 +206,10 @@ A few important things to consider when using the EAGLE based draft models:
be able to be loaded and used directly by vLLM after [PR 12304](https://github.com/vllm-project/vllm/pull/12304).
be able to be loaded and used directly by vLLM after [PR 12304](https://github.com/vllm-project/vllm/pull/12304).
If you are using vllm version before [PR 12304](https://github.com/vllm-project/vllm/pull/12304), please use the
If you are using vllm version before [PR 12304](https://github.com/vllm-project/vllm/pull/12304), please use the
[script](https://gist.github.com/abhigoyal1997/1e7a4109ccb7704fbc67f625e86b2d6d) to convert the speculative model,
[script](https://gist.github.com/abhigoyal1997/1e7a4109ccb7704fbc67f625e86b2d6d) to convert the speculative model,
and specify `speculative_model="path/to/modified/eagle/model"`. If weight-loading problems still occur when using
and specify `"model": "path/to/modified/eagle/model"` in `speculative_config`. If weight-loading problems still occur when using the latest version of vLLM, please leave a comment or raise an issue.
the latest version of vLLM, please leave a comment or raise an issue.
2. The EAGLE based draft models need to be run without tensor parallelism
2. The EAGLE based draft models need to be run without tensor parallelism
(i.e. speculative_draft_tensor_parallel_size is set to 1), although
(i.e. draft_tensor_parallel_size is set to 1 in `speculative_config`), although
it is possible to run the main model using tensor parallelism (see example above).
it is possible to run the main model using tensor parallelism (see example above).
3. When using EAGLE-based speculators with vLLM, the observed speedup is lower than what is
3. When using EAGLE-based speculators with vLLM, the observed speedup is lower than what is
@@ -16,7 +16,7 @@ The following parameters are supported, which must be added as extra parameters:
...
@@ -16,7 +16,7 @@ The following parameters are supported, which must be added as extra parameters:
-`guided_json`: the output will follow the JSON schema.
-`guided_json`: the output will follow the JSON schema.
-`guided_grammar`: the output will follow the context free grammar.
-`guided_grammar`: the output will follow the context free grammar.
-`guided_whitespace_pattern`: used to override the default whitespace pattern for guided json decoding.
-`guided_whitespace_pattern`: used to override the default whitespace pattern for guided json decoding.
-`guided_decoding_backend`: used to select the guided decoding backend to use.
-`guided_decoding_backend`: used to select the guided decoding backend to use. Additional backend-specific options can be supplied in a comma separated list following a colon after the backend name. For example `"xgrammar:no-fallback"` will not allow vLLM to fallback to a different backend on error.
You can see the complete list of supported parameters on the [OpenAI-Compatible Server](#openai-compatible-server)page.
You can see the complete list of supported parameters on the [OpenAI-Compatible Server](#openai-compatible-server)page.
@@ -209,6 +209,15 @@ AI21's Jamba-1.5 models are supported.
...
@@ -209,6 +209,15 @@ AI21's Jamba-1.5 models are supported.
Flags: `--tool-call-parser jamba`
Flags: `--tool-call-parser jamba`
### Qwen Models
For Qwen2.5, the chat template in tokenizer_config.json has already included support for the Hermes-style tool use. Therefore, you can use the `hermes` parser to enable tool calls for Qwen models. For more detailed information, please refer to the official [Qwen documentation](https://qwen.readthedocs.io/en/latest/framework/function_call.html#vllm)
* `Qwen/Qwen2.5-*`
* `Qwen/QwQ-32B`
Flags: `--tool-call-parser hermes`
### Models with Pythonic Tool Calls (`pythonic`)
### Models with Pythonic Tool Calls (`pythonic`)
A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models.
A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models.
