The community frequently requests the ability to extend vLLM with custom features. To facilitate this, vLLM includes a plugin system that allows users to add custom features without modifying the vLLM codebase. This document explains how plugins work in vLLM and how to create a plugin for vLLM.
How Plugins Work in vLLM
------------------------
Plugins are user-registered code that vLLM executes. Given vLLM's architecture (see :ref:`arch_overview`), multiple processes may be involved, especially when using distributed inference with various parallelism techniques. To enable plugins successfully, every process created by vLLM needs to load the plugin. This is done by the `load_general_plugins <https://github.com/vllm-project/vllm/blob/c76ac49d266e27aa3fea84ef2df1f813d24c91c7/vllm/plugins/__init__.py#L16>`__ function in the ``vllm.plugins`` module. This function is called for every process created by vLLM before it starts any work.
How vLLM Discovers Plugins
--------------------------
vLLM's plugin system uses the standard Python ``entry_points`` mechanism. This mechanism allows developers to register functions in their Python packages for use by other packages. An example of a plugin:
if "MyLlava" not in ModelRegistry.get_supported_archs():
ModelRegistry.register_model("MyLlava",
"vllm_add_dummy_model.my_llava:MyLlava")
For more information on adding entry points to your package, please check the `official documentation <https://setuptools.pypa.io/en/latest/userguide/entry_point.html>`__.
Every plugin has three parts:
1. **Plugin group**: The name of the entry point group. vLLM uses the entry point group ``vllm.general_plugins`` to register general plugins. This is the key of ``entry_points`` in the ``setup.py`` file. Always use ``vllm.general_plugins`` for vLLM's general plugins.
2. **Plugin name**: The name of the plugin. This is the value in the dictionary of the ``entry_points`` dictionary. In the example above, the plugin name is ``register_dummy_model``. Plugins can be filtered by their names using the ``VLLM_PLUGINS`` environment variable. To load only a specific plugin, set ``VLLM_PLUGINS`` to the plugin name.
3. **Plugin value**: The fully qualified name of the function to register in the plugin system. In the example above, the plugin value is ``vllm_add_dummy_model:register``, which refers to a function named ``register`` in the ``vllm_add_dummy_model`` module.
What Can Plugins Do?
--------------------
Currently, the primary use case for plugins is to register custom, out-of-the-tree models into vLLM. This is done by calling ``ModelRegistry.register_model`` to register the model. In the future, the plugin system may be extended to support more features, such as swapping in custom implementations for certain classes in vLLM.
Guidelines for Writing Plugins
------------------------------
- **Being re-entrant**: The function specified in the entry point should be re-entrant, meaning it can be called multiple times without causing issues. This is necessary because the function might be called multiple times in some processes.
Compatibility Guarantee
-----------------------
vLLM guarantees the interface of documented plugins, such as ``ModelRegistry.register_model``, will always be available for plugins to register models. However, it is the responsibility of plugin developers to ensure their plugins are compatible with the version of vLLM they are targeting. For example, ``"vllm_add_dummy_model.my_llava:MyLlava"`` should be compatible with the version of vLLM that the plugin targets. The interface for the model may change during vLLM's development.
vLLM has been adapted to work on ARM64 CPUs with NEON support, leveraging the CPU backend initially developed for the x86 platform. This guide provides installation instructions specific to ARM. For additional details on supported features, refer to the x86 platform documentation covering:
* CPU backend inference capabilities
* Relevant runtime environment variables
* Performance optimization tips
ARM CPU backend currently supports Float32, FP16 and BFloat16 datatypes.
Contents:
1. :ref:`Requirements <arm_backend_requirements>`
2. :ref:`Quick Start with Dockerfile <arm_backend_quick_start_dockerfile>`
3. :ref:`Building from Source <build_arm_backend_from_source>`
.. _arm_backend_requirements:
Requirements
------------
* **Operating System**: Linux or macOS
* **Compiler**: gcc/g++ >= 12.3.0 (optional, but recommended)
* **Instruction Set Architecture (ISA)**: NEON support is required
To build vLLM from source on Ubuntu 22.04 or other Linux distributions, follow a similar process as with x86. Testing has been conducted on AWS Graviton3 instances for compatibility.
vLLM initially supports basic model inferencing and serving on x86 CPU platform, with data types FP32 and BF16.
vLLM initially supports basic model inferencing and serving on x86 CPU platform, with data types FP32, FP16 and BF16. vLLM CPU backend supports the following vLLM features:
- BF16 is the default data type in the current CPU backend (that means the backend will cast FP16 to BF16), and is compatible will all CPUs with AVX512 ISA support.
- AVX512_BF16 is an extension ISA provides native BF16 data type conversion and vector product instructions, will brings some performance improvement compared with pure AVX512. The CPU backend build script will check the host CPU flags to determine whether to enable AVX512_BF16.
- If you want to force enable AVX512_BF16 for the cross-compilation, please set environment variable VLLM_CPU_AVX512BF16=1 before the building.
...
...
@@ -155,5 +145,20 @@ Performance tips
- If using vLLM CPU backend on a multi-socket machine with NUMA, be aware to set CPU cores using ``VLLM_CPU_OMP_THREADS_BIND`` to avoid cross NUMA node memory access.
CPU Backend Considerations
--------------------------
- The CPU backend significantly differs from the GPU backend since the vLLM architecture was originally optimized for GPU use. A number of optimizations are needed to enhance its performance.
- Decouple the HTTP serving components from the inference components. In a GPU backend configuration, the HTTP serving and tokenization tasks operate on the CPU, while inference runs on the GPU, which typically does not pose a problem. However, in a CPU-based setup, the HTTP serving and tokenization can cause significant context switching and reduced cache efficiency. Therefore, it is strongly recommended to segregate these two components for improved performance.
- On CPU based setup with NUMA enabled, the memory access performance may be largely impacted by the `topology <https://github.com/intel/intel-extension-for-pytorch/blob/main/docs/tutorials/performance_tuning/tuning_guide.md#non-uniform-memory-access-numa>`_. For NUMA architecture, two optimizations are to recommended: Tensor Parallel or Data Parallel.
* Using Tensor Parallel for a latency constraints deployment: following GPU backend design, a Megatron-LM's parallel algorithm will be used to shard the model, based on the number of NUMA nodes (e.g. TP = 2 for a two NUMA node system). With `TP feature on CPU <https://github.com/vllm-project/vllm/pull/6125>`_ merged, Tensor Parallel is supported for serving and offline inferencing. In general each NUMA node is treated as one GPU card. Below is the example script to enable Tensor Parallel = 2 for serving:
* Using Data Parallel for maximum throughput: to launch an LLM serving endpoint on each NUMA node along with one additional load balancer to dispatch the requests to those endpoints. Common solutions like `Nginx <../serving/deploying_with_nginx.html>`_ or HAProxy are recommended. Anyscale Ray project provides the feature on LLM `serving <https://docs.ray.io/en/latest/serve/index.html>`_. Here is the example to setup a scalable LLM serving with `Ray Serve <https://github.com/intel/llm-on-ray/blob/main/docs/setup.md>`_.
This document outlines some debugging strategies you can consider. If you think you've discovered a bug, please `search existing issues <https://github.com/vllm-project/vllm/issues?q=is%3Aissue>`_ first to see if it has already been reported. If not, please `file a new issue <https://github.com/vllm-project/vllm/issues/new/choose>`_, providing as much relevant information as possible.
.. note::
Once you've debugged a problem, remember to turn off any debugging environment variables defined, or simply start a new shell to avoid being affected by lingering debugging settings. Otherwise, the system might be slow with debugging functionalities left activated.
Hangs downloading a model
----------------------------------------
If the model isn't already downloaded to disk, vLLM will download it from the internet which can take time and depend on your internet connection.
It's recommended to download the model first using the `huggingface-cli <https://huggingface.co/docs/huggingface_hub/en/guides/cli>`_ and passing the local path to the model to vLLM. This way, you can isolate the issue.
When an vLLM instance hangs or crashes, it is very difficult to debug the issue. But wait a minute, it is also possible that vLLM is doing something that indeed takes a long time:
Hangs loading a model from disk
----------------------------------------
If the model is large, it can take a long time to load it from disk. Pay attention to where you store the model. Some clusters have shared filesystems across nodes, e.g. a distributed filesystem or a network filesystem, which can be slow.
It'd be better to store the model in a local disk. Additionally, have a look at the CPU memory usage, when the model is too large it might take a lot of CPU memory, slowing down the operating system because it needs to frequently swap between disk and memory.
- **Downloading a model**: Do you have the model already downloaded in your disk? If not, vLLM will download the model from the internet, which can take a long time. Be sure to check the internet connection. It would be better to download the model first using `huggingface-cli <https://huggingface.co/docs/huggingface_hub/en/guides/cli>`_ and then use the local path to the model. This way, you can isolate the issue.
- **Loading the model from disk**: If the model is large, it can take a long time to load the model from disk. Please take care of the location you store the model. Some clusters have shared filesystems across nodes, e.g. distributed filesystem or network filesystem, which can be slow. It would be better to store the model in a local disk. In addition, please also watch the CPU memory usage. When the model is too large, it might take much CPU memory, which can slow down the operating system because it needs to frequently swap memory between the disk and the memory.
- **Tensor parallel inference**: If the model is too large to fit in a single GPU, you might want to use tensor parallelism to split the model across multiple GPUs. In that case, every process will read the whole model and split it into chunks, which makes the disk reading time even longer (proportional to the size of tensor parallelism). You can convert the model checkpoint to a sharded checkpoint using `the provided script <https://docs.vllm.ai/en/latest/getting_started/examples/save_sharded_state.html>`_ . The conversion process might take some time, but later you can load the sharded checkpoint much faster. The model loading time should remain constant regardless of the size of tensor parallelism.
.. note::
If you have already taken care of the above issues, but the vLLM instance still hangs, with CPU and GPU utilization at near zero, it is likely that the vLLM instance is stuck somewhere. Here are some tips to help debug the issue:
To isolate the model downloading and loading issue, you can use the ``--load-format dummy`` argument to skip loading the model weights. This way, you can check if the model downloading and loading is the bottleneck.
- Set the environment variable ``export VLLM_LOGGING_LEVEL=DEBUG`` to turn on more logging.
- Set the environment variable ``export CUDA_LAUNCH_BLOCKING=1`` to know exactly which CUDA kernel is causing the trouble.
- Set the environment variable ``export NCCL_DEBUG=TRACE`` to turn on more logging for NCCL.
- Set the environment variable ``export VLLM_TRACE_FUNCTION=1``. All the function calls in vLLM will be recorded. Inspect these log files, and tell which function crashes or hangs.
Model is too large
----------------------------------------
If the model is too large to fit in a single GPU, you might want to `consider tensor parallelism <https://docs.vllm.ai/en/latest/serving/distributed_serving.html#distributed-inference-and-serving>`_ to split the model across multiple GPUs. In that case, every process will read the whole model and split it into chunks, which makes the disk reading time even longer (proportional to the size of tensor parallelism). You can convert the model checkpoint to a sharded checkpoint using `this example <https://docs.vllm.ai/en/latest/getting_started/examples/save_sharded_state.html>`_ . The conversion process might take some time, but later you can load the sharded checkpoint much faster. The model loading time should remain constant regardless of the size of tensor parallelism.
With more logging, hopefully you can find the root cause of the issue.
Enable more logging
----------------------------------------
If other strategies don't solve the problem, it's likely that the vLLM instance is stuck somewhere. You can use the following environment variables to help debug the issue:
If it crashes, and the error trace shows somewhere around ``self.graph.replay()`` in ``vllm/worker/model_runner.py``, it is a cuda error inside cudagraph. To know the particular cuda operation that causes the error, you can add ``--enforce-eager`` to the command line, or ``enforce_eager=True`` to the :class:`~vllm.LLM` class, to disable the cudagraph optimization. This way, you can locate the exact cuda operation that causes the error.
- ``export VLLM_LOGGING_LEVEL=DEBUG`` to turn on more logging.
- ``export CUDA_LAUNCH_BLOCKING=1`` to identify which CUDA kernel is causing the problem.
- ``export NCCL_DEBUG=TRACE`` to turn on more logging for NCCL.
- ``export VLLM_TRACE_FUNCTION=1`` to record all function calls for inspection in the log files to tell which function crashes or hangs.
Here are some common issues that can cause hangs:
Incorrect network setup
----------------------------------------
The vLLM instance cannot get the correct IP address if you have a complicated network config. You can find a log such as ``DEBUG 06-10 21:32:17 parallel_state.py:88] world_size=8 rank=0 local_rank=0 distributed_init_method=tcp://xxx.xxx.xxx.xxx:54641 backend=nccl`` and the IP address should be the correct one.
If it's not, override the IP address using the environment variable ``export VLLM_HOST_IP=<your_ip_address>``.
- **Incorrect network setup**: The vLLM instance cannot get the correct IP address if you have complicated network config. You can find the log such as ``DEBUG 06-10 21:32:17 parallel_state.py:88] world_size=8 rank=0 local_rank=0 distributed_init_method=tcp://xxx.xxx.xxx.xxx:54641 backend=nccl``. The IP address should be the correct one. If not, override the IP address by setting the environment variable ``export VLLM_HOST_IP=your_ip_address``. You might also need to set ``export NCCL_SOCKET_IFNAME=your_network_interface`` and ``export GLOO_SOCKET_IFNAME=your_network_interface`` to specify the network interface for the IP address.
- **Incorrect hardware/driver**: GPU/CPU communication cannot be established. You can run the following sanity check script to see if the GPU/CPU communication is working correctly.
You might also need to set ``export NCCL_SOCKET_IFNAME=<your_network_interface>`` and ``export GLOO_SOCKET_IFNAME=<your_network_interface>`` to specify the network interface for the IP address.
Error near ``self.graph.replay()``
----------------------------------------
If vLLM crashes and the error trace captures it somewhere around ``self.graph.replay()`` in ``vllm/worker/model_runner.py``, it is a CUDA error inside CUDAGraph.
To identify the particular CUDA operation that causes the error, you can add ``--enforce-eager`` to the command line, or ``enforce_eager=True`` to the :class:`~vllm.LLM` class to disable the CUDAGraph optimization and isolate the exact CUDA operation that causes the error.
Incorrect hardware/driver
----------------------------------------
If GPU/CPU communication cannot be established, you can use the following Python script and follow the instructions below to confirm whether the GPU/CPU communication is working correctly.
.. code-block:: python
...
...
@@ -54,11 +79,13 @@ Here are some common issues that can cause hangs:
print("PyTorch GLOO is successful!")
if world_size <= 1:
exit()
# Test vLLM NCCL, with cuda graph
from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
@@ -84,33 +111,87 @@ Here are some common issues that can cause hangs:
dist.destroy_process_group(gloo_group)
dist.destroy_process_group()
.. tip::
If you are testing with a single node, adjust ``--nproc-per-node`` to the number of GPUs you want to use:
.. code-block:: console
Save the script as ``test.py``.
If you are testing in a single-node, run it with ``NCCL_DEBUG=TRACE torchrun --nproc-per-node=8 test.py``, adjust ``--nproc-per-node`` to the number of GPUs you want to use.
If you are testing with multi-nodes, run it with ``NCCL_DEBUG=TRACE torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=$MASTER_ADDR test.py``. Adjust ``--nproc-per-node`` and ``--nnodes`` according to your setup. Make sure ``MASTER_ADDR``:
If the script runs successfully, you should see the message ``sanity check is successful!``.
If you are testing with multi-nodes, adjust ``--nproc-per-node`` and ``--nnodes`` according to your setup and set ``MASTER_ADDR`` to the correct IP address of the master node, reachable from all nodes. Then, run:
Note that multi-node environment is more complicated than single-node. If you see errors such as ``torch.distributed.DistNetworkError``, it is likely that the network/DNS setup is incorrect. In that case, you can manually assign node rank and specify the IP via command line arguments:
If the script runs successfully, you should see the message ``sanity check is successful!``.
If the test script hangs or crashes, usually it means the hardware/drivers are broken in some sense. You should try to contact your system administrator or hardware vendor for further assistance. As a common workaround, you can try to tune some NCCL environment variables, such as ``export NCCL_P2P_DISABLE=1`` to see if it helps. Please check `their documentation <https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html>`__ for more information. Please only use these environment variables as a temporary workaround, as they might affect the performance of the system. The best solution is still to fix the hardware/drivers so that the test script can run successfully.
.. note::
A multi-node environment is more complicated than a single-node one. If you see errors such as ``torch.distributed.DistNetworkError``, it is likely that the network/DNS setup is incorrect. In that case, you can manually assign node rank and specify the IP via command line arguments:
- In the first node, run ``NCCL_DEBUG=TRACE torchrun --nnodes 2 --nproc-per-node=2 --node-rank 0 --master_addr $MASTER_ADDR test.py``.
- In the second node, run ``NCCL_DEBUG=TRACE torchrun --nnodes 2 --nproc-per-node=2 --node-rank 1 --master_addr $MASTER_ADDR test.py``.
Adjust ``--nproc-per-node``, ``--nnodes``, and ``--node-rank`` according to your setup. The difference is that you need to execute different commands (with different ``--node-rank``) on different nodes.
Adjust ``--nproc-per-node``, ``--nnodes``, and ``--node-rank`` according to your setup, being sure to execute different commands (with different ``--node-rank``) on different nodes.
Python multiprocessing
----------------------
`RuntimeError` Exception
^^^^^^^^^^^^^^^^^^^^^^^^
If you have seen a warning in your logs like this:
If the problem persists, feel free to `open an issue on GitHub <https://github.com/vllm-project/vllm/issues/new/choose>`_, with a detailed description of the issue, your environment, and the logs.
.. code-block:: console
Some known issues:
WARNING 12-11 14:50:37 multiproc_worker_utils.py:281] CUDA was previously
initialized. We must use the `spawn` multiprocessing start method. Setting
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.
This probably means that you are not using fork to start your
child processes and you have forgotten to use the proper idiom
in the main module:
if __name__ == '__main__':
freeze_support()
...
The "freeze_support()" line can be omitted if the program
is not going to be frozen to produce an executable.
To fix this issue, refer to the "Safe importing of main module"
section in https://docs.python.org/3/library/multiprocessing.html
then you must update your Python code to guard usage of ``vllm`` behind a ``if
__name__ == '__main__':`` block. For example, instead of this:
.. code-block:: python
import vllm
llm = vllm.LLM(...)
try this instead:
.. code-block:: python
- In ``v0.5.2``, ``v0.5.3``, and ``v0.5.3.post1``, there is a bug caused by `zmq <https://github.com/zeromq/pyzmq/issues/2000>`_ , which can cause hangs at a low probability (once in about 20 times, depending on the machine configuration). The solution is to upgrade to the latest version of ``vllm`` to include the `fix <https://github.com/vllm-project/vllm/pull/6759>`_ .
if __name__ == '__main__':
import vllm
.. warning::
llm = vllm.LLM(...)
After you find the root cause and solve the issue, remember to turn off all the debugging environment variables defined above, or simply start a new shell to avoid being affected by the debugging settings. If you don't do this, the system might be slow because many debugging functionalities are turned on.
Known Issues
----------------------------------------
- In ``v0.5.2``, ``v0.5.3``, and ``v0.5.3.post1``, there is a bug caused by `zmq <https://github.com/zeromq/pyzmq/issues/2000>`_ , which can occasionally cause vLLM to hang depending on the machine configuration. The solution is to upgrade to the latest version of ``vllm`` to include the `fix <https://github.com/vllm-project/vllm/pull/6759>`_.
If you're observing the following error: ``docker: Error response from daemon: Unknown runtime specified habana.``, please refer to "Install Using Containers" section of `Intel Gaudi Software Stack and Driver Installation <https://docs.habana.ai/en/v1.18.0/Installation_Guide/Bare_Metal_Fresh_OS.html>`__. Make sure you have ``habana-container-runtime`` package installed and that ``habana`` container runtime is registered.
Build from source
~~~~~~~~~~~~~~~~~
Environment verification
^^^^^^^^^^^^^^^^^^^^^^^^
To verify that the Intel Gaudi software was correctly installed, run:
.. code:: console
$ hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible
$ apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core, habanalabs-thunk and habanalabs-container-runtime are installed
$ pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed
$ pip list | grep neural # verify that neural_compressor is installed
Currently, the latest features and performance optimizations are developed in Gaudi's `vLLM-fork <https://github.com/HabanaAI/vllm-fork>`__ and we periodically upstream them to vLLM main repo. To install latest `HabanaAI/vLLM-fork <https://github.com/HabanaAI/vllm-fork>`__, run the following:
with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
Performance Tuning
------------------
Execution modes
~~~~~~~~~~~~~~~
Currently in vLLM for HPU we support four execution modes, depending on selected HPU PyTorch Bridge backend (via ``PT_HPU_LAZY_MODE`` environment variable), and ``--enforce-eager`` flag.
.. list-table:: vLLM execution modes
:widths: 25 25 50
:header-rows: 1
* - ``PT_HPU_LAZY_MODE``
- ``enforce_eager``
- execution mode
* - 0
- 0
- torch.compile
* - 0
- 1
- PyTorch eager mode
* - 1
- 0
- HPU Graphs
* - 1
- 1
- PyTorch lazy mode
.. warning::
In 1.18.0, all modes utilizing ``PT_HPU_LAZY_MODE=0`` are highly experimental and should be only used for validating functional correctness. Their performance will be improved in the next releases. For obtaining the best performance in 1.18.0, please use HPU Graphs, or PyTorch lazy mode.
Bucketing mechanism
~~~~~~~~~~~~~~~~~~~
Intel Gaudi accelerators work best when operating on models with fixed tensor shapes. `Intel Gaudi Graph Compiler <https://docs.habana.ai/en/latest/Gaudi_Overview/Intel_Gaudi_Software_Suite.html#graph-compiler-and-runtime>`__ is responsible for generating optimized binary code that implements the given model topology on Gaudi. In its default configuration, the produced binary code may be heavily dependent on input and output tensor shapes, and can require graph recompilation when encountering differently shaped tensors within the same topology. While the resulting binaries utilize Gaudi efficiently, the compilation itself may introduce a noticeable overhead in end-to-end execution.
In a dynamic inference serving scenario, there is a need to minimize the number of graph compilations and reduce the risk of graph compilation occurring during server runtime. Currently it is achieved by "bucketing" model's forward pass across two dimensions - ``batch_size`` and ``sequence_length``.
.. note::
Bucketing allows us to reduce the number of required graphs significantly, but it does not handle any graph compilation and device code generation - this is done in warmup and HPUGraph capture phase.
Bucketing ranges are determined with 3 parameters - ``min``, ``step`` and ``max``. They can be set separately for prompt and decode phase, and for batch size and sequence length dimension. These parameters can be observed in logs during vLLM startup:
``min`` determines the lowest value of the bucket. ``step`` determines the interval between buckets, and ``max`` determines the upper bound of the bucket. Furthermore, interval between ``min`` and ``step`` has special handling - ``min`` gets multiplied by consecutive powers of two, until ``step`` gets reached. We call this the ramp-up phase and it is used for handling lower batch sizes with minimum wastage, while allowing larger padding on larger batch sizes.
In the logged scenario, 24 buckets were generated for prompt (prefill) runs, and 48 buckets for decode runs. Each bucket corresponds to a separate optimized device binary for a given model with specified tensor shapes. Whenever a batch of requests is processed, it is padded across batch and sequence length dimension to the smallest possible bucket.
.. warning::
If a request exceeds maximum bucket size in any dimension, it will be processed without padding, and its processing may require a graph compilation, potentially significantly increasing end-to-end latency. The boundaries of the buckets are user-configurable via environment variables, and upper bucket boundaries can be increased to avoid such scenario.
As an example, if a request of 3 sequences, with max sequence length of 412 comes in to an idle vLLM server, it will be padded executed as ``(4, 512)`` prefill bucket, as ``batch_size`` (number of sequences) will be padded to 4 (closest batch_size dimension higher than 3), and max sequence length will be padded to 512 (closest sequence length dimension higher than 412). After prefill stage, it will be executed as ``(4, 512)`` decode bucket and will continue as that bucket until either batch dimension changes (due to request being finished) - in which case it will become a ``(2, 512)`` bucket, or context length increases above 512 tokens, in which case it will become ``(4, 640)`` bucket.
.. note::
Bucketing is transparent to a client - padding in sequence length dimension is never returned to the client, and padding in batch dimension does not create new requests.
Warmup
~~~~~~
Warmup is an optional, but highly recommended step occurring before vLLM server starts listening. It executes a forward pass for each bucket with dummy data. The goal is to pre-compile all graphs and not incur any graph compilation overheads within bucket boundaries during server runtime. Each warmup step is logged during vLLM startup:
.. code-block::
INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][1/24] batch_size:4 seq_len:1024 free_mem:79.16 GiB
INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][2/24] batch_size:4 seq_len:896 free_mem:55.43 GiB
INFO 08-01 22:26:48 hpu_model_runner.py:1066] [Warmup][Prompt][3/24] batch_size:4 seq_len:768 free_mem:55.43 GiB
...
INFO 08-01 22:26:59 hpu_model_runner.py:1066] [Warmup][Prompt][24/24] batch_size:1 seq_len:128 free_mem:55.43 GiB
INFO 08-01 22:27:00 hpu_model_runner.py:1066] [Warmup][Decode][1/48] batch_size:4 seq_len:2048 free_mem:55.43 GiB
INFO 08-01 22:27:00 hpu_model_runner.py:1066] [Warmup][Decode][2/48] batch_size:4 seq_len:1920 free_mem:55.43 GiB
INFO 08-01 22:27:01 hpu_model_runner.py:1066] [Warmup][Decode][3/48] batch_size:4 seq_len:1792 free_mem:55.43 GiB
...
INFO 08-01 22:27:16 hpu_model_runner.py:1066] [Warmup][Decode][47/48] batch_size:2 seq_len:128 free_mem:55.43 GiB
INFO 08-01 22:27:16 hpu_model_runner.py:1066] [Warmup][Decode][48/48] batch_size:1 seq_len:128 free_mem:55.43 GiB
This example uses the same buckets as in *Bucketing mechanism* section. Each output line corresponds to execution of a single bucket. When bucket is executed for the first time, its graph is compiled and can be reused later on, skipping further graph compilations.
.. tip::
Compiling all the buckets might take some time and can be turned off with ``VLLM_SKIP_WARMUP=true`` environment variable. Keep in mind that if you do that, you may face graph compilations once executing a given bucket for the first time. It is fine to disable warmup for development, but it's highly recommended to enable it in deployment.
HPU Graph capture
~~~~~~~~~~~~~~~~~
`HPU Graphs <https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html>`__ are currently the most performant execution method of vLLM on Intel Gaudi. When HPU Graphs are enabled, execution graphs will be traced (recorded) ahead of time (after performing warmup), to be later replayed during inference, significantly reducing host overheads. Recording can take large amounts of memory, which needs to be taken into account when allocating KV cache. Enabling HPU Graphs will impact the number of available KV cache blocks, but vLLM provides user-configurable variables to control memory management.
When HPU Graphs are being used, they share the common memory pool ("usable memory") as KV cache, determined by ``gpu_memory_utilization`` flag (``0.9`` by default).
Before KV cache gets allocated, model weights are loaded onto the device, and a forward pass of the model is executed on dummy data, to estimate memory usage.
Only after that, ``gpu_memory_utilization`` flag is utilized - at its default value, will mark 90% of free device memory at that point as usable.
Next, KV cache gets allocated, model is warmed up, and HPU Graphs are captured.
Environment variable ``VLLM_GRAPH_RESERVED_MEM`` defines the ratio of memory reserved for HPU Graphs capture.
With its default value (``VLLM_GRAPH_RESERVED_MEM=0.1``), 10% of usable memory will be reserved for graph capture (later referred to as "usable graph memory"), and the remaining 90% will be utilized for KV cache.
Environment variable ``VLLM_GRAPH_PROMPT_RATIO`` determines the ratio of usable graph memory reserved for prefill and decode graphs. By default (``VLLM_GRAPH_PROMPT_RATIO=0.3``), both stages have equal memory constraints.
Lower value corresponds to less usable graph memory reserved for prefill stage, e.g. ``VLLM_GRAPH_PROMPT_RATIO=0.2`` will reserve 20% of usable graph memory for prefill graphs, and 80% of usable graph memory for decode graphs.
.. note::
``gpu_memory_utilization`` does not correspond to the absolute memory usage across HPU. It specifies the memory margin after loading the model and performing a profile run. If device has 100 GiB of total memory, and 50 GiB of free memory after loading model weights and executing profiling run, ``gpu_memory_utilization`` at its default value will mark 90% of 50 GiB as usable, leaving 5 GiB of margin, regardless of total device memory.
User can also configure the strategy for capturing HPU Graphs for prompt and decode stages separately. Strategy affects the order of capturing graphs. There are two strategies implemented:
- ``max_bs`` - graph capture queue will sorted in descending order by their batch sizes. Buckets with equal batch sizes are sorted by sequence length in ascending order (e.g. ``(64, 128)``, ``(64, 256)``, ``(32, 128)``, ``(32, 256)``, ``(1, 128)``, ``(1,256)``), default strategy for decode
- ``min_tokens`` - graph capture queue will be sorted in ascending order by the number of tokens each graph processes (``batch_size*sequence_length``), default strategy for prompt
When there's large amount of requests pending, vLLM scheduler will attempt to fill the maximum batch size for decode as soon as possible. When a request is finished, decode batch size decreases. When that happens, vLLM will attempt to schedule a prefill iteration for requests in the waiting queue, to fill the decode batch size to its previous state. This means that in a full load scenario, decode batch size is often at its maximum, which makes large batch size HPU Graphs crucial to capture, as reflected by ``max_bs`` strategy. On the other hand, prefills will be executed most frequently with very low batch sizes (1-4), which is reflected in ``min_tokens`` strategy.
.. note::
``VLLM_GRAPH_PROMPT_RATIO`` does not set a hard limit on memory taken by graphs for each stage (prefill and decode). vLLM will first attempt to use up entirety of usable prefill graph memory (usable graph memory * ``VLLM_GRAPH_PROMPT_RATIO``) for capturing prefill HPU Graphs, next it will attempt do the same for decode graphs and usable decode graph memory pool. If one stage is fully captured, and there is unused memory left within usable graph memory pool, vLLM will attempt further graph capture for the other stage, until no more HPU Graphs can be captured without exceeding reserved memory pool. The behavior on that mechanism can be observed in the example below.
Each described step is logged by vLLM server, as follows (negative values correspond to memory being released):
INFO 08-02 17:37:52 hpu_model_runner.py:430] Pre-loading model weights on hpu:0 took 14.97 GiB of device memory (14.97 GiB/94.62 GiB used) and 2.95 GiB of host memory (475.2 GiB/1007 GiB used)
INFO 08-02 17:37:52 hpu_model_runner.py:438] Wrapping in HPU Graph took 0 B of device memory (14.97 GiB/94.62 GiB used) and -252 KiB of host memory (475.2 GiB/1007 GiB used)
INFO 08-02 17:37:52 hpu_model_runner.py:442] Loading model weights took in total 14.97 GiB of device memory (14.97 GiB/94.62 GiB used) and 2.95 GiB of host memory (475.2 GiB/1007 GiB used)
INFO 08-02 17:37:54 hpu_worker.py:134] Model profiling run took 504 MiB of device memory (15.46 GiB/94.62 GiB used) and 180.9 MiB of host memory (475.4 GiB/1007 GiB used)
INFO 08-02 17:37:54 hpu_worker.py:158] Free device memory: 79.16 GiB, 39.58 GiB usable (gpu_memory_utilization=0.5), 15.83 GiB reserved for HPUGraphs (VLLM_GRAPH_RESERVED_MEM=0.4), 23.75 GiB reserved for KV cache
INFO 08-02 17:37:54 hpu_executor.py:85] # HPU blocks: 1519, # CPU blocks: 0
INFO 08-02 17:37:54 hpu_worker.py:190] Initializing cache engine took 23.73 GiB of device memory (39.2 GiB/94.62 GiB used) and -1.238 MiB of host memory (475.4 GiB/1007 GiB used)
INFO 08-02 17:37:54 hpu_model_runner.py:1066] [Warmup][Prompt][1/24] batch_size:4 seq_len:1024 free_mem:55.43 GiB
...
INFO 08-02 17:38:22 hpu_model_runner.py:1066] [Warmup][Decode][48/48] batch_size:1 seq_len:128 free_mem:55.43 GiB
INFO 08-02 17:38:22 hpu_model_runner.py:1159] Using 15.85 GiB/55.43 GiB of free device memory for HPUGraphs, 7.923 GiB for prompt and 7.923 GiB for decode (VLLM_GRAPH_PROMPT_RATIO=0.3)
INFO 08-02 17:38:22 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][1/24] batch_size:1 seq_len:128 free_mem:55.43 GiB
...
INFO 08-02 17:38:26 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][11/24] batch_size:1 seq_len:896 free_mem:48.77 GiB
INFO 08-02 17:38:27 hpu_model_runner.py:1066] [Warmup][Graph/Decode][1/48] batch_size:4 seq_len:128 free_mem:47.51 GiB
...
INFO 08-02 17:38:41 hpu_model_runner.py:1066] [Warmup][Graph/Decode][48/48] batch_size:1 seq_len:2048 free_mem:47.35 GiB
INFO 08-02 17:38:41 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][12/24] batch_size:4 seq_len:256 free_mem:47.35 GiB
INFO 08-02 17:38:42 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][13/24] batch_size:2 seq_len:512 free_mem:45.91 GiB
INFO 08-02 17:38:42 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][14/24] batch_size:1 seq_len:1024 free_mem:44.48 GiB
INFO 08-02 17:38:43 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][15/24] batch_size:2 seq_len:640 free_mem:43.03 GiB
INFO 08-02 17:38:43 hpu_model_runner.py:1206] Warmup finished in 49 secs, allocated 14.19 GiB of device memory
INFO 08-02 17:38:43 hpu_executor.py:91] init_cache_engine took 37.92 GiB of device memory (53.39 GiB/94.62 GiB used) and 57.86 MiB of host memory (475.4 GiB/1007 GiB used)
Recommended vLLM Parameters
~~~~~~~~~~~~~~~~~~~~~~~~~~~
- We recommend running inference on Gaudi 2 with ``block_size`` of 128
for BF16 data type. Using default values (16, 32) might lead to
sub-optimal performance due to Matrix Multiplication Engine
- For max throughput on Llama 7B, we recommend running with batch size
of 128 or 256 and max context length of 2048 with HPU Graphs enabled.
If you encounter out-of-memory issues, see troubleshooting section.
Environment variables
~~~~~~~~~~~~~~~~~~~~~
**Diagnostic and profiling knobs:**
- ``VLLM_PROFILER_ENABLED``: if ``true``, high level profiler will be enabled. Resulting JSON traces can be viewed in `perfetto.habana.ai <https://perfetto.habana.ai/#!/viewer>`__. Disabled by default.
- ``VLLM_HPU_LOG_STEP_GRAPH_COMPILATION``: if ``true``, will log graph compilations per each vLLM engine step, only when there was any - highly recommended to use alongside ``PT_HPU_METRICS_GC_DETAILS=1``. Disabled by default.
- ``VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL``: if ``true``, will log graph compilations per each vLLM engine step, always, even if there were none. Disabled by default.
- ``VLLM_HPU_LOG_STEP_CPU_FALLBACKS``: if ``true``, will log cpu fallbacks per each vLLM engine step, only when there was any. Disabled by default.
- ``VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL``: if ``true``, will log cpu fallbacks per each vLLM engine step, always, even if there were none. Disabled by default.
**Performance tuning knobs:**
- ``VLLM_SKIP_WARMUP``: if ``true``, warmup will be skipped, ``false`` by default
- ``VLLM_GRAPH_RESERVED_MEM``: percentage of memory dedicated for HPUGraph capture, ``0.1`` by default
- ``VLLM_GRAPH_PROMPT_RATIO``: percentage of reserved graph memory dedicated for prompt graphs, ``0.3`` by default
- ``VLLM_GRAPH_PROMPT_STRATEGY``: strategy determining order of prompt graph capture, ``min_tokens`` or ``max_bs``, ``min_tokens`` by default
- ``VLLM_GRAPH_DECODE_STRATEGY``: strategy determining order of decode graph capture, ``min_tokens`` or ``max_bs``, ``max_bs`` by default
- ``VLLM_{phase}_{dim}_BUCKET_{param}`` - collection of 12 environment variables configuring ranges of bucketing mechanism
- ``{phase}`` is either ``PROMPT`` or ``DECODE``
- ``{dim}`` is either ``BS``, ``SEQ`` or ``BLOCK``
- ``{param}`` is either ``MIN``, ``STEP`` or ``MAX``
- Default values:
- Prompt:
- batch size min (``VLLM_PROMPT_BS_BUCKET_MIN``): ``1``
@@ -46,98 +47,168 @@ You can install vLLM using pip:
Therefore, it is recommended to install vLLM with a **fresh new** conda environment. If either you have a different CUDA version or you want to use an existing PyTorch installation, you need to build vLLM from source. See below for instructions.
.. note::
vLLM also publishes a subset of wheels (Python 3.10, 3.11 with CUDA 12) for every commit since v0.5.3. You can download them with the following command:
.. _install-the-latest-code:
.. code-block:: console
Install the latest code
=======================
LLM inference is a fast-evolving field, and the latest code may contain bug fixes, performance improvements, and new features that are not released yet. To allow users to try the latest code without waiting for the next release, vLLM provides wheels for Linux running on a x86 platform with CUDA 12 for every commit since ``v0.5.3``. You can download and install it with the following command:
Note that the wheels are built with Python 3.8 ABI (see `PEP 425 <https://peps.python.org/pep-0425/>`_ for more details about ABI), so **they are compatible with Python 3.8 and later**. The version string in the wheel file name (``1.0.0.dev``) is just a placeholder to have a unified URL for the wheels. The actual versions of wheels are contained in the wheel metadata. Although we don't support Python 3.8 any more (because PyTorch 2.5 dropped support for Python 3.8), the wheels are still built with Python 3.8 ABI to keep the same wheel name as before.
Another way to access the latest code is to use the docker images:
.. code-block:: console
$ export VLLM_VERSION=0.6.1.post1 # vLLM's main branch version is currently set to latest released tag
These docker images are used for CI and testing only, and they are not intended for production use. They will be expired after several days.
The latest code can contain bugs and may not be stable. Please use it with caution.
.. _build_from_source:
Build from source
-----------------
=================
.. _python-only-build:
You can also build and install vLLM from source:
Python-only build (without compilation)
---------------------------------------
If you only need to change Python code, you can build and install vLLM without compilation. Using `pip's ``--editable`` flag <https://pip.pypa.io/en/stable/topics/local-project-installs/#editable-installs>`_, changes you make to the code will be reflected when you run vLLM:
This will download the latest nightly wheel and use the compiled libraries from there in the install.
The ``VLLM_PRECOMPILED_WHEEL_LOCATION`` environment variable can be used instead of ``VLLM_USE_PRECOMPILED`` to specify a custom path or URL to the wheel file. For example, to use the `0.6.1.post1 PyPi wheel <https://pypi.org/project/vllm/#files>`_:
You can find more information about vLLM's wheels `above <#install-the-latest-code>`_.
.. note::
This will uninstall existing PyTorch, and install the version required by vLLM. If you want to use an existing PyTorch installation, there need to be some changes:
There is a possibility that your source code may have a different commit ID compared to the latest vLLM wheel, which could potentially lead to unknown errors.
It is recommended to use the same commit ID for the source code as the vLLM wheel you have installed. Please refer to `the section above <#install-the-latest-code>`_ for instructions on how to install a specified wheel.
If you want to modify C++ or CUDA code, you'll need to build vLLM from source. This can take several minutes:
The differences are:
.. code-block:: console
- ``python use_existing_torch.py``: This script will remove all the PyTorch versions in the requirements files, so that the existing PyTorch installation will be used.
- ``pip install -r requirements-build.txt``: You need to manually install the requirements for building vLLM.
- ``pip install -e . --no-build-isolation``: You need to disable build isolation, so that the build system can use the existing PyTorch installation.
This is especially useful when the PyTorch dependency cannot be easily installed via pip, e.g.:
.. tip::
- build vLLM with PyTorch nightly or a custom PyTorch build.
- build vLLM with aarch64 and cuda (GH200), where the PyTorch wheels are not available on PyPI. Currently, only PyTorch nightly has wheels for aarch64 with CUDA. You can run ``pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124`` to install PyTorch nightly, and then build vLLM on top of it.
Building from source requires a lot of compilation. If you are building from source repeatedly, it's more efficient to cache the compilation results.
.. note::
For example, you can install `ccache <https://github.com/ccache/ccache>`_ using ``conda install ccache`` or ``apt install ccache`` .
As long as ``which ccache`` command can find the ``ccache`` binary, it will be used automatically by the build system. After the first build, subsequent builds will be much faster.
vLLM can fully run only on Linux, but you can still build it on other systems (for example, macOS). This build is only for development purposes, allowing for imports and a more convenient dev environment. The binaries will not be compiled and not work on non-Linux systems. You can create such a build with the following commands:
`sccache <https://github.com/mozilla/sccache>`_ works similarly to ``ccache``, but has the capability to utilize caching in remote storage environments.
The following environment variables can be set to configure the vLLM ``sccache`` remote: ``SCCACHE_BUCKET=vllm-build-sccache SCCACHE_REGION=us-west-2 SCCACHE_S3_NO_CREDENTIALS=1``. We also recommend setting ``SCCACHE_IDLE_TIMEOUT=0``.
.. code-block:: console
$ export VLLM_TARGET_DEVICE=empty
$ pip install -e .
Use an existing PyTorch installation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
There are scenarios where the PyTorch dependency cannot be easily installed via pip, e.g.:
* Building vLLM with PyTorch nightly or a custom PyTorch build.
* Building vLLM with aarch64 and CUDA (GH200), where the PyTorch wheels are not available on PyPI. Currently, only the PyTorch nightly has wheels for aarch64 with CUDA. You can run ``pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124`` to `install PyTorch nightly <https://pytorch.org/get-started/locally/>`_, and then build vLLM on top of it.
.. tip::
To build vLLM using an existing PyTorch installation:
Building from source requires quite a lot compilation. If you are building from source for multiple times, it is beneficial to cache the compilation results. For example, you can install `ccache <https://github.com/ccache/ccache>`_ via either ``conda install ccache`` or ``apt install ccache`` . As long as ``which ccache`` command can find the ``ccache`` binary, it will be used automatically by the build system. After the first build, the subsequent builds will be much faster.
.. code-block:: console
.. tip::
To avoid your system being overloaded, you can limit the number of compilation jobs
to be run simultaneously, via the environment variable ``MAX_JOBS``. For example:
Currently, before starting the build process, vLLM fetches cutlass code from GitHub. However, there may be scenarios where you want to use a local version of cutlass instead.
To achieve this, you can set the environment variable VLLM_CUTLASS_SRC_DIR to point to your local cutlass directory.
This is especially useful when you are building on less powerful machines. For example, when you use WSL, it only `gives you half of the memory by default <https://learn.microsoft.com/en-us/windows/wsl/wsl-config>`_, and you'd better use ``export MAX_JOBS=1`` to avoid compiling multiple files simultaneously and running out of memory. The side effect is that the build process will be much slower. If you only touch the Python code, slow compilation is okay, as you are building in an editable mode: you can just change the code and run the Python script without any re-compilation or re-installation.
.. code-block:: console
.. tip::
If you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image.
$ # Use `--ipc=host` to make sure the shared memory is large enough.
$ docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:23.10-py3
Troubleshooting
~~~~~~~~~~~~~~~
If you don't want to use docker, it is recommended to have a full installation of CUDA Toolkit. You can download and install it from `the official website <https://developer.nvidia.com/cuda-toolkit-archive>`_. After installation, set the environment variable ``CUDA_HOME`` to the installation path of CUDA Toolkit, and make sure that the ``nvcc`` compiler is in your ``PATH``, e.g.:
To avoid your system being overloaded, you can limit the number of compilation jobs
to be run simultaneously, via the environment variable ``MAX_JOBS``. For example:
.. code-block:: console
.. code-block:: console
$ export CUDA_HOME=/usr/local/cuda
$ export PATH="${CUDA_HOME}/bin:$PATH"
$ export MAX_JOBS=6
$ pip install -e .
Here is a sanity check to verify that the CUDA Toolkit is correctly installed:
This is especially useful when you are building on less powerful machines. For example, when you use WSL it only `assigns 50% of the total memory by default <https://learn.microsoft.com/en-us/windows/wsl/wsl-config#main-wsl-settings>`_, so using ``export MAX_JOBS=1`` can avoid compiling multiple files simultaneously and running out of memory.
A side effect is a much slower build process.
.. code-block:: console
Additionally, if you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image.
.. code-block:: console
$ # Use `--ipc=host` to make sure the shared memory is large enough.
$ docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:23.10-py3
If you don't want to use docker, it is recommended to have a full installation of CUDA Toolkit. You can download and install it from `the official website <https://developer.nvidia.com/cuda-toolkit-archive>`_. After installation, set the environment variable ``CUDA_HOME`` to the installation path of CUDA Toolkit, and make sure that the ``nvcc`` compiler is in your ``PATH``, e.g.:
.. code-block:: console
$ export CUDA_HOME=/usr/local/cuda
$ export PATH="${CUDA_HOME}/bin:$PATH"
Here is a sanity check to verify that the CUDA Toolkit is correctly installed:
.. code-block:: console
$ nvcc --version # verify that nvcc is in your PATH
$ ${CUDA_HOME}/bin/nvcc --version # verify that nvcc is in your CUDA_HOME
Unsupported OS build
--------------------
vLLM can fully run only on Linux but for development purposes, you can still build it on other systems (for example, macOS), allowing for imports and a more convenient development environment. The binaries will not be compiled and won't work on non-Linux systems.
Simply disable the ``VLLM_TARGET_DEVICE`` environment variable before installing:
.. code-block:: console
$ nvcc --version # verify that nvcc is in your PATH
$ ${CUDA_HOME}/bin/nvcc --version # verify that nvcc is in your CUDA_HOME
* Accelerator: NeuronCore_v2 (in trn1/inf2 instances)
* Pytorch 2.0.1/2.1.1
* AWS Neuron SDK 2.16/2.17 (Verified on python 3.8)
...
...
@@ -27,6 +27,10 @@ Installation steps:
.. _build_from_source_neuron:
.. note::
The currently supported version of Pytorch for Neuron installs `triton` version `2.1.0`. This is incompatible with vLLM >= 0.5.3. You may see an error `cannot import name 'default_dump_dir...`. To work around this, run a `pip install --upgrade triton==3.0.0` after installing the vLLM wheel.
vLLM powered by OpenVINO supports all LLM models from :doc:`vLLM supported models list <../models/supported_models>` and can perform optimal model serving on all x86-64 CPUs with, at least, AVX2 support. OpenVINO vLLM backend supports the following advanced vLLM features:
vLLM powered by OpenVINO supports all LLM models from :doc:`vLLM supported models list <../models/supported_models>` and can perform optimal model serving on all x86-64 CPUs with, at least, AVX2 support, as well as on both integrated and discrete Intel® GPUs (`the list of supported GPUs <https://docs.openvino.ai/2024/about-openvino/release-notes-openvino/system-requirements.html#gpu>`_). OpenVINO vLLM backend supports the following advanced vLLM features:
- [Optional] To use vLLM OpenVINO backend with a GPU device, ensure your system is properly set up. Follow the instructions provided here: `https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html <https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html>`_.
.. _openvino_backend_performance_tips:
Performance tips
----------------
vLLM OpenVINO backend uses the following environment variables to control behavior:
vLLM OpenVINO backend environment variables
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- ``VLLM_OPENVINO_DEVICE`` to specify which device utilize for the inference. If there are multiple GPUs in the system, additional indexes can be used to choose the proper one (e.g, ``VLLM_OPENVINO_DEVICE=GPU.1``). If the value is not specified, CPU device is used by default.
- ``VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON`` to enable U8 weights compression during model loading stage. By default, compression is turned off. You can also export model with different compression techniques using `optimum-cli` and pass exported folder as `<model_id>`
CPU performance tips
~~~~~~~~~~~~~~~~~~~~
CPU uses the following environment variables to control behavior:
- ``VLLM_OPENVINO_KVCACHE_SPACE`` to specify the KV Cache size (e.g, ``VLLM_OPENVINO_KVCACHE_SPACE=40`` means 40 GB space for KV cache), larger setting will allow vLLM running more requests in parallel. This parameter should be set based on the hardware configuration and memory management pattern of users.
- ``VLLM_OPENVINO_CPU_KV_CACHE_PRECISION=u8`` to control KV cache precision. By default, FP16 / BF16 is used depending on platform.
- ``VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON`` to enable U8 weights compression during model loading stage. By default, compression is turned off. You can also export model with different compression techniques using `optimum-cli` and pass exported folder as `<model_id>`
To enable better TPOT / TTFT latency, you can use vLLM's chunked prefill feature (``--enable-chunked-prefill``). Based on the experiments, the recommended batch size is ``256`` (``--max-num-batched-tokens``)
GPU device implements the logic for automatic detection of available GPU memory and, by default, tries to reserve as much memory as possible for the KV cache (taking into account ``gpu_memory_utilization`` option). However, this behavior can be overridden by explicitly specifying the desired amount of memory for the KV cache using ``VLLM_OPENVINO_KVCACHE_SPACE`` environment variable (e.g, ``VLLM_OPENVINO_KVCACHE_SPACE=8`` means 8 GB space for KV cache).
Currently, the best performance using GPU can be achieved with the default vLLM execution parameters for models with quantized weights (8 and 4-bit integer data types are supported) and `preemption-mode=swap`.
You can install vLLM using pip. It's recommended to use `conda <https://docs.conda.io/projects/conda/en/latest/user-guide/getting-started.html>`_ to create and manage Python environments.
.. code-block:: console
By default, vLLM downloads model from `HuggingFace <https://huggingface.co/>`_. If you would like to use models from `ModelScope <https://www.modelscope.cn>`_ in the following examples, please set the environment variable:
$ conda create -n myenv python=3.10 -y
$ conda activate myenv
$ pip install vllm
.. code-block:: shell
Please refer to the :ref:`installation documentation <installation>` for more details on installing vLLM.
export VLLM_USE_MODELSCOPE=True
.. _offline_batched_inference:
Offline Batched Inference
-------------------------
We first show an example of using vLLM for offline batched inference on a dataset. In other words, we use vLLM to generate texts for a list of input prompts.
With vLLM installed, you can start generating texts for list of input prompts (i.e. offline batch inferencing). The example script for this section can be found `here <https://github.com/vllm-project/vllm/blob/main/examples/offline_inference.py>`__.
The first line of this example imports the classes :class:`~vllm.LLM` and :class:`~vllm.SamplingParams`:
Import :class:`~vllm.LLM` and :class:`~vllm.SamplingParams` from vLLM.
The :class:`~vllm.LLM` class is the main class for running offline inference with vLLM engine.
The :class:`~vllm.SamplingParams` class specifies the parameters for the sampling process.
- :class:`~vllm.LLM` is the main class for running offline inference with vLLM engine.
- :class:`~vllm.SamplingParams` specifies the parameters for the sampling process.
.. code-block:: python
from vllm import LLM, SamplingParams
Define the list of input prompts and the sampling parameters for generation. The sampling temperature is set to 0.8 and the nucleus sampling probability is set to 0.95. For more information about the sampling parameters, refer to the `class definition <https://github.com/vllm-project/vllm/blob/main/vllm/sampling_params.py>`_.
The next section defines a list of input prompts and sampling parameters for text generation. The `sampling temperature <https://arxiv.org/html/2402.05201v1>`_ is set to ``0.8`` and the `nucleus sampling probability <https://en.wikipedia.org/wiki/Top-p_sampling>`_ is set to ``0.95``. You can find more information about the sampling parameters `here <https://docs.vllm.ai/en/stable/dev/sampling_params.html>`__.
.. code-block:: python
...
...
@@ -44,46 +56,46 @@ Define the list of input prompts and the sampling parameters for generation. The
Initialize vLLM's engine for offline inference with the :class:`~vllm.LLM` class and the `OPT-125M model <https://arxiv.org/abs/2205.01068>`_. The list of supported models can be found at :ref:`supported models <supported_models>`.
The :class:`~vllm.LLM` class initializes vLLM's engine and the `OPT-125M model <https://arxiv.org/abs/2205.01068>`_ for offline inference. The list of supported models can be found :ref:`here <supported_models>`.
.. code-block:: python
llm = LLM(model="facebook/opt-125m")
Call ``llm.generate`` to generate the outputs. It adds the input prompts to vLLM engine's waiting queue and executes the vLLM engine to generate the outputs with high throughput. The outputs are returned as a list of ``RequestOutput`` objects, which include all the output tokens.
.. note::
By default, vLLM downloads models from `HuggingFace <https://huggingface.co/>`_. If you would like to use models from `ModelScope <https://www.modelscope.cn>`_, set the environment variable ``VLLM_USE_MODELSCOPE`` before initializing the engine.
Now, the fun part! The outputs are generated using ``llm.generate``. It adds the input prompts to the vLLM engine's waiting queue and executes the vLLM engine to generate the outputs with high throughput. The outputs are returned as a list of ``RequestOutput`` objects, which include all of the output tokens.
The code example can also be found in `examples/offline_inference.py <https://github.com/vllm-project/vllm/blob/main/examples/offline_inference.py>`_.
.. _openai_compatible_server:
OpenAI-Compatible Server
------------------------
vLLM can be deployed as a server that implements the OpenAI API protocol. This allows vLLM to be used as a drop-in replacement for applications using OpenAI API.
By default, it starts the server at ``http://localhost:8000``. You can specify the address with ``--host`` and ``--port`` arguments. The server currently hosts one model at a time (OPT-125M in the command below) and implements `list models <https://platform.openai.com/docs/api-reference/models/list>`_, `create chat completion <https://platform.openai.com/docs/api-reference/chat/completions/create>`_, and `create completion <https://platform.openai.com/docs/api-reference/completions/create>`_ endpoints. We are actively adding support for more endpoints.
By default, it starts the server at ``http://localhost:8000``. You can specify the address with ``--host`` and ``--port`` arguments. The server currently hosts one model at a time and implements endpoints such as `list models <https://platform.openai.com/docs/api-reference/models/list>`_, `create chat completion <https://platform.openai.com/docs/api-reference/chat/completions/create>`_, and `create completion <https://platform.openai.com/docs/api-reference/completions/create>`_ endpoints.
Start the server:
Run the following command to start the vLLM server with the `Qwen2.5-1.5B-Instruct <https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct>`_ model:
.. code-block:: console
$ vllm serve facebook/opt-125m
$ vllm serve Qwen/Qwen2.5-1.5B-Instruct
By default, the server uses a predefined chat template stored in the tokenizer. You can override this template by using the ``--chat-template`` argument:
By default, the server uses a predefined chattemplate stored in the tokenizer. You can learn about overriding it `here <https://github.com/vllm-project/vllm/blob/main/docs/source/serving/openai_compatible_server.md#chat-template>`__.
This server can be queried in the same format as OpenAI API. For example, list the models:
This server can be queried in the same format as OpenAI API. For example, to list the models:
.. code-block:: console
...
...
@@ -91,17 +103,17 @@ This server can be queried in the same format as OpenAI API. For example, list t
You can pass in the argument ``--api-key`` or environment variable ``VLLM_API_KEY`` to enable the server to check for API key in the header.
Using OpenAI Completions API with vLLM
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
OpenAI Completions API with vLLM
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Query the model with input prompts:
Once your server is started, you can query the model with input prompts:
.. code-block:: console
$ curl http://localhost:8000/v1/completions \
$ -H "Content-Type: application/json" \
$ -d '{
$ "model": "facebook/opt-125m",
$ "model": "Qwen/Qwen2.5-1.5B-Instruct",
$ "prompt": "San Francisco is a",
$ "max_tokens": 7,
$ "temperature": 0
...
...
@@ -120,36 +132,32 @@ Since this server is compatible with OpenAI API, you can use it as a drop-in rep
For a more detailed client example, refer to `examples/openai_completion_client.py <https://github.com/vllm-project/vllm/blob/main/examples/openai_completion_client.py>`_.
Using OpenAI Chat API with vLLM
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
A more detailed client example can be found `here <https://github.com/vllm-project/vllm/blob/main/examples/openai_completion_client.py>`__.
The vLLM server is designed to support the OpenAI Chat API, allowing you to engage in dynamic conversations with the model. The chat interface is a more interactive way to communicate with the model, allowing back-and-forth exchanges that can be stored in the chat history. This is useful for tasks that require context or more detailed explanations.
OpenAI Chat Completions API with vLLM
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Querying the model using OpenAI Chat API:
vLLM is designed to also support the OpenAI Chat Completions API. The chat interface is a more dynamic, interactive way to communicate with the model, allowing back-and-forth exchanges that can be stored in the chat history. This is useful for tasks that require context or more detailed explanations.
You can use the `create chat completion <https://platform.openai.com/docs/api-reference/chat/completions/create>`_ endpoint to communicate with the model in a chat-like interface:
You can use the `create chat completion <https://platform.openai.com/docs/api-reference/chat/completions/create>`_ endpoint to interact with the model:
In order for you to use Cloud TPUs you need to have TPU quota granted to your
Google Cloud Platform project. TPU quotas specify how many TPUs you can use in a
GPC project and are specified in terms of TPU version, the number of TPU you
want to use, and quota type. For more information, see `TPU quota <https://cloud.google.com/tpu/docs/quota#tpu_quota>`_.
For TPU pricing information, see `Cloud TPU pricing <https://cloud.google.com/tpu/pricing>`_.
You may need additional persistent storage for your TPU VMs. For more
information, see `Storage options for Cloud TPU data <https://cloud.devsite.corp.google.com/tpu/docs/storage-options>`_.
Requirements
------------
* Google Cloud TPU VM (single & multi host)
* TPU versions: v5e, v5p, v4
* Python: 3.10
Installation options:
* Google Cloud TPU VM
* TPU versions: v6e, v5e, v5p, v4
* Python: 3.10 or newer
1. :ref:`Build a docker image with Dockerfile <build_docker_tpu>`.
2. :ref:`Build from source <build_from_source_tpu>`.
Provision Cloud TPUs
====================
.. _build_docker_tpu:
You can provision Cloud TPUs using the `Cloud TPU API <https://cloud.google.com/tpu/docs/reference/rest>`_
or the `queued resources <https://cloud.google.com/tpu/docs/queued-resources>`_
API. This section shows how to create TPUs using the queued resource API. For
more information about using the Cloud TPU API, see `Create a Cloud TPU using the Create Node API <https://cloud.google.com/tpu/docs/managing-tpus-tpu-vm#create-node-api>`_.
Queued resources enable you to request Cloud TPU resources in a queued manner.
When you request queued resources, the request is added to a queue maintained by
the Cloud TPU service. When the requested resource becomes available, it's
assigned to your Google Cloud project for your immediate exclusive use.
Build a docker image with :code:`Dockerfile.tpu`
------------------------------------------------
.. note::
In all of the following commands, replace the ALL CAPS parameter names with
appropriate values. See the parameter descriptions table for more information.
`Dockerfile.tpu <https://github.com/vllm-project/vllm/blob/main/Dockerfile.tpu>`_ is provided to build a docker image with TPU support.
Provision a Cloud TPU with the queued resource API
Since TPU relies on XLA which requires static shapes, vLLM bucketizes the possible input shapes and compiles an XLA graph for each different shape.
The compilation time may take 20~30 minutes in the first run.
However, the compilation time reduces to ~5 minutes afterwards because the XLA graphs are cached in the disk (in :code:`VLLM_XLA_CACHE_PATH` or :code:`~/.cache/vllm/xla_cache` by default).
$ # Make sure to add `--privileged --net host --shm-size=16G`.
$ docker run --privileged --net host --shm-size=16G -it vllm-tpu
.. note::
Since TPU relies on XLA which requires static shapes, vLLM bucketizes the
possible input shapes and compiles an XLA graph for each shape. The
compilation time may take 20~30 minutes in the first run. However, the
compilation time reduces to ~5 minutes afterwards because the XLA graphs are
cached in the disk (in :code:`VLLM_XLA_CACHE_PATH` or :code:`~/.cache/vllm/xla_cache` by default).
.. tip::
...
...
@@ -90,10 +188,11 @@ Next, build vLLM from source. This will only take a few seconds:
.. code-block:: console
from torch._C import * # noqa: F403
ImportError: libopenblas.so.0: cannot open shared object file: No such file or directory
ImportError: libopenblas.so.0: cannot open shared object file: No such
file or directory
Please install OpenBLAS with the following command:
- FP16 is the default data type in the current XPU backend. The BF16 data
type will be supported in the future.
Distributed inference and serving
---------------------------------
XPU platform supports tensor-parallel inference/serving and also supports pipeline parallel as a beta feature for online serving. We requires Ray as the distributed runtime backend. For example, a reference execution likes following:
.. code-block:: console
$ python -m vllm.entrypoints.openai.api_server \
$ --model=facebook/opt-13b \
$ --dtype=bfloat16 \
$ --device=xpu \
$ --max_model_len=1024 \
$ --distributed-executor-backend=ray \
$ --pipeline-parallel-size=2 \
$ -tp=8
By default, a ray instance will be launched automatically if no existing one is detected in system, with ``num-gpus`` equals to ``parallel_config.world_size``. We recommend properly starting a ray cluster before execution, referring helper `script <https://github.com/vllm-project/vllm/tree/main/examples/run_cluster.sh>`_.
