Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evolved into a community-driven project with contributions from both academia and industry.
- Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
安装依赖:
- Continuous batching of incoming requests
```shell
- Fast model execution with CUDA/HIP graph
pip install-r requirements/rocm.txt
- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516), INT4, INT8, and FP8
```
- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer
- 提供2种源码编译方式(进入vllm目录):
- Speculative decoding
```
- Chunked prefill
如果使用vllm基础镜像，需要先下载vllm: pip uninstall vllm
vLLM is flexible and easy to use with:
- Seamless integration with popular Hugging Face models
- High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
- Tensor, pipeline, data and expert parallelism support for distributed inference
- Streaming outputs
- OpenAI-compatible API server
- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, Arm CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
- Prefix caching support
- Multi-LoRA support
vLLM seamlessly supports most popular open-source models on HuggingFace, including:
- Transformer-like LLMs (e.g., Llama)
- Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
- Embedding Models (e.g., E5-Mistral)
- Multi-modal LLMs (e.g., LLaVA)
1. 编译whl包并安装
Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html).
Please check out [Contributing to vLLM](https://docs.vllm.ai/en/latest/contributing/index.html) for how to get involved.
## Citation
If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs/2309.06180):
```bibtex
@inproceedings{kwon2023efficient,
title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
author={Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica},
booktitle={Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles},
@@ -47,6 +47,10 @@ You can tune the performance by adjusting `max_num_batched_tokens`:
...
@@ -47,6 +47,10 @@ You can tune the performance by adjusting `max_num_batched_tokens`:
- For optimal throughput, we recommend setting `max_num_batched_tokens > 8192` especially for smaller models on large GPUs.
- For optimal throughput, we recommend setting `max_num_batched_tokens > 8192` especially for smaller models on large GPUs.
- If `max_num_batched_tokens` is the same as `max_model_len`, that's almost the equivalent to the V0 default scheduling policy (except that it still prioritizes decodes).
- If `max_num_batched_tokens` is the same as `max_model_len`, that's almost the equivalent to the V0 default scheduling policy (except that it still prioritizes decodes).
!!! warning
When chunked prefill is disabled, `max_num_batched_tokens` must be greater than `max_model_len`.
In that case, if `max_num_batched_tokens < max_model_len`, vLLM may crash at server start‑up.
@@ -43,9 +43,54 @@ Further update the model as follows:
...
@@ -43,9 +43,54 @@ Further update the model as follows:
)
)
```
```
- Implement [embed_multimodal][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_multimodal] that returns the embeddings from running the multimodal inputs through the multimodal tokenizer of the model. Below we provide a boilerplate of a typical implementation pattern, but feel free to adjust it to your own needs.
- Remove the embedding part from the [forward][torch.nn.Module.forward] method:
- Move the multi-modal embedding to [embed_multimodal][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_multimodal].
- The text embedding and embedding merge are handled automatically by a default implementation of [embed_input_ids][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_input_ids]. It does not need to be overridden in most cases.
Below we provide a boilerplate of a typical implementation pattern of [embed_multimodal][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_multimodal], but feel free to adjust it to your own needs.
@@ -10,7 +10,7 @@ receives a request for a LoRA adapter that hasn't been loaded yet, the resolver
...
@@ -10,7 +10,7 @@ receives a request for a LoRA adapter that hasn't been loaded yet, the resolver
to locate and load the adapter from their configured storage locations. This enables:
to locate and load the adapter from their configured storage locations. This enables:
-**Dynamic LoRA Loading**: Load adapters on-demand without server restarts
-**Dynamic LoRA Loading**: Load adapters on-demand without server restarts
-**Multiple Storage Backends**: Support for filesystem, S3, and custom backends. The built-in `lora_filesystem_resolver` requires a local storage path, but custom resolvers can be implemented to fetch from any source.
-**Multiple Storage Backends**: Support for filesystem, S3, and custom backends. The built-in `lora_filesystem_resolver` requires a local storage path, while the built-in `hf_hub_resolver` will pull LoRA adapters from Huggingface Hub and proceed in an identical manner. In general, custom resolvers can be implemented to fetch from any source.
-**Automatic Discovery**: Seamless integration with existing LoRA workflows
-**Automatic Discovery**: Seamless integration with existing LoRA workflows
-**Scalable Deployment**: Centralized adapter management across multiple vLLM instances
-**Scalable Deployment**: Centralized adapter management across multiple vLLM instances
| pplx | batched | fp8,int8 | G,A,T | Y | Y | [`PplxPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.pplx_prepare_finalize.PplxPrepareAndFinalize] |
| pplx | batched | fp8,int8 | G,A,T | Y | Y | [`PplxPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.pplx_prepare_finalize.PplxPrepareAndFinalize] |
| deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
| deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
| deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
| deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
| flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferAllToAllMoEPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize.FlashInferAllToAllMoEPrepareAndFinalize] |
| flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferA2APrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize.FlashInferA2APrepareAndFinalize] |
| flashinfer<sup>4</sup> | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferCutlassMoEPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize.FlashInferCutlassMoEPrepareAndFinalize] |
| MoEPrepareAndFinalizeNoEP<sup>5</sup> | standard | fp8,int8 | G,A,T | N | Y | [`MoEPrepareAndFinalizeNoEP`][vllm.model_executor.layers.fused_moe.prepare_finalize.MoEPrepareAndFinalizeNoEP] |
| MoEPrepareAndFinalizeNoEP<sup>5</sup> | standard | fp8,int8 | G,A,T | N | Y | [`MoEPrepareAndFinalizeNoEP`][vllm.model_executor.layers.fused_moe.prepare_finalize.MoEPrepareAndFinalizeNoEP] |
| BatchedPrepareAndFinalize<sup>5</sup> | batched | fp8,int8 | G,A,T | N | Y | [`BatchedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedPrepareAndFinalize] |
| BatchedPrepareAndFinalize<sup>5</sup> | batched | fp8,int8 | G,A,T | N | Y | [`BatchedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedPrepareAndFinalize] |
@@ -159,10 +159,12 @@ Alternatively, you can use the LoRAResolver plugin to dynamically load LoRA adap
...
@@ -159,10 +159,12 @@ Alternatively, you can use the LoRAResolver plugin to dynamically load LoRA adap
You can set up multiple LoRAResolver plugins if you want to load LoRA adapters from different sources. For example, you might have one resolver for local files and another for S3 storage. vLLM will load the first LoRA adapter that it finds.
You can set up multiple LoRAResolver plugins if you want to load LoRA adapters from different sources. For example, you might have one resolver for local files and another for S3 storage. vLLM will load the first LoRA adapter that it finds.
You can either install existing plugins or implement your own. By default, vLLM comes with a [resolver plugin to load LoRA adapters from a local directory.](https://github.com/vllm-project/vllm/tree/main/vllm/plugins/lora_resolvers)
You can either install existing plugins or implement your own. By default, vLLM comes with a [resolver plugin to load LoRA adapters from a local directory, as well as a resolver plugin to load LoRA adapters from repositories on Hugging Face Hub](https://github.com/vllm-project/vllm/tree/main/vllm/plugins/lora_resolvers)
To enable this resolver, set `VLLM_ALLOW_RUNTIME_LORA_UPDATING` to True, set `VLLM_PLUGINS` to include `lora_filesystem_resolver`, and then set `VLLM_LORA_RESOLVER_CACHE_DIR` to a local directory. When vLLM receives a request using a LoRA adapter `foobar`,
To enable either of these resolvers, you must `set VLLM_ALLOW_RUNTIME_LORA_UPDATING` to True.
it will first look in the local directory for a directory `foobar`, and attempt to load the contents of that directory as a LoRA adapter. If successful, the request will complete as normal and
that adapter will then be available for normal use on the server.
- To leverage a local directory, set `VLLM_PLUGINS` to include `lora_filesystem_resolver` and set `VLLM_LORA_RESOLVER_CACHE_DIR` to a local directory. When vLLM receives a request using a LoRA adapter `foobar`,
it will first look in the local directory for a directory `foobar`, and attempt to load the contents of that directory as a LoRA adapter. If successful, the request will complete as normal and that adapter will then be available for normal use on the server.
- To leverage repositories on Hugging Face Hub, set `VLLM_PLUGINS` to include `lora_hf_hub_resolver` and set `VLLM_LORA_RESOLVER_HF_REPO_LIST` to a comma separated list of repository IDs on Hugging Face Hub. When vLLM receives a request for the LoRA adapter `my/repo/subpath`, it will download the adapter at the `subpath` of `my/repo` if it exists and contains an `adapter_config.json`, then build a request to the cached dir for the adapter, similar to the `lora_filesystem_resolver`. Please note that enabling remote downloads is insecure and not intended for use in production environments.
Alternatively, follow these example steps to implement your own plugin:
Alternatively, follow these example steps to implement your own plugin:
