Skip to content

v0.7.2+Gaudi-1.21.0
Compare
Choose a tag to compare
@github-actions github-actions released this 19 May 14:07
· 2918 commits to habana_main since this release
v0.7.2+Gaudi-1.21.0
0275ce4
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

vLLM with Intel® Gaudi® AI Accelerators

This README provides instructions on how to run vLLM with Intel Gaudi devices.

Requirements and Installation

To set up the execution environment, please follow the instructions in the Gaudi Installation Guide. To achieve the best performance on HPU, please follow the methods outlined in the Optimizing Training Platform Guide.

Requirements

  • Python 3.10
  • Intel Gaudi 2 and 3 AI accelerators
  • Intel Gaudi software version 1.21.0 and above

Quick Start Using Dockerfile

Set up the container with the latest Intel Gaudi Software Suite release using the Dockerfile.

Ubuntu

$ docker build -f Dockerfile.hpu -t vllm-hpu-env  .
$ docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --rm vllm-hpu-env

Tip

If you are facing the following error: docker: Error response from daemon: Unknown runtime specified habana., please refer to the "Install Optional Packages" section of Install Driver and Software and "Configure Container Runtime" section of Docker Installation. Make sure you have habanalabs-container-runtime package installed and that habana container runtime is registered.

Red Hat Enterprise Linux for Use with Red Hat OpenShift AI

$ docker build -f Dockerfile.hpu.ubi -t vllm-hpu-env  .
$ docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --rm vllm-hpu-env

Build from Source

Environment Verification

To verify that the Intel Gaudi software was correctly installed, run the following:

$ 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

Refer to System Verification and Final Tests for more details.

Run Docker Image

It is highly recommended to use the latest Docker image from the Intel Gaudi vault. Refer to the Intel Gaudi documentation for more details.

Use the following commands to run a Docker image. Make sure to update the versions below as listed in the Support Matrix:

$ docker pull vault.habana.ai/gaudi-docker/1.21.0/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
$ docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.21.0/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest

Build and Install vLLM

Currently, multiple ways are provided which can be used to install vLLM with Intel® Gaudi®, pick one option:

1. Build and Install the stable version

vLLM releases are being performed periodically to align with Intel® Gaudi® software releases. The stable version is released with a tag, and supports fully validated features and performance optimizations in Gaudi's vLLM-fork. To install the stable release from HabanaAI/vLLM-fork, run the following:

$ git clone https://github.com/HabanaAI/vllm-fork.git
$ cd vllm-fork
$ git checkout v0.7.2+Gaudi-1.21.0
$ pip install -r requirements-hpu.txt
$ python setup.py develop

2. Build and Install the latest from vLLM-fork

Currently, the latest features and performance optimizations are being developed in Gaudi's vLLM-fork and periodically upstreamed to the vLLM main repository. To install latest HabanaAI/vLLM-fork, run the following:

$ git clone https://github.com/HabanaAI/vllm-fork.git
$ cd vllm-fork
$ git checkout habana_main
$ pip install --upgrade pip
$ pip install -r requirements-hpu.txt
$ python setup.py develop

3. Build and Install from the vLLM main source

If you prefer to build and install directly from the main vLLM source, where periodically we are upstreaming new features, run the following:

$ git clone https://github.com/vllm-project/vllm.git
$ cd vllm
$ pip install -r requirements-hpu.txt
$ python setup.py develop

Supported Features

Feature Description References
Offline batched inference Offline inference using LLM class from vLLM Python API Quickstart
Example
Online inference via OpenAI-Compatible Server Online inference using HTTP server that implements OpenAI Chat and Completions API Documentation
Example
HPU autodetection HPU users do not need to specify the target platform, it will be detected automatically upon vLLM startup N/A
Paged KV cache with algorithms enabled for Intel Gaudi accelerators vLLM HPU backend contains a custom Paged Attention and cache operators implementations optimized for Gaudi devices. N/A
Custom Intel Gaudi operator implementations vLLM HPU backend provides optimized implementations of operators such as prefill attention, Root Mean Square Layer Normalization, Rotary Positional Encoding. N/A
Tensor parallel inference (single or multi-node multi-HPU) vLLM HPU backend supports multi-HPU inference across multiple nodes with tensor parallelism with multiprocessing or Ray and HCCL. Documentation
Example
HCCL reference
Pipeline parallel inference (single or multi-node multi-HPU) vLLM HPU backend supports multi-HPU inference across single or multi-node with pipeline parallelism. Documentation
Running Pipeline Parallelism
Inference with HPU Graphs vLLM HPU backend uses HPU Graphs by default for optimal performance. When HPU Graphs are enabled, execution graphs will be recorded ahead of time and replayed later during inference, significantly reducing host overheads. Documentation
vLLM HPU backend execution modes
Optimization guide
Inference with torch.compile vLLM HPU backend supports inference with torch.compile. vLLM HPU backend execution modes
INC quantization vLLM HPU backend supports FP8 model and KV cache quantization and calibration with Intel Neural Compressor (INC). (Not fully supported with torch.compile execution mode) Documentation
AutoAWQ quantization vLLM HPU backend supports inference with models quantized using AutoAWQ library. Library
AutoGPTQ quantization vLLM HPU backend supports inference with models quantized using AutoGPTQ library. Library
LoRA/MultiLoRA support vLLM HPU backend includes support for LoRA and MultiLoRA on supported models. Documentation
Example
vLLM supported models
Multi-step scheduling support vLLM HPU backend includes multi-step scheduling support for host overhead reduction, configurable by standard --num-scheduler-seqs parameter. Feature RFC
Automatic prefix caching vLLM HPU backend includes automatic prefix caching (APC) support for more efficient prefills, configurable by standard --enable-prefix-caching parameter. Documentation
Details
Speculative decoding (functional release) vLLM HPU backend includes experimental speculative decoding support for improving inter-token latency in some scenarios, configurable via standard --speculative_model and --num_speculative_tokens parameters. (Not fully supported with torch.compile execution mode) Documentation
Example
Multiprocessing backend Multiprocessing is the default distributed runtime in vLLM. The vLLM HPU backend supports it alongside Ray. Documentation
Multimodal vLLM HPU backend supports the inference for multi-modal models. (Not fully supported with t.compile execution mode) Documentation
Multinode support vLLM HPU backend supports distributed, multiple-node inference with Ray.
vLLM v1 architecture (early release) V1 architecture is now available for the HPU backend, and will gradually enable it for every use case we plan to support. Documentation
Guided decode vLLM HPU supports a guided decoding backend for generating structured outputs. Documentation
Delayed Sampling (experimental) vLLM HPU supports delayed sampling scheduling for asynchronous execution, enabled by VLLM_DELAYED_SAMPLING=true environment variable. N/A
Exponential bucketing (experimental) vLLM HPU supports exponential bucketing spacing instead of linear to automate configuration of bucketing mechanism, enabled by VLLM_EXPONENTIAL_BUCKETING=true environment variable. N/A

Note

All specified features are also supported with the -- enforce-eager flag.

Unsupported Features

  • Beam search
  • Prefill chunking (mixed-batch inferencing)

Validated Models and Configurations

The following configurations have been validated to function with Gaudi 2 or Gaudi 3 devices with random or greedy sampling. Configurations that are not listed may or may not work.

Model Tensor Parallelism [x HPU] Datatype Validated on
meta-llama/Llama-2-7b 1, 2, 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Llama-2-7b-chat-hf 1, 2, 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Llama-2-70b 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Llama-2-70b-chat-hf 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3-8B 1, 2, 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3-8B-Instruct 1, 2, 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3-70B 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3-70B-Instruct 8 BF16 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3.1-8B 1 BF16, FP8, INT4, FP16 (Gaudi 2) Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3.1-8B-Instruct 1 BF16, FP8 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3.1-70B 2, 4, 8 BF16, FP8, INT4 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3.1-70B-Instruct 2, 4, 8 BF16, FP8, FP16 (Gaudi 2) Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3.1-405B 8 BF16, FP8 Gaudi 3
meta-llama/Meta-Llama-3.1-405B-Instruct 8 BF16, FP8 Gaudi 3
meta-llama/Llama-3.2-11B-Vision 1 BF16, FP8 Gaudi 2, Gaudi 3
meta-llama/Llama-3.2-90B-Vision 4, 8 (min. for Gaudi 2) BF16, FP8 Gaudi 2, Gaudi 3
meta-llama/Llama-3.2-90B-Vision-Instruct 4, 8 (min. for Gaudi 2) BF16 Gaudi 2, Gaudi 3
meta-llama/Meta-Llama-3.3-70B 4 BF16, FP8 Gaudi 3
meta-llama/Granite-3B-code-instruct-128k 1 BF16 Gaudi 3
meta-llama/Granite-3.0-8B-instruct 1 BF16, FP8 Gaudi 2, Gaudi 3
meta-llama/Granite-20B-code-instruct-8k 1 BF16, FP8 Gaudi 2, Gaudi 3
meta-llama/Granite-34B-code-instruc-8k 1 BF16 Gaudi 3
mistralai/Mistral-Large-Instruct-2407 1, 4 BF16 Gaudi 3
mistralai/Mistral-7B-Instruct-v0.3 1, 2 BF16 Gaudi 2
mistralai/Mixtral-8x7B-Instruct-v0.1 2 FP8, BF16 Gaudi 2, Gaudi 3
llava-hf/llava-1.5-7b-hf 1, 8 BF16 Gaudi 2, Gaudi 3
princeton-nlp/gemma-2-9b-it-SimPO 1 BF16 Gaudi 2, Gaudi 3
Qwen/Qwen2-72B-Instruct 8 BF16 Gaudi 2
Qwen/Qwen2.5-72B-Instruct 8 BF16 Gaudi 2
meta-llama/CodeLlama-34b-Instruct-hf 1 BF16 Gaudi 3
deepseek-ai/DeepSeek-R1
quick start scripts		 8 FP8, BF16 Gaudi 2, Gaudi 3

Performance Tuning

Execution Modes

Currently, vLLM for HPU supports four execution modes, determined by the selected HPU PyTorch Bridge backend (via the PT_HPU_LAZY_MODE environment variable) and the --enforce-eager flag.

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

Note

Starting with the 1.21.0 Intel Gaudi software release, the torch.compile execution mode is the default for vLLM. HPU Graphs mode remains supported to ensure backward compatibility. Please verify the compatibility of the torch.compile mode with the information in the Supported Features table.

Tip

We recommend experimenting with the PT_HPU_LAZY_MODE environment variable to determine whether HPU Graphs or torch.compile mode performs better for your specific use case. While both modes generally deliver comparable performance, certain edge cases may favor one over the other.

Bucketing Mechanism

Intel Gaudi accelerators perform best when operating on models with fixed tensor shapes. Intel Gaudi Graph Compiler generates optimized binary code that implements the given model topology on Gaudi. In its default configuration, the produced binary code may be highly dependent on input and output tensor shapes, requiring graph recompilation when encountering tensors with different shapes within the same topology. While these binaries efficiently utilize Gaudi, the compilation process itself can introduce noticeable overhead in end-to-end execution. In dynamic inference serving scenarios, minimizing the number of graph compilations and reducing the risk of graph compilation occurring during server runtime is important. Currently, this is achieved by "bucketing" the model's forward pass across two dimensions: batch_size and sequence_length.

Note

Bucketing helps significantly reduce the number of required graphs, but does not handle graph compilation or device code generation. These tasks are performed during the warmup and HPUGraph capture phase.

Bucketing ranges are determined with 3 parameters - min, step, and max. They can be set separately for the prompt and decode phase, and batch size and sequence length dimensions. These parameters can be observed in logs during vLLM startup:

INFO 08-01 21:37:59 hpu_model_runner.py:493] Prompt bucket config (min, step, max_warmup) bs:[1, 32, 4], seq:[128, 128, 1024]
INFO 08-01 21:37:59 hpu_model_runner.py:499] Generated 24 prompt buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024)]
INFO 08-01 21:37:59 hpu_model_runner.py:504] Decode bucket config (min, step, max_warmup) bs:[1, 128, 4], seq:[128, 128, 2048]
INFO 08-01 21:37:59 hpu_model_runner.py:509] Generated 48 decode buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (1, 1152), (1, 1280), (1, 1408), (1, 1536), (1, 1664), (1, 1792), (1, 1920), (1, 2048), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (2, 1152), (2, 1280), (2, 1408), (2, 1536), (2, 1664), (2, 1792), (2, 1920), (2, 2048), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024), (4, 1152), (4, 1280), (4, 1408), (4, 1536), (4, 1664), (4, 1792), (4, 1920), (4, 2048)]

min determines the lowest value of the bucket. step determines the interval between buckets, and max determines the upper bound of the bucket. Furthermore, the interval between min and step has special handling - min gets multiplied by consecutive powers of two, until the multiplier is less than or equal to step. 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.

Example with ramp-up

min = 2, step = 32, max = 64
=> ramp_up = (2, 4, 8, 16)
=> stable = (32, 64)
=> buckets = ramp_up + stable => (2, 4, 8, 16, 32, 64)

Example without ramp-up

min = 128, step = 128, max = 512
=> ramp_up = ()
=> stable = (128, 256, 384, 512)
=> buckets = ramp_up + stable => (128, 256, 384, 512)

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 the 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.

For example, if a request with 3 sequences, each having a maximum sequence length of 412, is sent to an idle vLLM server, it will be padded and executed as a (4, 512) prefill bucket. This is because the batch_size (number of sequences) will be padded to 4 (the nearest batch size dimension higher than 3), and the maximum sequence length will be padded to 512 (the nearest sequence length dimension higher than 412). After the prefill stage, it will be executed as a (4, 512) decode bucket and will remain in this bucket until either the batch dimension changes (e.g., due to a request being completed), in which case it will become a (2, 512) bucket, or the context length increases beyond 512 tokens. It will become a (4, 640) bucket at that point.

Note

Bucketing is transparent to the user – padding in the sequence length dimension is never returned, and padding in the batch dimension does not create new requests.

Warmup

Warmup is an optional but highly recommended step that occurs before the vLLM server starts listening. It executes a forward pass for each bucket using dummy data. The goal is to pre-compile all graphs and avoid any graph compilation overhead within bucket boundaries during server runtime. Each warmup step is logged during vLLM startup.

This example uses the same buckets as those in the Bucketing Mechanism section. Each output line corresponds to the execution of a single bucket. When a bucket is executed for the first time, its graph is compiled and can be reused later, avoiding further graph compilations.

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

Tip

Compiling all the buckets may take some time and can be disabled by setting the VLLM_SKIP_WARMUP=true environment variable. Remember that if you do this, you may encounter graph compilations when executing a given bucket for the first time.

Warning

Disabling warmup is fine for development, but it is highly recommended to enable it in deployment.

HPU Graph Capture

HPU Graphs 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 used, they share the common memory pool ("usable memory") with the KV cache, as determined by the gpu_memory_utilization flag (default value is 0.9). Before the KV cache is allocated, the 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, the gpu_memory_utilization flag is applied. At its default value, it marks 90% of the free device memory at that point as usable. Next, the KV cache is allocated, the model is warmed up, and HPU Graphs are captured. The VLLM_GRAPH_RESERVED_MEM environment variable defines the ratio of memory reserved for HPU Graph capture. With its default value (VLLM_GRAPH_RESERVED_MEM=0.1), 10% of the usable memory will be reserved for graph capture (referred to as "usable graph memory"), and the remaining 90% will be used for the KV cache. The 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 share equal memory constraints. A lower value corresponds to less usable graph memory reserved for the prefill stage. For example, setting VLLM_GRAPH_PROMPT_RATIO=0.2 reserves 20% of usable graph memory for prefill graphs, while 80% is allocated for decode graphs.

Note

gpu_memory_utilization does not represent the absolute memory usage across the HPU. Instead, it specifies the memory margin after loading the model and running a profile. For example, if a device has 100 GiB of total memory and 50 GiB of free memory after loading the model weights and executing the profiling run, the default value of gpu_memory_utilization will mark 90% of the 50 GiB as usable, leaving 5 GiB as a margin, regardless of the total device memory.

You can also configure the strategy for capturing HPU graphs separately for the prompt and decode stages. The strategy affects the order in which graphs are captured. Two strategies are implemented:

  • max_bs - The graph capture queue is sorted in descending order by batch size. 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)), which is the default strategy for decode.
  • min_tokens - The graph capture queue is sorted in ascending order by the number of tokens each graph processes (batch_size*sequence_length), which is the default strategy for prompt.

When many requests are pending, the vLLM scheduler attempts to fill the maximum batch size for decoding as quickly as possible. Once a request is finished, the decode batch size decreases. When this happens, vLLM attempts to schedule a prefill iteration for requests in the waiting queue to restore the decode batch size to its previous state. In a fully loaded scenario, the decode batch size is often at its maximum, making large-batch HPU graphs critical to capture, as indicated by the max_bs strategy. Conversely, prefill iterations will typically be executed with very low batch sizes (1-4), as reflected in the min_tokens strategy.

Note

VLLM_GRAPH_PROMPT_RATIO does not set a hard limit on the memory allocated for graphs in each stage (prefill and decode). vLLM first attempts to use the entire usable prefill graph memory (usable graph memory * VLLM_GRAPH_PROMPT_RATIO) to capture prefilled HPU graphs. It will then attempt to do the same for decode graphs and the usable decode graph memory pool. If one stage is fully captured and there is unused memory remaining in the usable graph memory pool, vLLM will attempt to capture more graphs for the other stage, until no more HPU Graphs can be captured without exceeding the reserved memory pool. The behavior of this mechanism is illustrated in the example below.

Each step outlined is logged by the vLLM server, with negative values indicating memory release:

INFO 08-02 17:37:44 hpu_model_runner.py:493] Prompt bucket config (min, step, max_warmup) bs:[1, 32, 4], seq:[128, 128, 1024]
INFO 08-02 17:37:44 hpu_model_runner.py:499] Generated 24 prompt buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024)]
INFO 08-02 17:37:44 hpu_model_runner.py:504] Decode bucket config (min, step, max_warmup) bs:[1, 128, 4], seq:[128, 128, 2048]
INFO 08-02 17:37:44 hpu_model_runner.py:509] Generated 48 decode buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (1, 1152), (1, 1280), (1, 1408), (1, 1536), (1, 1664), (1, 1792), (1, 1920), (1, 2048), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (2, 1152), (2, 1280), (2, 1408), (2, 1536), (2, 1664), (2, 1792), (2, 1920), (2, 2048), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024), (4, 1152), (4, 1280), (4, 1408), (4, 1536), (4, 1664), (4, 1792), (4, 1920), (4, 2048)]
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, 4.755 GiB for prompt and 11.095 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:1128] Graph/Prompt captured:15 (62.5%) used_mem:14.03 GiB buckets:[(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (4, 128), (4, 256)]
INFO 08-02 17:38:43 hpu_model_runner.py:1128] Graph/Decode captured:48 (100.0%) used_mem:161.9 MiB buckets:[(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (1, 1152), (1, 1280), (1, 1408), (1, 1536), (1, 1664), (1, 1792), (1, 1920), (1, 2048), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (2, 1152), (2, 1280), (2, 1408), (2, 1536), (2, 1664), (2, 1792), (2, 1920), (2, 2048), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024), (4, 1152), (4, 1280), (4, 1408), (4, 1536), (4, 1664), (4, 1792), (4, 1920), (4, 2048)]
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

  • It is recommended to run inference on Gaudi 2 with block_size of 128 for BF16 data type. Using the default values (16, 32) may result in suboptimal performance due to underutilization of the Matrix Multiplication Engine (see Gaudi Architecture).
  • To achieve maximum throughput on Llama 7B, it is recommended to use a batch size of 128 or 256 and a maximum context length of 2048 with HPU Graphs enabled. If you experience out-of-memory issues, please refer to the Troubleshooting section below.

Environment Variables

Diagnostic and Profiling Knobs:

  • VLLM_PROFILER_ENABLED: if true - enables high-level profiler. Resulting JSON traces can be viewed at perfetto.habana.ai. Disabled by default.
  • VLLM_HPU_LOG_STEP_GRAPH_COMPILATION: if true - logs graph compilations for each vLLM engine step, but only if any compilation occurs. It is highly recommended to use this in conjunction with PT_HPU_METRICS_GC_DETAILS=1. Disabled by default.
  • VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL: if true - logs graph compilations for every vLLM engine step, even if no compilation occurs. Disabled by default.
  • VLLM_HPU_LOG_STEP_CPU_FALLBACKS: if true - logs CPU fallbacks for each vLLM engine step, but only if any fallback occurs. Disabled by default.
  • VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL: if true - logs CPU fallbacks for each vLLM engine step, even if no fallback occurs. Disabled by default.
  • VLLM_T_COMPILE_FULLGRAPH: if true - PyTorch compile function raises an error if any graph breaks happen during compilation. This allows for the easy detection of existing graph breaks, which usually reduce performance. Disabled by default.

Performance Tuning Knobs:

  • VLLM_SKIP_WARMUP: if true, warmup is skipped. The default is false.
  • VLLM_GRAPH_RESERVED_MEM: percentage of memory dedicated to HPUGraph capture. The default is 0.1.
  • VLLM_GRAPH_PROMPT_RATIO: percentage of reserved graph memory dedicated to prompt graphs. The default is 0.3.
  • VLLM_GRAPH_PROMPT_STRATEGY: strategy determining order of prompt graph capture, min_tokens or max_bs. The default is min_tokens.
  • VLLM_GRAPH_DECODE_STRATEGY: strategy determining order of decode graph capture, min_tokens or max_bs. The default is max_bs.
  • VLLM_EXPONENTIAL_BUCKETING: if true, enables exponential bucket spacing instead of linear (experimental).
  • VLLM_{phase}_{dim}_BUCKET_{param}: collection of 12 environment variables configuring ranges of bucketing mechanism (linear bucketing only).
    • {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
        • batch size step (VLLM_PROMPT_BS_BUCKET_STEP): min(max_num_seqs, 32)
        • batch size max (VLLM_PROMPT_BS_BUCKET_MAX): min(max_num_seqs, 64)
        • sequence length min (VLLM_PROMPT_SEQ_BUCKET_MIN): block_size
        • sequence length step (VLLM_PROMPT_SEQ_BUCKET_STEP): block_size
        • sequence length max (VLLM_PROMPT_SEQ_BUCKET_MAX): 1024

      • Decode:

        • batch size min (VLLM_DECODE_BS_BUCKET_MIN): 1
        • batch size step (VLLM_DECODE_BS_BUCKET_STEP): min(max_num_seqs, 32)
        • batch size max (VLLM_DECODE_BS_BUCKET_MAX): max_num_seqs
        • block size min (VLLM_DECODE_BLOCK_BUCKET_MIN): block_size
        • block size step (VLLM_DECODE_BLOCK_BUCKET_STEP): block_size
        • block size max (VLLM_DECODE_BLOCK_BUCKET_MAX): max(128, (max_num_seqs*2048)/block_size)
    • Recommended Values:

      • Prompt:

        • sequence length max (VLLM_PROMPT_SEQ_BUCKET_MAX): max_model_len

      • Decode:

        • block size max (VLLM_DECODE_BLOCK_BUCKET_MAX): max(128, (max_num_seqs*max_model_len/block_size)

Note

If the model config reports a high max_model_len, set it to max input_tokens+output_tokens rounded up to a multiple of block_size as per actual requirements.

Tip

When a deployed workload does not utilize the full context that a model can handle, it is good practice to limit the maximum values upfront based on the input and output token lengths that will be generated after serving the vLLM server.

Example:

Let's assume that we want to deploy text generation model Qwen2.5-1.5B, which has a defined max_position_embeddings of 131072 (our max_model_len). At the same time, we know that our workload pattern will not use the full context length because we expect a maximum input token size of 1K and predict generating a maximum of 2K tokens as output. In this case, starting the vLLM server to be ready for the full context length is unnecessary. Instead, we should limit it upfront to achieve faster service preparation and decrease warmup time. The recommended values in this example should be:

  • --max_model_len: 3072 - the sum of input and output sequences (1+2)*1024.
  • VLLM_PROMPT_SEQ_BUCKET_MAX: 1024 - the maximum input token size that we expect to handle.
  • VLLM_HANDLE_TOPK_DUPLICATES: if true - handles duplicates outside top-k. The default is false.
  • VLLM_CONFIG_HIDDEN_LAYERS: configures how many hidden layers to run in a HPUGraph for model splitting among hidden layers when TP is 1. It helps to improve throughput by reducing inter-token latency limitations in some models. The default is 1.

Additionally, there are HPU PyTorch Bridge environment variables impacting vLLM execution:

  • PT_HPU_LAZY_MODE: if 0, PyTorch Eager backend for Gaudi will be used. If 1, PyTorch Lazy backend for Gaudi will be used. The default is 0.

  • PT_HPU_ENABLE_LAZY_COLLECTIVES: must be set to true for tensor parallel inference with HPU Graphs. The default is true.

  • PT_HPUGRAPH_DISABLE_TENSOR_CACHE: must be set to false for LLaVA, qwen, and RoBERTa models. The default is false.

  • VLLM_PROMPT_USE_FLEX_ATTENTION: enabled only for the Llama model, allowing usage of torch.nn.attention.flex_attention instead of FusedSDPA. Requires VLLM_PROMPT_USE_FUSEDSDPA=0. The default is false.

Quantization, FP8 Inference and Model Calibration Process

Note

Measurement files are required to run quantized models with vLLM on Gaudi accelerators. The FP8 model calibration procedure is described in detail in docs.habana.ai vLLM Inference Section. An end-to-end example tutorial for quantizing a BF16 Llama 3.1 model to FP8 and then inferencing is provided in this Gaudi-tutorials repository.

Once you have completed the model calibration process and collected the measurements, you can run FP8 inference with vLLM using the following command:

export QUANT_CONFIG=/path/to/quant/config/inc/meta-llama-3.1-405b-instruct/maxabs_quant_g3.json
vllm serve meta-llama/Llama-3.1-405B-Instruct --dtype bfloat16 --max-model-len  2048 --block-size 128 --max-num-seqs 32 --quantization inc --kv-cache-dtype fp8_inc --weights-load-device cpu --tensor-parallel-size 8

QUANT_CONFIG is an environment variable that points to the measurement or quantization configuration file. The measurement configuration file is used during the calibration procedure to collect measurements for a given model. The quantization configuration is used during inference.

Tip

If you are prototyping or testing your model with FP8, you can use the VLLM_SKIP_WARMUP=true environment variable to disable the warmup stage, which is time-consuming. However, disabling this feature in production environments is not recommended, as it can lead to a significant performance decrease.

Tip

If you are benchmarking an FP8 model with scale_format=const, setting VLLM_DISABLE_MARK_SCALES_AS_CONST=true can help speed up the warmup stage.

Tip

When using FP8 models, you may experience timeouts caused by the long compilation time of FP8 operations. To mitigate this, set the following environment variables:

  • VLLM_ENGINE_ITERATION_TIMEOUT_S - to adjust the vLLM server timeout. You can set the value in seconds, e.g., 600 equals 10 minutes.
  • VLLM_RPC_TIMEOUT - to adjust the RPC protocol timeout used by the OpenAI-compatible API. This value is in microseconds, e.g., 600000 equals 10 minutes.

Long Context Support

Long context feature enables support for a token context window exceeding 128K tokens. It is supported by the following models:

Environment Variables Settings

Set the following environment variables to avoid OOM/functional issues. Additional environment variable settings depend on context length:

  • VLLM_ENGINE_ITERATION_TIMEOUT_S=3600
  • VLLM_RPC_TIMEOUT=100000
  • VLLM_PROMPT_USE_FUSEDSDPA=1
  • PT_HPU_ENABLE_LAZY_COLLECTIVES=true
  • PT_HPUGRAPH_DISABLE_TENSOR_CACHE=1
  • VLLM_ALLOW_LONG_MAX_MODEL_LEN=1

32K context length flags examples:

  • VLLM_GRAPH_RESERVED_MEM: The value depends on the model and context length settings. Use VLLM_GRAPH_RESERVED_MEM=0.02 for Llama3.1-8B or VLLM_GRAPH_RESERVED_MEM=0.1 for Llama3.1-70B.
  • VLLM_PROMPT_BS_BUCKET_MIN=1: Suggested value, depends on the model. You can increase it until you reach an OOM error or decrease it if OOM occurs.
  • VLLM_PROMPT_BS_BUCKET_STEP=16: Suggested value, depends on the model. Increasing the step value results in fewer buckets. If an OOM error occurs, the value should be increased.
  • VLLM_PROMPT_BS_BUCKET_MAX=16: Suggested value, depends on the model. You can increase it until you reach an OOM error or decrease it if OOM occurs.
  • VLLM_PROMPT_SEQ_BUCKET_MIN=24576: Suggested value, depends on warmup results.
  • VLLM_PROMPT_SEQ_BUCKET_STEP=2048: Suggested value, depends on warmup results. It is recommended to increase it to a higher value for faster warmup. VLLM_PROMPT_SEQ_BUCKET_STEP=16384 - Suggested value for Intel Gaudi 3.
  • VLLM_PROMPT_SEQ_BUCKET_MAX=32768: Value for context length of 32K. Use 16384 for 16K.
  • VLLM_DECODE_BLOCK_BUCKET_MIN=1024: Suggested value, depends on warmup results.
  • VLLM_DECODE_BLOCK_BUCKET_STEP=1024: Suggested value, depends on warmup results.
  • VLLM_DECODE_BLOCK_BUCKET_MAX=33792: max_num_seqs * max_decode_seq // self.block_size, where max_decode_seq represents the sum of input and output sequences. For example:
    • 128 *((32 + 1)* 1024) / 128
    • 32 *((32 + 1)* 1024) / 128

Batch Size Settings

The default batch_size=256 is not optimal for long contexts (8K+). Recompilations may occur if there is not enough KV cache space for some sequence groups.

If recompilation or next recomputation warnings appear during inference, reduce batch_size to improve stability.

Recompilation message example:

Configuration: (prompt, 1, 36864) was not warmed-up!

Warning message example:

Sequence group cmpl-3cbf19b0c6d74b3f90b5d5db2ed2385e-0 is preempted by PreemptionMode.RECOMPUTE mode because there is not enough KV cache space. This can affect the end-to-end performance. Increase gpu_memory_utilization or tensor_parallel_size to provide more KV cache memory.

Multi-Step Scheduling Feature Usage

Enabling Multi-Step Scheduling is recommended for better decode performance. Refer to vllm-project#6854 for more details.

Pipeline Parallelism

Pipeline parallelism is a distributed model parallelization technique that splits the model vertically across its layers, distributing different parts of the model across multiple devices. With this feature, when running a model that does not fit on a single node with tensor parallelism and requires a multi-node solution, we can split the model vertically across its layers and distribute the slices across available nodes. For example, if we have two nodes, each with 8 HPUs, we no longer have to use tensor_parallel_size=16 but can instead set tensor_parallel_size=8 with pipeline_parallel_size=2. Because pipeline parallelism runs a pp_size number of virtual engines on each device, we have to lower max_num_seqs accordingly, since it acts as a micro batch for each virtual engine.

Running Pipeline Parallelism

The following example shows how to use Pipeline parallelism with vLLM on HPU:

vllm serve <model_path> --device hpu --tensor-parallel-size 8 --pipeline_parallel_size 2 --distributed-executor-backend ray

Note

Currently, pipeline parallelism on Lazy mode requires the PT_HPUGRAPH_DISABLE_TENSOR_CACHE=0 flag.

Multi-node Support

vLLM works with a multi-node environment setup via Ray. To run models on multiple nodes, follow the procedure below.

Prerequisites

Perform the following on all nodes:

  • Install the latest vllm-fork.

  • Check if all Gaudi NIC ports are up by running:

Note

Following commands should be run on the host and NOT inside the container.

cd /opt/habanalabs/qual/gaudi2/bin 
./manage_network_ifs.sh --status 
# All the ports should be in 'up' state. Try flipping the state
./manage_network_ifs.sh --down 
./manage_network_ifs.sh --up
# Give it a minute for the NIC's to flip and check the status again
  • Set the following flags:
# Check the network interface for outbound/inbound comms. Command 'ip a' or 'ifconfig' should list all the interfaces
export GLOO_SOCKET_IFNAME=eth0
export HCCL_SOCKET_IFNAME=eth0

1. Start Ray on the head node:

ray start --head --port=6379

2. Add workers to the Ray cluster:

ray start --address='<ip-of-ray-head-node>:6379'

3. Start the vLLM server:

vllm serve meta-llama/Llama-3.1-405B-Instruct --dtype bfloat16 --max-model-len  2048 --block-size 128 --max-num-seqs 32 --tensor-parallel-size 16 --distributed-executor-backend ray

Note

Running FP8 models with a multi-node setup is described in the documentation of FP8 calibration procedure: README.

Other Online Serving Examples

Please refer to this collection of static-batched online serving example scripts designed to help the user reproduce performance numbers with vLLM on Gaudi for various types of models and varying context lengths. Below is a list of the models and example scripts provided for 2K and 4K context length scenarios:

  • deepseek-r1-distill-llama-70b_gaudi3_1.20_contextlen-2k
  • deepseek-r1-distill-llama-70b_gaudi3_1.20_contextlen-4k
  • llama-3.1-70b-instruct_gaudi3_1.20_contextlen-2k
  • llama-3.1-70b-instruct_gaudi3_1.20_contextlen-4k
  • llama-3.1-8b-instruct_gaudi3_1.20_contextlen-2k
  • llama-3.1-8b-instruct_gaudi3_1.20_contextlen-4k
  • llama-3.3-70b-instruct_gaudi3_1.20_contextlen-2k
  • llama-3.3-70b-instruct_gaudi3_1.20_contextlen-4k

Troubleshooting

The following steps address Out of Memory related errors:

  • Increase gpu_memory_utilization - This addresses insufficient overall memory. The vLLM pre-allocates HPU cache by using gpu_memory_utilization% of device memory. By default, gpu_memory_utilization is set to 0.9. By increasing this utilization, you can provide more KV cache space.
  • Decrease max_num_seqs or max_num_batched_tokens - This may reduce the number of concurrent requests in a batch, thereby requiring less KV cache space when overall usable memory is limited.
  • Increase tensor_parallel_size - This approach shards model weights, so each GPU has more memory available for KV cache.
  • To maximize the memory available for the KV cache, you can disable HPUGraph completely. With HPU Graphs disabled, you are trading latency and throughput at lower batches for potentially higher throughput on higher batches. You can do that by adding --enforce-eager flag to the server (for online inference), or by passing enforce_eager=True argument to LLM constructor (for offline inference).

What's Changed

New Contributors

Full Changelog: v0.6.2...v0.7.2+Gaudi-1.21.0

Contributors

CAC, xuechendi, and 76 other contributors
Assets 2
Loading