Skip to content

Latest commit

 

History

History
63 lines (41 loc) · 4.04 KB

File metadata and controls

63 lines (41 loc) · 4.04 KB

llm-d Accelerators

llm-d supports multiple accelerator vendors and we are expanding our coverage.

Support

Maintainers for each accelerator type are listed below. See our well-lit path guides for details of deploying on each hardware type.

Vendor Models Maintainers Supported Well-lit Paths
AMD ROCm Kenny Roche (Kenny.Roche@amd.com) Coming soon
Google TPU Edwin Hernandez (@Edwinhr716), Cong Liu (@liu-cong, congliu.thu@gmail.com) Inference Scheduling, Prefill/Decode Disaggregation
Intel XPU Yuan Wu (@yuanwu2017, yuan.wu@intel.com) Inference Scheduling, Prefill/Decode Disaggregation
NVIDIA GPU Will Eaton (weaton@redhat.com), Greg (grpereir@redhat.com) All

Requirements

We welcome contributions from accelerator vendors. To be referenced as a supported hardware vendor we require at minimum a publicly available container image that launches vLLM in the recommended configuration.

For integration into the well-lit paths our standard for contribution is higher, requiring:

  • A named maintainer responsible for keeping guide contents up to date
  • Manual or automated verification of the guide deployment for each release

Note

We aim to increase our requirements to have active CI coverage for all hardware guide variants in a future release.

[!NOTE] The community can assist but is not responsible for keeping hardware guide variants updated. We reserve the right to remove stale examples and documentation with regard to hardware support.

Intel XPU

Intel accelerators are supported via the well-lit paths (see the Intel row in the table above). For cluster prerequisites and image expectations, see the infrastructure prereq.

Accelerator Resource Management

To enable llm-d accelerators to access hardware devices, the devices must be exposed to containers. Kubernetes provides two mechanisms to accomplish this:

  1. Device Plugins
  2. Dynamic Resource Allocation (DRA)

Typically, clusters use one mechanism or the other to expose accelerator devices. While it's possible to use both mechanisms simultaneously, this requires special configuration not covered in this document.

Device Plugins

Each vendor provides Device Plugins for their accelerators. The following plugins are available by vendor:

Dynamic Resource Allocation

Each vendor provides DRA resource drivers for their accelerators. The following drivers and setup documentation are available by vendor:

Since DRA is a newer Kubernetes feature, some feature gates may be required. Consult your vendor and cluster provider documentation for specific requirements.