-
Epic
-
Resolution: Done
-
Undefined
-
None
-
None
-
None
I spoke with Wil Weaton and Manoj Kumar today regarding Spyre on p and
z-series and got a brain dump from them. In order to support Spyre on any
of the architectures we must support torch-2.5.1 (which may or may not be
a fork of torch), and provide a compatible 'empty' vllm
build. The 'empty' build is similar to the cpu build of vllm – it is
light weight and fast to compile. Spyre has a vllm plugin (which we know
about) that is provided to support Spyre. All of this
code, and several proprietary RPMs from IBM are required to support Spyre
on x86 on RHOAI. There are some additional install time issues to resolve as well.
Right now, RHOAI is running on packages directly from IBM – these are not
built in Konflux or any build system at Red Hat. They shipped a 'Tech
Preview' of this code, in which they pointed customers to a quay.io
container that had the torch-2.5.1, 'empty' vllm build, Spyre vllm plugin,
and IBM RPMs.
There are significant concerns with this approach, least of which is
maintaining CVE reports against this code.
- blocks
-
AIPCC-638 x86_64: IBM Spyre AIU Accelerator support in RHEL9.4
-
- Closed
-