Re: [PATCH v2 0/4] vLLM and the vLLM production stack

On Sat, Oct 04, 2025 at 12:55:36PM -0400, Chuck Lever wrote:
> On 10/4/25 12:38 PM, Luis Chamberlain wrote:
> > This adds initial vLLM and vLLM production stack support on kdevops.
> > 
> > This v2 series augments vLLM support for real CPUs on bare metal using
> > the DECLARE_HOSTS and also goes tested against a real GPU on the cloud,
> > showing that essentially now anyone can use the vLLM production stack on
> > any cloud provider we support in a flash. All we need are the instances
> > which have GPUs added, and for that we expect growth soon using dynamic
> > kconfig support.
> 
> As an update/road-map on that:
> 
> I think Lambda has GPU support already,

Yes, this goest tested with that.

> and AWS has enough dynamic menu
> support now that GPU-enabled instance types are available there with the
> default menus in the git tree. Please let me know if that's missing
> something.

Oh! I hadn't seen that and had been waiting for this! I'll test in a
couple of days!

Exciting times!

> I haven't done the follow-up work yet to integrate GPU-enabled AMIs into
> the AWS Compute menu. That seems like it should be the top priority. I
> need to go back and look at what you did to generate those in your
> prototype, to close those gaps.

Oh yes that's needed. I also need a patch to disable VPCs and enable
public IPs. While at it, so that AWS won't eat my corporate expenditures
I added a slack cloud-bill support too. The whole "static" stuff can be
ignored, it doesn't work, I was just trying to add static instances
to see if I could get some larger GPU instnaces to work but it didn't
work and I gave up. But the rest of the changes are legit, please feel
free to cherry pick what you see useful from here:

https://github.com/linux-kdevops/kdevops/tree/ci-testing/mcgrof/20251004-cloud-bill

> When that is complete, my next steps are to ask Claude to "copy" the
> scripts from terraform/aws/scripts to the other three major cloud
> providers that kdevops supports... an NFS bake-a-thon is this coming
> week, so there will be some delay.

Nice!!

> In the medium term, adding support for enabling RDMA fabrics in these
> environments is on my to-do list. I believe that will allow testing
> things like GPU direct with NVMe-o-F devices.

Oh my, that would be dreamy!

 Luis