Closed andy108369 closed 4 months ago
the best provider to test this now is provider.medc1.com
(akash1ffpcy473xqs37yvv4jhhh2hsuv786nzs4xt0dj
) which got t4
, a100
and rtx4090
(a40
probably soon)
Thanks Andrey - I ran into this yesterday and wasn't aware of this issue - so, appreciated! - going to addd this to backlog
May 14th, 2024:
akash node/network 0.30.0, and 0.32.2 provider-services version 0.4.8, and 0.5.13
Update (Apr 30, 2024): TL;DR The issue is basically this:
model: *
set in the SDL eventually gets through the bid engine which then gets passed to the bid price script (or whatever reads thestdin
to process theresources
); What's expected is that the actually available GPU which got selected by the K8s engine.The current version of the bid price script automatically sets the highest price when a specific GPU is not designated in the SDL. This approach is adopted because the script is unable to determine which GPU is being requested or which one is available; therefore, it defaults to the highest price. This serves as a temporary solution to ensure that the provider does not inadvertently offer a high-end GPU at the cost of a lower-end model. However, it causes an issue when only the lower-end models are available and the client does not explicitly specify the model in the SDL - this way he gets lower-end models for the highest price of the highest-end model.
Provider needs to pass the GPU model it picked to the bid price script before it Bids so it can price it (when the GPU
model
is not set in SDL deployment manifest).