Viewing profile — ssheng

To get a full A10, one must also get 36 vCPU and 440 GB of memory. I must be missing something.

Quality loss with quantization is expected. It seems like with GPTQ the loss is within acceptable range based on the perplexity score shown.

How does Exllama rank among these? Heard good things about it.

Creative idea. Any data on how much it takes to load the LoRAs and how much latency it adds to the generation speed?

Hello! We're the authors of this blog post. Please let us know if there are other models and inference backends you'd like us to benchmark next.

The user feedback seems really encouraging. https://www.linkedin.com/posts/eric-riddoch_its-early-to-say...

FastAPI is great building block but can't expect it to work for model serving out of box.