ToTheBeginning
I have a problem related to pulid (in comfy) but I think Only pulid creator can HELP (So I am writing it here)
Hello @ToTheBeginning , My problem is that, I was only able to run pulid flux thought these instructions (https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) But for some reason it CANNOT run on comfy.
The reason might be related to the following, and I hope you can find an idea how to solve this:
- Either it is because my Card (Quadro RTX 6000) Here is what chatgpt says: (he Quadro RTX 6000 has Tensor Cores capable of BF16 computations, but its Tensor Core performance is below the 7 TFLOPS mark, which you correctly identified as the approximate minimal threshold for BF16-intensive workloads to run smoothly)
- or something about my comfy going to f16 instead of using bf16 (which can still be related to the first point)
for instance it shows:
And someone had this observation to say about it:
My question to you @ToTheBeginning , is why do you think I was able (with my card) to run pulid locally (with a local gradio from that page you provided) but not on comfy? IF you can understand something in the code that could be responsible for this, could you think of a solution or a change I can do in the code of pulid please?
For instance the error I get is: RuntimeError: expected scalar type Half but found BFloat16
What do you think?
