modules.sd_hijack_optimization.py:

[moonryul]

line 336 - 342:

if mem_required > mem_free_total:
steps = 2 ** (math.ceil(math.log(mem_required / mem_free_total, 2))) #MJ: 20, 22, 2**3, ...
print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")

    if steps > 64: #MJ: 2.5:  for every 2.5 bytes of free memory, only 1 byte is effectively usable for the tensor. The divisor "2.5" might be derived from some heuristic or specific memory management consideration  
        max_res = math.floor( math.sqrt(   math.sqrt(mem_free_total / 2.5)  ) / 8 ) * 64
        
        =>
        
        Can you explain to me 
        1) What does "steps" refer to ?
        2) what is the logic behind the formula for the computation of steps
        3) what is the logic behind the computation of "max_res"? I guess that number 8 refers to the compression factor by which the image is compressed into the latent space, where the attention matrix is computed (The code snippet seems to compute the memory needed to compute the self attention matrix for the latent tensor.
        
        Sincerely
        Moon R. Jung
        
        Professor, Dept of Art & Technology
        Sogang Univ, Seoul, Korea