Reduce default VRAM used by sub-quad attention

This does not seem to cause a significant change in performance, but should significantly lower the amount of RAM used in many cases.

Author: brkirch

modules/sd_hijack_optimizations.py

@@ -279,7 +279,7 @@ def sub_quad_attention(q, k, v, q_chunk_size=1024, kv_chunk_size=None, kv_chunk_
     qk_matmul_size_bytes = batch_x_heads * bytes_per_token * q_tokens * k_tokens
 
     if chunk_threshold is None:
-        chunk_threshold_bytes = max(int(get_available_vram() * 0.9), 1073741824) if q.device.type == 'mps' else int(get_available_vram() * 0.7)
+        chunk_threshold_bytes = 536870912
     elif chunk_threshold == 0:
         chunk_threshold_bytes = None
     else: