Speed up inference on nvidia 10 series on Linux.

comfy/model_management.py

@@ -370,8 +370,10 @@ def offloaded_memory(loaded_models, device):
             offloaded_mem += m.model_offloaded_memory()
     return offloaded_mem
 
+WINDOWS = any(platform.win32_ver())
+
 EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
-if any(platform.win32_ver()):
+if WINDOWS:
     EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
 
 if args.reserve_vram is not None:
@@ -1002,7 +1004,10 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"]
     for x in nvidia_10_series:
         if x in props.name.lower():
-            return True
+            if WINDOWS or manual_cast:
+                return True
+            else:
+                return False #weird linux behavior where fp32 is faster
 
     if manual_cast:
         free_model_memory = maximum_vram_for_weights(device)