Fix nan issue when quantizing fp16 tensor. (#11213)

Author: comfyanonymous

comfy/quant_ops.py

@@ -399,7 +399,10 @@ def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_roun
         orig_dtype = tensor.dtype
 
         if isinstance(scale, str) and scale == "recalculate":
-            scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max
+            scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(dtype).max
+            if tensor.dtype not in [torch.float32, torch.bfloat16]:  # Prevent scale from being too small
+                tensor_info = torch.finfo(tensor.dtype)
+                scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max))
 
         if scale is not None:
             if not isinstance(scale, torch.Tensor):