I found a case where this is needed (#10875)

Author: comfyanonymous

comfy/quant_ops.py

@@ -405,8 +405,8 @@ def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn):
 
         tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype)
         # TODO: uncomment this if it's actually needed because the clamp has a small performance penality'
-        # lp_amax = torch.finfo(dtype).max
-        # torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled)
+        lp_amax = torch.finfo(dtype).max
+        torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled)
         qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format)
 
         layout_params = {