float8 ops on cpu broken in inductor

[vkuzo]

🐛 Describe the bug

If we have some code with device=cpu which uses float8, such as below, things are broken in inductor codegen:

def foo(x):                       
    x = x.to(torch.float8_e4m3fn) 
    return x                      
                                  
foo = torch.compile(foo)          
x = torch.randn(2, 2)             
x = foo(x)                        
print(x)                          

logs: https://gist.github.com/vkuzo/ae434bfe8a48c083e74377c313263a90

can we fall back to eager mode on cpu?

Versions

>>> torch.__version__
'2.2.0a0+git967863d'

[vkuzo]

cc @int3 , @ipiszy

[zhuhaozhe]

Enabled FP8 cast on CPU backend #117737

[zhuhaozhe]

Closed as #117737 merged. Feel free to re-open if there are still some issues.