Right now we have sufficient fixed overhead for model run (#279). 

**no cudagraph**
Below: run of an empty model, no cudagraph times in ms. 

Before. 
Inductor overhead is 0.052 = 0.032 + 0.02 where 0.02 is overhead before
entering in compiler and 0.032 directly inductor overhead.
Hidet overhead is 0.205 = 0.185 + 0.02

After.
Hidet overhead is 0.068 = 0.048 + 0.02

Overhead reduced from 0.185ms -> 0.048ms or by 3.85x


**cudagraph**
Before 
0.162ms

After
0.124ms

Inductor 
0.88ms

For cudagraph there is one more room for improvement(left TODO in code).

Right now we have sufficient fixed overhead for model run (#279). 

**no cudagraph**
Below: run of an empty model, no cudagraph times in ms. 

Before. 
Inductor overhead is 0.052 = 0.032 + 0.02 where 0.02 is overhead before
entering in compiler and 0.032 directly inductor overhead.
Hidet overhead is 0.205 = 0.185 + 0.02

After.
Hidet overhead is 0.068 = 0.048 + 0.02

Overhead reduced from 0.185ms -> 0.048ms or by 3.85x


**cudagraph**
Before 
0.162ms

After
0.124ms

Inductor 
0.88ms

For cudagraph there is one more room for improvement(left TODO in code).

Right now we have sufficient fixed overhead for model run (#279). 

**no cudagraph**
Below: run of an empty model, no cudagraph times in ms. 

Before. 
Inductor overhead is 0.052 = 0.032 + 0.02 where 0.02 is overhead before
entering in compiler and 0.032 directly inductor overhead.
Hidet overhead is 0.205 = 0.185 + 0.02

After.
Hidet overhead is 0.068 = 0.048 + 0.02

Overhead reduced from 0.185ms -> 0.048ms or by 3.85x


**cudagraph**
Before 
0.162ms

After
0.124ms

Inductor 
0.88ms

For cudagraph there is one more room for improvement(left TODO in code).