torch compile and cuda alloc improvements (#1755)

torch compile and cuda alloc improvements (#1755) #133