Question about the new 40X speed improvement in MDM (~0.4 sec/sample)

[XGraphKhipu]

Hi! I noticed that MDM is now around 40 times faster, with sample generation taking just ~0.4 seconds. Amazing work! 🎉

I was wondering if you could share some insights on how this impressive speedup was achieved. I understand that the specific code hasn't been shared, but I'd be very interested to know what methods, optimizations, or general approaches you followed to reach this level of efficiency. Did you explore particular model architecture changes, optimization techniques, or hardware accelerations? Any information on the path you took would be greatly appreciated!