ConceptMarlin kernel / weight-only Marlin
weight-only Marlin kernel
A GPU kernel that stores weights in low precision only and dequantizes them back to 16-bit during computation, allowing GPUs without native 4-bit compute capability to benefit from the memory and bandwidth savings of reduced weight size.
Timeline (last 90 days)
Oct 18
On the RTX 4070 Ti SUPER, NVFP4 runs via the weight-only Marlin kernel, with weights kept in low-precision storage but the actual computation still performed at 16 bits.
Unverified50%