Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Affiliation:
4
0
7
2
This paper asks how many FP4 GEMMs are needed for one integer matrix product, and proposes Adaptive Weight Encoding (AWE): the limbs take freely chosen integer weights, the stored planes are linear combinations of limbs, and the exact product is the sum of the FP4 GEMMs scaled by reconstruction coefficients.
Deopt-Reopt can significantly accelerate CUDA translation for certain HPC kernels, but its advantages are kernel-dependent and not universally applicable.
LLMs generating code from a shared prefix exhibit surprisingly high MoE routing similarity even when producing different tokens, but this similarity varies significantly across layers, revealing potential optimization points.
Forget synthetic benchmarks – this work trains LLMs to write faster HPC code by directly rewarding GFLOPS measured on a real supercomputer.