[Architectural Proof-of-Concept] Saliency-Boosted GPTQ & High-Entropy Routing

[Subramanyam6]

This PR is submitted as an architectural proof-of-concept and an explicit pitch for the $500
RunPod Advanced Competitor Grant.

Due to a metric scaling bug (bytes_per_token=3.5), we initially logged a pre-quantization BPB of
0.9756, which corrects to a true pre-quantization BPB of 1.402 (starved of convergence due to a 10x
throughput deficit without Flash Attention 3).

However, the architectural innovations introduced here—Optimizer Saliency-Boosted GPTQ (a
zero-byte architectural shield) and High-Entropy Routing (QK-Gain 4.0)—are highly novel. We are
requesting the compute grant to debug the Int6 outlier clamping and build the Latent-PABU CUDA kernel,
clearing our path to the top leaderboard spot.