Publications & Research Library

Definitive retrieval-systems teardown of ColBERTv2 (Santhanam et al., arXiv:2112.01488): bridging single-vector bi-encoder compression loss and cross-encoder latency via token-level MaxSim late interaction, centroid-residual b-bit quantization, and PLAID inverted-list pruning.

Definitive mathematical and distributed-systems teardown of DeepSeekMath (arXiv:2402.03300): eliminating the PPO value critic via Group Relative Policy Optimization (GRPO), intra-group advantage normalization, unbiased per-token KL divergence estimation, and outcome vs. process supervision.

Definitive mathematical and systems teardown of QLoRA (Dettmers et al., arXiv:2305.14314): information-theoretically optimal 4-bit NormalFloat (NF4) quantile binning, blockwise k-bit quantization, FP8 Double Quantization of scale constants, and CUDA Unified Memory Paged Optimizers.

Definitive distributed-systems and arithmetic-intensity teardown of Ring Attention (Liu et al., arXiv:2310.01889): overcoming single-GPU HBM limits via ring-topology P2P KV block circulation, overlapped communication and FlashAttention compute, and exact online softmax rescaling.

Definitive mathematical and kernel-level teardown of RoFormer (Su et al., arXiv:2104.09864): encoding relative token distances via multiplicative 2D complex-plane rotations, sparse element-wise slice kernels, long-term decay bounds, and NTK-aware / YaRN context extension.

Definitive mathematical and systems teardown of Speculative Decoding (Leviathan et al., arXiv:2211.17192): overcoming the memory-bandwidth roofline via small draft speculation, parallel target verification, lossless modified rejection sampling, and residual distribution renormalization.

Technical paper teardown of Attention Is All You Need (Vaswani et al., NIPS 2017) detailing Scaled Dot-Product Attention math, Multi-Head Attention projections, Sinusoidal Positional Encoding, and KV cache memory bounds.

Definitive technical paper breakdown of DeepSeek-V3 and R1 detailing Multi-Head Latent Attention (MLA) low-rank KV compression, auxiliary-loss-free MoE load balancing, and DualPipe pipeline parallelism.

Definitive paper teardown of FlashAttention-3 detailing producer-consumer warp specialization, asynchronous TMA memory loads, FP8 GEMM MMA execution, and inter-warp communication on Hopper GPUs.

Definitive paper teardown of vLLM's PagedAttention architecture detailing virtual memory block translation, dynamic copy-on-write sequence forks, and prefix caching.

Predictive processing offers a useful account of hierarchical inference and error correction, but it is not a shortcut from brain metaphor to system architecture.