Research arm 01
Compression
Reports on pipeline-parallel activation compression, weight compression, and related methods for reducing communication, memory use, and serving cost.
Initial Activation Boundary Quantization
Quantizing only the tensors that cross pipeline-parallel stage boundaries, leaving weights, KV cache and compute untouched. Six wire formats measured at three pipeline depths on a 35B mixture of experts model and a 122B model with FP8 weights, scored by teacher-forced WikiText NLL, GSM8K accuracy, and a 1,024 step decode-position probe. Covers which 8-bit format holds once the model is large, why per-token INT4 collapses while microscaled 4-bit survives, and why the scale format rather than the block size separates MXFP4 from NVFP4.
2026-08-24 · Literature reviewActivation Compression at LLM Inference
An interactive, grouped map of 116 cross-checked entries: wire and communication compression, weight and activation quantization, rotations, sparsity and low-rank methods, modern MoE and hybrid architectures (DeepSeek, Kimi, Qwen), hardware datatype support from RTX 4090 to NVL72, SGLang and vLLM framework flags, and the off-the-shelf toolchain. Filter by what is compressed, deployment cost, or entry type.
Source boundary. Current implementation work and evidence live in the separate SGLang repository. This index holds only self-contained published reports.