viva_tensor: Memory Multiplication

Gabriel Maia · VIVA Research · 2026

Abstract

Pure Gleam tensor library achieving 8x memory multiplication through mathematical compression.

graph LR
    A["24 GB Physical"] -->|"NF4"| B["192 GB Effective"]

Problem

LLMs are memory-bound, not compute-bound.

Solution

flowchart TB
    subgraph Quantization
        INT8["INT8: 4x"]
        NF4["NF4: 7.5x"]
        AWQ["AWQ: 7.7x"]
    end

    subgraph Result
        M["Memory × 8"]
    end

    Quantization --> Result

Algorithms

INT8

Linear quantization. Fast, simple.

scale = 127 / max|x|
q = round(x × scale)

NF4 (QLoRA)

16 levels from normal distribution quantiles. Optimal for Gaussian weights.

AWQ (MLSys 2024 Best Paper)

Key insight: 1% of weights are salient — identified by activation magnitude.

flowchart LR
    A[Activations] --> S[Stats]
    S --> T["Top 1% Salient"]
    T --> U["Scale UP"]
    U --> Q[Quantize]

Results

Why Gleam?

graph TB
    subgraph BEAM
        P1[Process 1]
        P2[Process 2]
        P3[Process N]
    end

    subgraph Properties
        I[Immutable]
        F[Fault-tolerant]
        C[Concurrent]
    end

    BEAM --> Properties

References

1. Lin et al. “AWQ” MLSys 2024 Best Paper
2. Dettmers et al. “QLoRA” NeurIPS 2023
3. NVIDIA Blackwell Architecture 2024