[](https://gleam.run/) [](./test) [](./LICENSE) **The fastest tensor library on the BEAM**

Performance

GPU Tensor Cores (RTX 4090)

GPU 2:4 Structured Sparsity

CPU (Intel MKL)

Xeon 24-core (AVX2), MKL dgemm FP64, compact affinity, MADV_HUGEPAGE. All numbers verified with CUDA events and IQR outlier removal.

Install

gleam add viva_tensor

Architecture

graph TB
    subgraph "Gleam Layer (44 modules, 67K lines)"
        A[viva_tensor API]
        B[core/ - tensor, ops, shape, ffi]
        C[quant/ - INT8, NF4, AWQ]
        D[nn/ - autograd, layers, flash_attention]
    end

    subgraph "Erlang Layer"
        E[viva_tensor_zig.erl - NIF wrapper]
    end

    subgraph "Native Layer (13K+ lines C/CUDA)"
        F[nif_entry.c - dispatch]
        G[nif_cpu_ops.c - AVX2 SIMD]
        H[nif_cuda_fp32/fp16/int8.c - Tensor Cores]
        I[nif_sparse.c - 2:4 sparsity]
        J[nif_specialized.c - fused GEMM]
    end

    subgraph "Backend Libraries"
        K[Intel MKL]
        L[CUDA cuBLAS/cuBLASLt]
        M[cuSPARSELt]
        N[CUTLASS]
    end

    A --> B & C & D
    B --> E
    E --> F
    F --> G & H & I & J
    G --> K
    H --> L
    I --> M & N
    J --> L

    style A fill:#FFAFF3
    style K fill:#0071C5,color:#fff
    style L fill:#76B900,color:#fff
    style M fill:#76B900,color:#fff
    style N fill:#76B900,color:#fff

Quick Start

import viva_tensor as t

// Create tensors
let a = t.zeros([1000, 1000])
let b = t.random_uniform([1000, 1000])

// Matrix multiplication (auto-selects best backend)
let c = t.matmul(a, b)

// Activations
let activated = t.relu(c) |> t.sigmoid()

Features

mindmap
  root((viva_tensor))
    Core Ops
      add/sub/mul/div
      matmul/transpose
      sum/mean/max/min
      dot/outer/broadcast
    GPU Backends
      FP32/TF32 cuBLAS
      FP16 Tensor Cores
      INT8 IMMA
      FP8 E4M3 CUTLASS
    Sparsity
      INT4 2:4 CUTLASS
      INT8 2:4 cuSPARSELt
      FP8/FP16 Sparse
    Quantization
      INT8 4x compress
      NF4 7.5x compress
      AWQ 7.7x compress
    Neural Networks
      autograd
      linear layers
      flash attention
      fused GEMM+act
    CNN
      conv2d
      max/avg pool2d
      global_avg_pool2d

Quantization

Build

# Pure Gleam (no native deps)
make build && make test

# With NIF acceleration (Intel MKL + CUDA)
make zig && make build

# Full build
make build-all

Requirements

• Gleam 1.14.0+
• OTP 27+
• Zig 0.14+ (for NIF build)
• Intel MKL (CPU BLAS)
• CUDA 13+ with cuBLAS, cuBLASLt (GPU)
• cuSPARSELt 0.8.1+ (sparse ops)
• CUTLASS 4.3+ (FP8, INT4 sparse)

GPU Benchmark Suite

# Individual benchmarks (Erlang escripts)
./bench/bench_gpu_peak.erl       # FP32/TF32
./bench/bench_fp16_imma.erl      # FP16 Tensor Cores
./bench/bench_int8_imma.erl      # INT8 IMMA
./bench/bench_fp8_peak.erl       # FP8 E4M3
./bench/bench_sparse_peak.erl    # 2:4 Sparsity
./bench/bench_fused_peak.erl     # Fused GEMM+activation
./bench/bench_batched_peak.erl   # Batched GEMM

```mermaid flowchart LR G[Gleam] --> Z[Zig NIF] --> M[Intel MKL] Z --> C[CUDA Tensor Cores] Z --> S[cuSPARSELt] Z --> CU[CUTLASS] style G fill:#FFAFF3,color:#000 style Z fill:#F7A41D,color:#000 style M fill:#0071C5,color:#fff style C fill:#76B900,color:#fff style S fill:#76B900,color:#fff style CU fill:#76B900,color:#fff ``` **Built with love for the BEAM**