tropical-gemm

Fast tropical matrix multiplication with automatic differentiation support.

Installation

# From PyPI
pip install tropical-gemm

# With PyTorch support (for automatic differentiation)
pip install tropical-gemm[torch]

# For GPU support (requires CUDA toolkit)
pip install maturin
git clone https://github.com/TensorBFS/tropical-gemm
cd tropical-gemm/crates/tropical-gemm-python
maturin develop --features cuda

Quick Start

import numpy as np
import tropical_gemm

# Create matrices
a = np.array([[1.0, 2.0, 3.0],
              [4.0, 5.0, 6.0]], dtype=np.float32)
b = np.array([[1.0, 2.0],
              [3.0, 4.0],
              [5.0, 6.0]], dtype=np.float32)

# MaxPlus tropical matmul: C[i,j] = max_k(A[i,k] + B[k,j])
c = tropical_gemm.maxplus_matmul(a, b)
print("MaxPlus result:", c)

# MinPlus tropical matmul: C[i,j] = min_k(A[i,k] + B[k,j])
c = tropical_gemm.minplus_matmul(a, b)
print("MinPlus result:", c)

# MaxMul tropical matmul: C[i,j] = max_k(A[i,k] * B[k,j])
c = tropical_gemm.maxmul_matmul(a, b)
print("MaxMul result:", c)

# With argmax for backpropagation
c, argmax = tropical_gemm.maxplus_matmul_with_argmax(a, b)
print("Result:", c)
print("Argmax:", argmax)

# GPU acceleration (if compiled with CUDA)
if tropical_gemm.cuda_available():
    c = tropical_gemm.maxplus_matmul_gpu(a, b)
    c = tropical_gemm.minplus_matmul_gpu(a, b)
    c = tropical_gemm.maxmul_matmul_gpu(a, b)

PyTorch Integration

The package includes a pytorch submodule with pre-built autograd functions:

import torch
from tropical_gemm.pytorch import (
    # CPU operations
    tropical_maxplus_matmul,
    tropical_minplus_matmul,
    tropical_maxmul_matmul,
    # GPU operations (requires CUDA)
    tropical_maxplus_matmul_gpu,
    tropical_minplus_matmul_gpu,
    tropical_maxmul_matmul_gpu,
    GPU_AVAILABLE,
)

# Create tensors with gradient tracking
a = torch.randn(100, 50, requires_grad=True)
b = torch.randn(50, 80, requires_grad=True)

# Forward pass
c = tropical_maxplus_matmul(a, b)

# Backward pass - gradients computed automatically
loss = c.sum()
loss.backward()

print(f"grad_a shape: {a.grad.shape}")  # (100, 50)
print(f"grad_b shape: {b.grad.shape}")  # (50, 80)

# Use GPU for larger matrices
if GPU_AVAILABLE:
    c = tropical_maxplus_matmul_gpu(a, b)

Gradient Semantics

The gradient computation depends on the semiring type:

MaxPlus/MinPlus (additive rule):

• grad_A[i,k] = grad_C[i,j] if k == argmax[i,j]
• grad_B[k,j] = grad_C[i,j] if k == argmax[i,j]

MaxMul (multiplicative rule):

• grad_A[i,k] = grad_C[i,j] * B[k,j] if k == argmax[i,j]
• grad_B[k,j] = grad_C[i,j] * A[i,k] if k == argmax[i,j]

API Reference

Core Functions

GPU Functions (requires CUDA)

Data Types

All functions support:

• f32 (default): maxplus_matmul, etc.
• f64: maxplus_matmul_f64, etc.
• i32: maxplus_matmul_i32, etc.
• i64: maxplus_matmul_i64, etc.

License

MIT