subquadratic-ops-torch-cu13 0.1.1

subquadratic-ops-torch-cu13 - GPU Accelerated Torch Extensions for Subquadratic Operations

subquadratic_ops_torch

Introduction

subquadratic_ops_torch provides CUDA kernels for subquadratic operations e.g. long and short convolution. It contains pytorch bindings to optimized kernels.

Installation

Please install using pip install subquadratic-ops-torch-cu[12,13]

Documentation

For detailed usage information of the kernels, please refer to the doc-strings in their respective function.

Usage

You can import the library from python:

import subquadratic_ops_torch as subq

Kernels are primarily exposed as function call underlying torch.ops, which also provide a lower-level interface as torch.library operators. This allows you to export models using this operations using torch.export, and running inference on them using TensorRT.

Support and Feedback

Please contact the developers for any issues you might encounter.

Requirements

• CUDA-compatible NVIDIA GPU (Ampere+)
• CUDA Toolkit 12.0 or higher
• Python 3.11-13

Modules

B2B CausalConv1d

Operation

Back-to-back causal conv1d for the striped hyena 2 architecture used in Evo2 model. The operation is performed in a causal manner, meaning each position only attends to previous positions in the sequence. In code terms,

in_dim = 8192

width_proj = 8
width_mixer = 128

dtype = torch.float32

class Conv1DModel(nn.Module):
    def __init__(self, in_dim, width, dtype, skip_bias=False):
        super(Conv1DModel, self).__init__()

        self.conv = nn.Conv1d(
            in_dim,
            in_dim,
            width,
            padding=width - 1,
            groups=in_dim,
            bias=False,
            dtype=dtype,
            device="cuda:0",
        )
        self.width = width
        self.weight = self.conv.weight.reshape(-1, width)
        if skip_bias:
            self.skip_bias = nn.Parameter(torch.zeros(in_dim, dtype=dtype, device="cuda:0").reshape(1, -1, 1))
        else:
            self.skip_bias = None

    def forward(self, x):
        seqlen = x.shape[-1]
        out = self.conv(x)
        return out[..., :seqlen]

def model(x, conv1d_proj, conv1d_mixer):
    xv = conv1d_proj(x)
    z = xv[:,1::3, :] * xv[:, 2::3, :]
    y = conv1d_mixer(z) + conv1d_mixer.skip_bias * z
    return y * xv[:, ::3, :]
x = torch.randn(batch_size, 3*in_dim, seq_dim)
conv1d_proj = Conv1DModel(3*in_dim, width_proj)
conv1d_mixer = Conv1DModel(in_dim, width_mixer, True)

y = model(x, conv1d_proj, conv1d_mixer)

is equivalent to,

weight_proj = torch.randn(3*in_dim, width_proj).to(dtype)
weight_mixer = torch.randn(in_dim, width_mixer).to(dtype)
skip_bias = torch.randn(in_dim).to(dtype)

b2b_causal_conv1d(x, weight_proj, weight_mixer, skip_bias)

Supported Kernel Sizes for B2B Causal Conv1d

Kernel Type	Supported Sizes
Projection	2, 3, 4, 8, 16, 32
Mixer	2, 3, 4, 5, 6, 7, 8, 16, 32, 64, 128, 256

CausalConv1d

Causal conv1d : the convolution operation is performed in a causal manner, meaning each position only attends to previous positions in the sequence. In code terms,

in_dim = 8192
width = 8
dtype = torch.float32
model = nn.Conv1d(
            in_dim,
            in_dim,
            width,
            padding=width - 1,
            groups=in_dim,
            bias=False,
            dtype=dtype,
            device="cuda:0",
        )
y = model(x)

is equivalent to,

weight = torch.randn((in_dim, width))
causal_conv1d(x, weight)

Supported Kernel Sizes for Causal Conv1d

Kernel Type	Supported Sizes
CausalConv1d	2, 3, 4, 5, 6, 7, 8, 16, 32, 64, 128, 256

FFT CausalConv1d

FFT Causal Conv1d : the convolution operation is performed in a causal manner, meaning each position only attends to previous positions in the sequence. It uses real FFT and IFFT instead of direct summation for convolution. In code terms,

in_dim = 8192
width = 1024
dtype = torch.float32
weight = torch.randn(1, in_dim, width)
def model(x, w):
    fft_size = x.shape[-1] * 2
    xf = torch.fft.rfft(x, n=fft_size, dim=-1)
    wf = torch.fft.rfft(w, n=fft_size, dim=-1)
    return torch.fft.irfft(xf*wf, n=fft_size, dim=-1)[..., x.shape[-1]]

y = model(x, weight)

is equivalent to,

weight = torch.randn((in_dim, width))
y = fft_causal_conv1d(x, weight)