CublasOps: High-Performance Linear Layers with cuBLAS and cuBLASLt

CublasOps is a PyTorch extension library that provides high-performance linear layers for half-precision (FP16) matrix multiplications using NVIDIA's cuBLAS and cuBLASLt libraries. It offers fast and efficient execution of A x B^T matrix multiplications with optional bias addition and activation functions (ReLU or GELU).

Features

• Fast half-precision (FP16) matrix multiplications using cuBLAS and cuBLASLt with half precision accumulation (2x speedup on 4090).
• Support for fused operations: matmul + bias + activation (ReLU or GELU)
• Easy-to-use linear layers: CublasLinear, CublasLinearRelu, and CublasLinearGelu
• Seamless integration with PyTorch models
• Batched and non-batched operations

For example: using the cublas linear with 4096, 4096 in/out features, and a (2, 4096, 4096) input tensor on my RTX 4090:

CUBLAS INFERENCE: 
FLOPS: 274877906944
TFLOP/s: 305.801

TORCH INFERENCE: 
FLOPS: 274877906944
TFLOP/s: 166.989

Installation

To install CublasOps, follow these steps:

1. Make sure you have PyTorch installed with CUDA support.
2. Clone the repository:

   git clone https://github.com/aredden/torch-cublas-hgemm.git

3. Navigate to the cloned repository:

   cd torch-cublas-hgemm

4. Build and install the extension:

   python -m pip install -U -v .

Usage

Here's a simple example of how to use CublasOps in your PyTorch code:

import torch
from cublas_ops import CublasLinear, CublasLinearGelu, CublasLinearRelu

in_features = 64
out_features = 64
bias = True  # or False

# (A x B^T + bias)
linear = CublasLinear(in_features, out_features, bias=bias, device='cuda', dtype=torch.float16)
input_tensor = torch.randn((2, 8, 64)).cuda().half()
# or...
input_tensor = torch.randn((8, 64)).cuda().half()

output_tensor = linear(input_tensor)

# For fused GELU: gelu(A x B^T + bias)
linear_gelu = CublasLinearGelu(in_features, out_features, bias=bias, device='cuda', dtype=torch.float16)

# For fused ReLU: relu(A x B^T + bias)
linear_relu = CublasLinearRelu(in_features, out_features, bias=bias, device='cuda', dtype=torch.float16)

API Reference

Linear Layers

• CublasLinear(in_features, out_features, bias=True, device=None, dtype=torch.float16, epilogue_str="NONE")
  • A linear layer that performs A x B^T + bias matrix multiplication with optional bias addition.
• CublasLinearGelu(in_features, out_features, bias=True, device=None, dtype=torch.float16)
  • A linear layer with fused GELU activation: gelu(A x B^T + bias).
• CublasLinearRelu(in_features, out_features, bias=True, device=None, dtype=torch.float16)
  • A linear layer with fused ReLU activation: relu(A x B^T + bias).

Low-Level Functions

• cublas_half_matmul_simple(a: torch.Tensor, b: torch.Tensor)
  • Performs a simple A x B^T matrix multiplication using cuBLAS.
• cublas_half_matmul_batched_simple(a: torch.Tensor, b: torch.Tensor)
  • Performs a batched A x B^T batched matrix multiplication using cuBLAS. At least one of A/B should have 3 dimensions, with the other having 2 or 3.
• cublaslt_fused_half_matmul_simple(a: torch.Tensor, b: torch.Tensor, bias: Optional[torch.Tensor] = None, epilogue_str: Optional[Literal["NONE", "RELU", "GELU"]] = "NONE")
  • Performs a fused optional_activation(A x B^T + optional(bias)) matrix multiplication with optional bias addition and activation using cuBLASLt.
• def cublaslt_fused_half_matmul_batched_simple(a: torch.Tensor, b: torch.Tensor, bias: Optional[torch.Tensor] = None, epilogue_str: Optional[Literal["NONE", "RELU", "GELU"]] = "NONE")
  • Performs a fused optional_activation(A x B^T + optional(bias)) batched matrix multiplication with optional bias addition and activation using cuBLASLt. At least one of A/B should have 3 dimensions, with the other having 2 or 3.

Contributing

Contributions to CublasOps are welcome! If you encounter any issues, have suggestions for improvements, or want to add new features, please open an issue or submit a pull request on the GitHub repository.

Acknowledgments

CublasOps is built upon the powerful cuBLAS and cuBLASLt libraries provided by NVIDIA. We would like to thank the NVIDIA team for their excellent work on these libraries.