0.0.3

flash attention module

0.0.2

Changelog

1. Flash Multi-head Attention Integration

   • Initially, the code was using torch.nn.MultiheadAttention. It has been changed to use FlashMultiHeadAttention from the flash_attn library. This change is expected to improve the efficiency of the model's attention mechanism.

2. GPU Support

   • All computations are now done on a specified GPU device, defined at the beginning of the script. This significantly improves the model's speed and efficiency.

3. Use of 16-bit Floating Point Precision

   • The data type for computations was changed from default (32-bit floating-point) to 16-bit floating-point (torch.float16). This reduces memory usage and improves the speed of computations on modern GPUs.

4. Added Dropout

   • A dropout layer has been added after the attention operation in the DilatedAttention class. Dropout is a regularization technique that prevents overfitting by randomly setting a fraction of input units to 0 at each update during training time.

   The changes were made in the following lines of code:

   # Initialize dropout layer in the constructor
   self.dropout = nn.Dropout(dropout)
   
   # Apply dropout after performing attention in the forward function
   attn_output = self.dropout(attn_output)

5. Added Unit Tests and Benchmarks

   • Unit tests and benchmarking code have been added to ensure the correctness and efficiency of the DilatedAttention class.

6. Documentation and Example Updates

   • Updated the documentation and usage examples for the DilatedAttention class to reflect the above changes.

7. Twitter Thread

   • Created a Twitter thread in the style of Richard Feynman to promote the project and its new updates.

0.0.1

workflow