[Python] How to implement SwiGLU activation? Why does SwiGLU takes in two tensors?

The SwiGLU variant introduced in https://arxiv.org/pdf/2002.05202 is simply "divine benevolence" and the implementation on Flash-Attention just works out of the box https://github.com/Dao-AILab/flash-attention/tree/main

Comparing the the implementation:

import torch


swiglu_fwd_codestring = """
template <typename T> T swiglu_fwd(T x, T y) {
return float(x) * float(y) / (1.0f + ::exp(-float(x)));
}
"""

swiglu_bwd_codestring = """
template <typename T> T swiglu_bwd(T x, T y, T g, T& dx, T& dy) {
float x_sigmoid = 1.0f / (1.0f + ::exp(-float(x)));
dx = x_sigmoid * (1 + float(x) * (1.0f - x_sigmoid)) * float(g) * float(y);
dy = float(x) * x_sigmoid * float(g);
}
"""
swiglu_fwd = torch.cuda.jiterator._create_jit_fn(swiglu_fwd_codestring)
swiglu_bwd = torch.cuda.jiterator._create_multi_output_jit_fn(swiglu_bwd_codestring, num_outputs=2)


class SwiGLUFunction(torch.autograd.Function):

@staticmethod
def forward(ctx, x, y):
ctx.save_for_backward(x, y)
return swiglu_fwd(x, y)

@staticmethod
def backward(ctx, dout):
x, y = ctx.saved_tensors
return swiglu_bwd(x, y, dout)

swiglu = SwiGLUFunction.apply

swiglu(torch.tensor(1.0).to('cuda'), torch.tensor(0.5).to('cuda'))


and

import torch.nn.functional as F
swiglu_torch = lambda x, y: F.silu(x) * y

swiglu_torch(torch.tensor(1.0).to('cuda'), torch.tensor(0.5).to('cuda'))


They both give the same output:

tensor(0.3655, device='cuda:0')


The paper description, the input to the activation is a single x tensor,



but both the implementation is taking x and y, what does the y corresponds to when we only have the x tensor as in the input in the formula? Is that x = W1.x and y=W2.x?

Continue reading...