How can I do equivalent of torch.chunk in ggml tensor ?

[mnlife]

I want using chatglm3-6b's model in llama.cpp, how can I do equivalent of torch.chunk in ggml tensor ?


# Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
def swiglu(x):
    x = torch.chunk(x, 2, dim=-1)
    return F.silu(x[0]) * x[1]

this code from https://hf-mirror.com/THUDM/chatglm3-6b/blob/main/modeling_chatglm.py

[balisujohn]

You can accomplish this using ggml_view_1d

ggml_tensor * conditioning_scale = ggml_view_1d(ctx0, model.diffusion_conditioning_latent, 1024, 0); 
ggml_tensor * conditioning_shift = ggml_view_1d(ctx0, model.diffusion_conditioning_latent, 1024, ggml_element_size(model.diffusion_conditioning_latent) * 1024);

For example splitting a tensor with 2048 elements in half.

The first int argument is the number of elements, the second argument is the byte offset.