llamafu_image_process allocates malloc(n_embd * sizeof(float)) and then passes
that buffer to clip_image_batch_encode, which writes n_output_tokens × n_embd
floats — for a typical vision tower that's hundreds of times the allocation.
cheers, thanks for the initial!
llamafu_image_process allocates malloc(n_embd * sizeof(float)) and then passes
that buffer to clip_image_batch_encode, which writes n_output_tokens × n_embd
floats — for a typical vision tower that's hundreds of times the allocation.
cheers, thanks for the initial!