Because of the lack of embeddings extraction to create a vector space database, I need to increase the context length up to 32k or more if possible.
I see that some models such as Qwen2 support up to 128k context. How can I increase it.
Should I have to update the rknn_api.h:
define RKNN_MAX_DIMS 32 / increased maximum dimension of tensor. /
define RKNN_MAX_NUM_CHANNEL 20 / increased maximum channel number of input tensor. /
define RKNN_MAX_NAME_LEN 512 / increased maximum name length of tensor. /
define RKNN_MAX_DYNAMIC_SHAPE_NUM 32768 / increased maximum number of dynamic shape for each input. /
or is it enough to set param.max_context_len = 32768; ?
Because of the lack of embeddings extraction to create a vector space database, I need to increase the context length up to 32k or more if possible. I see that some models such as Qwen2 support up to 128k context. How can I increase it.
Should I have to update the rknn_api.h:
define RKNN_MAX_DIMS 32 / increased maximum dimension of tensor. /
define RKNN_MAX_NUM_CHANNEL 20 / increased maximum channel number of input tensor. /
define RKNN_MAX_NAME_LEN 512 / increased maximum name length of tensor. /
define RKNN_MAX_DYNAMIC_SHAPE_NUM 32768 / increased maximum number of dynamic shape for each input. /
or is it enough to set param.max_context_len = 32768; ?
@airockchip @waydong