You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I wanna add some params to the section "message ModelConf{} ", and I notice that the modelConf in lightseq/docs/export_model.md and lightseq/proto/transformer.proto differs:
lightseq/docs/export_model.md
message ModelConf {
int32 head_num = 1; // head number for multi-head attention
int32 beam_size = 2; // beam size of beam search
int32 extra_decode_length =3; // extra decode length compared with source length
float length_penalty = 4; // length penalty of beam search
int32 src_padding_id = 5; // source padding id
int32 trg_start_id = 6; // target start id
}
lightseq/proto/transformer.proto
`message ModelConf {
int32 head_num = 1; // head number for multi-head attention
int32 beam_size = 2; // beam size of beam search
int32 extra_decode_length =
3; // extra decode length compared with source length
float length_penalty = 4; // length penalty of beam search
int32 src_padding_id = 5; // source padding id
int32 trg_start_id = 6; // target start id
float diverse_lambda = 7; // diverse beam search lambda
string sampling_method = 8; // choice of beam_search, topk, topp, topk_greedy
float topp = 9; // parameter for topp sampling
int32 topk = 10; // parameter for topk sampling
int32 trg_end_id = 11; // eos of target embedding
bool is_post_ln = 12; // Pre-LN or Post-LN
bool no_scale_embedding = 13; // whether to scale embedding by sqrt(emb_dim)
bool use_gelu = 14; // use gelu for activation otherwise relu
// Whether it is a multilingual model.
// If it is set to true, lang_emb and trg_vocab_mask should be non-empty.
bool is_multilingual = 15;
}`
So I wonder that is message ModelConf{} in gpt.proto editable? Can you provide a set of addable parameters?
Thanks a lot for your time😊
The text was updated successfully, but these errors were encountered:
@Sueying Sorry about the confusing docs, the former needs to be updated, files in proto folder will always be the latest version. You can add parameter based on it, but remain those already existing parameters and their numbers to keep compatibility
Hi! I'm a newcomer of LightSeq.
I wanna add some params to the section "message ModelConf{} ", and I notice that the modelConf in
lightseq/docs/export_model.md
andlightseq/proto/transformer.proto
differs:So I wonder that is message ModelConf{} in gpt.proto editable? Can you provide a set of addable parameters?
Thanks a lot for your time😊
The text was updated successfully, but these errors were encountered: