root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-31 22:04:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

llama : aboud ggml_repeat during classification

Browse Source
This commit is contained in:
Georgi Gerganov 2024-09-23 20:20:38 +03:00
parent 6235c62ac9
commit 6916ed1606
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 14 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
src/llama.cpp
View File

@ -10243,9 +10243,6 @@ struct llm_build_context {
cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls, inp), model.cls_b);
cur = ggml_tanh(ctx0, cur);
cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls_out, cur), model.cls_out_b);
// broadcast across the embedding size to make it compatible with the llama_get_embeddings API
cur = ggml_repeat(ctx0, cur, inp);
} break;
default:
{
@ -16997,7 +16994,6 @@ static int llama_decode_internal(
case LLAMA_POOLING_TYPE_MEAN:
case LLAMA_POOLING_TYPE_CLS:
case LLAMA_POOLING_TYPE_LAST:
case LLAMA_POOLING_TYPE_RANK:
{
// extract sequence embeddings (cleared before processing each batch)
auto & embd_seq_out = lctx.embd_seq;
@ -17011,6 +17007,20 @@ static int llama_decode_internal(
ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
}
} break;
case LLAMA_POOLING_TYPE_RANK:
{
// extract the rank score - a single float per sequence
auto & embd_seq_out = lctx.embd_seq;
for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
const llama_seq_id seq_id = ubatch.seq_id[s][0];
if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
continue;
}
embd_seq_out[seq_id].resize(1);
ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float));
}
} break;
case LLAMA_POOLING_TYPE_UNSPECIFIED:
{
GGML_ABORT("unknown pooling type");