| 1179 | } |
| 1180 | |
| 1181 | void common_batch_add( |
| 1182 | struct llama_batch & batch, |
| 1183 | llama_token id, |
| 1184 | llama_pos pos, |
| 1185 | const std::vector<llama_seq_id> & seq_ids, |
| 1186 | bool logits) { |
| 1187 | GGML_ASSERT(batch.seq_id[batch.n_tokens] && "llama_batch size exceeded"); |
| 1188 | |
| 1189 | batch.token [batch.n_tokens] = id; |
| 1190 | batch.pos [batch.n_tokens] = pos; |
| 1191 | batch.n_seq_id[batch.n_tokens] = seq_ids.size(); |
| 1192 | for (size_t i = 0; i < seq_ids.size(); ++i) { |
| 1193 | batch.seq_id[batch.n_tokens][i] = seq_ids[i]; |
| 1194 | } |
| 1195 | batch.logits [batch.n_tokens] = logits; |
| 1196 | |
| 1197 | batch.n_tokens++; |
| 1198 | } |
| 1199 | |
| 1200 | // |
| 1201 | // Token utils |