diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-09-29 13:11:00 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-09-29 13:11:00 +0200 |
| commit | 37cbaef9ceb828c667da8f398027ba5415ee7475 (patch) | |
| tree | 6103e92a9030766e78a35df00f7a1155d0d0ecf5 /liblangdet/forward.c | |
| parent | cf7d8169ec12cb3adf90d4ff78b3257e74985302 (diff) | |
| download | langdet-37cbaef9ceb828c667da8f398027ba5415ee7475.tar.gz langdet-37cbaef9ceb828c667da8f398027ba5415ee7475.zip | |
Rename CPU library API and sources to langdet
Diffstat (limited to 'liblangdet/forward.c')
| -rw-r--r-- | liblangdet/forward.c | 96 |
1 files changed, 96 insertions, 0 deletions
diff --git a/liblangdet/forward.c b/liblangdet/forward.c new file mode 100644 index 0000000..14b1327 --- /dev/null +++ b/liblangdet/forward.c @@ -0,0 +1,96 @@ +#include "model_internal.h" +#include "kernels.h" + +#include <limits.h> +#include <math.h> +#include <stdio.h> +#include <string.h> + +static void add_bias_rows(float *x, const float *bias, int rows, int dim) { + for (int row = 0; row < rows; row++) { + float *dst = x + (size_t)row * dim; + for (int d = 0; d < dim; d++) + dst[d] += bias[d]; + } +} + +int langdet_model_encode_into(const langdet_model_t *model, + langdet_workspace_t *ws, + const int *token_ids, + int n_tokens, + float *out_embedding) { + if (!model || !ws || ws->model != model || !token_ids || n_tokens <= 0 || !out_embedding) + return -1; + + const langdet_config_t *cfg = &model->config; + const langdet_weights_t *w = &model->weights; + int hidden = cfg->hidden_size; + int n_heads = cfg->n_heads; + int head_dim = cfg->head_dim; + int inter = cfg->intermediate_size; + float eps = cfg->layer_norm_eps; + + if (n_tokens > cfg->max_position_embeddings - cfg->position_id_offset) { + fprintf(stderr, + "langdet: sequence length %d plus position offset %d exceeds max_position_embeddings " + "%d\n", + n_tokens, cfg->position_id_offset, cfg->max_position_embeddings); + return -1; + } + if (n_tokens > INT_MAX / hidden || n_tokens > INT_MAX / inter || + ensure_buffers(ws, cfg, n_tokens) != 0) + return -1; + int offsets[2] = {0, n_tokens}; + ensure_attention_scores(ws, offsets, 1); + + float *x = ws->x; + float *q_buf = ws->q, *k_buf = ws->k, *v_buf = ws->v; + float *attn_out = ws->attn_out, *proj_out = ws->proj_out, *ffn_up = ws->ffn_up; + + for (int i = 0; i < n_tokens; i++) { + int id = token_ids[i]; + if (id < 0 || id >= cfg->vocab_size) { + fprintf(stderr, "langdet: invalid token id %d at position %d\n", id, i); + return -1; + } + float *row = x + (size_t)i * hidden; + memcpy(row, w->embed_tokens + (size_t)id * hidden, (size_t)hidden * sizeof(float)); + int pos = cfg->position_id_offset + i; + add_inplace(row, w->position_embeddings + (size_t)pos * hidden, hidden); + add_inplace(row, w->token_type_embeddings, hidden); + } + layer_norm(x, x, w->langdet_ln_w, w->langdet_ln_b, n_tokens, hidden, eps); + + float scale = 1.0f / sqrtf((float)head_dim); + + for (int layer = 0; layer < cfg->n_layers; layer++) { + const langdet_layer_t *l = &w->layers[layer]; + + linear_nobias(q_buf, x, l->wq, n_tokens, hidden, hidden); + linear_nobias(k_buf, x, l->wk, n_tokens, hidden, hidden); + linear_nobias(v_buf, x, l->wv, n_tokens, hidden, hidden); + add_bias_rows(q_buf, l->q_bias, n_tokens, hidden); + add_bias_rows(k_buf, l->k_bias, n_tokens, hidden); + add_bias_rows(v_buf, l->v_bias, n_tokens, hidden); + + bidirectional_gqa_attention_packed_with_scratch( + attn_out, q_buf, k_buf, v_buf, offsets, 1, n_heads, n_heads, head_dim, scale, + ws->attn_scores, ws->attn_scores_bytes); + + linear_nobias(proj_out, attn_out, l->wo, n_tokens, hidden, hidden); + add_bias_rows(proj_out, l->o_bias, n_tokens, hidden); + add_inplace(x, proj_out, n_tokens * hidden); + layer_norm(x, x, l->input_norm, l->attn_ln_bias, n_tokens, hidden, eps); + + linear_nobias(ffn_up, x, l->up_proj, n_tokens, hidden, inter); + add_bias_rows(ffn_up, l->ffn_inter_bias, n_tokens, inter); + gelu_inplace(ffn_up, n_tokens * inter); + linear_nobias(proj_out, ffn_up, l->down_proj, n_tokens, inter, hidden); + add_bias_rows(proj_out, l->ffn_out_bias, n_tokens, hidden); + add_inplace(x, proj_out, n_tokens * hidden); + layer_norm(x, x, l->post_attn_norm, l->ffn_ln_bias, n_tokens, hidden, eps); + } + /* The classifier consumes <s> after the encoder's final LayerNorm, without L2 normalization. */ + memcpy(out_embedding, x, (size_t)hidden * sizeof(float)); + return 0; +} |
