llama : add n_enc_output field in llm_build_context containing the nu…

…mber of embeddings generated by the encoder
fairydreaming · Jun 20, 2024 · eb4c17e · eb4c17e
1 parent 684160a
commit eb4c17e
Showing 1 changed file with 7 additions and 6 deletions.
diff --git a/llama.cpp b/llama.cpp
@@ -7347,6 +7347,7 @@ struct llm_build_context {
     const int32_t n_tokens;
     const int32_t n_kv;     // size of KV cache to consider (n_kv <= kv_self.size)
     const int32_t n_outputs;
+    const int32_t n_enc_outputs;
     const int32_t kv_head;  // index of where we store new KV data in the cache
     const int32_t n_ctx_orig;
 
@@ -7396,6 +7397,7 @@ struct llm_build_context {
         n_tokens         (batch.n_tokens),
         n_kv             (worst_case ? kv_self.size : kv_self.n),
         n_outputs        (worst_case ? n_tokens : lctx.n_outputs),
+        n_enc_outputs    (worst_case ? n_tokens : lctx.encoder_output.size() / hparams.n_embd),
         kv_head          (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head),
         n_ctx_orig       (cparams.n_ctx_orig_yarn),
         flash_attn       (cparams.flash_attn),
@@ -7660,14 +7662,14 @@ struct llm_build_context {
 
     struct ggml_tensor * llm_build_inp_enc_output() {
         const int64_t n_embd = hparams.n_embd;
-        lctx.inp_enc_output = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, lctx.encoder_output.size() == 0 ? 512 : lctx.encoder_output.size() / n_embd);
+        lctx.inp_enc_output = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc_outputs);
         ggml_set_input(lctx.inp_enc_output);
         cb(lctx.inp_enc_output, "enc_output", -1);
         return lctx.inp_enc_output;
     }
 
     struct ggml_tensor * llm_build_inp_cross_KQ_mask() {
-        lctx.inp_cross_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, lctx.encoder_output.size() == 0 ? 512 : lctx.encoder_output.size() / n_embd, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        lctx.inp_cross_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_enc_outputs, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
         ggml_set_input(lctx.inp_cross_KQ_mask);
         cb(lctx.inp_cross_KQ_mask, "enc_mask", -1);
         return lctx.inp_cross_KQ_mask;
@@ -11717,7 +11719,6 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        const int32_t n_enc_output = lctx.encoder_output.size() == 0 ? 512 : lctx.encoder_output.size() / n_embd;
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -11926,7 +11927,7 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
 
                     Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_enc_output);
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_enc_outputs);
 
                     struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
                     struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
@@ -11937,10 +11938,10 @@ struct llm_build_context {
                     kq = ggml_soft_max_ext(ctx0, kq, enc_KQ_mask, 1.0f, hparams.f_max_alibi_bias);
                     cb(kq, "kq_soft_max_ext", il);
 
-                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_enc_output)));
+                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_enc_outputs)));
                     cb(v, "v", il);
 
-                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_enc_output, n_embd_head, n_head_kv), kq);
+                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_enc_outputs, n_embd_head, n_head_kv), kq);
                     cb(kqv, "kqv", il);
 
                     struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);