diff options
Diffstat (limited to 'llama.cpp')
| -rw-r--r-- | llama.cpp | 24 |
1 files changed, 20 insertions, 4 deletions
@@ -659,6 +659,7 @@ struct llama_model_loader { LLAMA_ASSERT(lt.ne.size() == 1); tensor = ggml_new_tensor_1d(ggml_ctx, lt.type, lt.ne.at(0)); } + ggml_set_name(tensor, lt.name.c_str()); LLAMA_ASSERT(lt.ggml_tensor == NULL); // if this fails, we called get_tensor twice on the same tensor lt.ggml_tensor = tensor; num_ggml_tensors_created++; @@ -798,6 +799,8 @@ static bool kv_cache_init( cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements); cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements); + ggml_set_name(cache.k, "cache_k"); + ggml_set_name(cache.v, "cache_v"); return true; } @@ -1084,6 +1087,7 @@ static bool llama_eval_internal( gf.n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads; struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + ggml_set_name(embd, "embd"); memcpy(embd->data, tokens, N*ggml_element_size(embd)); struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd); @@ -1110,6 +1114,8 @@ static bool llama_eval_internal( // compute Q and K and RoPE them struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); + ggml_set_name(Qcur, "Qcur"); + ggml_set_name(Kcur, "Kcur"); // store key and value to memory { @@ -1130,6 +1136,7 @@ static bool llama_eval_internal( ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + ggml_set_name(Q, "Q"); struct ggml_tensor * K = ggml_permute(ctx0, @@ -1137,21 +1144,26 @@ static bool llama_eval_internal( ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd), n_embd/n_head, n_head, n_past + N), 0, 2, 1, 3); + ggml_set_name(K, "K"); // K * Q struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + ggml_set_name(KQ, "KQ"); // KQ_scaled = KQ / sqrt(n_embd/n_head) - struct ggml_tensor * KQ_scaled = - ggml_scale(ctx0, - KQ, - ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); + struct ggml_tensor * KQ_scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)); + ggml_set_name(KQ_scale, "1/sqrt(n_embd/n_head)"); + + struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); + ggml_set_name(KQ_scaled, "KQ_scaled"); // KQ_masked = mask_past(KQ_scaled) struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + ggml_set_name(KQ_masked, "KQ_masked"); // KQ = soft_max(KQ_masked) struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); + ggml_set_name(KQ_soft_max, "KQ_soft_max"); // split cached V into n_head heads struct ggml_tensor * V = @@ -1160,9 +1172,11 @@ static bool llama_eval_internal( n_ctx*ggml_element_size(kv_self.v), n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head, il*n_ctx*ggml_element_size(kv_self.v)*n_embd); + ggml_set_name(V, "V"); #if 1 struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + ggml_set_name(KQV, "KQV"); #else // make V contiguous in memory to speed up the matmul, however we waste time on the copy // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation @@ -1173,11 +1187,13 @@ static bool llama_eval_internal( // KQV_merged = KQV.permute(0, 2, 1, 3) struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + ggml_set_name(KQV_merged, "KQV_merged"); // cur = KQV_merged.contiguous().view(n_embd, N) cur = ggml_cpy(ctx0, KQV_merged, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + ggml_set_name(cur, "KQV_merged_contiguous"); // projection (no bias) cur = ggml_mul_mat(ctx0, |