diff --git a/ggml-metal.metal b/ggml-metal.metal index f45b1490f5b49..7f1a148689fff 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -118,7 +118,7 @@ kernel void kernel_soft_max( device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; // parallel max - float lmax = psrc0[tpitg[0]]; + float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY; for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) { lmax = MAX(lmax, psrc0[i00]); } @@ -158,7 +158,7 @@ kernel void kernel_soft_max_4( device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); // parallel max - float4 lmax4 = psrc4[tpitg[0]]; + float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY; for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) { lmax4 = fmax(lmax4, psrc4[i00]); } diff --git a/llama.cpp b/llama.cpp index 7776211bd8d43..21eebfbd755d6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3446,7 +3446,9 @@ static struct ggml_cgraph * llm_build_starcoder( const int64_t n_layer = hparams.n_layer; const int64_t n_ctx = hparams.n_ctx; const int64_t n_head = hparams.n_head; + const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head = hparams.n_embd_head(); + const int64_t n_embd_gqa = hparams.n_embd_gqa(); GGML_ASSERT(n_embd_head == hparams.n_rot); @@ -3508,28 +3510,44 @@ static struct ggml_cgraph * llm_build_starcoder( position = ggml_get_rows(ctx0, model.pos_embeddings, inp_positions); } + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + ggml_allocr_alloc(lctx.alloc, KQ_scale); + if (!ggml_allocr_is_measure(lctx.alloc)) { + ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); + } + ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); + inpL = ggml_add(ctx0, token, position); + ggml_set_name(inpL, "inpL"); for (int il = 0; il < n_layer; ++il) { { // Norm cur = ggml_norm(ctx0, inpL, norm_eps); cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b); - } { // Self Attention cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv); - struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd); - struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd); - struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd); + struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd); + struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd); + struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd); - // store key and value to memory - if (N >= 1) { - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); - struct ggml_tensor * v = ggml_view_1d(ctx0, kv_self.v, N*n_embd, (ggml_element_size(kv_self.v)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * Qcur = tmpq; + struct ggml_tensor * Kcur = tmpk; + + { + struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, N)); + ggml_set_name(Vcur, "Vcur"); + + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past)); + ggml_set_name(k, "k"); + + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v)); ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); @@ -3541,56 +3559,62 @@ static struct ggml_cgraph * llm_build_starcoder( Qcur, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)), 0, 2, 1, 3); + ggml_set_name(Q, "Q"); struct ggml_tensor * K = - ggml_permute(ctx0, - ggml_reshape_3d(ctx0, - 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); //TODO: need to be tiled + ggml_view_3d(ctx0, kv_self.k, + n_embd_head, n_past + N, n_head_kv, + ggml_element_size(kv_self.k)*n_embd_gqa, + ggml_element_size(kv_self.k)*n_embd_head, + ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); + ggml_set_name(K, "K"); // K * Q - // [n_past + N, N, 12] struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + ggml_set_name(KQ, "KQ"); - // KQ_scaled = KQ / sqrt(n_embd/n_head) - // [n_past + N, N, 12] - struct ggml_tensor * KQ_scaled = - ggml_scale_inplace(ctx0, - KQ, - ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)) - ); + // KQ_scaled = KQ / sqrt(n_embd_head) + // KQ_scaled shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); + ggml_set_name(KQ_scaled, "KQ_scaled"); // KQ_masked = mask_past(KQ_scaled) - // [n_past + N, N, 12] struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); + ggml_set_name(KQ_masked, "KQ_masked"); // KQ = soft_max(KQ_masked) - // [n_past + N, N, 12] struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); + ggml_set_name(KQ_soft_max, "KQ_soft_max"); - // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous() - // [n_past + N, 64, 12] - struct ggml_tensor * V_trans = - ggml_cpy(ctx0, - ggml_permute(ctx0, - ggml_reshape_3d(ctx0, - ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.v)*n_embd), - n_embd/n_head, n_head, n_past + N), - 1, 2, 0, 3), - ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head)); - - // KQV = transpose(V) * KQ_soft_max - // [64, N, 12] - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max); + // split cached V into n_head heads + struct ggml_tensor * V = + ggml_view_3d(ctx0, kv_self.v, + n_past + N, n_embd_head, n_head_kv, + ggml_element_size(kv_self.v)*n_ctx, + ggml_element_size(kv_self.v)*n_ctx*n_embd_head, + ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); + 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 + // is there a better way? + struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd_head, n_head)); + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); +#endif // KQV_merged = KQV.permute(0, 2, 1, 3) - // [64, 12, N] 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