whispercpp 1.3.3 → 1.3.5

data/ext/sources/examples/talk-llama/models/plamo2.cpp

@@ -0,0 +1,316 @@
+#include "models.h"
+
+llm_build_plamo2::llm_build_plamo2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+    cb(inpL, "embedding_output", -1);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_hybrid = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * residual = inpL;
+
+        // ggml_graph_add_node(gf, model.layers[il].attn_norm);
+        // cb(model.layers[il].attn_norm, "attn_norm", il);
+
+        // pre_mixer_norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+
+        // check if this layer is Mamba or Attention
+        bool is_mamba_layer = hparams.is_recurrent(il);
+
+        if (is_mamba_layer) {
+            // PLaMo-2 Mamba layer
+            cur = build_plamo2_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);
+        } else {
+            // PLaMo-2 Attention layer
+            cur = build_plamo2_attn_layer(inp_hybrid->get_attn(), inp_pos, cur, model, il);
+        }
+
+        // post_mixer_norm
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        // residual connection
+        cur = ggml_add(ctx0, cur, residual);
+        cb(cur, "attn_residual", il);
+        residual = cur;
+
+        // pre-ffn norm
+        cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_pre_norm", il);
+
+        // feed-forward network
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                NULL, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        // post ffn norm
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
+            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+        }
+
+        // residual connection
+        cur = ggml_add(ctx0, cur, residual);
+        cb(cur, "ffn_residual", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // final norm
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+
+    // Explicitly mark as output tensor to ensure proper backend assignment
+    ggml_set_output(cur);
+
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_plamo2::build_plamo2_attn_layer(llm_graph_input_attn_kv * inp,
+                                                        ggml_tensor *             inp_pos,
+                                                        ggml_tensor *             cur,
+                                                        const llama_model &       model,
+                                                        int                       il) {
+    // self-attention
+    {
+        // PLaMo-2 uses combined QKV tensor
+        ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
+        cb(qkv, "wqkv", il);
+
+        // split QKV tensor into Q, K, V
+        const int64_t n_embd_head_q = hparams.n_embd_head_k;
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+        const int64_t n_embd_head_v = hparams.n_embd_head_v;
+        int32_t       n_head        = hparams.n_head(il);
+        int32_t       n_head_kv     = hparams.n_head_kv(il);
+
+        const int64_t q_offset = 0;
+        const int64_t k_offset = n_embd_head_q * n_head;
+        const int64_t v_offset = k_offset + n_embd_head_k * n_head_kv;
+
+        ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head, n_tokens, n_embd_head_q * sizeof(float),
+                                          qkv->nb[1], q_offset * ggml_element_size(qkv));
+        ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float),
+                                          qkv->nb[1], k_offset * ggml_element_size(qkv));
+        ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv, n_tokens, n_embd_head_v * sizeof(float),
+                                          qkv->nb[1], v_offset * ggml_element_size(qkv));
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+        cb(Qcur, "Qcur_normed", il);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+        cb(Kcur, "Kcur_normed", il);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        cur = build_attn(inp,
+            model.layers[il].wo, NULL,
+            Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f / sqrtf(float(n_embd_head_v)), il);
+    }
+
+    cb(cur, "attn_out", il);
+
+    return cur;
+}
+
+ggml_tensor * llm_build_plamo2::build_plamo2_mamba_layer(llm_graph_input_rs * inp,
+                                                         ggml_tensor *        cur,
+                                                         const llama_model &  model,
+                                                         const llama_ubatch & ubatch,
+                                                         int                  il) {
+    const auto * mctx_cur = inp->mctx;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const int64_t d_conv   = hparams.ssm_d_conv;
+    const int64_t d_inner  = hparams.ssm_d_inner;
+    const int64_t d_state  = hparams.ssm_d_state;
+    const int64_t n_heads  = hparams.ssm_dt_rank;
+    const int64_t head_dim = d_inner / n_heads;
+    const int64_t n_group  = hparams.ssm_n_group;
+    const int64_t n_seqs   = ubatch.n_seqs;
+
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+    ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
+    conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs);
+
+    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    // in_proj: {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * zx = build_lora_mm(model.layers[il].ssm_in, cur);
+    cb(zx, "mamba_in_proj", il);
+    // {8192, 5, 1, 1} -> {8192, 1, 5, 1}
+    zx = ggml_permute(ctx0, zx, 0, 2, 1, 3);
+    zx = ggml_cont_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs);
+    cb(zx, "mamba_in_proj_out", il);
+
+    // split into z and x
+    // => {head_dim * n_heads, n_seq_tokens, n_seqs}
+    ggml_tensor * x = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3],
+                                   head_dim * ggml_element_size(zx));
+    x               = ggml_cont_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs);
+    // x = ggml_permute(ctx0, x, 0, 2, 1, 3);
+    cb(x, "mamba_x_split", il);
+
+    ggml_tensor * z =
+        ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], 0);
+    cb(z, "mamba_z_split", il);
+
+    // conv1d
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
+        cb(conv_x, "mamba_conv1d_input", il);
+
+        // copy last (d_conv - 1) columns back into the state cache
+        ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2],
+                                               n_seq_tokens * (conv_x->nb[0]));
+
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv,
+                                               ggml_view_1d(ctx0, conv_states_all,
+                                                            (d_conv - 1) * (d_inner + 2 * n_group * d_state) * (n_seqs),
+                                                            kv_head * (d_conv - 1) * (d_inner + 2 * n_group * d_state) *
+                                                                ggml_element_size(conv_states_all))));
+        cb(conv_states_all, "mamba_conv1d_state", il);
+
+        // 1D convolution
+        x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
+        cb(x, "mamba_conv1d", il);
+
+        x = ggml_silu(ctx0, x);
+        cb(x, "mamba_conv1d_silu", il);
+    }
+
+    // SSM
+    {
+        // bcdt_proj: {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}
+        ggml_tensor * x_bcdt = build_lora_mm(model.layers[il].ssm_x, x);
+        cb(x_bcdt, "mamba_bcdt_proj", il);
+
+        // split into dt, B, C
+        const int64_t dt_dim = std::max(64, int(hparams.n_embd / 16));
+        ggml_tensor * B  = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], 0);
+        ggml_tensor * C  = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2],
+                                        ggml_element_size(x_bcdt) * d_state);
+        ggml_tensor * dt = ggml_view_3d(ctx0, x_bcdt, dt_dim, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2],
+                                        ggml_element_size(x_bcdt) * (2 * d_state));
+        cb(B, "mamba_B_raw", il);
+        cb(C, "mamba_C_raw", il);
+        cb(dt, "mamba_dt_raw", il);
+
+        // Apply RMS norm to dt, B, C (PLaMo-2 specific)
+        B  = build_norm(B, model.layers[il].ssm_b_norm, NULL, LLM_NORM_RMS, il);
+        C  = build_norm(C, model.layers[il].ssm_c_norm, NULL, LLM_NORM_RMS, il);
+        dt = build_norm(dt, model.layers[il].ssm_dt_norm, NULL, LLM_NORM_RMS, il);
+        cb(B, "mamba_B_normed", il);
+        cb(C, "mamba_C_normed", il);
+        cb(dt, "mamba_dt_normed", il);
+
+        // dt_proj: {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
+        dt = build_lora_mm(model.layers[il].ssm_dt, dt);
+        dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b);
+        cb(dt, "mamba_dt_proj", il);
+
+        ggml_tensor * A = ggml_reshape_2d(ctx0, model.layers[il].ssm_a, 1, n_heads);
+        cb(A, "mamba_A", il);
+
+        x = ggml_view_4d(ctx0, x, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x),
+                         head_dim * n_heads * ggml_element_size(x),
+                         head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
+        B = ggml_view_4d(ctx0, B, d_state, 1, n_seq_tokens, n_seqs, d_state * B->nb[0], B->nb[1], B->nb[2], 0);
+        C = ggml_view_4d(ctx0, C, d_state, 1, n_seq_tokens, n_seqs, d_state * C->nb[0], C->nb[1], C->nb[2], 0);
+
+        // use the states and the indices provided by build_recurrent_state
+        // (this is necessary in order to properly use the states before they are overwritten,
+        //  while avoiding to make unnecessary copies of the states)
+        auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+            ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_heads, mctx_cur->get_size());
+
+            // Custom operator to optimize the parallel associative scan
+            // as described in the Annex D of the Mamba paper.
+            // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+            return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+        };
+
+        ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+        cb(y_ssm, "mamba_ssm_scan", il);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(
+                    ctx0,
+                    ggml_view_1d(ctx0, y_ssm, n_heads * head_dim * d_state * n_seqs,
+                                 n_heads * head_dim * n_seq_tokens * n_seqs * ggml_element_size(y_ssm)),
+                    ggml_view_1d(ctx0, ssm_states_all, n_heads * head_dim * d_state * n_seqs,
+                                 kv_head * n_seqs * n_heads * head_dim * d_state * ggml_element_size(ssm_states_all))));
+        cb(ssm_states_all, "mamba_ssm_states", il);
+
+        ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs,
+                                       head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x),
+                                       head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
+        cb(y, "mamba_y_view", il);
+
+        // Add D parameter and apply gating with z
+        // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs}
+        ggml_tensor * D = ggml_reshape_2d(ctx0, model.layers[il].ssm_d, 1, n_heads);
+        y               = ggml_add(ctx0, y, ggml_mul(ctx0, x, D));
+        cb(y, "mamba_y_add_d", il);
+
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+        cb(y, "mamba_y_swiglu_z", il);
+
+        // out_proj: {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        y   = ggml_view_3d(ctx0, y, head_dim * n_heads, n_seq_tokens, n_seqs, y->nb[2], y->nb[3], 0);
+        cur = build_lora_mm(model.layers[il].ssm_out, y);
+        cb(cur, "mamba_out_proj", il);
+    }
+
+    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
+    cb(cur, "mamba_out", il);
+
+    return cur;
+}

data/ext/sources/examples/talk-llama/models/plamo3.cpp

@@ -0,0 +1,128 @@
+#include "models.h"
+
+template <bool iswa>
+llm_build_plamo3<iswa>::llm_build_plamo3(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t head_dim_q = hparams.n_embd_head_k;
+    const int64_t head_dim_v = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL = build_inp_embd(model.tok_embd);
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * residual = inpL;
+
+        float freq_base_l  = 0.0f;
+        float freq_scale_l = 0.0f;
+        if constexpr (iswa) {
+            freq_base_l  = model.get_rope_freq_base (cparams, il);
+            freq_scale_l = model.get_rope_freq_scale(cparams, il);
+        } else {
+            freq_base_l  = freq_base;
+            freq_scale_l = freq_scale;
+        }
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
+        cb(cur, "wqkv", il);
+
+        const int32_t n_head    = hparams.n_head(il);
+        const int32_t n_head_kv = hparams.n_head_kv(il);
+
+        const int64_t q_offset = 0;
+        const int64_t k_offset = head_dim_q * n_head;
+        const int64_t v_offset = k_offset + head_dim_q * n_head_kv;
+
+        ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head, n_tokens,
+                head_dim_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));
+        ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head_kv, n_tokens,
+                head_dim_q * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));
+        ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, head_dim_v, n_head_kv, n_tokens,
+                head_dim_v * sizeof(float), qkv->nb[1], v_offset * ggml_element_size(qkv));
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+        cb(Qcur, "attn_q_norm", il);
+        Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+        cb(Kcur, "attn_k_norm", il);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                ext_factor, attn_factor, beta_fast, beta_slow);
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                ext_factor, attn_factor, beta_fast, beta_slow);
+
+        const float attn_scale = 1.0f / sqrtf(float(head_dim_q));
+
+        cur = build_attn(inp_attn,
+                model.layers[il].wo, NULL,
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, attn_scale, il);
+        cb(cur, "attn_out", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
+            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+        }
+
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, residual);
+        cb(cur, "attn_residual", il);
+
+        residual = cur;
+
+        cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                NULL,                      NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, residual);
+        cb(cur, "ffn_residual", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_plamo3<false>;
+template struct llm_build_plamo3<true>;

data/ext/sources/examples/talk-llama/models/plm.cpp

@@ -0,0 +1,168 @@
+#include "models.h"
+
+llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k));
+
+    const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * q = NULL;
+            q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            cb(q, "q", il);
+
+            // split into {n_head * n_embd_head_qk_nope, n_tokens}
+            ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                    ggml_row_size(q->type, hparams.n_embd_head_k),
+                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_head * n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                    ggml_row_size(q->type, hparams.n_embd_head_k),
+                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
+                    kv_pe_compresseed->nb[1],
+                    0);
+            cb(kv_compressed, "kv_compressed", il);
+
+            // and {n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
+                    kv_pe_compresseed->nb[1],
+                    kv_pe_compresseed->nb[1],
+                    ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            kv_compressed = build_norm(kv_compressed,
+                    model.layers[il].attn_kv_a_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(kv_compressed, "kv_compressed", il);
+
+            // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
+            ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
+            cb(kv, "kv", il);
+
+            // split into {n_head * n_embd_head_qk_nope, n_tokens}
+            ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
+                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    0);
+            cb(k_nope, "k_nope", il);
+
+            // and {n_head * n_embd_head_v, n_tokens}
+            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope)));
+            cb(v_states, "v_states", il);
+
+            v_states = ggml_cont(ctx0, v_states);
+            cb(v_states, "v_states", il);
+
+            v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
+                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
+                    0);
+            cb(v_states, "v_states", il);
+
+            q_pe = ggml_rope_ext(
+                    ctx0, q_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+            cb(q_pe, "q_pe", il);
+
+            // shared RoPE key
+            k_pe = ggml_rope_ext(
+                    ctx0, k_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+            cb(k_pe, "k_pe", il);
+
+            ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+            cb(q_states, "q_states", il);
+
+            ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+            cb(k_states, "k_states", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                NULL, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}