whispercpp 1.3.4 → 1.3.5

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

@@ -0,0 +1,211 @@
+#include "models.h"
+
+
+llm_build_granite::llm_build_granite(
+    const llama_model & model,
+    const llm_graph_params & params)
+    : llm_graph_context(params) {
+
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - built only if rope enabled
+    ggml_tensor * inp_pos = nullptr;
+    if (hparams.rope_finetuned) {
+        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
+        cur = build_attention_layer(
+            cur, inp_pos, inp_attn,
+            model, n_embd_head, 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);
+        }
+        // ffn
+        cur = build_layer_ffn(cur, inpSA, model, 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;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // For Granite architectures - scale logits
+    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_granite::build_attention_layer(
+          ggml_tensor             * cur,
+          ggml_tensor             * inp_pos,
+          llm_graph_input_attn_kv * inp_attn,
+    const llama_model             & model,
+    const int64_t                 n_embd_head,
+    const int                     il) {
+
+    // compute Q and K and (optionally) RoPE them
+    ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur, "Qcur", il);
+    if (model.layers[il].bq) {
+        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+        cb(Qcur, "Qcur", il);
+    }
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+    if (model.layers[il].bk) {
+        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+        cb(Kcur, "Kcur", il);
+    }
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+    if (model.layers[il].bv) {
+        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+        cb(Vcur, "Vcur", il);
+    }
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+    const bool use_rope = hparams.rope_finetuned;
+    if (use_rope) {
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+        Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+        Kcur = ggml_rope_ext(
+                ctx0, Kcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+    }
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, model.layers[il].bo,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llm_build_granite::build_layer_ffn(
+          ggml_tensor       * cur,
+          ggml_tensor       * inpSA,
+    const llama_model       & model,
+    const int                 il) {
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "ffn_inp", il);
+
+    // feed-forward network (non-MoE)
+    if (model.layers[il].ffn_gate_inp == nullptr) {
+
+        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,   model.layers[il].ffn_up_b,   NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+
+    } else {
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // For Granite MoE Shared
+        if (hparams.n_ff_shexp > 0) {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                model.layers[il].ffn_up_shexp,   NULL, NULL,
+                model.layers[il].ffn_gate_shexp, NULL, NULL,
+                model.layers[il].ffn_down_shexp, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+    }
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "ffn_out", il);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}

data/ext/sources/examples/talk-llama/models/graph-context-mamba.cpp

@@ -0,0 +1,283 @@
+#include "models.h"
+
+llm_graph_context_mamba::llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {}
+
+ggml_tensor * llm_graph_context_mamba::build_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 auto & layer = model.layers[il];
+
+    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 dt_rank        = hparams.ssm_dt_rank;
+    const int64_t n_head         = d_inner;
+    const int64_t head_dim       = 1;
+    const int64_t n_seqs         = ubatch.n_seqs;
+    // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers)
+    const bool    ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms;
+
+    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, 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);
+
+    // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur);
+    // split the above in two
+    // => {d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * x  = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0);
+    ggml_tensor * z =
+        ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner * ggml_element_size(xz));
+
+    // conv
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
+
+        // 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) * (n_seqs),
+                                      kv_head * (d_conv - 1) * (d_inner) *ggml_element_size(conv_states_all))));
+
+        // 1D convolution
+        // The equivalent is to make a self-overlapping view of conv_x
+        // over d_conv columns at each stride in the 3rd dimension,
+        // then element-wise multiply that with the conv1d weight,
+        // then sum the elements of each row,
+        // (the last two steps are a dot product over rows (also doable with mul_mat))
+        // then permute away the ne[0] dimension,
+        // and then you're left with the resulting x tensor.
+        // For simultaneous sequences, all sequences need to have the same length.
+        x = ggml_ssm_conv(ctx0, conv_x, layer.ssm_conv1d);
+
+        // bias
+        x = ggml_add(ctx0, x, layer.ssm_conv1d_b);
+
+        x = ggml_silu(ctx0, x);
+    }
+
+    // ssm
+    {
+        // {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_db = build_lora_mm(layer.ssm_x, x);
+        // split
+        ggml_tensor * dt   = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0);
+        ggml_tensor * B =
+            ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1],
+                         x_db->nb[2], ggml_element_size(x_db) * dt_rank);
+        ggml_tensor * C =
+            ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1],
+                         x_db->nb[2], ggml_element_size(x_db) * (dt_rank + d_state));
+
+        // Some Mamba variants (e.g. FalconMamba, Jamba) apply RMS norm in B, C & Dt layers
+        if (ssm_dt_b_c_rms || (layer.ssm_dt_norm && layer.ssm_b_norm && layer.ssm_c_norm)) {
+            dt = build_norm(dt, layer.ssm_dt_norm, NULL, LLM_NORM_RMS, il);
+            B  = build_norm(B, layer.ssm_b_norm, NULL, LLM_NORM_RMS, il);
+            C  = build_norm(C, layer.ssm_c_norm, NULL, LLM_NORM_RMS, il);
+        }
+
+        // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
+        dt = build_lora_mm(layer.ssm_dt, dt);
+        dt = ggml_add(ctx0, dt, layer.ssm_dt_b);
+
+        cur = x;
+        x   = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs);
+
+        ggml_tensor * A = layer.ssm_a;
+
+        // 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_head, 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);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, x->nb[3] * x->ne[3]),
+                         ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs,
+                                      kv_head * d_state * d_inner * ggml_element_size(ssm_states_all))));
+
+        ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0);
+
+        // TODO: skip computing output earlier for unused tokens
+
+        y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, layer.ssm_d));
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+
+        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        cur = build_lora_mm(layer.ssm_out, y);
+    }
+
+    // {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);
+
+    return cur;
+}
+
+ggml_tensor * llm_graph_context_mamba::build_mamba2_layer(llm_graph_input_rs * inp,
+                                                          ggml_tensor *        cur,
+                                                          const llama_model &  model,
+                                                          const llama_ubatch & ubatch,
+                                                          int                  il) const {
+    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_head   = hparams.ssm_dt_rank;
+    const int64_t head_dim = d_inner / n_head;
+    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);
+
+    // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
+
+    // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
+    ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur);
+
+    // split the above in three
+    ggml_tensor * z   = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * zxBCdt->nb[0],
+                                     zxBCdt->nb[1], zxBCdt->nb[2], 0);
+    ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2 * n_group * d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1],
+                                     zxBCdt->nb[2], d_inner * ggml_element_size(zxBCdt));
+    ggml_tensor * dt  = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2],
+                                     (2 * d_inner + 2 * n_group * d_state) * ggml_element_size(zxBCdt));
+
+    // conv
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0);
+
+        // 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 + 2 * n_group * d_state, 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))));
+
+        // 1D convolution
+        // The equivalent is to make a self-overlapping view of conv_x
+        // over d_conv columns at each stride in the 3rd dimension,
+        // then element-wise multiply that with the conv1d weight,
+        // then sum the elements of each row,
+        // (the last two steps are a dot product over rows (also doable with mul_mat))
+        // then permute away the ne[0] dimension,
+        // and then you're left with the resulting x tensor.
+        // For simultaneous sequences, all sequences need to have the same length.
+        xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
+
+        // bias
+        xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b);
+
+        xBC = ggml_silu(ctx0, xBC);
+    }
+
+    // ssm
+    {
+        // These correspond to V K Q in SSM/attention duality
+        ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], 0);
+        ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], d_inner * ggml_element_size(xBC));
+        ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], (d_inner + n_group * d_state) * ggml_element_size(xBC));
+
+        // {n_head, n_seq_tokens, n_seqs}
+        dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b);
+
+        ggml_tensor * A = model.layers[il].ssm_a;
+
+        // 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_head, mctx_cur->get_size());
+
+            // TODO: use semistructured matrices to implement state-space duality
+            // => {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);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, ggml_nelements(x) * x->nb[0]),
+                         ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs,
+                                      kv_head * d_state * d_inner * ggml_element_size(ssm_states_all))));
+
+        ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head * x->nb[1],
+                                       n_seq_tokens * n_head * x->nb[1], 0);
+
+        // TODO: skip computing output earlier for unused tokens
+
+        y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
+        cb(y, "mamba2_y_add_d", il);
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+
+        // grouped RMS norm
+        if (model.layers[il].ssm_norm) {
+            y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs);
+            y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
+        }
+
+        y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
+
+        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        cur = build_lora_mm(model.layers[il].ssm_out, y);
+    }
+
+    // {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/grok.cpp

@@ -0,0 +1,159 @@
+#include "models.h"
+
+llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    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
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                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_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            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 = 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
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, 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);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_out_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_out_norm", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // MoE branch
+        ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_GELU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        if (model.layers[il].ffn_up) {
+            ggml_tensor * ffn_out = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(ffn_out, "ffn_out", il);
+
+            cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+        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, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        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;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
+
+    // final logit soft-capping
+    if (hparams.f_final_logit_softcapping) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}