Felipe97/llama-cpp-compiled

1#include "models.h"2 3#include "llama-memory-recurrent.h"4 5llm_build_rwkv6_base::llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) :6    llm_graph_context(params),7    model(model) {}8 9ggml_tensor * llm_build_rwkv6_base::build_rwkv6_channel_mix(const llama_layer * layer,10                                                            ggml_tensor *       cur,11                                                            ggml_tensor *       x_prev,12                                                            llm_arch            arch) const {13    ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);14    switch (arch) {15        case LLM_ARCH_RWKV6:16            {17                ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur);18                ggml_tensor * xr = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_r), cur);19 20                ggml_tensor * r = ggml_sigmoid(ctx0, build_lora_mm(layer->channel_mix_receptance, xr));21                ggml_tensor * k = ggml_sqr(ctx0, ggml_relu(ctx0, build_lora_mm(layer->channel_mix_key, xk)));22                cur             = ggml_mul(ctx0, r, build_lora_mm(layer->channel_mix_value, k));23            }24            break;25        default:26            GGML_ABORT("fatal error");27    }28    return cur;29}30 31ggml_tensor * llm_build_rwkv6_base::build_rwkv6_time_mix(llm_graph_input_rs * inp,32                                                         ggml_tensor *        cur,33                                                         ggml_tensor *        x_prev,34                                                         const llama_ubatch & ubatch,35                                                         int                  il) const {36    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);37 38    const auto n_tokens     = ubatch.n_tokens;39    const auto n_seqs       = ubatch.n_seqs;40    const auto n_seq_tokens = ubatch.n_seq_tokens;41    const auto n_embd       = hparams.n_embd;42    const auto head_size    = hparams.wkv_head_size;43    const auto n_head       = n_embd / head_size;44    const auto n_head_kv    = hparams.n_head_kv(il);45 46    const auto kv_head = mctx_cur->get_head();47 48    const auto & layer = model.layers[il];49 50    bool is_qrwkv = layer.time_mix_first == nullptr;51 52    ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);53 54    sx  = ggml_reshape_2d(ctx0, sx, n_embd, n_tokens);55    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);56 57    ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_x), cur);58 59    xxx = ggml_reshape_4d(ctx0, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xxx)),60                          layer.time_mix_w1->ne[1] / 5, 1, 5, n_tokens);61 62    xxx = ggml_cont(ctx0, ggml_permute(ctx0, xxx, 0, 1, 3, 2));63 64    xxx = ggml_mul_mat(65        ctx0, ggml_reshape_4d(ctx0, layer.time_mix_w2, layer.time_mix_w2->ne[0], layer.time_mix_w2->ne[1], 1, 5), xxx);66 67    ggml_tensor *xw, *xk, *xv, *xr, *xg;68    if (layer.time_mix_lerp_fused) {69        // fusing these weights makes some performance improvement70        sx  = ggml_reshape_3d(ctx0, sx, n_embd, 1, n_tokens);71        cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);72        xxx = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xxx, layer.time_mix_lerp_fused), sx), cur);73        xw  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);74        xk  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));75        xv  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));76        xr  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));77        xg  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));78    } else {79        // for backward compatibility80        xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);81        xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));82        xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));83        xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));84        xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));85 86        xw = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xw, layer.time_mix_lerp_w), sx), cur);87        xk = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xk, layer.time_mix_lerp_k), sx), cur);88        xv = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xv, layer.time_mix_lerp_v), sx), cur);89        xr = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xr, layer.time_mix_lerp_r), sx), cur);90        xg = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xg, layer.time_mix_lerp_g), sx), cur);91    }92    ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr);93    ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk);94    ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv);95    if (layer.time_mix_receptance_b) {96        r = ggml_add(ctx0, r, layer.time_mix_receptance_b);97    }98    if (layer.time_mix_key_b) {99        k = ggml_add(ctx0, k, layer.time_mix_key_b);100    }101    if (layer.time_mix_value_b) {102        v = ggml_add(ctx0, v, layer.time_mix_value_b);103    }104    ggml_tensor * g = build_lora_mm(layer.time_mix_gate, xg);105    if (is_qrwkv) {106        g = ggml_sigmoid(ctx0, g);107    } else {108        g = ggml_silu(ctx0, g);109    }110    if (n_head_kv != 0 && n_head_kv != n_head) {111        GGML_ASSERT(n_head % n_head_kv == 0);112        k                 = ggml_reshape_4d(ctx0, k, head_size, 1, n_head_kv, n_tokens);113        v                 = ggml_reshape_4d(ctx0, v, head_size, 1, n_head_kv, n_tokens);114        ggml_tensor * tmp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_size, n_head / n_head_kv, n_head_kv, n_tokens);115        k                 = ggml_repeat(ctx0, k, tmp);116        v                 = ggml_repeat(ctx0, v, tmp);117    }118    k = ggml_reshape_3d(ctx0, k, head_size, n_head, n_tokens);119    v = ggml_reshape_3d(ctx0, v, head_size, n_head, n_tokens);120    r = ggml_reshape_3d(ctx0, r, head_size, n_head, n_tokens);121 122    ggml_tensor * w =123        ggml_mul_mat(ctx0, layer.time_mix_decay_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_decay_w1, xw)));124 125    w = ggml_add(ctx0, w, layer.time_mix_decay);126    w = ggml_exp(ctx0, ggml_neg(ctx0, ggml_exp(ctx0, w)));127    w = ggml_reshape_3d(ctx0, w, head_size, n_head, n_tokens);128 129    if (is_qrwkv) {130        // k = k * (1 - w)131        k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w));132    }133    ggml_tensor * wkv_state = build_rs(inp, mctx_cur->get_s_l(il), hparams.n_embd_s(), n_seqs);134 135    ggml_tensor * wkv_output;136    if (is_qrwkv) {137        wkv_output = ggml_gated_linear_attn(ctx0, k, v, r, w, wkv_state, pow(head_size, -0.5f));138    } else {139        wkv_output = ggml_rwkv_wkv6(ctx0, k, v, r, layer.time_mix_first, w, wkv_state);140    }141    cur       = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0);142    wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));143 144    ggml_build_forward_expand(145        gf, ggml_cpy(ctx0, wkv_state,146                     ggml_view_1d(ctx0, mctx_cur->get_s_l(il), hparams.n_embd_s() * n_seqs,147                                  hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il)))));148 149    if (!is_qrwkv) {150        // group norm with head_count groups151        cur = ggml_reshape_3d(ctx0, cur, n_embd / n_head, n_head, n_tokens);152        cur = ggml_norm(ctx0, cur, 64e-5f);153 154        // Convert back to regular vectors.155        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);156        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b);157    } else {158        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);159    }160    cur = ggml_mul(ctx0, cur, g);161    cur = build_lora_mm(layer.time_mix_output, cur);162 163    return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs);164}165