Brunobkr/llama.cpp_AlgMor24_github

1#include "llama-hparams.h"2 3#include "ggml.h"4 5#include <algorithm>6#include <cassert>7 8void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {9    if (dense_first) {10        for (uint32_t il = 0; il < n_layer(); ++il) {11            is_swa_impl[il] = n_pattern == 0 || (il % n_pattern != 0);12        }13    } else {14        for (uint32_t il = 0; il < n_layer(); ++il) {15            is_swa_impl[il] = n_pattern == 0 || (il % n_pattern < (n_pattern - 1));16        }17    }18 19    for (uint32_t il = n_layer(); il < n_layer_all; ++il) {20        is_swa_impl[il] = false;21    }22}23 24void llama_hparams::set_recr_pattern(uint32_t n_pattern, bool dense_first) {25    if (dense_first) {26        for (uint32_t il = 0; il < n_layer(); ++il) {27            is_recr_impl[il] = n_pattern == 0 || (il % n_pattern != 0);28        }29    } else {30        for (uint32_t il = 0; il < n_layer(); ++il) {31            is_recr_impl[il] = n_pattern == 0 || (il % n_pattern < (n_pattern - 1));32        }33    }34 35    for (uint32_t il = n_layer(); il < n_layer_all; ++il) {36        is_recr_impl[il] = false;37    }38}39 40bool llama_hparams::is_swa_any() const {41    for (uint32_t il = 0; il < n_layer_all; ++il) {42        if (is_swa_impl[il]) {43            return true;44        }45    }46 47    return false;48}49 50uint32_t llama_hparams::n_head(uint32_t il) const {51    if (il < n_layer_all) {52        return n_head_arr[il];53    }54 55    GGML_ABORT("fatal error");56}57 58uint32_t llama_hparams::n_head_kv(uint32_t il) const {59    if (il < n_layer_all) {60        return n_head_kv_arr[il];61    }62 63    GGML_ABORT("fatal error");64}65 66uint32_t llama_hparams::n_ff(uint32_t il) const {67    if (il < n_layer_all) {68        return n_ff_arr[il];69    }70 71    GGML_ABORT("fatal error");72}73 74uint32_t llama_hparams::n_gqa(uint32_t il) const {75    const uint32_t n_head    = this->n_head(il);76    const uint32_t n_head_kv = this->n_head_kv(il);77 78    if (n_head_kv == 0) {79        return 0;80    }81 82    return n_head/n_head_kv;83}84 85uint32_t llama_hparams::n_rot(uint32_t il) const {86    if (il < n_layer_all) {87        return is_swa(il) ? n_rot_swa : n_rot_full;88    }89 90    GGML_ABORT("fatal error");91}92 93uint32_t llama_hparams::n_embd_inp() const {94    if (n_embd_inp_impl > 0) {95        return n_embd_inp_impl;96    }97 98    uint32_t n_embd_inp = n_embd;99 100    if (n_deepstack_layers > 0) {101        n_embd_inp += n_embd * n_deepstack_layers;102    }103 104    return n_embd_inp;105}106 107uint32_t llama_hparams::n_embd_inp_enc() const {108    return n_embd_inp_enc_impl > 0 ? n_embd_inp_enc_impl : n_embd_inp();109}110 111uint32_t llama_hparams::n_embd_out() const {112    return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;113}114 115uint32_t llama_hparams::n_embd_head_k(uint32_t il) const {116    if (il < n_layer_all) {117        return is_swa(il) ? n_embd_head_k_swa : n_embd_head_k_full;118    }119 120    GGML_ABORT("fatal error");121}122 123uint32_t llama_hparams::n_embd_head_v(uint32_t il) const {124    if (il < n_layer_all) {125        return is_swa(il) ? n_embd_head_v_swa : n_embd_head_v_full;126    }127 128    GGML_ABORT("fatal error");129}130 131uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {132    const uint32_t n_head_kv = this->n_head_kv(il);133 134    return n_embd_head_k(il) * n_head_kv;135}136 137uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {138    const uint32_t n_head_kv = this->n_head_kv(il);139 140    return n_embd_head_v(il) * n_head_kv;141}142 143bool llama_hparams::is_n_embd_k_gqa_variable() const {144    const uint32_t val = n_embd_k_gqa();145    for (uint32_t il = 0; il < n_layer_all; ++il) {146        if (val != n_embd_k_gqa(il)) {147            return true;148        }149    }150 151    return false;152}153 154bool llama_hparams::is_n_embd_v_gqa_variable() const {155    const uint32_t val = n_embd_v_gqa();156    for (uint32_t il = 0; il < n_layer_all; ++il) {157        if (val != n_embd_v_gqa(il)) {158            return true;159        }160    }161 162    return false;163}164 165uint32_t llama_hparams::n_embd_k_gqa_max() const {166    uint32_t val = n_embd_k_gqa();167    for (uint32_t il = 0; il < n_layer_all; ++il) {168        val = std::max(val, n_embd_k_gqa(il));169    }170 171    return val;172}173 174uint32_t llama_hparams::n_embd_v_gqa_max() const {175    uint32_t val = n_embd_v_gqa();176    for (uint32_t il = 0; il < n_layer_all; ++il) {177        val = std::max(val, n_embd_v_gqa(il));178    }179 180    return val;181}182 183uint32_t llama_hparams::n_embd_r() const {184    if (wkv_head_size != 0) {185        // for RWKV models186        return token_shift_count * n_embd;187    }188 189    if (n_shortconv_l_cache != 0) {190        // for LFM2 models191        return n_embd * (n_shortconv_l_cache - 1);192    }193 194    if (n_embd_head_kda != 0) {195        // for Kimi KDA layers196        // Conv state for Q, K, V: 3 * (d_conv - 1) * n_head * head_dim197        const uint32_t d_inner = n_head() * n_embd_head_kda;  // 32 * 128 = 4096198        return 3 * (ssm_d_conv > 0 ? ssm_d_conv - 1 : 3) * d_inner;199    }200 201    // TODO: maybe support other convolution strides than 1202    // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed203    // Corresponds to Mamba's conv_states size204    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * (ssm_d_inner + 2*ssm_n_group*ssm_d_state);205}206 207uint32_t llama_hparams::n_embd_s() const {208    if (wkv_head_size != 0) {209        // corresponds to RWKV's wkv_states size210        return n_embd * wkv_head_size;211    }212 213    if (n_embd_head_kda != 0) {214        // for Kimi KDA layers215        // Full recurrent state: head_dim * head_dim * n_head216        // h tensor shape for delta attention: [head_dim, head_dim, n_head]217        return n_embd_head_kda * n_embd_head_kda * n_head();  // 128 * 128 * 32 = 524288218    }219 220    // corresponds to Mamba's ssm_states size221    return ssm_d_state * ssm_d_inner;222}223 224bool llama_hparams::is_recr(uint32_t il) const {225    if (il < n_layer_all) {226        return is_recr_impl[il];227    }228 229    GGML_ABORT("%s: il (%u) out of bounds (n_layer_all: %u)\n", __func__, il, n_layer_all);230}231 232uint32_t llama_hparams::n_pos_per_embd() const {233    return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;234}235 236bool llama_hparams::is_swa(uint32_t il) const {237    if (il < n_layer_all) {238        return is_swa_impl[il];239    }240 241    GGML_ABORT("%s: il (%u) out of bounds (n_layer_all: %u)\n", __func__, il, n_layer_all);242}243 244bool llama_hparams::is_mla() const {245    assert((n_embd_head_k_mla_impl == 0 && n_embd_head_v_mla_impl == 0) ||246           (n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0));247 248    return n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0;249}250 251bool llama_hparams::is_indexer_full(uint32_t il) const {252    if (il < n_layer()) {253        return is_indexer_full_impl[il];254    }255 256    GGML_ABORT("%s: il (%u) out of bounds (n_layer: %u)\n", __func__, il, n_layer());257}258 259uint32_t llama_hparams::n_embd_head_k_mla() const {260    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k();261}262 263uint32_t llama_hparams::n_embd_head_v_mla() const {264    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v();265}266 267bool llama_hparams::has_kv(uint32_t il) const {268    if (n_layer_kv_from_start >= 0) {269        if (il < (uint32_t) n_layer_kv_from_start) {270            return true;271        }272 273        return false;274    }275 276    // by default, all layers have kv277    return true;278}279 280bool llama_hparams::has_rope(uint32_t il) const {281    // the router layer stores adapter routing signal, not positional info,282    // so it must not be RoPE-shifted283    if (router_layer >= 0 && (int32_t) il == router_layer) {284        return false;285    }286 287    return true;288}289 290uint32_t llama_hparams::n_layer() const {291    return n_layer_all - n_layer_nextn;292}293 294bool llama_hparams::use_mrope() const {295    return rope_sections[0] > 0 && rope_sections[1] > 0;296}297