char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Georgi Gerganovhparams : refactor `hparams.n_layer` (#24060)7acb4e8cd

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
9.1 KiB202 linesraw
1#include "models.h"
2
3void llama_model_arwkv7::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,                hparams.f_norm_eps, false);
5    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,            hparams.f_norm_rms_eps, false);
6    ml.get_key(LLM_KV_WKV_HEAD_SIZE,                          hparams.wkv_head_size);
7    ml.get_key(LLM_KV_ATTENTION_DECAY_LORA_RANK,              hparams.n_lora_decay);
8    ml.get_key(LLM_KV_ATTENTION_ICLR_LORA_RANK,               hparams.n_lora_iclr);
9    ml.get_key(LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, hparams.n_lora_value_res_mix);
10    ml.get_key(LLM_KV_ATTENTION_GATE_LORA_RANK,               hparams.n_lora_gate, false);
11    ml.get_key(LLM_KV_TOKEN_SHIFT_COUNT,                      hparams.token_shift_count, false);
12
13    switch (hparams.n_layer()) {
14        case 12:
15            switch (hparams.n_embd) {
16                case 768: type = LLM_TYPE_190M; break;
17                default: type = LLM_TYPE_UNKNOWN;
18            } break;
19        case 24:
20            switch (hparams.n_embd) {
21                case 1024: type = LLM_TYPE_450M; break;
22                case 2048: type = LLM_TYPE_1_5B; break;
23                default: type = LLM_TYPE_UNKNOWN;
24            } break;
25        case 28:
26            switch (hparams.n_embd) {
27                case 1536: type = LLM_TYPE_1_5B; break;
28                case 3584: type = LLM_TYPE_7B; break;
29                default: type = LLM_TYPE_UNKNOWN;
30            } break;
31        case 32:
32            switch (hparams.n_embd) {
33                case 2560: type = LLM_TYPE_2_9B; break;
34                case 4096: type = LLM_TYPE_7B; break;
35                default: type = LLM_TYPE_UNKNOWN;
36            } break;
37        case 61:
38            switch (hparams.n_embd) {
39                case 4096: type = LLM_TYPE_14B; break;
40                default: type = LLM_TYPE_UNKNOWN;
41            } break;
42        default: type = LLM_TYPE_UNKNOWN;
43    }
44}
45
46void llama_model_arwkv7::load_arch_tensors(llama_model_loader &) {
47    LLAMA_LOAD_LOCALS;
48
49    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
50
51    // output
52    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
53    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
54
55    const int n_lora_decay = hparams.n_lora_decay;
56    const int n_lora_iclr = hparams.n_lora_iclr;
57    const int n_lora_value_res_mix = hparams.n_lora_value_res_mix;
58    const int n_lora_gate = hparams.n_lora_gate;
59    const int attn_hidden_size = n_embd;
60
61    for (int i = 0; i < n_layer; ++i) {
62        auto & layer = layers[i];
63
64        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
65
66        layer.time_mix_w0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W0, "weight", i), {n_embd}, 0);
67        layer.time_mix_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W1, "weight", i), {n_embd, n_lora_decay}, 0);
68        layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {n_lora_decay, n_embd}, 0);
69
70        layer.time_mix_a0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A0, "weight", i), {n_embd}, 0);
71        layer.time_mix_a1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A1, "weight", i), {n_embd, n_lora_iclr}, 0);
72        layer.time_mix_a2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A2, "weight", i), {n_lora_iclr, n_embd}, 0);
73
74        if (i == 0) {
75            // actually not used
76            layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0);
77            layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_iclr}, 0);
78            layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_iclr, n_embd}, 0);
79        } else {
80            layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0);
81            layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_value_res_mix}, 0);
82            layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0);
83        }
84
85        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED);
86        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED);
87
88        try {
89            layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0);
90        } catch(std::runtime_error & e) {
91            // ARWKV models may not have gate tensors
92            layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0);
93        }
94
95        layer.time_mix_k_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_K, "weight", i), {attn_hidden_size}, 0);
96        layer.time_mix_k_a = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_A, "weight", i), {attn_hidden_size}, 0);
97        layer.time_mix_r_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_R_K, "weight", i), {attn_hidden_size}, 0);
98
99        layer.time_mix_key = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "weight", i), {attn_hidden_size, n_embd}, 0);
100        layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0);
101        layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
102
103        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
104        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
105        layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
106
107        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
108
109        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
110        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
111        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
112    }
113
114}
115
116std::unique_ptr<llm_graph_context> llama_model_arwkv7::build_arch_graph(const llm_graph_params & params) const {
117    return std::make_unique<graph>(*this, params);
118}
119
120llama_model_arwkv7::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
121    GGML_ASSERT(n_embd == hparams.n_embd_r());
122
123    ggml_tensor * cur;
124    ggml_tensor * inpL;
125    ggml_tensor * v_first = nullptr;
126
127    inpL = build_inp_embd(model.tok_embd);
128
129    auto * rs_inp = build_rs_inp();
130
131    const auto n_embd = hparams.n_embd;
132    const auto n_seq_tokens = ubatch.n_seq_tokens;
133    const auto n_seqs = ubatch.n_seqs;
134
135    ggml_tensor * inp_out_ids = build_inp_out_ids();
136
137    for (int il = 0; il < n_layer; ++il) {
138        const llama_layer * layer = &model.layers[il];
139        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
140
141        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
142
143        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
144        cb(att_norm, "attn_norm", il);
145
146        ggml_tensor * x_prev = ggml_concat(
147                ctx0,
148                token_shift,
149                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
150                1
151                );
152
153        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
154
155        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
156        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
157
158        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
159        cb(ffn_inp, "ffn_inp", il);
160
161        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
162        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
163
164        if (il == n_layer - 1 && inp_out_ids) {
165            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
166            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
167        }
168        // feed-forward network
169        cur = build_norm(ffn_inp,
170                model.layers[il].ffn_norm, NULL,
171                LLM_NORM_RMS, il);
172        cb(cur, "ffn_norm", il);
173
174        cur = build_ffn(cur,
175                model.layers[il].ffn_up,   NULL, NULL,
176                model.layers[il].ffn_gate, NULL, NULL,
177                model.layers[il].ffn_down, NULL, NULL,
178                NULL,
179                LLM_FFN_SILU, LLM_FFN_PAR, il);
180        cb(cur, "ffn_out", il);
181
182        cur = ggml_add(ctx0, cur, ffn_inp);
183
184        cur = build_cvec(cur, il);
185        cb(cur, "l_out", il);
186
187        // input for next layer
188        inpL = cur;
189    }
190    cur = inpL;
191    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
192
193    cb(cur, "result_norm", -1);
194    res->t_embd = cur;
195
196    cur = build_lora_mm(model.output, cur, model.output_s);
197
198    cb(cur, "result_output", -1);
199    res->t_logits = cur;
200
201    ggml_build_forward_expand(gf, cur);
202}