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 →
7.7 KiB191 linesraw
1#include "models.h"
2
3void llama_model_deci::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    switch (hparams.n_layer()) {
7        case 32: type = LLM_TYPE_7B; break;
8        case 80: type = LLM_TYPE_70B; break;
9        case 162: type = LLM_TYPE_405B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12}
13
14void llama_model_deci::load_arch_tensors(llama_model_loader &) {
15    LLAMA_LOAD_LOCALS;
16
17    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
18
19    // output
20    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
21    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
22
23    // if output is NULL, init from the input tok embed
24    if (output == NULL) {
25        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
26    }
27
28    for (int i = 0; i < n_layer; ++i) {
29        auto & layer = layers[i];
30        const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa(i);
31        const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa(i);
32        const int64_t n_ff          = hparams.n_ff(i);
33        const int64_t n_head        = hparams.n_head(i);
34        const int64_t n_head_kv     = hparams.n_head_kv(i);
35
36        if (n_head_kv == 0 && n_head > 0) {
37            // linear attention for DeciLMCausalModel
38            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
39            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
40        }
41        else if (n_head_kv > 0) {
42            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
43
44            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
45            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
46        }
47
48        // optional bias tensors
49        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
50
51        if (n_ff > 0) {
52            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
53        }
54
55        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
56            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
57            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
58        }
59        else {
60            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
61        }
62
63        if (n_ff > 0) {
64            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
65            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
66            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
67        }
68
69        // optional MLP bias
70        layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
71        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
72        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
73    }
74}
75
76std::unique_ptr<llm_graph_context> llama_model_deci::build_arch_graph(const llm_graph_params & params) const {
77    return std::make_unique<graph>(*this, params);
78}
79
80llama_model_deci::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
81    const int64_t n_embd_head = hparams.n_embd_head_v();
82
83    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
84    GGML_ASSERT(n_embd_head == n_rot);
85
86    ggml_tensor * cur;
87    ggml_tensor * inpL;
88
89    inpL = build_inp_embd(model.tok_embd);
90
91    // inp_pos - contains the positions
92    ggml_tensor * inp_pos = build_inp_pos();
93
94    auto * inp_attn = build_attn_inp_kv();
95
96    const float kq_scale =
97        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
98
99    ggml_tensor * inp_out_ids = build_inp_out_ids();
100
101    for (int il = 0; il < n_layer; ++il) {
102        ggml_tensor * inpSA     = inpL;
103        const int64_t n_head_kv = hparams.n_head_kv(il);
104        const int64_t n_head    = hparams.n_head(il);
105        const int64_t n_ff      = hparams.n_ff(il);
106
107        if (n_head == 0) {
108            // attention-free layer of Llama-3_1-Nemotron-51B
109            cur = inpL;
110        } else {
111            // norm
112            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
113            cb(cur, "attn_norm", il);
114        }
115        if (n_head > 0 && n_head_kv == 0) {
116            // "linear attention" of Llama-3_1-Nemotron-51B
117            cur = build_lora_mm(model.layers[il].wo, cur);
118            cb(cur, "wo", il);
119        } else if (n_head > 0) {
120            // self-attention
121            // rope freq factors for llama3; may return nullptr for llama2 and other models
122            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
123
124            // compute Q and K and RoPE them
125            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
126                    n_embd_head, n_head, n_head_kv, il);
127
128            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
129                                 ext_factor, attn_factor, beta_fast, beta_slow);
130
131            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
132                                 ext_factor, attn_factor, beta_fast, beta_slow);
133
134            cb(Qcur, "Qcur", il);
135            cb(Kcur, "Kcur", il);
136            cb(Vcur, "Vcur", il);
137
138            cur = build_attn(inp_attn,
139                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
140                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
141        }
142        if (il == n_layer - 1 && inp_out_ids) {
143            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
144            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
145        }
146        // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
147        if (n_ff == 0) {
148            continue;
149        }
150        // modified to support attention-free layer of Llama-3_1-Nemotron-51B
151        ggml_tensor * ffn_inp = cur;
152        if (n_head > 0) {
153            ffn_inp = ggml_add(ctx0, cur, inpSA);
154            cb(ffn_inp, "ffn_inp", il);
155        }
156        // feed-forward network
157        if (model.layers[il].ffn_gate_inp == nullptr) {
158            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
159            cb(cur, "ffn_norm", il);
160
161            cur = build_ffn(cur,
162                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
163                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
164                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
165                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
166            cb(cur, "ffn_out", il);
167        }
168        cur = ggml_add(ctx0, cur, ffn_inp);
169        cb(cur, "ffn_out", il);
170
171        cur = build_cvec(cur, il);
172        cb(cur, "l_out", il);
173
174        // input for next layer
175        inpL = cur;
176    }
177    cur = inpL;
178
179    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
180
181    cb(cur, "result_norm", -1);
182    res->t_embd = cur;
183
184    // lm_head
185    cur = build_lora_mm(model.output, cur, model.output_s);
186
187    cb(cur, "result_output", -1);
188    res->t_logits = cur;
189
190    ggml_build_forward_expand(gf, cur);
191}