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 →
7.2 KiB196 linesraw
1#include "models.h"
2
3void llama_model_phi3::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 24: type = LLM_TYPE_1B; break;
8        case 32: type = LLM_TYPE_3B; break;
9        case 40: type = LLM_TYPE_14B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12
13    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
14
15    if (found_swa && hparams.n_swa > 0) {
16        LLAMA_LOG_WARN("%s: Phi SWA is currently disabled - results might be suboptimal for some models (see %s)\n",
17                __func__, "https://github.com/ggml-org/llama.cpp/pull/13676");
18
19        // TODO: fix conversion scripts to correctly populate `n_swa` and `n_swa_pattern`
20        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
21
22        hparams.n_swa         = 0;
23        hparams.set_swa_pattern(1);
24    }
25}
26
27void llama_model_phi3::load_arch_tensors(llama_model_loader &) {
28    LLAMA_LOAD_LOCALS;
29
30    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
31
32    // output
33    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
34    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
35
36    // if output is NULL, init from the input tok embed
37    if (output == NULL) {
38        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
39    }
40
41    for (int i = 0; i < n_layer; ++i) {
42        auto & layer = layers[i];
43
44        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
45
46        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, TENSOR_NOT_REQUIRED);
47        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);
48
49        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
50
51        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
52        layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff }, 0);
53
54        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
55        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
56    }
57}
58
59std::unique_ptr<llm_graph_context> llama_model_phi3::build_arch_graph(const llm_graph_params & params) const {
60    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
61        return std::make_unique<graph<true>> (*this, params);
62    } else {
63        return std::make_unique<graph<false>>(*this, params);
64    }
65}
66
67template<bool iswa>
68llama_model_phi3::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
69    const int64_t n_embd_head = hparams.n_embd_head_v();
70
71    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
72
73    ggml_tensor * cur;
74    ggml_tensor * inpL;
75
76    inpL = build_inp_embd(model.tok_embd);
77
78    // inp_pos - contains the positions
79    ggml_tensor * inp_pos = build_inp_pos();
80
81    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
82    inp_attn_type * inp_attn = nullptr;
83
84    if constexpr (iswa) {
85        inp_attn = build_attn_inp_kv_iswa();
86    } else {
87        inp_attn = build_attn_inp_kv();
88    }
89    ggml_tensor * inp_out_ids = build_inp_out_ids();
90
91    for (int il = 0; il < n_layer; ++il) {
92        auto * residual = inpL;
93
94        // self-attention
95        {
96            // rope freq factors for 128k context
97            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
98
99            ggml_tensor* attn_norm_output = build_norm(inpL,
100                    model.layers[il].attn_norm,
101                    model.layers[il].attn_norm_b,
102                    LLM_NORM_RMS, il);
103            cb(attn_norm_output, "attn_norm", il);
104
105            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], attn_norm_output,
106                    n_embd_head, n_head, n_head_kv, il);
107            Qcur = ggml_rope_ext(
108                    ctx0, Qcur, inp_pos, rope_factors,
109                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
110                    ext_factor, attn_factor, beta_fast, beta_slow
111                    );
112
113            Kcur = ggml_rope_ext(
114                    ctx0, Kcur, inp_pos, rope_factors,
115                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
116                    ext_factor, attn_factor, beta_fast, beta_slow
117                    );
118
119            cb(Qcur, "Qcur", il);
120            cb(Kcur, "Kcur", il);
121            cb(Vcur, "Vcur", il);
122
123            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
124            cb(Qcur, "Qcur", il);
125
126            cur = build_attn(inp_attn,
127                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
128                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
129        }
130        if (il == n_layer - 1 && inp_out_ids) {
131            cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
132            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
133        }
134        cur = ggml_add(ctx0, cur, residual);
135        residual = cur;
136
137        cur = build_norm(cur,
138                model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
139                LLM_NORM_RMS, il);
140        cb(cur, "ffn_norm", il);
141
142        // feed-forward network
143        if (model.layers[il].ffn_gate_inp == nullptr) {
144            cur = build_ffn(cur,
145                    model.layers[il].ffn_up,   NULL, NULL,
146                    NULL,                      NULL, NULL,
147                    model.layers[il].ffn_down, NULL, NULL,
148                    NULL,
149                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
150            cb(cur, "ffn_out", il);
151        } else {
152            // MoE branch
153            cur = build_moe_ffn(cur,
154                    model.layers[il].ffn_gate_inp,
155                    model.layers[il].ffn_up_exps,
156                    model.layers[il].ffn_gate_exps,
157                    model.layers[il].ffn_down_exps,
158                    nullptr,
159                    n_expert, n_expert_used,
160                    LLM_FFN_SILU, true,
161                    hparams.expert_weights_scale,
162                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
163                    il);
164            cb(cur, "ffn_moe_out", il);
165        }
166        cur = ggml_add(ctx0, residual, cur);
167
168        cur = build_cvec(cur, il);
169        cb(cur, "l_out", il);
170
171        // input for next layer
172        inpL = cur;
173    }
174    cur = build_norm(inpL,
175            model.output_norm,
176            model.output_norm_b,
177            LLM_NORM_RMS, -1);
178
179    cb(cur, "result_norm", -1);
180    res->t_embd = cur;
181
182    cur = build_lora_mm(model.output, cur, model.output_s);
183
184    if (model.output_b != nullptr) {
185        cb(cur, "result_output_no_bias", -1);
186        cur = ggml_add(ctx0, cur, model.output_b);
187    }
188    cb(cur, "result_output", -1);
189    res->t_logits = cur;
190
191    ggml_build_forward_expand(gf, cur);
192}
193
194// Explicit template instantiations
195template struct llama_model_phi3::graph<false>;
196template struct llama_model_phi3::graph<true>;