char-fork/llama.cpp

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

Yaniss Amazouzmodel: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)c61b98b87

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
7.2 KiB194 linesraw
1#include "models.h"
2
3void llama_model_qwen2moe::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
5    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
6
7    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
8
9    switch (hparams.n_layer()) {
10        case 24: type = LLM_TYPE_A2_7B; break;
11        case 28: type = LLM_TYPE_57B_A14B; break;
12        default: type = LLM_TYPE_UNKNOWN;
13    }
14}
15
16void llama_model_qwen2moe::load_arch_tensors(llama_model_loader &) {
17    LLAMA_LOAD_LOCALS;
18
19    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
20
21    // output
22    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
23    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
24
25    for (int i = 0; i < n_layer; ++i) {
26        auto & layer = layers[i];
27
28        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
29
30        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
31        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
32
33        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
34
35        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
36
37        if (n_expert == 0) {
38            throw std::runtime_error("n_expert must be > 0 for QWEN2MOE");
39        }
40        if (n_expert_used == 0) {
41            throw std::runtime_error("n_expert_used must be > 0 for QWEN2MOE");
42        }
43
44        // MoE branch
45        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
46
47        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
48        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
49        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
50
51        // Shared expert branch
52        const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
53
54        layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), {n_embd}, 0);
55        layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, n_ff_shexp}, 0);
56        layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp,     n_embd}, 0);
57        layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, n_ff_shexp}, 0);
58    }
59}
60
61std::unique_ptr<llm_graph_context> llama_model_qwen2moe::build_arch_graph(const llm_graph_params & params) const {
62    return std::make_unique<graph>(*this, params);
63}
64
65llama_model_qwen2moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
66    const int64_t n_embd_head = hparams.n_embd_head_v();
67
68    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
69    GGML_ASSERT(n_embd_head == n_rot);
70
71    ggml_tensor * cur;
72    ggml_tensor * inpL;
73
74    inpL = build_inp_embd(model.tok_embd);
75
76    // inp_pos - contains the positions
77    ggml_tensor * inp_pos = build_inp_pos();
78
79    auto * inp_attn = build_attn_inp_kv();
80
81    ggml_tensor * inp_out_ids = build_inp_out_ids();
82
83    for (int il = 0; il < n_layer; ++il) {
84        ggml_tensor * inpSA = inpL;
85
86        // norm
87        cur = build_norm(inpL,
88                model.layers[il].attn_norm, NULL,
89                LLM_NORM_RMS, il);
90        cb(cur, "attn_norm", il);
91
92        // self_attention
93        {
94            // compute Q and K and RoPE them
95            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
96                    n_embd_head, n_head, n_head_kv, il);
97
98            Qcur = ggml_rope_ext(
99                    ctx0, Qcur, inp_pos, nullptr,
100                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
101                    ext_factor, attn_factor, beta_fast, beta_slow
102                    );
103
104            Kcur = ggml_rope_ext(
105                    ctx0, Kcur, inp_pos, nullptr,
106                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
107                    ext_factor, attn_factor, beta_fast, beta_slow
108                    );
109
110            cb(Qcur, "Qcur", il);
111            cb(Kcur, "Kcur", il);
112            cb(Vcur, "Vcur", il);
113
114            cur = build_attn(inp_attn,
115                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
116                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
117        }
118        if (il == n_layer - 1 && inp_out_ids) {
119            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
120            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
121        }
122        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
123        cb(ffn_inp, "ffn_inp", il);
124
125        // MoE branch
126        cur = build_norm(ffn_inp,
127                model.layers[il].ffn_norm, NULL,
128                LLM_NORM_RMS, il);
129        cb(cur, "ffn_norm", il);
130
131        ggml_tensor * moe_out =
132            build_moe_ffn(cur,
133                    model.layers[il].ffn_gate_inp,
134                    model.layers[il].ffn_up_exps,
135                    model.layers[il].ffn_gate_exps,
136                    model.layers[il].ffn_down_exps,
137                    nullptr,
138                    n_expert, n_expert_used,
139                    LLM_FFN_SILU, false,
140                    hparams.expert_weights_scale,
141                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
142                    il);
143        cb(moe_out, "ffn_moe_out", il);
144
145        // FFN shared expert
146        {
147            ggml_tensor * cur_gate_inp = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
148            cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
149
150            // sigmoid
151            ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
152            cb(cur_gate, "ffn_shexp_gate", il);
153
154            ggml_tensor * cur_ffn = build_ffn(cur,
155                    model.layers[il].ffn_up_shexp,   NULL, NULL,
156                    model.layers[il].ffn_gate_shexp, NULL, NULL,
157                    model.layers[il].ffn_down_shexp, NULL, NULL,
158                    NULL,
159                    LLM_FFN_SILU, LLM_FFN_PAR, il);
160            cb(cur_ffn, "ffn_shexp", il);
161
162            ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
163            cb(ffn_shexp_out, "ffn_shexp_out", il);
164
165            moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
166            cb(moe_out, "ffn_out", il);
167
168            cur = moe_out;
169        }
170        cur = ggml_add(ctx0, cur, ffn_inp);
171
172        cur = build_cvec(cur, il);
173        cb(cur, "l_out", il);
174
175        // input for next layer
176        inpL = cur;
177    }
178    cur = inpL;
179
180    cur = build_norm(cur,
181            model.output_norm, NULL,
182            LLM_NORM_RMS, -1);
183
184    cb(cur, "result_norm", -1);
185    res->t_embd = cur;
186
187    // lm_head
188    cur = build_lora_mm(model.output, cur, model.output_s);
189
190    cb(cur, "result_output", -1);
191    res->t_logits = cur;
192
193    ggml_build_forward_expand(gf, cur);
194}