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 →
7.5 KiB190 linesraw
1#include "models.h"
2
3void llama_model_smallthinker::load_arch_hparams(llama_model_loader & ml) {
4    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
5
6    if (found_swa && hparams.n_swa > 0) {
7        hparams.swa_type    = LLAMA_SWA_TYPE_STANDARD;
8        hparams.n_swa       = 4096;
9        uint32_t swa_period = 4;
10        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
11        hparams.set_swa_pattern(swa_period, true);
12
13        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
14        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
15        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
16    } else {
17        hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
18        hparams.n_no_rope_layer_step = hparams.n_layer();
19    }
20
21    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
22    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
23    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
24
25    switch (hparams.n_layer()) {
26        case 32: type = LLM_TYPE_4B;  break;
27        case 52: type = LLM_TYPE_20B; break;
28        default: type = LLM_TYPE_UNKNOWN;
29    }
30}
31
32void llama_model_smallthinker::load_arch_tensors(llama_model_loader &) {
33    LLAMA_LOAD_LOCALS;
34
35    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
36
37    // output
38    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
39    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
40
41    // if output is NULL, init from the input tok embed
42    if (output == NULL) {
43        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
44    }
45
46    for (int i = 0; i < n_layer; ++i) {
47        auto & layer = layers[i];
48
49        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
50
51        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
52        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
53
54        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
55
56        GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for SMALLTHINKER");
57        GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for SMALLTHINKER");
58
59        // MoE branch
60        const int64_t n_ff_exp = hparams.n_ff_exp();
61        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0);
62        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
63        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
64        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
65    }
66}
67
68std::unique_ptr<llm_graph_context> llama_model_smallthinker::build_arch_graph(const llm_graph_params & params) const {
69    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
70        return std::make_unique<graph<true>> (*this, params);
71    } else {
72        return std::make_unique<graph<false>>(*this, params);
73    }
74}
75
76template <bool iswa>
77llama_model_smallthinker::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
78    const int64_t n_embd_head = hparams.n_embd_head_v();
79
80    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
81    GGML_ASSERT(n_embd_head == n_rot);
82
83    ggml_tensor * cur;
84    ggml_tensor * inpL;
85
86    inpL = build_inp_embd(model.tok_embd);
87
88    // inp_pos - contains the positions
89    ggml_tensor * inp_pos = build_inp_pos();
90
91    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
92    inp_attn_type * inp_attn = nullptr;
93
94    if constexpr (iswa) {
95        inp_attn = build_attn_inp_kv_iswa();
96    } else {
97        inp_attn = build_attn_inp_kv();
98    }
99    ggml_tensor * inp_out_ids = build_inp_out_ids();
100
101    for (int il = 0; il < n_layer; ++il) {
102        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
103        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
104
105        ggml_tensor * inpSA  = inpL;
106
107        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
108        const bool use_rope = hparams.n_no_rope_layer_step == n_layer ||
109                              il % hparams.n_no_rope_layer_step != 0;
110
111        ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
112        cb(probs, "ffn_moe_logits", il);
113
114        // norm
115        cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
116        cb(cur, "attn_norm", il);
117
118        // self_attention
119        {
120            // compute Q and K and RoPE them
121            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
122                    n_embd_head, n_head, n_head_kv, il);
123
124            if (use_rope) {
125                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
126                                    ext_factor, attn_factor, beta_fast, beta_slow);
127
128                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
129                                    ext_factor, attn_factor, beta_fast, beta_slow);
130            }
131            cb(Qcur, "Qcur", il);
132            cb(Kcur, "Kcur", il);
133
134            cur = build_attn(inp_attn,
135                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
136                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
137        }
138        if (il == n_layer - 1 && inp_out_ids) {
139            cur = ggml_get_rows(ctx0, cur, inp_out_ids);
140            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
141            probs = ggml_get_rows(ctx0, probs, inp_out_ids);
142        }
143        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
144        cb(ffn_inp, "ffn_inp", il);
145
146        // MoE branch
147        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
148        cb(cur, "ffn_norm", il);
149
150        ggml_tensor * ffn_out =
151            build_moe_ffn(cur,
152                    nullptr,
153                    model.layers[il].ffn_up_exps,
154                    model.layers[il].ffn_gate_exps,
155                    model.layers[il].ffn_down_exps,
156                    nullptr,
157                    n_expert, n_expert_used,
158                    LLM_FFN_RELU, true,
159                    hparams.expert_weights_scale,
160                    static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
161                    il, probs);
162
163        cb(ffn_out, "ffn_out", il);
164        cur = ffn_out;
165
166        cur = ggml_add(ctx0, cur, ffn_inp);
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 = inpL;
175
176    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
177    cb(cur, "result_norm", -1);
178    res->t_embd = cur;
179
180    // lm_head
181    cur = build_lora_mm(model.output, cur, model.output_s);
182    cb(cur, "result_output", -1);
183    res->t_logits = cur;
184
185    ggml_build_forward_expand(gf, cur);
186}
187
188// Explicit template instantiations
189template struct llama_model_smallthinker::graph<false>;
190template struct llama_model_smallthinker::graph<true>;