char-fork/llama.cpp

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

zqlmodels : support nanbeige4.2-3B (#27730)d7a207411

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.9 KiB185 linesraw
1#include "models.h"
2
3void llama_model_nanbeige::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    uint32_t n_loops_u = 1;
7    ml.get_key(LLM_KV_NUM_LOOPS, n_loops_u, false);
8    GGML_ASSERT(n_loops_u >= 1);
9
10    skip_loop_final_norm = false;
11    ml.get_key(LLM_KV_SKIP_LOOP_FINAL_NORM, skip_loop_final_norm, false);
12
13    n_layer_phys = (int) hparams.n_layer();
14
15    // Bound-check before casting: signed int mul can overflow and bypass the guard.
16    GGML_ASSERT((size_t) n_layer_phys * (size_t) n_loops_u <= (size_t) LLAMA_MAX_LAYERS);
17    n_loops = (int) n_loops_u;
18
19    // Expand logical layer count before load_tensors() allocates layers / KV.
20    if (n_loops > 1) {
21        for (int j = 1; j < n_loops; ++j) {
22            for (int i = 0; i < n_layer_phys; ++i) {
23                const int dst = i + j * n_layer_phys;
24                hparams.n_head_arr[dst]    = hparams.n_head_arr[i];
25                hparams.n_head_kv_arr[dst] = hparams.n_head_kv_arr[i];
26                hparams.n_ff_arr[dst]      = hparams.n_ff_arr[i];
27                hparams.is_swa_impl[dst]   = hparams.is_swa_impl[i];
28                hparams.is_recr_impl[dst]  = hparams.is_recr_impl[i];
29            }
30        }
31        hparams.n_layer_all = (uint32_t) ((size_t) n_layer_phys * (size_t) n_loops);
32    }
33
34    type = LLM_TYPE_UNKNOWN;
35}
36
37void llama_model_nanbeige::load_arch_tensors(llama_model_loader &) {
38    LLAMA_LOAD_LOCALS;
39
40    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
41
42    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
43    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
44    if (output == NULL) {
45        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
46    }
47
48    const int n_phys = n_layer_phys > 0 ? n_layer_phys : n_layer;
49    for (int i = 0; i < n_phys; ++i) {
50        auto & layer = layers[i];
51
52        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
53
54        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
55        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
56
57        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2},
58                TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
59
60        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
61        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
62        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
63        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
64    }
65
66    // Share physical weights across loops; each slot still has its own KV index.
67    if (n_loops > 1) {
68        for (int j = 1; j < n_loops; ++j) {
69            for (int i = 0; i < n_phys; ++i) {
70                layers[i + j * n_phys] = layers[i];
71            }
72        }
73    }
74}
75
76std::unique_ptr<llm_graph_context> llama_model_nanbeige::build_arch_graph(const llm_graph_params & params) const {
77    return std::make_unique<graph>(*this, params);
78}
79
80llama_model_nanbeige::graph::graph(const llama_model & model, const llm_graph_params & params) :
81        llm_graph_context(params) {
82    const auto & nb = static_cast<const llama_model_nanbeige &>(model);
83
84    const int64_t n_embd_head = hparams.n_embd_head_v();
85    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
86
87    const int n_phys  = nb.n_layer_phys > 0 ? nb.n_layer_phys : (int) n_layer;
88    const int n_loops = nb.n_loops > 0 ? nb.n_loops : 1;
89
90    ggml_tensor * cur;
91    ggml_tensor * inpL;
92
93    inpL = build_inp_embd(model.tok_embd);
94
95    ggml_tensor * inp_pos = build_inp_pos();
96
97    auto * inp_attn = build_attn_inp_kv();
98
99    const float kq_scale = hparams.f_attention_scale == 0.0f
100        ? 1.0f / sqrtf(float(n_embd_head))
101        : hparams.f_attention_scale;
102
103    ggml_tensor * inp_out_ids = build_inp_out_ids();
104
105    for (int il = 0; il < n_layer; ++il) {
106        res->t_layer_inp[il] = inpL;
107        ggml_tensor * inpSA = inpL;
108
109        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
110        cb(cur, "attn_norm", il);
111
112        {
113            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
114
115            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
116                    n_embd_head, n_head, n_head_kv, il);
117
118            Qcur = ggml_rope_ext(
119                    ctx0, Qcur, inp_pos, rope_factors,
120                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
121                    ext_factor, attn_factor, beta_fast, beta_slow);
122
123            Kcur = ggml_rope_ext(
124                    ctx0, Kcur, inp_pos, rope_factors,
125                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
126                    ext_factor, attn_factor, beta_fast, beta_slow);
127
128            cb(Qcur, "Qcur", il);
129            cb(Kcur, "Kcur", il);
130            cb(Vcur, "Vcur", il);
131
132            cur = build_attn(inp_attn,
133                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
134                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
135            cb(cur, "attn_out", il);
136        }
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        }
142
143        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
144        cb(ffn_inp, "ffn_inp", il);
145
146        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
147        cb(cur, "ffn_norm", il);
148
149        cur = build_ffn(cur,
150                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,
151                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,
152                model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
153                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
154        cb(cur, "ffn_out", il);
155
156        cur = ggml_add(ctx0, cur, ffn_inp);
157        cb(cur, "ffn_out", il);
158
159        cur = build_cvec(cur, il);
160        cb(cur, "l_out", il);
161
162        inpL = cur;
163
164        if (n_loops > 1 &&
165            ((il + 1) % n_phys) == 0 &&
166            (il + 1) < n_layer &&
167            !nb.skip_loop_final_norm) {
168            cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, il);
169            cb(cur, "loop_norm", il);
170            inpL = cur;
171        }
172    }
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    cur = build_lora_mm(model.output, cur, model.output_s);
181    cb(cur, "result_output", -1);
182    res->t_logits = cur;
183
184    ggml_build_forward_expand(gf, cur);
185}