char-fork/llama.cpp

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

PikaPikachuconvert : add `--fuse-qkv` flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (#22780)465e49b9c

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.0 KiB173 linesraw
1#include "models.h"
2
3void llama_model_olmoe::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 16: type = LLM_TYPE_A1_7B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10}
11
12void llama_model_olmoe::load_arch_tensors(llama_model_loader &) {
13    LLAMA_LOAD_LOCALS;
14
15    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
16
17    // output
18    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
19    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
20
21    for (int i = 0; i < n_layer; ++i) {
22        auto & layer = layers[i];
23
24        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
25
26        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
27        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
28        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, 0);
29        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, 0);
30
31        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
32
33        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
34
35        if (n_expert == 0) {
36            throw std::runtime_error("n_expert must be > 0");
37        }
38        if (n_expert_used == 0) {
39            throw std::runtime_error("n_expert_used must be > 0");
40        }
41
42        // MoE branch
43        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
44        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
45        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
46    }
47}
48
49std::unique_ptr<llm_graph_context> llama_model_olmoe::build_arch_graph(const llm_graph_params & params) const {
50    return std::make_unique<graph>(*this, params);
51}
52
53llama_model_olmoe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
54    const int64_t n_embd_head = hparams.n_embd_head_v();
55
56    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
57    GGML_ASSERT(n_embd_head == n_rot);
58
59    ggml_tensor * cur;
60    ggml_tensor * inpL;
61
62    inpL = build_inp_embd(model.tok_embd);
63
64    // inp_pos - contains the positions
65    ggml_tensor * inp_pos = build_inp_pos();
66
67    auto * inp_attn = build_attn_inp_kv();
68
69    ggml_tensor * inp_out_ids = build_inp_out_ids();
70
71    for (int il = 0; il < n_layer; ++il) {
72        ggml_tensor * inpSA = inpL;
73
74        // norm
75        cur = build_norm(inpL,
76                model.layers[il].attn_norm, NULL,
77                LLM_NORM_RMS, il);
78        cb(cur, "attn_norm", il);
79
80        // self_attention
81        {
82            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
83                    n_embd_head, n_head,
84                    n_embd_head, n_head_kv,
85                    n_embd_head, n_head_kv,
86                    il, false);
87            cb(Qcur, "Qcur", il);
88            cb(Kcur, "Kcur", il);
89            cb(Vcur, "Vcur", il);
90
91            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
92                    LLM_NORM_RMS, il);
93            cb(Qcur, "Qcur_normed", il);
94
95            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
96                    LLM_NORM_RMS, il);
97            cb(Kcur, "Kcur_normed", il);
98
99            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
100            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
101            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
102
103            Qcur = ggml_rope_ext(
104                    ctx0, Qcur, inp_pos, nullptr,
105                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
106                    ext_factor, attn_factor, beta_fast, beta_slow
107                    );
108
109            Kcur = ggml_rope_ext(
110                    ctx0, Kcur, inp_pos, nullptr,
111                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
112                    ext_factor, attn_factor, beta_fast, beta_slow
113                    );
114
115            cb(Qcur, "Qcur", il);
116            cb(Kcur, "Kcur", il);
117            cb(Vcur, "Vcur", il);
118
119            cur = build_attn(inp_attn,
120                    model.layers[il].wo, NULL, model.layers[il].wo_s,
121                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
122        }
123        if (il == n_layer - 1 && inp_out_ids) {
124            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
125            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
126        }
127        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
128        cb(ffn_inp, "ffn_inp", il);
129
130        // MoE branch
131        cur = build_norm(ffn_inp,
132                model.layers[il].ffn_norm, NULL,
133                LLM_NORM_RMS, il);
134        cb(cur, "ffn_norm", il);
135
136        cur = build_moe_ffn(cur,
137                model.layers[il].ffn_gate_inp,
138                model.layers[il].ffn_up_exps,
139                model.layers[il].ffn_gate_exps,
140                model.layers[il].ffn_down_exps,
141                nullptr,
142                n_expert, n_expert_used,
143                LLM_FFN_SILU, false,
144                hparams.expert_weights_scale,
145                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
146                il);
147        cb(cur, "ffn_moe_out", il);
148
149        cur = ggml_add(ctx0, cur, ffn_inp);
150
151        cur = build_cvec(cur, il);
152        cb(cur, "l_out", il);
153
154        // input for next layer
155        inpL = cur;
156    }
157    cur = inpL;
158
159    cur = build_norm(cur,
160            model.output_norm, NULL,
161            LLM_NORM_RMS, -1);
162
163    cb(cur, "result_norm", -1);
164    res->t_embd = cur;
165
166    // lm_head
167    cur = build_lora_mm(model.output, cur, model.output_s);
168
169    cb(cur, "result_output", -1);
170    res->t_logits = cur;
171
172    ggml_build_forward_expand(gf, cur);
173}