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 →
4.8 KiB137 linesraw
1#include "models.h"
2
3void llama_model_mamba::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
5    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
6    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
7    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
8    ml.get_key(LLM_KV_SSM_DT_B_C_RMS,     hparams.ssm_dt_b_c_rms, false);
9
10    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
11
12    switch (hparams.n_layer()) {
13        case 24:
14            switch (hparams.n_embd) {
15                case 768: type = LLM_TYPE_SMALL; break;
16                default: type = LLM_TYPE_UNKNOWN;
17            } break;
18        case 48:
19            switch (hparams.n_embd) {
20                case 1024: type = LLM_TYPE_MEDIUM; break;
21                case 1536: type = LLM_TYPE_LARGE; break;
22                case 2048: type = LLM_TYPE_XL; break;
23                default:   type = LLM_TYPE_UNKNOWN;
24            } break;
25        case 64:
26            switch (hparams.n_embd) {
27                case 2560: type = LLM_TYPE_3B; break;
28                default: type = LLM_TYPE_UNKNOWN;
29            } break;
30        default: type = LLM_TYPE_UNKNOWN;
31    }
32}
33
34void llama_model_mamba::load_arch_tensors(llama_model_loader &) {
35    LLAMA_LOAD_LOCALS;
36
37    const int64_t d_conv  = hparams.ssm_d_conv;
38    const int64_t d_inner = hparams.ssm_d_inner;
39    const int64_t d_state = hparams.ssm_d_state;
40    const int64_t dt_rank = hparams.ssm_dt_rank;
41
42    // only an expansion factor of 2 is supported for now
43    if (2 * n_embd != d_inner) {
44        throw std::runtime_error("only an expansion factor of 2 is supported for now");
45    }
46
47    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
48
49    // output
50    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
51
52    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
53    // if output is NULL, init from the input tok embed, duplicated to allow offloading
54    if (output == NULL) {
55        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
56    }
57
58    for (int i = 0; i < n_layer; ++i) {
59        auto & layer = layers[i];
60
61        // norm
62        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
63
64        layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, 2*d_inner}, 0);
65
66        layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner}, 0);
67        layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner}, 0);
68
69        layer.ssm_x = create_tensor(tn(LLM_TENSOR_SSM_X, "weight", i), {d_inner, dt_rank + 2*d_state}, 0);
70
71        layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "weight", i), {dt_rank, d_inner}, 0);
72        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}, 0);
73
74        // no "weight" suffix for these
75        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}, 0);
76        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {d_inner}, 0);
77
78        // out_proj
79        layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
80    }
81}
82
83std::unique_ptr<llm_graph_context> llama_model_mamba::build_arch_graph(const llm_graph_params & params) const {
84    return std::make_unique<graph>(*this, params);
85}
86
87llama_model_mamba::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
88    ggml_tensor * cur;
89    ggml_tensor * inpL;
90
91    // {n_embd, n_tokens}
92    inpL = build_inp_embd(model.tok_embd);
93
94    auto * rs_inp = build_rs_inp();
95
96    ggml_tensor * inp_out_ids = build_inp_out_ids();
97
98    for (int il = 0; il < n_layer; ++il) {
99        // norm
100        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
101        cb(cur, "attn_norm", il);
102
103        if (model.arch == LLM_ARCH_MAMBA2) {
104            cur = build_mamba2_layer(rs_inp, cur, model, ubatch, il);
105        } else {
106            cur = build_mamba_layer(rs_inp, cur, model, ubatch, il);
107        }
108
109        if (il == n_layer - 1 && inp_out_ids) {
110            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
111            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
112        }
113
114        // residual
115        cur = ggml_add(ctx0, cur, inpL);
116
117        cur = build_cvec(cur, il);
118        cb(cur, "l_out", il);
119
120        // input for next layer
121        inpL = cur;
122    }
123
124    // final rmsnorm
125    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
126
127    cb(cur, "result_norm", -1);
128    res->t_embd = cur;
129
130    // lm_head
131    cur = build_lora_mm(model.output, cur, model.output_s);
132
133    cb(cur, "result_output", -1);
134    res->t_logits = cur;
135
136    ggml_build_forward_expand(gf, cur);
137}