char-fork/llama.cpp

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

Arsen Arutunanmamba2: remove hardcoded 2x expansion factor and invalid d_inner % d_state check (#23082)960d628f4

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.4 KiB88 linesraw
1#include "models.h"
2
3void llama_model_mamba2::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_GROUP_COUNT,    hparams.ssm_n_group);
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                case 4096: type = LLM_TYPE_7B; break;
29                default: type = LLM_TYPE_UNKNOWN;
30            } break;
31        default: type = LLM_TYPE_UNKNOWN;
32    }
33}
34
35void llama_model_mamba2::load_arch_tensors(llama_model_loader &) {
36    LLAMA_LOAD_LOCALS;
37
38    const int64_t d_conv  = hparams.ssm_d_conv;
39    const int64_t d_inner = hparams.ssm_d_inner;
40    const int64_t d_state = hparams.ssm_d_state;
41    const int64_t n_group = hparams.ssm_n_group;
42    const int64_t dt_rank  = hparams.ssm_dt_rank;
43
44    const int64_t conv_dim = d_inner + 2 * n_group * d_state;
45    const int64_t d_in_proj = d_inner + conv_dim + dt_rank;
46
47
48    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
49
50    // output
51    {
52        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
53
54        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
55        // if output is NULL, init from the input tok embed, duplicated to allow offloading
56        if (output == NULL) {
57            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
58        }
59    }
60
61    for (int i = 0; i < n_layer; ++i) {
62        auto & layer = layers[i];
63
64        // norm
65        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
66
67        layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
68
69        layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
70        layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, 0);
71
72        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {dt_rank}, 0);
73
74        // no "weight" suffix for these
75        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, dt_rank}, 0);
76        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, dt_rank}, 0);
77
78        layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
79
80        // out_proj
81        layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
82    }
83}
84
85std::unique_ptr<llm_graph_context> llama_model_mamba2::build_arch_graph(const llm_graph_params & params) const {
86    return std::make_unique<graph>(*this, params);
87}
88