char-fork/llama.cpp

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

Sigbjørn Skjæretmodel : load hparams.n_layer_nextn before n_layer() calls (#28159)9d817213a

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
7.6 KiB186 linesraw
1#include "models.h"
2
3void llama_model_glm4::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
5    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
6
7    switch (hparams.n_layer()) {
8        case 17: type = LLM_TYPE_1B; break; // GLM-OCR
9        case 40: type = LLM_TYPE_9B; break;
10        case 61: type = LLM_TYPE_32B; break;
11        default: type = LLM_TYPE_UNKNOWN;
12    }
13}
14
15void llama_model_glm4::load_arch_tensors(llama_model_loader &) {
16    LLAMA_LOAD_LOCALS;
17
18    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
19
20    // output
21    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
22    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
23    // if output is NULL, init from the input tok embed
24    if (output == NULL) {
25        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
26    }
27
28    for (int i = 0; i < n_layer_all; ++i) {
29        int flags = 0;
30        if (i >= n_layer) {
31            // skip all tensors in the NextN layers
32            flags |= TENSOR_SKIP;
33        }
34
35        auto & layer = layers[i];
36
37        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
38        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);
39
40        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);
41
42        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);
43
44        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
45        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
46        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, flags);
47
48        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);
49
50        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
51        if (i >= n_layer) {
52            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
53            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
54            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
55
56            // Optional tensors
57            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
58            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
59            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
60        }
61    }
62}
63
64std::unique_ptr<llm_graph_context> llama_model_glm4::build_arch_graph(const llm_graph_params & params) const {
65    return std::make_unique<graph>(*this, params);
66}
67
68llama_model_glm4::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
69    const int64_t n_embd_head = hparams.n_embd_head_v();
70
71    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
72
73    int sections[4];
74    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
75
76    ggml_tensor * cur;
77    ggml_tensor * inpL;
78
79    inpL = build_inp_embd(model.tok_embd);
80
81    bool use_mrope = hparams.use_mrope();
82    if (ubatch.embd && !use_mrope) {
83        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
84        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
85    }
86
87    // inp_pos - contains the positions
88    ggml_tensor * inp_pos = build_inp_pos();
89
90    auto * inp_attn = build_attn_inp_kv();
91
92    ggml_tensor * inp_out_ids = build_inp_out_ids();
93
94    // Only process up to last layer (skip final NextN layer)
95    // Final layer tensors are loaded but not processed in forward pass
96    for (int il = 0; il < n_layer; ++il) {
97        ggml_tensor * inpSA = inpL;
98
99        // Pre-attention norm
100        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
101        cb(cur, "attn_norm", il);
102
103        // self-attention
104        {
105            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
106                    n_embd_head, n_head, n_head_kv, il);
107
108            if (use_mrope) {
109                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
110                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
111                            ext_factor, attn_factor, beta_fast, beta_slow);
112
113                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
114                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
115                            ext_factor, attn_factor, beta_fast, beta_slow);
116            } else {
117                // Normal RoPE
118                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
119                                    rope_type, n_ctx_orig, freq_base, freq_scale,
120                                    ext_factor, attn_factor, beta_fast, beta_slow);
121
122                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
123                                    rope_type, n_ctx_orig, freq_base, freq_scale,
124                                    ext_factor, attn_factor, beta_fast, beta_slow);
125            }
126
127            cb(Qcur, "Qcur", il);
128            cb(Kcur, "Kcur", il);
129            cb(Vcur, "Vcur", il);
130
131            cur = build_attn(inp_attn,
132                    model.layers[il].wo, NULL, model.layers[il].wo_s,
133                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
134        }
135        if (il == n_layer - 1 && inp_out_ids) {
136            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
137            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
138        }
139        // Post-attention norm (new!)
140        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
141        cb(cur, "post_attn_norm", il);
142
143        // Add the input (residual connection after post-attention norm)
144        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
145        cb(ffn_inp, "ffn_inp", il);
146
147        // FF
148        {
149            // Pre-MLP norm
150            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
151            cb(cur, "ffn_norm", il);
152
153            // MLP
154            cur = build_ffn(cur,
155                    model.layers[il].ffn_up, NULL, NULL,
156                    NULL, NULL, NULL,
157                    model.layers[il].ffn_down, NULL, NULL,
158                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
159            cb(cur, "ffn_out", il);
160
161            // Post-MLP norm
162            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
163            cb(cur, "post_mlp_norm", il);
164        }
165        cur = ggml_add(ctx0, cur, ffn_inp);
166
167        cur = build_cvec(cur, il);
168        cb(cur, "l_out", il);
169
170        // input for next layer
171        inpL = cur;
172    }
173    // Final norm
174    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
175
176    cb(cur, "result_norm", -1);
177    res->t_embd = cur;
178
179    // Output projection
180    cur = build_lora_mm(model.output, cur, model.output_s);
181
182    cb(cur, "result_output", -1);
183    res->t_logits = cur;
184
185    ggml_build_forward_expand(gf, cur);
186}