char-fork/llama.cpp

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

Pedro Cuencamodel: Muse Glimmer Support (#26841)62bf73d25

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
8.6 KiB208 linesraw
1#include "models.h"
2
3void llama_model_muse_glimmer::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(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
6    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);
7    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
8
9    hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;
10    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
11
12    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
13    uint32_t swa_period = 4;
14    if (ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false)) {
15        hparams.set_swa_pattern(swa_period);
16    } else {
17        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());
18    }
19
20    switch (hparams.n_layer()) {
21        case 52: type = LLM_TYPE_30B; break;
22        default: type = LLM_TYPE_UNKNOWN;
23    }
24}
25
26void llama_model_muse_glimmer::load_arch_tensors(llama_model_loader &) {
27    LLAMA_LOAD_LOCALS;
28
29    tok_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
30    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
31    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
32
33    for (int i = 0; i < n_layer; ++i) {
34        auto & layer = layers[i];
35
36        // Pre/post-attention norms (Muse Glimmer's `weight + 1` applied at conversion time).
37        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);
38        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
39
40        // Q/K/V/O projections.
41        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
42        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
43
44        // QK-norm. Weights are synthesized at conversion time to absorb `qk_scale_factor`.
45        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
46        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
47
48        // Attention output gate: sigmoid(gate) * attn_out before o_proj (same as afmoe).
49        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
50
51        // Pre/post-FFN norms (FFN_PRE_NORM is aliased to LLM_TENSOR_FFN_NORM).
52        layer.ffn_norm      = create_tensor(tn(LLM_TENSOR_FFN_NORM,      "weight", i), {n_embd}, 0);
53        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
54
55        // Dense FFN (unlike afmoe, no MoE branches).
56        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
57        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
58        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
59    }
60}
61
62llama_model_muse_glimmer::graph::graph(const llama_model & model, const llm_graph_params & params)
63    : llm_graph_context(params) {
64    const int64_t n_embd_head = hparams.n_embd_head_v();
65    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
66
67    // Different to f_norm_rms_eps for post-attn / post-FFN norms
68    const float post_norm_eps = 1e-8f;
69
70    ggml_tensor * cur;
71    ggml_tensor * inpL;
72
73    inpL = build_inp_embd(model.tok_embd);
74    inpL = build_norm(inpL, nullptr, nullptr, LLM_NORM_RMS, -1);
75    cb(inpL, "embd_norm", -1);
76
77    ggml_tensor * inp_pos = build_inp_pos();
78    auto * inp_attn = build_attn_inp_kv_iswa();
79    ggml_tensor * inp_out_ids = build_inp_out_ids();
80
81    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
82
83    for (int il = 0; il < n_layer; ++il) {
84        // expose per-layer residual for speculative drafts (see LLM_KV_TARGET_LAYERS).
85        res->t_layer_inp[il] = inpL;
86
87        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
88        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
89
90        ggml_tensor * inpSA = inpL;
91
92        // RoPE runs on the SWA layers, NoPE on full ones.
93        const bool use_rope = hparams.is_swa(il);
94
95        // pre-attention norm (weight+1 folded at conversion time)
96        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
97        cb(cur, "attn_norm", il);
98
99        // self-attention: attention output gate around SDPA (afmoe.cpp:147-191)
100        {
101            ggml_tensor * attn_inp = cur;  // save input for gate computation
102
103            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
104                    n_embd_head, n_head, n_head_kv, il);
105
106            // gate = wqkv_gate @ attn_inp (from pre-attn hidden state)
107            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
108            cb(gate, "attn_gate_proj", il);
109
110            // QK-norm. attn_q_norm weight was synthesized at conversion to broadcast
111            // qk_scale_factor across head_dim; attn_k_norm is identity (ones).
112            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
113            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
114            cb(Qcur, "Qcur_normed", il);
115            cb(Kcur, "Kcur_normed", il);
116
117            if (use_rope) {
118                Qcur = ggml_rope_ext(
119                        ctx0, Qcur, inp_pos, nullptr,
120                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
121                        ext_factor, attn_factor, beta_fast, beta_slow);
122                cb(Qcur, "Qcur_rope", il);
123
124                Kcur = ggml_rope_ext(
125                        ctx0, Kcur, inp_pos, nullptr,
126                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
127                        ext_factor, attn_factor, beta_fast, beta_slow);
128                cb(Kcur, "Kcur_rope", il);
129            }
130
131            // SDPA. wo is deferred; the gate goes between attn_out and o_proj.
132            cur = build_attn(inp_attn,
133                    NULL, NULL, NULL,
134                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
135            cb(cur, "attn_out", il);
136
137            gate = ggml_sigmoid(ctx0, gate);
138            cb(gate, "attn_gate_sig", il);
139            cur = ggml_mul(ctx0, cur, gate);
140            cb(cur, "attn_gated", il);
141
142            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
143            cb(cur, "attn_o_proj", il);
144        }
145
146        cur = ggml_rms_norm(ctx0, cur, post_norm_eps);
147        cur = ggml_mul(ctx0, cur, model.layers[il].attn_post_norm);
148        cb(cur, "attn_post_norm", il);
149
150        if (il == n_layer - 1 && inp_out_ids) {
151            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
152            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
153        }
154
155        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
156        cb(ffn_inp, "ffn_inp", il);
157
158        // pre-FFN norm
159        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
160        cb(cur, "ffn_norm", il);
161
162        // SwiGLU dense FFN
163        cur = build_ffn(cur,
164                model.layers[il].ffn_up,   NULL, NULL,
165                model.layers[il].ffn_gate, NULL, NULL,
166                model.layers[il].ffn_down, NULL, NULL,
167                NULL,
168                LLM_FFN_SILU, LLM_FFN_PAR, il);
169        cb(cur, "ffn_out", il);
170
171        cur = ggml_rms_norm(ctx0, cur, post_norm_eps);
172        cur = ggml_mul(ctx0, cur, model.layers[il].ffn_post_norm);
173        cb(cur, "ffn_post_norm", il);
174
175        cur = ggml_add(ctx0, cur, ffn_inp);
176        cur = build_cvec(cur, il);
177        cb(cur, "l_out", il);
178
179        inpL = cur;
180    }
181
182    cur = inpL;
183
184    // final norm
185    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
186    cb(cur, "result_norm", -1);
187    res->t_embd = cur;
188
189    // lm_head, followed by output multiplier
190    cur = build_lora_mm(model.output, cur, model.output_s);
191    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
192
193    // Final logit tanh softcap (from gemma3.cpp).
194    if (hparams.f_final_logit_softcapping) {
195        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
196        cur = ggml_tanh(ctx0, cur);
197        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
198    }
199
200    cb(cur, "result_output", -1);
201    res->t_logits = cur;
202
203    ggml_build_forward_expand(gf, cur);
204}
205
206std::unique_ptr<llm_graph_context> llama_model_muse_glimmer::build_arch_graph(const llm_graph_params & params) const {
207    return std::make_unique<graph>(*this, params);
208}