/src/llama.cpp/src/models/glm4-moe.cpp

Source
#include "models.h"

llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
    const int64_t n_embd_head = hparams.n_embd_head_v;

    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);

    int sections[4];
    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);

    ggml_tensor * cur;
    ggml_tensor * inpL;

    inpL = build_inp_embd(model.tok_embd);

    bool use_mrope = hparams.use_mrope();
    if (ubatch.embd && !use_mrope) {
        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
    }

    // inp_pos - contains the positions
    ggml_tensor * inp_pos = build_inp_pos();

    auto * inp_attn = build_attn_inp_kv();

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    // Only process up to last layer (skip final NextN layer)
    // Final layer tensors are loaded but not processed in forward pass
    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
    for (int il = 0; il < n_transformer_layers; ++il) {
        ggml_tensor * inpSA = inpL;

        // Pre-attention norm
        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
        cb(cur, "attn_norm", il);

        // self-attention
        {
            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
            if (model.layers[il].bq) {
                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
            }
            cb(Qcur, "Qcur", il);

            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
            if (model.layers[il].bk) {
                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
            }
            cb(Kcur, "Kcur", il);

            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
            if (model.layers[il].bv) {
                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
            }
            cb(Vcur, "Vcur", il);

            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);

            // Apply Q/K norm if available (GLM-4.5 355B variant)
            if (model.layers[il].attn_q_norm) {
                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
                cb(Qcur, "Qcur_normed", il);
            }
            if (model.layers[il].attn_k_norm) {
                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
                cb(Kcur, "Kcur_normed", il);
            }

            if (use_mrope) {
                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                            ext_factor, attn_factor, beta_fast, beta_slow);

                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                            ext_factor, attn_factor, beta_fast, beta_slow);
            } else {
                // Normal RoPE
                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
                                    rope_type, n_ctx_orig, freq_base, freq_scale,
                                    ext_factor, attn_factor, beta_fast, beta_slow);

                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
                                    rope_type, n_ctx_orig, freq_base, freq_scale,
                                    ext_factor, attn_factor, beta_fast, beta_slow);
            }

            cb(Qcur, "Qcur", il);
            cb(Kcur, "Kcur", il);
            cb(Vcur, "Vcur", il);

            cur = build_attn(inp_attn,
                    model.layers[il].wo, NULL,
                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
        }
        if (il == n_transformer_layers - 1 && inp_out_ids) {
            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
        }
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
        cb(ffn_inp, "ffn_inp", il);

        // Post-attention norm
        cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
        cb(cur, "post_attn_norm", il);

        // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
            // Dense FFN layer
            cur = build_ffn(cur,
                    model.layers[il].ffn_up,   NULL, NULL,
                    model.layers[il].ffn_gate, NULL, NULL,
                    model.layers[il].ffn_down, NULL, NULL,
                    NULL,
                    LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(cur, "ffn_out", il);
        } else {
            // Process routed experts using existing MoE infrastructure
            ggml_tensor * routed_out = build_moe_ffn(cur,
                    model.layers[il].ffn_gate_inp,
                    model.layers[il].ffn_up_exps,
                    model.layers[il].ffn_gate_exps,
                    model.layers[il].ffn_down_exps,
                    model.layers[il].ffn_exp_probs_b,
                    n_expert, n_expert_used,
                    LLM_FFN_SILU, hparams.expert_weights_norm,
                    true, hparams.expert_weights_scale,
                    (llama_expert_gating_func_type) hparams.expert_gating_func,
                    il);
            cb(routed_out, "ffn_moe_out", il);

            // Process shared expert on original input
            ggml_tensor * shared_out = build_ffn(cur,
                    model.layers[il].ffn_up_shexp,   NULL, NULL,
                    model.layers[il].ffn_gate_shexp, NULL, NULL,
                    model.layers[il].ffn_down_shexp, NULL, NULL,
                    NULL,
                    LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(shared_out, "ffn_shexp_out", il);

            // Final output: routed_output + shared_output
            cur = ggml_add(ctx0, routed_out, shared_out);
            cb(cur, "ffn_out", il);
        }
        cur = ggml_add(ctx0, cur, ffn_inp);

        cur = build_cvec(cur, il);
        cb(cur, "l_out", il);

        // input for next layer
        inpL = cur;
    }
    cur = inpL;
    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);

    cb(cur, "result_norm", -1);
    res->t_embd = cur;

    // lm_head
    cur = build_lora_mm(model.output, cur);

    cb(cur, "result_output", -1);
    res->t_logits = cur;

    ggml_build_forward_expand(gf, cur);
}

Line	Count	Source
1		#include "models.h"
2
3	0	llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4	0	const int64_t n_embd_head = hparams.n_embd_head_v;
5
6	0	GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
7
8	0	int sections[4];
9	0	std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
10
11	0	ggml_tensor * cur;
12	0	ggml_tensor * inpL;
13
14	0	inpL = build_inp_embd(model.tok_embd);
15
16	0	bool use_mrope = hparams.use_mrope();
17	0	if (ubatch.embd && !use_mrope) {
18		// unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
19	0	GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
20	0	}
21
22		// inp_pos - contains the positions
23	0	ggml_tensor * inp_pos = build_inp_pos();
24
25	0	auto * inp_attn = build_attn_inp_kv();
26
27	0	ggml_tensor * inp_out_ids = build_inp_out_ids();
28
29		// Only process up to last layer (skip final NextN layer)
30		// Final layer tensors are loaded but not processed in forward pass
31	0	const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
32	0	for (int il = 0; il < n_transformer_layers; ++il) {
33	0	ggml_tensor * inpSA = inpL;
34
35		// Pre-attention norm
36	0	cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
37	0	cb(cur, "attn_norm", il);
38
39		// self-attention
40	0	{
41	0	ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
42	0	if (model.layers[il].bq) {
43	0	Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
44	0	}
45	0	cb(Qcur, "Qcur", il);
46
47	0	ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
48	0	if (model.layers[il].bk) {
49	0	Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
50	0	}
51	0	cb(Kcur, "Kcur", il);
52
53	0	ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
54	0	if (model.layers[il].bv) {
55	0	Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
56	0	}
57	0	cb(Vcur, "Vcur", il);
58
59	0	Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
60	0	Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
61	0	Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
62
63		// Apply Q/K norm if available (GLM-4.5 355B variant)
64	0	if (model.layers[il].attn_q_norm) {
65	0	Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
66	0	cb(Qcur, "Qcur_normed", il);
67	0	}
68	0	if (model.layers[il].attn_k_norm) {
69	0	Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
70	0	cb(Kcur, "Kcur_normed", il);
71	0	}
72
73	0	if (use_mrope) {
74	0	Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
75	0	n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
76	0	ext_factor, attn_factor, beta_fast, beta_slow);
77
78	0	Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
79	0	n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
80	0	ext_factor, attn_factor, beta_fast, beta_slow);
81	0	} else {
82		// Normal RoPE
83	0	Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
84	0	rope_type, n_ctx_orig, freq_base, freq_scale,
85	0	ext_factor, attn_factor, beta_fast, beta_slow);
86
87	0	Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
88	0	rope_type, n_ctx_orig, freq_base, freq_scale,
89	0	ext_factor, attn_factor, beta_fast, beta_slow);
90	0	}
91
92	0	cb(Qcur, "Qcur", il);
93	0	cb(Kcur, "Kcur", il);
94	0	cb(Vcur, "Vcur", il);
95
96	0	cur = build_attn(inp_attn,
97	0	model.layers[il].wo, NULL,
98	0	Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
99	0	}
100	0	if (il == n_transformer_layers - 1 && inp_out_ids) {
101	0	cur = ggml_get_rows(ctx0, cur, inp_out_ids);
102	0	inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
103	0	}
104	0	ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
105	0	cb(ffn_inp, "ffn_inp", il);
106
107		// Post-attention norm
108	0	cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
109	0	cb(cur, "post_attn_norm", il);
110
111		// Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
112	0	if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
113		// Dense FFN layer
114	0	cur = build_ffn(cur,
115	0	model.layers[il].ffn_up, NULL, NULL,
116	0	model.layers[il].ffn_gate, NULL, NULL,
117	0	model.layers[il].ffn_down, NULL, NULL,
118	0	NULL,
119	0	LLM_FFN_SILU, LLM_FFN_PAR, il);
120	0	cb(cur, "ffn_out", il);
121	0	} else {
122		// Process routed experts using existing MoE infrastructure
123	0	ggml_tensor * routed_out = build_moe_ffn(cur,
124	0	model.layers[il].ffn_gate_inp,
125	0	model.layers[il].ffn_up_exps,
126	0	model.layers[il].ffn_gate_exps,
127	0	model.layers[il].ffn_down_exps,
128	0	model.layers[il].ffn_exp_probs_b,
129	0	n_expert, n_expert_used,
130	0	LLM_FFN_SILU, hparams.expert_weights_norm,
131	0	true, hparams.expert_weights_scale,
132	0	(llama_expert_gating_func_type) hparams.expert_gating_func,
133	0	il);
134	0	cb(routed_out, "ffn_moe_out", il);
135
136		// Process shared expert on original input
137	0	ggml_tensor * shared_out = build_ffn(cur,
138	0	model.layers[il].ffn_up_shexp, NULL, NULL,
139	0	model.layers[il].ffn_gate_shexp, NULL, NULL,
140	0	model.layers[il].ffn_down_shexp, NULL, NULL,
141	0	NULL,
142	0	LLM_FFN_SILU, LLM_FFN_PAR, il);
143	0	cb(shared_out, "ffn_shexp_out", il);
144
145		// Final output: routed_output + shared_output
146	0	cur = ggml_add(ctx0, routed_out, shared_out);
147	0	cb(cur, "ffn_out", il);
148	0	}
149	0	cur = ggml_add(ctx0, cur, ffn_inp);
150
151	0	cur = build_cvec(cur, il);
152	0	cb(cur, "l_out", il);
153
154		// input for next layer
155	0	inpL = cur;
156	0	}
157	0	cur = inpL;
158	0	cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
159
160	0	cb(cur, "result_norm", -1);
161	0	res->t_embd = cur;
162
163		// lm_head
164	0	cur = build_lora_mm(model.output, cur);
165
166	0	cb(cur, "result_output", -1);
167	0	res->t_logits = cur;
168
169	0	ggml_build_forward_expand(gf, cur);
170	0	}

Coverage Report

Created: 2025-12-28 06:25