/src/llama.cpp/src/models/lfm2.cpp

Source
#include "models.h"

#include "../llama-memory-hybrid-iswa.h"
#include "../llama-memory-hybrid.h"

template <bool iswa>
llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
    llm_graph_context(params) {
    using inp_hybrid_type = std::conditional_t<iswa, llm_graph_input_mem_hybrid_iswa,  llm_graph_input_mem_hybrid>;
    using inp_attn_type   = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa,     llm_graph_input_attn_kv>;
    using mem_hybrid_ctx  = std::conditional_t<iswa, llama_memory_hybrid_iswa_context, llama_memory_hybrid_context>;

    // lambda helpers for readability
    auto build_dense_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {
        GGML_ASSERT(!model.layers[il].ffn_up_b);
        GGML_ASSERT(!model.layers[il].ffn_gate_b);
        GGML_ASSERT(!model.layers[il].ffn_down_b);
        return 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);
    };
    auto build_moe_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {
        return 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, true,
                hparams.expert_weights_scale,
                static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
                il);
    };
    auto build_attn_block = [&model, this](ggml_tensor *   cur,
                                           ggml_tensor *   inp_pos,
                                           inp_attn_type * inp_attn,
                                           int             il) -> ggml_tensor * {
        GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
        const auto n_embd_head = hparams.n_embd_head_v();
        const auto n_head_kv   = hparams.n_head_kv(il);

        auto * q = build_lora_mm(model.layers[il].wq, cur);
        cb(q, "model.layers.{}.self_attn.q_proj", il);
        auto * k = build_lora_mm(model.layers[il].wk, cur);
        cb(k, "model.layers.{}.self_attn.k_proj", il);
        auto * v = build_lora_mm(model.layers[il].wv, cur);
        cb(v, "model.layers.{}.self_attn.v_proj", il);

        q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens);
        k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
        v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);

        // qk norm
        q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
        cb(q, "model.layers.{}.self_attn.q_layernorm", il);
        k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
        cb(k, "model.layers.{}.self_attn.k_layernorm", il);

        // RoPE
        q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
                          attn_factor, beta_fast, beta_slow);
        k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
                          attn_factor, beta_fast, beta_slow);

        cur = build_attn(inp_attn,
                model.layers[il].wo, NULL,
                q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);

        cb(cur, "model.layers.{}.self_attn.out_proj", il);

        return cur;
    };
    auto build_shortconv_block = [&model, this](ggml_tensor *        cur,
                                                llm_graph_input_rs * inp_recr,
                                                int                  il) -> ggml_tensor * {
        const auto * mctx_cur = static_cast<const mem_hybrid_ctx *>(mctx)->get_recr();
        const uint32_t kv_head      = mctx_cur->get_head();
        const int64_t  n_seq_tokens = ubatch.n_seq_tokens;
        const int64_t  n_seqs       = ubatch.n_seqs;
        GGML_ASSERT(n_seqs != 0);
        GGML_ASSERT(ubatch.equal_seqs());
        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);

        GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
        const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;

        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);

        auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
        cb(bcx, "model.layers.{}.conv.in_proj", il);

        constexpr auto n_chunks = 3;
        GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
        const auto chunk_size = bcx->ne[0] / n_chunks;
        auto *     b          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
                                             0 * chunk_size * ggml_element_size(bcx));
        auto *     c          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
                                             1 * chunk_size * ggml_element_size(bcx));
        auto *     x          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
                                             2 * chunk_size * ggml_element_size(bcx));

        auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));

        // read conv state
        auto * conv_state = mctx_cur->get_r_l(il);
        auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
        auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);

        bx = ggml_concat(ctx0, conv, bx, 0);
        GGML_ASSERT(bx->ne[0] > conv->ne[0]);

        // last d_conv columns is a new conv state
        auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
                                       (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
        GGML_ASSERT(ggml_are_same_shape(conv, new_conv));

        // write new conv conv state
        ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
                                               ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
                                                            kv_head * d_conv * n_embd * ggml_element_size(new_conv))));

        auto * conv_kernel = model.layers[il].shortconv.conv;
        auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);
        cb(conv_out, "model.layers.{}.conv.conv", il);

        auto * y = ggml_mul(ctx0, c, conv_out);
        y        = build_lora_mm(model.layers[il].shortconv.out_proj, y);
        cb(y, "model.layers.{}.conv.out_proj", il);
        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
        y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);

        return y;
    };

    // actual graph construction starts here
    ggml_tensor * cur = build_inp_embd(model.tok_embd);
    cb(cur, "model.embed_tokens", -1);

    ggml_build_forward_expand(gf, cur);

    inp_hybrid_type * inp_hybrid = nullptr;
    if constexpr (iswa) {
        inp_hybrid = build_inp_mem_hybrid_iswa();
    } else {
        inp_hybrid = build_inp_mem_hybrid();
    }

    ggml_tensor * inp_pos     = build_inp_pos();
    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
        const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);

        auto * prev_cur = cur;
        cur             = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
        cb(cur, "model.layers.{}.operator_norm", il);

        cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
                                         build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);

        if (il == n_layer - 1 && inp_out_ids) {
            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
            prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
        }

        cur = ggml_add(ctx0, prev_cur, cur);

        auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
        cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);

        ggml_tensor * ffn_out =
            is_moe_layer ? build_moe_feed_forward(ffn_norm_out, il) : build_dense_feed_forward(ffn_norm_out, il);
        cb(ffn_norm_out, "model.layers.{}.ffn_out", il);

        cur = ggml_add(ctx0, cur, ffn_out);

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

    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
    cb(cur, "result_norm", -1);
    res->t_embd = cur;

    cur = build_lora_mm(model.output, cur);
    cb(cur, "result_output", -1);

    res->t_logits = cur;

    ggml_build_forward_expand(gf, cur);
}

// Explicit template instantiations
template struct llm_build_lfm2<true>;
template struct llm_build_lfm2<false>;

Line	Count	Source
1		#include "models.h"
2
3		#include "../llama-memory-hybrid-iswa.h"
4		#include "../llama-memory-hybrid.h"
5
6		template <bool iswa>
7		llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
8	0	llm_graph_context(params) {
9	0	using inp_hybrid_type = std::conditional_t<iswa, llm_graph_input_mem_hybrid_iswa, llm_graph_input_mem_hybrid>;
10	0	using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
11	0	using mem_hybrid_ctx = std::conditional_t<iswa, llama_memory_hybrid_iswa_context, llama_memory_hybrid_context>;
12
13		// lambda helpers for readability
14	0	auto build_dense_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {
15	0	GGML_ASSERT(!model.layers[il].ffn_up_b);
16	0	GGML_ASSERT(!model.layers[il].ffn_gate_b);
17	0	GGML_ASSERT(!model.layers[il].ffn_down_b);
18	0	return build_ffn(cur,
19	0	model.layers[il].ffn_up, NULL, NULL,
20	0	model.layers[il].ffn_gate, NULL, NULL,
21	0	model.layers[il].ffn_down, NULL, NULL,
22	0	NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
23	0	}; Unexecuted instantiation: llm_build_lfm2<true>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, int)#1}::operator()(ggml_tensor, int) const Unexecuted instantiation: llm_build_lfm2<false>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, int)#1}::operator()(ggml_tensor, int) const
24	0	auto build_moe_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {
25	0	return build_moe_ffn(cur,
26	0	model.layers[il].ffn_gate_inp,
27	0	model.layers[il].ffn_up_exps,
28	0	model.layers[il].ffn_gate_exps,
29	0	model.layers[il].ffn_down_exps,
30	0	model.layers[il].ffn_exp_probs_b,
31	0	n_expert, n_expert_used,
32	0	LLM_FFN_SILU, true,
33	0	hparams.expert_weights_scale,
34	0	static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
35	0	il);
36	0	}; Unexecuted instantiation: llm_build_lfm2<true>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, int)#2}::operator()(ggml_tensor, int) const Unexecuted instantiation: llm_build_lfm2<false>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, int)#2}::operator()(ggml_tensor, int) const
37	0	auto build_attn_block = [&model, this](ggml_tensor * cur,
38	0	ggml_tensor * inp_pos,
39	0	inp_attn_type * inp_attn,
40	0	int il) -> ggml_tensor * {
41	0	GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
42	0	const auto n_embd_head = hparams.n_embd_head_v();
43	0	const auto n_head_kv = hparams.n_head_kv(il);
44
45	0	auto * q = build_lora_mm(model.layers[il].wq, cur);
46	0	cb(q, "model.layers.{}.self_attn.q_proj", il);
47	0	auto * k = build_lora_mm(model.layers[il].wk, cur);
48	0	cb(k, "model.layers.{}.self_attn.k_proj", il);
49	0	auto * v = build_lora_mm(model.layers[il].wv, cur);
50	0	cb(v, "model.layers.{}.self_attn.v_proj", il);
51
52	0	q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens);
53	0	k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
54	0	v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
55
56		// qk norm
57	0	q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
58	0	cb(q, "model.layers.{}.self_attn.q_layernorm", il);
59	0	k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
60	0	cb(k, "model.layers.{}.self_attn.k_layernorm", il);
61
62		// RoPE
63	0	q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
64	0	attn_factor, beta_fast, beta_slow);
65	0	k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
66	0	attn_factor, beta_fast, beta_slow);
67
68	0	cur = build_attn(inp_attn,
69	0	model.layers[il].wo, NULL,
70	0	q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
71
72	0	cb(cur, "model.layers.{}.self_attn.out_proj", il);
73
74	0	return cur;
75	0	}; Unexecuted instantiation: llm_build_lfm2<true>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, ggml_tensor, llm_graph_input_attn_kv_iswa, int)#1}::operator()(ggml_tensor, ggml_tensor, llm_graph_input_attn_kv_iswa, int) const Unexecuted instantiation: llm_build_lfm2<false>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, ggml_tensor, llm_graph_input_attn_kv, int)#1}::operator()(ggml_tensor, ggml_tensor, llm_graph_input_attn_kv, int) const
76	0	auto build_shortconv_block = [&model, this](ggml_tensor * cur,
77	0	llm_graph_input_rs * inp_recr,
78	0	int il) -> ggml_tensor * {
79	0	const auto * mctx_cur = static_cast<const mem_hybrid_ctx *>(mctx)->get_recr();
80	0	const uint32_t kv_head = mctx_cur->get_head();
81	0	const int64_t n_seq_tokens = ubatch.n_seq_tokens;
82	0	const int64_t n_seqs = ubatch.n_seqs;
83	0	GGML_ASSERT(n_seqs != 0);
84	0	GGML_ASSERT(ubatch.equal_seqs());
85	0	GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
86
87	0	GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
88	0	const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
89
90		// {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
91	0	cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
92
93	0	auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
94	0	cb(bcx, "model.layers.{}.conv.in_proj", il);
95
96	0	constexpr auto n_chunks = 3;
97	0	GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
98	0	const auto chunk_size = bcx->ne[0] / n_chunks;
99	0	auto * b = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
100	0	0 * chunk_size * ggml_element_size(bcx));
101	0	auto * c = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
102	0	1 * chunk_size * ggml_element_size(bcx));
103	0	auto * x = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
104	0	2 * chunk_size * ggml_element_size(bcx));
105
106	0	auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
107
108		// read conv state
109	0	auto * conv_state = mctx_cur->get_r_l(il);
110	0	auto * conv_rs = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
111	0	auto * conv = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
112
113	0	bx = ggml_concat(ctx0, conv, bx, 0);
114	0	GGML_ASSERT(bx->ne[0] > conv->ne[0]);
115
116		// last d_conv columns is a new conv state
117	0	auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
118	0	(bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
119	0	GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
120
121		// write new conv conv state
122	0	ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
123	0	ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
124	0	kv_head * d_conv * n_embd * ggml_element_size(new_conv))));
125
126	0	auto * conv_kernel = model.layers[il].shortconv.conv;
127	0	auto * conv_out = ggml_ssm_conv(ctx0, bx, conv_kernel);
128	0	cb(conv_out, "model.layers.{}.conv.conv", il);
129
130	0	auto * y = ggml_mul(ctx0, c, conv_out);
131	0	y = build_lora_mm(model.layers[il].shortconv.out_proj, y);
132	0	cb(y, "model.layers.{}.conv.out_proj", il);
133		// {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
134	0	y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);
135
136	0	return y;
137	0	}; Unexecuted instantiation: llm_build_lfm2<true>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, llm_graph_input_rs, int)#1}::operator()(ggml_tensor, llm_graph_input_rs, int) const Unexecuted instantiation: llm_build_lfm2<false>::llm_build_lfm2(llama_model const&, llm_graph_params const&)::{lambda(ggml_tensor, llm_graph_input_rs, int)#1}::operator()(ggml_tensor, llm_graph_input_rs, int) const
138
139		// actual graph construction starts here
140	0	ggml_tensor * cur = build_inp_embd(model.tok_embd);
141	0	cb(cur, "model.embed_tokens", -1);
142
143	0	ggml_build_forward_expand(gf, cur);
144
145	0	inp_hybrid_type * inp_hybrid = nullptr;
146	0	if constexpr (iswa) {
147	0	inp_hybrid = build_inp_mem_hybrid_iswa();
148	0	} else {
149	0	inp_hybrid = build_inp_mem_hybrid();
150	0	}
151
152	0	ggml_tensor * inp_pos = build_inp_pos();
153	0	ggml_tensor * inp_out_ids = build_inp_out_ids();
154
155	0	for (int il = 0; il < n_layer; ++il) {
156	0	const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);
157
158	0	auto * prev_cur = cur;
159	0	cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
160	0	cb(cur, "model.layers.{}.operator_norm", il);
161
162	0	cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
163	0	build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);
164
165	0	if (il == n_layer - 1 && inp_out_ids) {
166	0	cur = ggml_get_rows(ctx0, cur, inp_out_ids);
167	0	prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
168	0	}
169
170	0	cur = ggml_add(ctx0, prev_cur, cur);
171
172	0	auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
173	0	cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
174
175	0	ggml_tensor * ffn_out =
176	0	is_moe_layer ? build_moe_feed_forward(ffn_norm_out, il) : build_dense_feed_forward(ffn_norm_out, il);
177	0	cb(ffn_norm_out, "model.layers.{}.ffn_out", il);
178
179	0	cur = ggml_add(ctx0, cur, ffn_out);
180
181	0	cur = build_cvec(cur, il);
182	0	cb(cur, "l_out", il);
183	0	}
184
185	0	cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
186	0	cb(cur, "result_norm", -1);
187	0	res->t_embd = cur;
188
189	0	cur = build_lora_mm(model.output, cur);
190	0	cb(cur, "result_output", -1);
191
192	0	res->t_logits = cur;
193
194	0	ggml_build_forward_expand(gf, cur);
195	0	} Unexecuted instantiation: llm_build_lfm2<true>::llm_build_lfm2(llama_model const&, llm_graph_params const&) Unexecuted instantiation: llm_build_lfm2<false>::llm_build_lfm2(llama_model const&, llm_graph_params const&)
196
197		// Explicit template instantiations
198		template struct llm_build_lfm2<true>;
199		template struct llm_build_lfm2<false>;

Coverage Report

Created: 2026-03-21 06:50