Skip to content

Commit 5a8463f

Browse files
pwilkinpwilkin
committed
Add missing LFM2 code
1 parent 64de434 commit 5a8463f

File tree

2 files changed

+135
-15
lines changed

2 files changed

+135
-15
lines changed

‎src/models/llm_build_lfm2.cpp‎

Lines changed: 129 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -1,37 +1,38 @@
1-
#include "../llama-model.h"
1+
#include "llm_build_lfm2.h"
2+
23
#include "../llama-graph.h"
4+
#include "../llama-model.h"
5+
#include "../llama-memory-hybrid.h"
36

4-
#include "llm_build_lfm2.h"
57
#include <cmath>
68

7-
llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
9+
llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
10+
llm_graph_context(params),
11+
model(model) {
812
ggml_tensor * cur = build_inp_embd(model.tok_embd);
913
cb(cur, "model.embed_tokens", -1);
1014

1115
ggml_tensor * inp_pos = build_inp_pos();
12-
auto * inp_hybrid = build_inp_mem_hybrid();
16+
auto * inp_hybrid = build_inp_mem_hybrid();
1317
ggml_tensor * inp_out_ids = build_inp_out_ids();
1418

1519
for (int il = 0; il < n_layer; ++il) {
1620
auto * prev_cur = cur;
17-
cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
21+
cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
1822
cb(cur, "model.layers.{}.operator_norm", il);
1923

20-
// TODO: implement recurrent/attention logic inline
21-
// cur = hparams.is_recurrent(il) ?
22-
// build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
23-
// build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il) ;
24+
cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
25+
build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);
2426

2527
if (il == n_layer - 1 && inp_out_ids) {
26-
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
28+
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
2729
prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
2830
}
29-
;
31+
3032
cur = ggml_add(ctx0, prev_cur, cur);
31-
// TODO: implement feed_forward inline
32-
// cur = ggml_add(ctx0, cur, build_feed_forward(cur, il));
33+
cur = ggml_add(ctx0, cur, build_feed_forward(cur, il));
3334
}
34-
;
35+
3536
cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1);
3637
cb(cur, "model.embedding_norm", -1);
3738
res->t_embd = cur;
@@ -43,4 +44,117 @@ llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params
4344

4445
ggml_build_forward_expand(gf, cur);
4546
}
46-
;
47+
48+
ggml_tensor * llm_build_lfm2::build_feed_forward(ggml_tensor * cur, int il) const {
49+
cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
50+
cb(cur, "model.layers.{}.ffn_norm", il);
51+
52+
GGML_ASSERT(!model.layers[il].ffn_up_b);
53+
GGML_ASSERT(!model.layers[il].ffn_gate_b);
54+
GGML_ASSERT(!model.layers[il].ffn_down_b);
55+
cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL,
56+
model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
57+
cb(cur, "model.layers.{}.feed_forward.w2", il);
58+
59+
return cur;
60+
}
61+
62+
ggml_tensor * llm_build_lfm2::build_attn_block(ggml_tensor * cur,
63+
ggml_tensor * inp_pos,
64+
llm_graph_input_attn_kv * inp_attn,
65+
int il) const {
66+
GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
67+
const auto n_embd_head = hparams.n_embd_head_v;
68+
const auto n_head_kv = hparams.n_head_kv(il);
69+
70+
auto * q = build_lora_mm(model.layers[il].wq, cur);
71+
cb(q, "model.layers.{}.self_attn.q_proj", il);
72+
auto * k = build_lora_mm(model.layers[il].wk, cur);
73+
cb(k, "model.layers.{}.self_attn.k_proj", il);
74+
auto * v = build_lora_mm(model.layers[il].wv, cur);
75+
cb(v, "model.layers.{}.self_attn.v_proj", il);
76+
77+
q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens);
78+
k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
79+
v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
80+
81+
// qk norm
82+
q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
83+
cb(q, "model.layers.{}.self_attn.q_layernorm", il);
84+
k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
85+
cb(k, "model.layers.{}.self_attn.k_layernorm", il);
86+
87+
// RoPE
88+
q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
89+
attn_factor, beta_fast, beta_slow);
90+
k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
91+
attn_factor, beta_fast, beta_slow);
92+
93+
cur = build_attn(inp_attn, model.layers[il].wo, NULL, q, k, v, nullptr, nullptr, nullptr,
94+
1.0f / sqrtf(float(n_embd_head)), il);
95+
96+
cb(cur, "model.layers.{}.self_attn.out_proj", il);
97+
98+
return cur;
99+
}
100+
101+
ggml_tensor * llm_build_lfm2::build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il) {
102+
const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
103+
const uint32_t kv_head = mctx_cur->get_head();
104+
const int64_t n_seq_tokens = ubatch.n_seq_tokens;
105+
const int64_t n_seqs = ubatch.n_seqs;
106+
GGML_ASSERT(n_seqs != 0);
107+
GGML_ASSERT(ubatch.equal_seqs());
108+
GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
109+
110+
GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
111+
const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
112+
113+
// {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
114+
cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
115+
116+
auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
117+
cb(bcx, "model.layers.{}.conv.in_proj", il);
118+
119+
constexpr auto n_chunks = 3;
120+
GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
121+
const auto chunk_size = bcx->ne[0] / n_chunks;
122+
auto * b = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
123+
0 * chunk_size * ggml_element_size(bcx));
124+
auto * c = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
125+
1 * chunk_size * ggml_element_size(bcx));
126+
auto * x = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
127+
2 * chunk_size * ggml_element_size(bcx));
128+
129+
auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
130+
131+
// read conv state
132+
auto * conv_state = mctx_cur->get_r_l(il);
133+
auto * conv_rs = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
134+
auto * conv = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
135+
136+
bx = ggml_concat(ctx0, conv, bx, 0);
137+
GGML_ASSERT(bx->ne[0] > conv->ne[0]);
138+
139+
// last d_conv columns is a new conv state
140+
auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
141+
(bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
142+
GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
143+
144+
// write new conv conv state
145+
ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
146+
ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
147+
kv_head * d_conv * n_embd * ggml_element_size(new_conv))));
148+
149+
auto * conv_kernel = model.layers[il].shortconv.conv;
150+
auto * conv_out = ggml_ssm_conv(ctx0, bx, conv_kernel);
151+
cb(conv_out, "model.layers.{}.conv.conv", il);
152+
153+
auto * y = ggml_mul(ctx0, c, conv_out);
154+
y = build_lora_mm(model.layers[il].shortconv.out_proj, y);
155+
cb(y, "model.layers.{}.conv.out_proj", il);
156+
// {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
157+
y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);
158+
159+
return y;
160+
}

‎src/models/llm_build_lfm2.h‎

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -6,5 +6,11 @@
66
#include <cmath>
77

88
struct llm_build_lfm2 : public llm_graph_context {
9+
const llama_model & model;
10+
911
llm_build_lfm2(const llama_model & model, const llm_graph_params & params);
12+
ggml_tensor * build_feed_forward(ggml_tensor * cur, int il) const;
13+
ggml_tensor * build_attn_block(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, int il) const;
14+
ggml_tensor * build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il);
15+
1016
};

0 commit comments

Comments
 (0)