public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-17 20:03:45 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2026-02-27 20:57:58 +02:00
parent 4734056067
commit 84f8db71d8
41 changed files with 1734 additions and 2227 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/talk-llama/models/deepseek2.cpp
+3 -1
View File
@@ -218,7 +218,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
LLM_FFN_SILU, hparams.expert_weights_norm,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
il,
nullptr,
model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// FFN shared expert
examples/talk-llama/models/delta-net-base.cpp
+376
View File
@@ -0,0 +1,376 @@
#include "models.h"
#define CHUNK_SIZE 64
// utility to get one slice from the third dimension
// input dim: [x, y, c, b]
// output dim: [x, y, 1, b]
static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
}
llm_build_delta_net_base::llm_build_delta_net_base(const llm_graph_params & params) : llm_graph_context(params) {}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
const bool kda = (g->ne[0] == S_k && g->ne[1] == H_k);
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT( g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(b, "b_in", il);
cb(g, "g_in", il);
q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
g = ggml_permute(ctx0, g, 0, 2, 1, 3); // [g_0, n_tokens, H_v, n_seqs]
b = ggml_permute(ctx0, b, 0, 2, 1, 3); // [ 1, n_tokens, H_v, n_seqs]
const int CS = CHUNK_SIZE;
const int pad = (CS - n_tokens % CS) % CS;
const int n_chunks = (n_tokens + pad) / CS;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, 0, pad, 0, 0);
b = ggml_pad(ctx0, b, 0, pad, 0, 0);
ggml_tensor * v_b = ggml_mul(ctx0, v, b);
ggml_tensor * k_b = ggml_mul(ctx0, k, b);
cb(v_b, "v_b", il);
cb(k_b, "k_b", il);
q = ggml_reshape_4d(ctx0, q, S_k, CS, n_chunks, H_k * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, CS, n_chunks, H_k * n_seqs);
k_b = ggml_reshape_4d(ctx0, k_b, S_k, CS, n_chunks, H_v * n_seqs);
v = ggml_reshape_4d(ctx0, v, S_v, CS, n_chunks, H_v * n_seqs);
v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs);
g = ggml_reshape_4d(ctx0, g, g->ne[0], CS, n_chunks, H_v * n_seqs);
b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs);
// [CS, g_0, n_chunks, H_v * n_seqs]
// TODO: extend ggml_cumsum with axis parameter to avoid transpose
ggml_tensor * g_cs = ggml_cumsum(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, g)));
cb(g_cs, "g_cs", il);
ggml_tensor * kb = nullptr;
ggml_tensor * kq = nullptr;
if (kda) {
const int64_t CHB = n_chunks * H_k * n_seqs;
ggml_tensor * g_cs_i = ggml_reshape_4d(ctx0, g_cs, CS, 1, S_k, CHB); // [chunk_size, 1, S_k, CHB]
ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, S_k, CHB); // [1, chunk_size, S_k, CHB]
g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, S_k, CHB); // [1, chunk_size, S_k, CHB] -> [chunk_size, chunk_size, S_k, CHB]
// decay_mask [chunk_size,chunk_size,S_k,CHB]
ggml_tensor * decay_mask;
decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
decay_mask = ggml_exp(ctx0, decay_mask);
cb(decay_mask, "decay_mask", il);
// decay_mask [S_k,BT_j,BT_i,CHB] *Note* second and third chunk_sizes are switched
decay_mask = ggml_cont_4d(ctx0, ggml_permute(ctx0, decay_mask, 2, 1, 0, 3), S_k, CS, CS, CHB);
ggml_tensor * k_b_i = ggml_reshape_4d(ctx0, k_b, S_k, CS, 1, CHB);
ggml_tensor * k_j = ggml_reshape_4d(ctx0, k, S_k, 1, CS, CHB);
ggml_tensor * q_i = ggml_reshape_4d(ctx0, q, S_k, CS, 1, CHB);
ggml_tensor * decay_k_b_i = ggml_mul(ctx0, decay_mask, k_b_i);
ggml_tensor * decay_q_i = ggml_mul(ctx0, decay_mask, q_i);
// decay_k_b_i [S,BT,BT,CHB] @ k_j [S,1,BT,CHB] = Akk [BT,1,BT,CHB]
kb = ggml_mul_mat(ctx0, decay_k_b_i, k_j);
kq = ggml_mul_mat(ctx0, decay_q_i, k_j);
kb = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kb, CS, CS, n_chunks, H_v * n_seqs)));
kq = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kq, CS, CS, n_chunks, H_v * n_seqs)));
} else {
ggml_tensor * g_cs_i = g_cs;
ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
// [CS, CS, n_chunks, H_v * n_seqs]
ggml_tensor * decay_mask;
decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
decay_mask = ggml_exp(ctx0, decay_mask);
cb(decay_mask, "decay_mask", il);
// [CS, CS, n_chunks, H_k * n_seqs]
kb = ggml_mul_mat(ctx0, k, k_b);
kb = ggml_mul (ctx0, kb, decay_mask);
// [CS, CS, n_chunks, H_k * n_seqs]
kq = ggml_mul_mat(ctx0, k, q);
kq = ggml_mul(ctx0, kq, decay_mask);
}
kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
cb(kq, "kq", il);
// [CS, CS, n_chunks, H_k * n_seqs]
ggml_tensor * attn;
attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
cb(attn, "attn", il);
ggml_tensor * identity;
identity = ggml_view_1d(ctx0, attn, CS, 0);
identity = ggml_fill (ctx0, identity, 1.0f);
identity = ggml_diag (ctx0, identity);
ggml_tensor * lhs = ggml_add(ctx0, attn, identity);
cb(lhs, "dnet_add_ch_lhs", il);
attn = ggml_neg(ctx0, attn);
cb(attn, "attn_pre_solve", il);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_add(ctx0, lin_solve, identity);
cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs]
// [S_v, CS, n_chunks, H_v * n_seqs]
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);
// [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);
k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
// [CS, S_k, n_chunks, H_k * n_seqs]
ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp);
cb(kbg, "k_beta_g_exp", il);
// [S_k, CS, n_chunks, H_k * n_seqs]
ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
cb(k_cd, "k_cumdecay", il);
// [1, CS, n_chunks, H_k * n_seqs] KDA: [S_k, CS, n_chunks, H_k * n_seqs]
ggml_tensor * g_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_exp));
ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t);
// vectorized calculation of key_gdiff
// improved from the chunked version:
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
// key_gdiff = key * g_diff.unsqueeze(-1)
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along CS dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
// [1, 1, n_chunks, H_v * n_seqs] KDA: [1, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, g_cs->ne[1], g_cs->ne[2], g_cs->ne[3],
g_cs->nb[1],
g_cs->nb[2],
g_cs->nb[3],
ggml_row_size(g_cs->type, g_cs->ne[0] - 1));
cb(g_last, "g_last", il);
// TODO: remove this cont when CUDA supports non-cont unary ops
g_last = ggml_cont(ctx0, g_last);
// [1, 1, n_chunks, H_v * n_seqs] KDA: [S_k, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_last_exp_t = ggml_transpose(ctx0, ggml_exp(ctx0, g_last));
cb(g_last_exp_t, "g_last_exp_t", il);
// [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last));
cb(g_diff, "g_diff", il);
ggml_tensor * g_diff_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_exp(ctx0, g_diff)));
// [S_k, CS, n_chunks, H_v * n_seqs]
ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
cb(kg, "key_gdiff", il);
// [CS, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * kg_t = ggml_cont(ctx0, ggml_transpose(ctx0, kg));
cb(kg_t, "key_gdiff_t", il);
ggml_tensor * s_t = ggml_transpose(ctx0, s);
s_t = ggml_cont_4d(ctx0, s_t, S_v, S_v, 1, H_v * n_seqs);
cb(s_t, "dnet_add_ch_state", il);
// [CS, S_v, n_chunks, H_v * n_seqs]
ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v));
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
ggml_tensor * ch_k_cd = get_slice_2d(ctx0, k_cd, chunk); // [S_k, CS, 1, H_k * n_seqs]
ggml_tensor * ch_v_t = get_slice_2d(ctx0, v_t, chunk); // [ CS, S_v, 1, H_v * n_seqs]
ggml_tensor * ch_kq = get_slice_2d(ctx0, kq, chunk); // [ CS, CS, 1, H_k * n_seqs]
ggml_tensor * ch_q_g_exp = get_slice_2d(ctx0, q_g_exp, chunk); // [S_k, CS, 1, H_k * n_seqs]
ggml_tensor * ch_kg_t = get_slice_2d(ctx0, kg_t, chunk); // [ CS, S_k, 1, H_v * n_seqs]
// [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s_t);
cb(v_t_p, "v_prime", il);
// [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p);
cb(v_t_new, "v_t_new", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq);
cb(v_attn, "v_attn", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s_t, ch_q_g_exp);
cb(attn_inter, "attn_inter", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * o_ch = ggml_add(ctx0, attn_inter, v_attn);
cb(o_ch, "dnet_add_ch_attn_out", il);
v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]);
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// TODO: head broadcast might not work here - probably will need a transpose
ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * ch_g_last_exp_t = get_slice_2d(ctx0, g_last_exp_t, chunk);
s_t = ggml_mul(ctx0, s_t, ch_g_last_exp_t);
s_t = ggml_add(ctx0, s_t, kgv);
cb(s_t, "dnet_add_ch_state", il);
}
s_t = ggml_reshape_4d(ctx0, s_t, S_v, S_v, H_v, n_seqs);
// truncate padded tokens
ggml_tensor * o = ggml_view_4d(ctx0, v,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(v->type, S_v),
ggml_row_size(v->type, S_v * CS * n_chunks),
ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0);
o = ggml_permute (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
s = ggml_transpose(ctx0, s_t);
cb(s, "output_state", il);
return {o, s};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b, // beta
ggml_tensor * s, // state
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1);
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT( g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(b, "b_in", il);
cb(g, "g_in", il);
// GDA: [1, 1, H_v, n_seqs]
// KDA: [1, S_k, H_v, n_seqs]
g = ggml_reshape_4d(ctx0, g, 1, g->ne[0], H_v, n_seqs);
b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
// [S_v, S_v, H_v, n_seqs]
g = ggml_exp(ctx0, g);
s = ggml_mul(ctx0, s, g);
ggml_tensor * s_t = ggml_cont(ctx0, ggml_transpose(ctx0, s));
// [1, S_v, H_v, n_seqs]
ggml_tensor * sk;
sk = ggml_mul (ctx0, s_t, k);
sk = ggml_sum_rows(ctx0, sk);
// [S_v, 1, H_v, n_seqs]
ggml_tensor * d;
d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk));
d = ggml_mul(ctx0, d, b);
// [1, S_v, H_v, n_seqs]
ggml_tensor * d_t;
d_t = ggml_transpose(ctx0, d);
// [S_v, S_v, H_v, n_seqs]
ggml_tensor * kd;
k = ggml_repeat(ctx0, k, s);
kd = ggml_mul (ctx0, k, d_t);
s_t = ggml_add(ctx0, s_t, kd);
cb(s_t, "dnet_add_ar_state", il);
ggml_tensor * s_q = ggml_mul (ctx0, s_t, q);
ggml_tensor * o = ggml_sum_rows(ctx0, s_q);
o = ggml_permute (ctx0, o, 2, 0, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs]
return {o, s};
}
examples/talk-llama/models/eurobert.cpp
+97
View File
@@ -0,0 +1,97 @@
#include "models.h"
llm_build_eurobert::llm_build_eurobert(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);
ggml_tensor * cur;
ggml_tensor * inpL;
ggml_tensor * inp_pos = build_inp_pos();
inpL = build_inp_embd(model.tok_embd);
cb(inpL, "inp_embd", -1);
auto * inp_attn = build_attn_inp_no_cache();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * cur = inpL;
cur = build_norm(inpL,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, il);
{
ggml_tensor * Qcur;
ggml_tensor * Kcur;
ggml_tensor * Vcur;
Qcur = build_lora_mm(model.layers[il].wq, cur);
Kcur = build_lora_mm(model.layers[il].wk, cur);
Vcur = build_lora_mm(model.layers[il].wv, cur);
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);
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, nullptr,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
cb(cur, "kqv_out", il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
}
cur = ggml_add(ctx0, cur, inpL);
ggml_tensor * ffn_inp = cur;
cb(ffn_inp, "ffn_inp", il);
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
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);
cur = ggml_add(ctx0, cur, ffn_inp);
inpL = cur;
}
cur = inpL;
cur = build_norm(cur,
model.output_norm, NULL,
LLM_NORM_RMS, -1);
cb(cur, "result_embd", -1);
res->t_embd = cur;
ggml_build_forward_expand(gf, cur);
}
examples/talk-llama/models/falcon-h1.cpp
+1 -3
View File
@@ -1,9 +1,7 @@
#include "models.h"
llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params) {
llm_build_mamba_base(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
ggml_tensor * cur;
examples/talk-llama/models/glm4.cpp
+12 -5
View File
@@ -29,7 +29,10 @@ llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
// 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
@@ -100,7 +103,7 @@ llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {
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);
}
@@ -130,9 +133,13 @@ llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params
cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "post_mlp_norm", il);
}
// Add residual connection after post-MLP norm
inpL = ggml_add(ctx0, cur, ffn_inp);
cb(inpL, "l_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;
}
// Final norm
cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
examples/talk-llama/models/granite-hybrid.cpp
+1 -1
View File
@@ -2,7 +2,7 @@
llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params) {
llm_build_mamba_base(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
examples/talk-llama/models/jais2.cpp
+123
View File
@@ -0,0 +1,123 @@
#include "models.h"
// JAIS-2 model graph builder
// Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings
llm_build_jais2::llm_build_jais2(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);
GGML_ASSERT(n_embd_head == hparams.n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
// KV input for attention
auto * inp_attn = build_attn_inp_kv();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
// Pre-attention LayerNorm
cur = build_norm(inpL,
model.layers[il].attn_norm,
model.layers[il].attn_norm_b,
LLM_NORM, il);
cb(cur, "attn_norm", il);
// Self-attention with separate Q, K, V projections
{
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur_bias", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur_bias", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur_bias", il);
// Reshape for attention
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 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_rope", il);
cb(Kcur, "Kcur_rope", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
}
// Residual connection
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
cb(ffn_inp, "ffn_inp", il);
// Pre-FFN LayerNorm
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm,
model.layers[il].ffn_norm_b,
LLM_NORM, il);
cb(cur, "ffn_norm", il);
// FFN with relu2 activation (ReLU squared) - no gate projection
// up -> relu2 -> down
cur = build_ffn(cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
NULL, NULL, NULL, // no gate
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
NULL,
LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
cb(cur, "ffn_out", il);
// Residual connection
inpL = ggml_add(ctx0, cur, ffn_inp);
inpL = build_cvec(inpL, il);
cb(inpL, "l_out", il);
}
// Final LayerNorm
cur = build_norm(inpL,
model.output_norm,
model.output_norm_b,
LLM_NORM, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// Output projection
cur = build_lora_mm(model.output, cur);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
examples/talk-llama/models/jamba.cpp
+1 -1
View File
@@ -1,6 +1,6 @@
#include "models.h"
llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
ggml_tensor * cur;
examples/talk-llama/models/kimi-linear.cpp
+20 -403
View File
@@ -1,7 +1,7 @@
#include "models.h"
#include "ggml.h"
#define CHUNK_SIZE 64
#include "llama-memory-recurrent.h"
// Causal Conv1d function for Q,K,V
// When qkv is 0, it is Q, 1 is K, 2 is V
@@ -65,7 +65,7 @@ static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_t
}
llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params), model(model) {
llm_build_delta_net_base(params), model(model) {
ggml_tensor * cur;
ggml_tensor * inpL;
@@ -84,17 +84,6 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
// Output ids for selecting which tokens to output
ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * chunked_causal_mask =
ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
GGML_TRI_TYPE_LOWER);
ggml_tensor * chunked_identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
ggml_tensor * chunked_diag_mask = ggml_add(ctx0, chunked_causal_mask, chunked_identity);
ggml_build_forward_expand(gf, chunked_causal_mask);
ggml_build_forward_expand(gf, chunked_identity);
ggml_build_forward_expand(gf, chunked_diag_mask);
// Kimi dimension constants
const int64_t n_head = hparams.n_head();
const int64_t head_dim = hparams.n_embd_head_kda;
@@ -127,6 +116,8 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Check layer type by checking which tensors exist
// KDA layers have ssm_a_log tensor, MLA layers have wkv_a_mqa tensor
bool is_kda = (layer.ssm_a != nullptr);
@@ -160,27 +151,35 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
g1 = ggml_mul(ctx0, g1, A);
cb(g1, "kda_g1", il);
g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head, n_seq_tokens, n_seqs);
// Compute beta (mixing coefficient)
ggml_tensor * beta = ggml_mul_mat(ctx0, layer.ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, n_head, 1, n_seq_tokens, n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, n_head, n_seq_tokens, n_seqs);
cb(beta, "kda_beta", il);
beta = ggml_sigmoid(ctx0, beta);
// Reshape for KDA recurrence
// {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);
g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head, n_seq_tokens, n_seqs);
// Get SSM state and compute KDA recurrence using ggml_kda_scan
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head, n_seqs);
// Choose between build_kda_chunking and build_kda_recurrent based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?
build_kda_autoregressive(Qcur, Kcur, Vcur, g1, beta, state, il) :
build_kda_chunking(Qcur, Kcur, Vcur, g1, beta, state, chunked_causal_mask, chunked_identity, chunked_diag_mask, il);
ggml_tensor * output = attn_out.first;
const float eps_norm = hparams.f_norm_rms_eps;
Qcur = ggml_l2_norm(ctx0, Qcur, eps_norm);
Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
// Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?
build_delta_net_autoregressive(Qcur, Kcur, Vcur, g1, beta, state, il) :
build_delta_net_chunking(Qcur, Kcur, Vcur, g1, beta, state, il);
ggml_tensor * output = ggml_cont(ctx0, attn_out.first);
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
cb(new_state, "new_state", il);
@@ -391,385 +390,3 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
ggml_build_forward_expand(gf, cur);
}
/*
This is a ggml implementation of the naive_chunk_kda function of
https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/kda/naive.py
*/
std::pair<ggml_tensor *, ggml_tensor *> llm_build_kimi_linear::build_kda_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * gk,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
GGML_ASSERT(ggml_is_contiguous(state));
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(gk->ne[0] == S_v && gk->ne[1] == H_v && gk->ne[2] == n_tokens && gk->ne[3] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v && state->ne[2] == H_v && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
// TODO: can this ever be false?
const bool use_qk_l2norm = true;
if (use_qk_l2norm) {
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
}
const float scale = 1.0f / sqrtf(S_v);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(gk, "gk_in", il);
q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_k, n_tokens, H_k, n_seqs);
k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_k, n_tokens, H_k, n_seqs);
v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
gk = ggml_cont_4d(ctx0, ggml_permute(ctx0, gk, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
cb(q, "q_perm", il);
cb(k, "k_perm", il);
cb(v, "v_perm", il);
cb(beta, "beta_perm", il);
cb(gk, "gk_perm", il);
cb(state, "state_in", il);
GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
// Do padding
const int64_t chunk_size = CHUNK_SIZE;
const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
const int64_t n_chunks = (n_tokens + pad) / chunk_size;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
gk = ggml_pad(ctx0, gk, 0, pad, 0, 0);
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
cb(q, "q_pad", il);
cb(k, "k_pad", il);
cb(v, "v_pad", il);
cb(beta, "beta_pad", il);
cb(gk, "gk_pad", il);
ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
cb(v_beta, "v_beta", il);
cb(k_beta, "k_beta", il);
const int64_t HB = H_k * n_seqs;
q = ggml_cont_4d(ctx0, q, S_k, chunk_size, n_chunks, HB);
k = ggml_cont_4d(ctx0, k, S_k, chunk_size, n_chunks, HB);
k_beta = ggml_cont_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, HB);
v = ggml_cont_4d(ctx0, v, S_v, chunk_size, n_chunks, HB);
v_beta = ggml_cont_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, HB);
gk = ggml_cont_4d(ctx0, gk, S_k, chunk_size, n_chunks, HB);
beta = ggml_cont_4d(ctx0, beta, 1, chunk_size, n_chunks, HB);
// switch for cumsum
gk = ggml_cont_4d(ctx0, ggml_permute(ctx0, gk, 1, 0, 2, 3), chunk_size, S_k, n_chunks, HB);
cb(gk, "gk", il);
ggml_tensor * gk_cumsum = ggml_cumsum(ctx0, gk);
cb(gk_cumsum, "gk_cumsum", il);
/*
Compute Akk and Aqk loop together
Akk loop:
for i in range(BT):
k_i = k[..., i, :] # k_i [B,H,NT,S]
g_i = g[..., i:i+1, :] # g_i [B,H,NT,1,S]
A[..., i] = torch.einsum('... c d, ... d -> ... c', k * (g - g_i).exp(), k_i)
Aqk loop:
for j in range(BT):
k_j = k[:, :, i, j]
g_j = g[:, :, i, j:j+1, :]
A[..., j] = torch.einsum('... c d, ... d -> ... c', q_i * (g_i - g_j).exp(), k_j)
*/
const int64_t CHB = n_chunks * H_k * n_seqs;
ggml_tensor * gkcs_i = ggml_reshape_4d(ctx0, gk_cumsum, chunk_size, 1, S_k, CHB); // [chunk_size, 1, S_k, CHB]
ggml_tensor * gkcs_j = ggml_reshape_4d(ctx0, gkcs_i, 1, chunk_size, S_k, CHB); // [1, chunk_size, S_k, CHB]
ggml_tensor * gkcs_j_bc = ggml_repeat_4d(ctx0, gkcs_j, chunk_size, chunk_size, S_k, CHB); // [1, chunk_size, S_k, CHB] -> [chunk_size, chunk_size, S_k, CHB]
// decay_mask [chunk_size,chunk_size,S_k,CHB]
ggml_tensor * decay_mask = ggml_sub(ctx0, gkcs_j_bc, gkcs_i);
cb(decay_mask, "decay_mask", il);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
cb(decay_mask, "decay_masked", il);
decay_mask = ggml_exp(ctx0, decay_mask);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
// decay_mask [S_k,BT_j,BT_i,CHB] *Note* second and third chunk_sizes are switched
decay_mask = ggml_cont_4d(ctx0, ggml_permute(ctx0, decay_mask, 2, 1, 0, 3), S_k, chunk_size, chunk_size, CHB);
ggml_tensor * k_i = ggml_reshape_4d(ctx0, k, S_k, chunk_size, 1, CHB);
ggml_tensor * k_j = ggml_reshape_4d(ctx0, k, S_k, 1, chunk_size, CHB);
ggml_tensor * q_i = ggml_reshape_4d(ctx0, q, S_k, chunk_size, 1, CHB);
ggml_tensor * decay_k_i = ggml_mul(ctx0, decay_mask, k_i);
ggml_tensor * decay_q_i = ggml_mul(ctx0, decay_mask, q_i);
// decay_k_i [S.BT,BT,CHB] @ k_j [S,1,BT,CHB] = Akk [BT,1,BT,CHB]
ggml_tensor * Akk = ggml_mul_mat(ctx0, decay_k_i, k_j);
ggml_tensor * Aqk = ggml_mul_mat(ctx0, decay_q_i, k_j);
Akk = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, Akk, chunk_size, chunk_size, n_chunks, HB)));
Aqk = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, Aqk, chunk_size, chunk_size, n_chunks, HB)));
cb(Akk, "Akk", il);
cb(Aqk, "Aqk", il);
Akk = ggml_mul(ctx0, Akk, beta);
Akk = ggml_neg(ctx0, ggml_mul(ctx0, Akk, causal_mask));
cb(Akk, "attn_pre_solve", il);
Aqk = ggml_mul(ctx0, Aqk, diag_mask);
Aqk = ggml_scale(ctx0, Aqk, scale); // scale q
cb(Aqk, "Aqk_masked", il);
// for i in range(1, chunk_size):
// row = attn[..., i, :i].clone()
// sub = attn[..., :i, :i].clone()
// attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
// attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
//
// We reduce this to a linear triangular solve: AX = B, where B = attn, A = I - tril(A)
ggml_tensor * attn_lower = ggml_mul(ctx0, Akk, causal_mask);
ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, Akk, true, true, false);
Akk = ggml_mul(ctx0, lin_solve, causal_mask);
Akk = ggml_add(ctx0, Akk, identity);
cb(Akk, "attn_solved", il);
// switch back for downstream
gk_cumsum = ggml_cont_4d(ctx0, ggml_permute(ctx0, gk_cumsum, 1, 0, 2, 3), S_k, chunk_size, n_chunks, HB);
ggml_tensor * gkexp = ggml_exp(ctx0, gk_cumsum);
cb(gk_cumsum, "gk_cumsum", il);
// u = (A*beta[..., None, :]) @ v aka U_[t]
ggml_tensor * vb = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), Akk);
ggml_tensor * kbeta_gkexp = ggml_mul(ctx0, k_beta, gkexp);
cb(kbeta_gkexp, "kbeta_gkexp", il);
ggml_tensor * k_cumdecay = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gkexp)), Akk);
cb(k_cumdecay, "k_cumdecay", il);
ggml_tensor * core_attn_out = nullptr;
ggml_tensor * new_state = ggml_dup(ctx0, state);
cb(new_state, "new_state", il);
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
// extract one chunk worth of data
auto chunkify = [=](ggml_tensor * t) {
return ggml_cont(ctx0, ggml_view_4d(ctx0, t, t->ne[0], chunk_size, 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk));
};
auto chunkify_A = [=](ggml_tensor * t) {
return ggml_cont(ctx0, ggml_view_4d(ctx0, t, chunk_size, chunk_size, 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk));
};
// k [S,BT,NT,H*B] => k_chunk [S,BT,1,H*B]
ggml_tensor * k_chunk = chunkify(k);
ggml_tensor * q_chunk = chunkify(q);
ggml_tensor * vb_chunk = chunkify(vb);
// gk_cumsum [S,BT,NT,H*B] => gk_cs_chunk [S,BT,1,H*B]
ggml_tensor * gk_cs_chunk = chunkify(gk_cumsum);
ggml_tensor * k_cumdecay_chunk = chunkify(k_cumdecay);
ggml_tensor * gkexp_chunk = ggml_exp(ctx0, gk_cs_chunk);
ggml_tensor * Aqk_chunk = chunkify_A(Aqk);
ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
// new_state [S,S,1,H*B] k_cumdecay_chunk [S,BT,1,H*B]
// v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state or W_[t] @ S_[t]
ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
// v_new = v_i - v_prime or U_[t] - W_[t]*S_[t]
ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, vb_chunk, v_prime), v_prime);
ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
// q_chunk [S,BT,1,H*B] gkexp_chunk [S,BT,1,H*B]
// attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
// or Gamma_[t]*Q_]t] @ S
ggml_tensor * q_gk_exp = ggml_mul(ctx0, q_chunk, gkexp_chunk);
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_gk_exp);
attn_inter = ggml_scale(ctx0, attn_inter, scale); // scale q
// v_new_t [S,BT,1,H*B] Aqk [BT,BT,1,H*B]
// core_attn_out[:, :, i] = attn_inter + attn @ v_new or A' @ (U_[t] - W_[t]*S_[t])
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, Aqk_chunk);
// o[:, :, i] = (q_i * g_i.exp()) @ S + A @ v_i
ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
core_attn_out = core_attn_out == nullptr ? core_attn_out_chunk : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 1);
ggml_tensor * gk_cum_last =
ggml_cont(ctx0, ggml_view_4d(ctx0, gk_cs_chunk, gk_cs_chunk->ne[0], 1, gk_cs_chunk->ne[2], gk_cs_chunk->ne[3],
gk_cs_chunk->nb[1], gk_cs_chunk->nb[2], gk_cs_chunk->nb[3],
gk_cs_chunk->nb[1] * (gk_cs_chunk->ne[1] - 1)));
ggml_tensor * gkexp_last = ggml_exp(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, gk_cum_last)));
ggml_tensor * gk_diff = ggml_neg(ctx0, ggml_sub(ctx0, gk_cs_chunk, gk_cum_last));
ggml_tensor * gk_diff_exp = ggml_exp(ctx0, gk_diff);
ggml_tensor * key_gkdiff = ggml_mul(ctx0, k_chunk, gk_diff_exp);
// rearrange((g_i[:,:,-1:] - g_i).exp()*k_i, 'b h c k -> b h k c') @ (U_[t] - W_[t] @ S)
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gkdiff)));
new_state = ggml_add(ctx0,
ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gkexp_last, gkexp_last->ne[0], gkexp_last->ne[1], H_v, n_seqs)),
ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
}
core_attn_out = ggml_cont_4d(ctx0, core_attn_out, S_v, chunk_size * n_chunks, H_v, n_seqs);
// truncate padded tokens
ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(core_attn_out->type, S_v),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
output_tokens = ggml_cont(ctx0, output_tokens);
// permute back to (S_v, H_v, n_tokens, n_seqs)
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
output_tokens = ggml_cont(ctx0, output_tokens);
cb(new_state, "output_state", il);
return {output_tokens, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_kimi_linear::build_kda_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * gk,
ggml_tensor * beta,
ggml_tensor * state,
int il) {
GGML_ASSERT(ggml_is_contiguous(v));
GGML_ASSERT(ggml_is_contiguous(gk));
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1);
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(gk->ne[0] == S_k && gk->ne[1] == H_k && gk->ne[2] == n_tokens && gk->ne[3] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_k && state->ne[2] == H_v && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(gk, "gk_in", il);
// g [H,1,B,1] g_t [1,H,B,1] => [1,1,H,B]
// gk [S,H,1,B] => [S,1,H,B] gk_t [1,S,H,B]
// beta [H,1,1,B] beta_t [1,H,1,B] => [1,1,H,B]
gk = ggml_reshape_4d(ctx0, gk, S_k, 1, H_k, n_seqs);
ggml_tensor * gk_t = ggml_cont(ctx0, ggml_transpose(ctx0, gk));
ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
// Apply exponential to gk_t
gk_t = ggml_exp(ctx0, gk_t);
// Apply the gated delta rule for the single timestep
// last_recurrent_state = last_recurrent_state * gk_t
// S = S * g_i[..., None].exp()
state = ggml_mul(ctx0, state, gk_t);
ggml_tensor * state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state));
// state [S,S,H,B] k [S,1,H,B] k_state [S_v,1,H,B]
k = ggml_reshape_4d(ctx0, k, S_k, 1, H_k, n_seqs);
ggml_tensor * k_state = ggml_mul_mat(ctx0, state_t, k);
// v_i - (k_i[..., None] * S).sum(-2)
v = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
ggml_tensor * v_diff = ggml_sub(ctx0, v, k_state);
// b_i[..., None] * k_i
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta_t);
// S = S + torch.einsum('b h k, b h v -> b h k v', b_i[..., None] * k_i, v_i - (k_i[..., None] * S).sum(-2))
// v_diff_t [1,S_v,H,B] k_beta_t [1,S_k,H,B] state [S_v,S_k,H,B]
state = ggml_add(ctx0, state, ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_diff)), ggml_cont(ctx0, ggml_transpose(ctx0, k_beta))));
q = ggml_reshape_4d(ctx0, q, S_k, 1, H_k, n_seqs);
state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state));
ggml_tensor * core_attn_out = ggml_mul_mat(ctx0, state_t, q);
// core_attn_out should be [S_v, 1, H_v, n_seqs] after this
cb(core_attn_out, "output_tokens", il);
cb(state, "new_state", il);
return {core_attn_out, state};
}
examples/talk-llama/models/lfm2.cpp
+138 -123
View File
@@ -1,18 +1,149 @@
#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>;
llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params),
model(model) {
// 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, false, 0.0,
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();
auto * inp_hybrid = build_inp_mem_hybrid();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
@@ -54,122 +185,6 @@ llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params
ggml_build_forward_expand(gf, cur);
}
ggml_tensor * llm_build_lfm2::build_moe_feed_forward(ggml_tensor * cur, int il) const {
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, false, 0.0,
static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func), il);
}
ggml_tensor * llm_build_lfm2::build_dense_feed_forward(ggml_tensor * cur, int il) const {
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);
}
ggml_tensor * llm_build_lfm2::build_attn_block(ggml_tensor * cur,
ggml_tensor * inp_pos,
llm_graph_input_attn_kv * inp_attn,
int il) const {
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;
}
ggml_tensor * llm_build_lfm2::build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il) {
const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(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;
}
// Explicit template instantiations
template struct llm_build_lfm2<true>;
template struct llm_build_lfm2<false>;
examples/talk-llama/models/graph-context-mamba.cpp → examples/talk-llama/models/mamba-base.cpp
+5 -3
View File
@@ -1,8 +1,10 @@
#include "models.h"
llm_graph_context_mamba::llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {}
#include "llama-memory-recurrent.h"
ggml_tensor * llm_graph_context_mamba::build_mamba_layer(llm_graph_input_rs * inp,
llm_build_mamba_base::llm_build_mamba_base(const llm_graph_params & params) : llm_graph_context(params) {}
ggml_tensor * llm_build_mamba_base::build_mamba_layer(llm_graph_input_rs * inp,
ggml_tensor * cur,
const llama_model & model,
const llama_ubatch & ubatch,
@@ -143,7 +145,7 @@ ggml_tensor * llm_graph_context_mamba::build_mamba_layer(llm_graph_input_rs * in
return cur;
}
ggml_tensor * llm_graph_context_mamba::build_mamba2_layer(llm_graph_input_rs * inp,
ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
ggml_tensor * cur,
const llama_model & model,
const llama_ubatch & ubatch,
examples/talk-llama/models/mamba.cpp
+1 -2
View File
@@ -1,7 +1,6 @@
#include "models.h"
llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
ggml_tensor * cur;
ggml_tensor * inpL;
examples/talk-llama/models/models.h
+66 -99
View File
@@ -1,23 +1,51 @@
#pragma once
#include "../llama-model.h"
#include "../llama-graph.h"
#include "llama-model.h"
#include "llama-graph.h"
// TODO: remove in follow-up PR - move to .cpp files
#include "../llama-memory-recurrent.h"
// note: almost all graphs require atleast sqrtf, so include cmath globally
#include <cmath>
struct llm_graph_context_mamba : public llm_graph_context {
llm_graph_context_mamba(const llm_graph_params & params);
//
// base classes
//
virtual ~llm_graph_context_mamba() = default;
struct llm_build_mamba_base : public llm_graph_context {
llm_build_mamba_base(const llm_graph_params & params);
virtual ~llm_build_mamba_base() = default;
ggml_tensor * build_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il);
ggml_tensor * build_mamba2_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il) const;
};
// Base class for RWKV-related models
struct llm_build_delta_net_base : public llm_graph_context {
llm_build_delta_net_base(const llm_graph_params & params);
virtual ~llm_build_delta_net_base() = default;
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il);
};
struct llm_build_rwkv6_base : public llm_graph_context {
const llama_model & model;
@@ -58,6 +86,10 @@ struct llm_build_rwkv7_base : public llm_graph_context {
int il) const;
};
//
// models
//
struct llm_build_afmoe : public llm_graph_context {
llm_build_afmoe(const llama_model & model, const llm_graph_params & params);
};
@@ -158,6 +190,10 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_paddleocr : public llm_graph_context {
llm_build_paddleocr(const llama_model & model, const llm_graph_params & params);
};
template <bool iswa>
struct llm_build_exaone4 : public llm_graph_context {
llm_build_exaone4(const llama_model & model, const llm_graph_params & params);
@@ -175,7 +211,7 @@ struct llm_build_falcon : public llm_graph_context {
llm_build_falcon(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_falcon_h1 : public llm_graph_context_mamba {
struct llm_build_falcon_h1 : public llm_build_mamba_base {
llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params);
};
@@ -253,7 +289,7 @@ private:
const int il);
};
struct llm_build_granite_hybrid : public llm_graph_context_mamba {
struct llm_build_granite_hybrid : public llm_build_mamba_base {
llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params);
ggml_tensor * build_layer_ffn(ggml_tensor * cur, ggml_tensor * inpSA, const llama_model & model, const int il);
ggml_tensor * build_attention_layer(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn,
@@ -284,11 +320,15 @@ struct llm_build_jais : public llm_graph_context {
llm_build_jais(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_jamba : public llm_graph_context_mamba {
struct llm_build_jais2 : public llm_graph_context {
llm_build_jais2(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_jamba : public llm_build_mamba_base {
llm_build_jamba(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_kimi_linear : public llm_graph_context_mamba {
struct llm_build_kimi_linear : public llm_build_delta_net_base {
llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params);
std::pair<ggml_tensor *, ggml_tensor *> build_kda_autoregressive(
@@ -315,15 +355,9 @@ struct llm_build_kimi_linear : public llm_graph_context_mamba {
const llama_model & model;
};
template <bool iswa>
struct llm_build_lfm2 : public llm_graph_context {
const llama_model & model;
llm_build_lfm2(const llama_model & model, const llm_graph_params & params);
ggml_tensor * build_moe_feed_forward(ggml_tensor * cur, int il) const;
ggml_tensor * build_dense_feed_forward(ggml_tensor * cur, int il) const;
ggml_tensor * build_attn_block(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, int il) const;
ggml_tensor * build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il);
};
struct llm_build_llada : public llm_graph_context {
@@ -347,7 +381,7 @@ struct llm_build_maincoder : public llm_graph_context {
llm_build_maincoder(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_mamba : public llm_graph_context_mamba {
struct llm_build_mamba : public llm_build_mamba_base {
llm_build_mamba(const llama_model & model, const llm_graph_params & params);
};
@@ -379,17 +413,21 @@ struct llm_build_nemotron : public llm_graph_context {
llm_build_nemotron(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_nemotron_h : public llm_graph_context_mamba {
struct llm_build_nemotron_h : public llm_build_mamba_base {
llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params);
ggml_tensor * build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il);
ggml_tensor * build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il);
ggml_tensor * build_attention_layer(ggml_tensor * cur, llm_graph_input_attn_kv * inp_attn,
const llama_model & model, const int64_t n_embd_head, const int il);
const llama_model & model, int64_t n_embd_head, int il);
};
struct llm_build_neo_bert : public llm_graph_context {
llm_build_neo_bert(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_eurobert : public llm_graph_context {
llm_build_eurobert(const llama_model & model, const llm_graph_params & params);
};
template <bool iswa>
struct llm_build_olmo2 : public llm_graph_context {
llm_build_olmo2(const llama_model & model, const llm_graph_params & params);
@@ -428,7 +466,7 @@ struct llm_build_phi3 : public llm_graph_context {
llm_build_phi3(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_plamo2 : public llm_graph_context_mamba {
struct llm_build_plamo2 : public llm_build_mamba_base {
llm_build_plamo2(const llama_model & model, const llm_graph_params & params);
private:
ggml_tensor * build_plamo2_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il);
@@ -477,7 +515,7 @@ struct llm_build_qwen3vlmoe : public llm_graph_context {
llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_qwen3next : public llm_graph_context_mamba {
struct llm_build_qwen3next : public llm_build_delta_net_base {
llm_build_qwen3next(const llama_model & model, const llm_graph_params & params);
private:
ggml_tensor * build_layer_attn(
@@ -495,26 +533,6 @@ private:
ggml_tensor * cur,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il);
ggml_tensor * build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
@@ -529,7 +547,7 @@ private:
const llama_model & model;
};
struct llm_build_qwen35 : public llm_graph_context_mamba {
struct llm_build_qwen35 : public llm_build_delta_net_base {
llm_build_qwen35(const llama_model & model, const llm_graph_params & params);
private:
ggml_tensor * build_layer_attn(
@@ -542,38 +560,12 @@ private:
ggml_tensor * build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
ggml_tensor * build_layer_ffn(
ggml_tensor * cur,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il);
ggml_tensor * build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
@@ -588,7 +580,8 @@ private:
const llama_model & model;
};
struct llm_build_qwen35moe : public llm_graph_context_mamba {
// TODO: derive llm_build_delta_net_base instead
struct llm_build_qwen35moe : public llm_build_delta_net_base {
llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params);
private:
ggml_tensor * build_layer_attn(
@@ -601,38 +594,12 @@ private:
ggml_tensor * build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
ggml_tensor * build_layer_ffn(
ggml_tensor * cur,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il);
ggml_tensor * build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
examples/talk-llama/models/modern-bert.cpp
-7
View File
@@ -104,13 +104,6 @@ llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const ll
LLM_NORM, -1);
cb(cur, "final_norm_out", -1);
if (hparams.pooling_type == LLAMA_POOLING_TYPE_CLS) {
// extracting cls token
cur = ggml_view_1d(ctx0, cur, hparams.n_embd, 0);
cb(cur, "cls_pooled_embd", -1);
}
cb(cur, "res_embd", -1);
res->t_embd = cur;
ggml_build_forward_expand(gf, cur);
}
examples/talk-llama/models/nemotron-h.cpp
+4 -6
View File
@@ -1,9 +1,7 @@
#include "models.h"
llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params) {
llm_build_mamba_base(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -65,8 +63,8 @@ llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_
ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor * cur,
llm_graph_input_attn_kv * inp_attn,
const llama_model & model,
const int64_t n_embd_head,
const int il) {
int64_t n_embd_head,
int il) {
// compute Q and K
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
@@ -106,7 +104,7 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
return cur;
}
ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il) {
if (model.layers[il].ffn_gate_inp == nullptr) {
cur = build_ffn(cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
examples/talk-llama/models/paddleocr.cpp
+122
View File
@@ -0,0 +1,122 @@
#include "models.h"
llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
// NOTE: same with qwen2vl.cpp, but bias tensors are optional
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
// 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();
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
// 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);
cb(Qcur, "Qcur", il);
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);
cb(Kcur, "Kcur", il);
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);
cb(Vcur, "Vcur", il);
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);
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
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
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);
// feed-forward network
{
cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
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);
}
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);
}
examples/talk-llama/models/plamo2.cpp
+3 -1
View File
@@ -1,7 +1,9 @@
#include "models.h"
#include "llama-memory-recurrent.h"
llm_build_plamo2::llm_build_plamo2(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params) {
llm_build_mamba_base(params) {
ggml_tensor * cur;
ggml_tensor * inpL;
examples/talk-llama/models/qwen35.cpp
+55 -409
View File
@@ -1,10 +1,9 @@
#include "ggml.h"
#include "models.h"
#define CHUNK_SIZE 64
#include "llama-memory-recurrent.h"
llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params), model(model) {
llm_build_delta_net_base(params), model(model) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -24,27 +23,18 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
ggml_tensor * inp_pos = build_inp_pos();
ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * causal_mask =
ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
GGML_TRI_TYPE_LOWER);
ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
ggml_build_forward_expand(gf, causal_mask);
ggml_build_forward_expand(gf, identity);
ggml_build_forward_expand(gf, diag_mask);
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
cur = build_layer_attn_linear(inp->get_recr(), cur, il);
} else {
// Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
@@ -94,361 +84,6 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
ggml_build_forward_expand(gf, cur);
}
// utility to get one slice from the third dimension
// input dim: [x, y, c, b]
// output dim: [x, y, 1, b]
static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
cb(q, "q_perm", il);
cb(k, "k_perm", il);
cb(v, "v_perm", il);
cb(beta, "beta_perm", il);
cb(g, "g_perm", il);
cb(state, "state_in", il);
GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
// Do padding
const int64_t chunk_size = CHUNK_SIZE;
const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
const int64_t n_chunks = (n_tokens + pad) / chunk_size;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, pad, 0, 0, 0);
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
cb(q, "q_pad", il);
cb(k, "k_pad", il);
cb(v, "v_pad", il);
cb(beta, "beta_pad", il);
cb(g, "g_pad", il);
ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
cb(v_beta, "v_beta", il);
cb(k_beta, "k_beta", il);
q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs);
k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs);
v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j_broadcast =
ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
decay_mask = ggml_exp(ctx0, decay_mask);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_mul(ctx0, lin_solve, causal_mask);
attn = ggml_add(ctx0, attn, identity);
cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * k_cumdecay =
ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
// vectorized calculation of key_gdiff
// improved from the chunked version:
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
// key_gdiff = key * g_diff.unsqueeze(-1)
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along chunk_size dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
(g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
g_last = ggml_cont(ctx0, g_last);
cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
1, chunk_size, n_chunks, g_diff_exp->ne[3]);
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
// state to be updated per chunk
ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
// shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * core_attn_out = nullptr;
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
// shape: (S_k, chunk_size, 1, H_k * n_seqs)
ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
// shape: (S_v, chunk_size, 1, H_v * n_seqs)
ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
// shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
// shape: (chunk_size, 1, H_v * n_seqs)
ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
// attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
// replaced by precomputed attn_kq
ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
cb(attn_chunk, "attn_chunk", il);
ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
// v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
// v_new = v_i - v_prime
ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
cb(v_new, "v_new_chunk", il);
// attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
cb(attn_inter, "attn_inter_chunk", il);
// core_attn_out[:, :, i] = attn_inter + attn @ v_new
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
cb(v_attn, "v_attn_chunk", il);
ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
core_attn_out = core_attn_out == nullptr
? core_attn_out_chunk
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
new_state = ggml_add(ctx0,
ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
}
// truncate padded tokens
ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(core_attn_out->type, S_v),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
output_tokens = ggml_cont(ctx0, output_tokens);
cb(output_tokens, "output_tokens", il);
// permute back to (S_v, H_v, n_tokens, n_seqs)
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
output_tokens = ggml_cont(ctx0, output_tokens);
return {output_tokens, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
// Apply exponential to g_t
g_t = ggml_exp(ctx0, g_t);
// Apply the gated delta rule for the single timestep
// last_recurrent_state = last_recurrent_state * g_t
state = ggml_mul(ctx0, state, g_t);
// kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed);
// we need to sum over dim=-2, so we transpose, sum, then transpose again
kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
// v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
// delta = (v_t - kv_mem) * beta_t
ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs]
ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
// last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
state = ggml_add(ctx0, state, k_t_delta);
// Compute the attention output
// core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t
ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
// again, since it's over dim = -2, transpose, sum, transpose back
ggml_tensor * core_attn_out =
ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
// core_attn_out should be [S_v, 1, H_v, n_seqs] after this
cb(core_attn_out, "output_tokens", il);
cb(state, "new_state", il);
return {core_attn_out, state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_qkvz(
ggml_tensor * input,
int il) {
@@ -560,9 +195,6 @@ ggml_tensor * llm_build_qwen35::build_layer_attn(
ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const auto * mctx_cur = inp->mctx;
@@ -586,8 +218,11 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
ggml_tensor * z = qkvz.second;
ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
cb(beta, "beta", il);
beta = ggml_sigmoid(ctx0, beta);
ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
cb(alpha, "alpha", il);
@@ -595,15 +230,16 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
cb(alpha_softplus, "a_softplus", il);
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il);
gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
// Get convolution states from cache
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
// bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
// Build the convolution states tensor
ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
cb(conv_states, "conv_states", il);
@@ -612,11 +248,12 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;
const int64_t conv_kernel_size = conv_kernel->ne[0];
const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
cb(conv_states, "conv_states_reshaped", il);
qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
cb(qkv_mixed, "qkv_mixed_permuted", il);
qkv_mixed = ggml_transpose(ctx0, qkv_mixed);
cb(qkv_mixed, "qkv_mixed_transposed", il);
ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
cb(conv_input, "conv_input", il);
@@ -634,9 +271,11 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
// Apply SSM convolution
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
cb(state, "state_predelta", il);
ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
cb(conv_output_proper, "conv_output_raw", il);
@@ -650,31 +289,41 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
// Extract the convolved Q, K, V from conv_output
ggml_tensor * q_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
ggml_row_size(conv_qkv_mix->type, head_k_dim),
nb1_qkv,
nb1_qkv * n_seq_tokens,
0);
ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
ggml_row_size(conv_qkv_mix->type, head_k_dim),
nb1_qkv,
nb1_qkv * n_seq_tokens,
head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,
ggml_row_size(conv_qkv_mix->type, head_v_dim),
nb1_qkv,
nb1_qkv * n_seq_tokens,
ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));
cb(q_conv, "q_conv", il);
ggml_tensor * k_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(k_conv, "k_conv", il);
ggml_tensor * v_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(v_conv, "v_conv", il);
// Unsqueeze them
q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
const float eps_norm = hparams.f_norm_rms_eps;
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
cb(state, "state_predelta", il);
q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
// if head keys and value keys are different, repeat Q/K to match V's head count
// V heads are in tiled order (from conversion), so simple tiled repeat works
//q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
//k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
// TODO: try to avoid these explicit repeats by utilizing op broadcast
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
@@ -688,7 +337,7 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
@@ -697,19 +346,15 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
// Update the recurrent states
ggml_build_forward_expand(gf,
ggml_cpy(ctx0, new_state,
ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
// Reshape both attn_out_final and z to 2D tensors for normalization
// attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
ggml_cpy(ctx0, new_state,
ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
// z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// Apply gated normalization: self.norm(core_attn_out, z)
ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
@@ -720,7 +365,8 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions
cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}
examples/talk-llama/models/qwen35moe.cpp
+57 -411
View File
@@ -1,10 +1,9 @@
#include "ggml.h"
#include "models.h"
#define CHUNK_SIZE 64
#include "llama-memory-recurrent.h"
llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params), model(model) {
llm_build_delta_net_base(params), model(model) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -24,27 +23,18 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
ggml_tensor * inp_pos = build_inp_pos();
ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * causal_mask =
ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
GGML_TRI_TYPE_LOWER);
ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
ggml_build_forward_expand(gf, causal_mask);
ggml_build_forward_expand(gf, identity);
ggml_build_forward_expand(gf, diag_mask);
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
cur = build_layer_attn_linear(inp->get_recr(), cur, il);
} else {
// Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
@@ -94,362 +84,6 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
ggml_build_forward_expand(gf, cur);
}
// utility to get one slice from the third dimension
// input dim: [x, y, c, b]
// output dim: [x, y, 1, b]
static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
cb(q, "q_perm", il);
cb(k, "k_perm", il);
cb(v, "v_perm", il);
cb(beta, "beta_perm", il);
cb(g, "g_perm", il);
cb(state, "state_in", il);
GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
// Do padding
const int64_t chunk_size = CHUNK_SIZE;
const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
const int64_t n_chunks = (n_tokens + pad) / chunk_size;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, pad, 0, 0, 0);
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
cb(q, "q_pad", il);
cb(k, "k_pad", il);
cb(v, "v_pad", il);
cb(beta, "beta_pad", il);
cb(g, "g_pad", il);
ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
cb(v_beta, "v_beta", il);
cb(k_beta, "k_beta", il);
q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs);
k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs);
v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j_broadcast =
ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
decay_mask = ggml_exp(ctx0, decay_mask);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_mul(ctx0, lin_solve, causal_mask);
attn = ggml_add(ctx0, attn, identity);
cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * k_cumdecay =
ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
// vectorized calculation of key_gdiff
// improved from the chunked version:
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
// key_gdiff = key * g_diff.unsqueeze(-1)
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along chunk_size dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
(g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
g_last = ggml_cont(ctx0, g_last);
cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
1, chunk_size, n_chunks, g_diff_exp->ne[3]);
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
// state to be updated per chunk
ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
// shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * core_attn_out = nullptr;
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
// shape: (S_k, chunk_size, 1, H_k * n_seqs)
ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
// shape: (S_v, chunk_size, 1, H_v * n_seqs)
ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
// shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
// shape: (chunk_size, 1, H_v * n_seqs)
ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
// attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
// replaced by precomputed attn_kq
ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
cb(attn_chunk, "attn_chunk", il);
ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
// v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
// v_new = v_i - v_prime
ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
cb(v_new, "v_new_chunk", il);
// attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
cb(attn_inter, "attn_inter_chunk", il);
// core_attn_out[:, :, i] = attn_inter + attn @ v_new
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
cb(v_attn, "v_attn_chunk", il);
ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
core_attn_out = core_attn_out == nullptr
? core_attn_out_chunk
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
new_state = ggml_add(ctx0,
ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
}
// truncate padded tokens
ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(core_attn_out->type, S_v),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
output_tokens = ggml_cont(ctx0, output_tokens);
cb(output_tokens, "output_tokens", il);
// permute back to (S_v, H_v, n_tokens, n_seqs)
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
output_tokens = ggml_cont(ctx0, output_tokens);
return {output_tokens, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
// Apply exponential to g_t
g_t = ggml_exp(ctx0, g_t);
// Apply the gated delta rule for the single timestep
// last_recurrent_state = last_recurrent_state * g_t
state = ggml_mul(ctx0, state, g_t);
// kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed);
// we need to sum over dim=-2, so we transpose, sum, then transpose again
kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
// v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
// delta = (v_t - kv_mem) * beta_t
ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs]
ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
// last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
state = ggml_add(ctx0, state, k_t_delta);
// Compute the attention output
// core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t
ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
// again, since it's over dim = -2, transpose, sum, transpose back
ggml_tensor * core_attn_out =
ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
// core_attn_out should be [S_v, 1, H_v, n_seqs] after this
cb(core_attn_out, "output_tokens", il);
cb(state, "new_state", il);
return {core_attn_out, state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_qkvz(
ggml_tensor * input,
int il) {
@@ -561,9 +195,6 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const auto * mctx_cur = inp->mctx;
@@ -587,8 +218,11 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
ggml_tensor * z = qkvz.second;
ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
cb(beta, "beta", il);
beta = ggml_sigmoid(ctx0, beta);
ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
cb(alpha, "alpha", il);
@@ -596,15 +230,16 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
cb(alpha_softplus, "a_softplus", il);
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il);
gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
// Get convolution states from cache
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
// bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
// Build the convolution states tensor
ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
cb(conv_states, "conv_states", il);
@@ -613,11 +248,12 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;
const int64_t conv_kernel_size = conv_kernel->ne[0];
const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
cb(conv_states, "conv_states_reshaped", il);
qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
cb(qkv_mixed, "qkv_mixed_permuted", il);
qkv_mixed = ggml_transpose(ctx0, qkv_mixed);
cb(qkv_mixed, "qkv_mixed_transposed", il);
ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
cb(conv_input, "conv_input", il);
@@ -635,9 +271,11 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
// Apply SSM convolution
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
cb(state, "state_predelta", il);
ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
cb(conv_output_proper, "conv_output_raw", il);
@@ -651,31 +289,41 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
// Extract the convolved Q, K, V from conv_output
ggml_tensor * q_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
ggml_row_size(conv_qkv_mix->type, head_k_dim),
nb1_qkv,
nb1_qkv * n_seq_tokens,
0);
ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
ggml_row_size(conv_qkv_mix->type, head_k_dim),
nb1_qkv,
nb1_qkv * n_seq_tokens,
head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,
ggml_row_size(conv_qkv_mix->type, head_v_dim),
nb1_qkv,
nb1_qkv * n_seq_tokens,
ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));
cb(q_conv, "q_conv", il);
ggml_tensor * k_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(k_conv, "k_conv", il);
ggml_tensor * v_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(v_conv, "v_conv", il);
// Unsqueeze them
q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
const float eps_norm = hparams.f_norm_rms_eps;
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
cb(state, "state_predelta", il);
q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
// if head keys and value keys are different, repeat Q/K to match V's head count
// V heads are in tiled order (from conversion), so simple tiled repeat works
//q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
//k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
// TODO: try to avoid these explicit repeats by utilizing op broadcast
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
@@ -689,7 +337,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
@@ -698,19 +346,15 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
// Update the recurrent states
ggml_build_forward_expand(gf,
ggml_cpy(ctx0, new_state,
ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
// Reshape both attn_out_final and z to 2D tensors for normalization
// attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
ggml_cpy(ctx0, new_state,
ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
// z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// Apply gated normalization: self.norm(core_attn_out, z)
ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
@@ -721,7 +365,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions
cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}
@@ -735,7 +380,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int
model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used, LLM_FFN_SILU,
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
nullptr, model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// Add shared experts if present - following Qwen3Next reference implementation
examples/talk-llama/models/qwen3next.cpp
+9 -327
View File
@@ -1,10 +1,9 @@
#include "ggml.h"
#include "models.h"
#define CHUNK_SIZE 64
#include "llama-memory-recurrent.h"
llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params), model(model) {
llm_build_delta_net_base(params), model(model) {
ggml_tensor * cur;
ggml_tensor * inpL;
@@ -22,6 +21,8 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
@@ -83,326 +84,6 @@ static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(b, "b_in", il);
cb(g, "g_in", il);
q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
g = ggml_permute(ctx0, g, 2, 1, 3, 0); // [ 1, n_tokens, H_v, n_seqs]
b = ggml_permute(ctx0, b, 2, 0, 1, 3); // [ 1, n_tokens, H_v, n_seqs]
const int CS = CHUNK_SIZE;
const int pad = (CS - n_tokens % CS) % CS;
const int n_chunks = (n_tokens + pad) / CS;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, 0, pad, 0, 0);
b = ggml_pad(ctx0, b, 0, pad, 0, 0);
ggml_tensor * v_b = ggml_mul(ctx0, v, b);
ggml_tensor * k_b = ggml_mul(ctx0, k, b);
cb(v_b, "v_b", il);
cb(k_b, "k_b", il);
q = ggml_reshape_4d(ctx0, q, S_k, CS, n_chunks, H_k * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, CS, n_chunks, H_k * n_seqs);
k_b = ggml_reshape_4d(ctx0, k_b, S_k, CS, n_chunks, H_v * n_seqs);
v = ggml_reshape_4d(ctx0, v, S_v, CS, n_chunks, H_v * n_seqs);
v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs);
g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs);
b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs);
// [CS, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_cs = ggml_cumsum(ctx0, g);
cb(g_cs, "g_cs", il);
ggml_tensor * g_cs_i = g_cs;
ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
// [CS, CS, n_chunks, H_v * n_seqs]
ggml_tensor * decay_mask;
decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
decay_mask = ggml_exp(ctx0, decay_mask);
cb(decay_mask, "decay_mask", il);
// [CS, CS, n_chunks, H_k * n_seqs]
ggml_tensor * kb;
kb = ggml_mul_mat(ctx0, k, k_b);
kb = ggml_mul (ctx0, kb, decay_mask);
// [CS, CS, n_chunks, H_k * n_seqs]
ggml_tensor * attn;
attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
ggml_tensor * identity;
identity = ggml_view_1d(ctx0, attn, CS, 0);
identity = ggml_fill (ctx0, identity, 1.0f);
identity = ggml_diag (ctx0, identity);
ggml_tensor * lhs = ggml_add(ctx0, attn, identity);
cb(lhs, "dnet_add_ch_lhs", il);
attn = ggml_neg(ctx0, attn);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_add(ctx0, lin_solve, identity);
cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs]
// [S_v, CS, n_chunks, H_v * n_seqs]
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);
// [CS, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);
k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
// [CS, S_k, n_chunks, H_k * n_seqs]
ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp);
cb(kbg, "k_beta_g_exp", il);
// [S_k, CS, n_chunks, H_k * n_seqs]
ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
cb(k_cd, "k_cumdecay", il);
// [S_k, CS, n_chunks, H_k * n_seqs]
ggml_tensor * g_exp_t = ggml_transpose(ctx0, g_exp);
ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t);
// [CS, CS, n_chunks, H_k * n_seqs]
ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
kq = ggml_mul(ctx0, kq, decay_mask);
kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
cb(kq, "kq", il);
// vectorized calculation of key_gdiff
// improved from the chunked version:
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
// key_gdiff = key * g_diff.unsqueeze(-1)
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along CS dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
// [1, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3],
g_cs->nb[1],
g_cs->nb[2],
g_cs->nb[3],
ggml_row_size(g_cs->type, g_cs->ne[0] - 1));
cb(g_last, "g_last", il);
// TODO: remove this cont when CUDA supports non-cont unary ops
g_last = ggml_cont(ctx0, g_last);
// [1, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
cb(g_last_exp, "g_last_exp", il);
// [CS, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last));
cb(g_diff, "g_diff", il);
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
ggml_tensor * g_diff_exp_t = ggml_transpose(ctx0, g_diff_exp);
// [S_k, CS, n_chunks, H_v * n_seqs]
ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
cb(kg, "key_gdiff", il);
// [CS, S_k, n_chunks, H_v * n_seqs]
ggml_tensor * kg_t = ggml_cont(ctx0, ggml_transpose(ctx0, kg));
cb(kg_t, "key_gdiff_t", il);
ggml_tensor * s_t = ggml_transpose(ctx0, s);
s_t = ggml_cont_4d(ctx0, s_t, S_v, S_v, 1, H_v * n_seqs);
cb(s_t, "dnet_add_ch_state", il);
// [CS, S_v, n_chunks, H_v * n_seqs]
ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v));
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
ggml_tensor * ch_k_cd = get_slice_2d(ctx0, k_cd, chunk); // [S_k, CS, 1, H_k * n_seqs]
ggml_tensor * ch_v_t = get_slice_2d(ctx0, v_t, chunk); // [ CS, S_v, 1, H_v * n_seqs]
ggml_tensor * ch_kq = get_slice_2d(ctx0, kq, chunk); // [ CS, CS, 1, H_k * n_seqs]
ggml_tensor * ch_q_g_exp = get_slice_2d(ctx0, q_g_exp, chunk); // [S_k, CS, 1, H_k * n_seqs]
ggml_tensor * ch_kg_t = get_slice_2d(ctx0, kg_t, chunk); // [ CS, S_k, 1, H_v * n_seqs]
// [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s_t);
cb(v_t_p, "v_prime", il);
// [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p);
cb(v_t_new, "v_t_new", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq);
cb(v_attn, "v_attn", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s_t, ch_q_g_exp);
cb(attn_inter, "attn_inter", il);
// [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * o_ch = ggml_add(ctx0, attn_inter, v_attn);
cb(o_ch, "dnet_add_ch_attn_out", il);
v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]);
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// TODO: head broadcast might not work here - probably will need a transpose
ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk);
s_t = ggml_mul(ctx0, s_t, ch_g_last_exp);
s_t = ggml_add(ctx0, s_t, kgv);
cb(s_t, "dnet_add_ch_state", il);
}
s_t = ggml_reshape_4d(ctx0, s_t, S_v, S_v, H_v, n_seqs);
// truncate padded tokens
ggml_tensor * o = ggml_view_4d(ctx0, v,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(v->type, S_v),
ggml_row_size(v->type, S_v * CS * n_chunks),
ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0);
o = ggml_permute (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs]
return {o, s};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b, // beta
ggml_tensor * s, // state
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1);
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(b, "b_in", il);
cb(g, "g_in", il);
g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs);
b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
// [S_v, S_v, H_v, n_seqs]
g = ggml_exp(ctx0, g);
s = ggml_mul(ctx0, s, g);
ggml_tensor * s_t = ggml_cont(ctx0, ggml_transpose(ctx0, s));
// [1, S_v, H_v, n_seqs]
ggml_tensor * sk;
sk = ggml_mul (ctx0, s_t, k);
sk = ggml_sum_rows(ctx0, sk);
// [S_v, 1, H_v, n_seqs]
ggml_tensor * d;
d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk));
d = ggml_mul(ctx0, d, b);
// [1, S_v, H_v, n_seqs]
ggml_tensor * d_t;
d_t = ggml_transpose(ctx0, d);
// [S_v, S_v, H_v, n_seqs]
ggml_tensor * kd;
k = ggml_repeat(ctx0, k, s);
kd = ggml_mul (ctx0, k, d_t);
s_t = ggml_add(ctx0, s_t, kd);
cb(s_t, "dnet_add_ar_state", il);
ggml_tensor * s_q = ggml_mul (ctx0, s_t, q);
ggml_tensor * o = ggml_sum_rows(ctx0, s_q);
o = ggml_permute (ctx0, o, 2, 0, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs]
return {o, s};
}
ggml_tensor * llm_build_qwen3next::build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
@@ -627,8 +308,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
ggml_tensor * beta = ggml_sigmoid(ctx0, b);
beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
// Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]
ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
@@ -639,6 +318,9 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il);
beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
// Get convolution states from cache
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
@@ -674,7 +356,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
@@ -798,7 +479,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used, LLM_FFN_SILU,
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
nullptr, model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// Add shared experts if present - following Qwen3Next reference implementation
examples/talk-llama/models/rwkv6-base.cpp
+2
View File
@@ -1,5 +1,7 @@
#include "models.h"
#include "llama-memory-recurrent.h"
llm_build_rwkv6_base::llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params),
model(model) {}
examples/talk-llama/models/rwkv7-base.cpp
+2
View File
@@ -1,5 +1,7 @@
#include "models.h"
#include "llama-memory-recurrent.h"
llm_build_rwkv7_base::llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params),
model(model) {}