nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 24 Packages Projects Releases Wiki Activity

models : optimize qwen3next graph (#19375)

Browse Source 
* models : optimizing qwen3next graph

* cont

* wip

* wip

* wip

* wip

* wip

* wip

* wip

* wip

* wip

* wip

* cont : remove redundant q, g chunking

* minor

* minor

* avoid passing masks around

* avoid concats during chunking

* naming + shapes

* update names and use prefix to disable CUDA graphs
This commit is contained in:
Georgi Gerganov
2026-02-14 12:57:36 +02:00
committed by GitHub
parent b7742cf321
commit 1725e316c1
4 changed files with 263 additions and 300 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/ggml-cuda.cu
+5 -1
View File
@@ -2872,6 +2872,7 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased"; const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased";
const std::string nemotron_h_block_out_prefix = "nemotron_h_block_out"; const std::string nemotron_h_block_out_prefix = "nemotron_h_block_out";
const std::string mamba2_y_add_d_prefix = "mamba2_y_add_d"; const std::string mamba2_y_add_d_prefix = "mamba2_y_add_d";
const std::string delta_net_prefix = "dnet_add";
for (int i = 0; i < cgraph->n_nodes; i++) { for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i]; ggml_tensor * node = cgraph->nodes[i];
@@ -2902,7 +2903,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 && strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 &&
strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0 && strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0 &&
strncmp(node->name, nemotron_h_block_out_prefix.c_str(), nemotron_h_block_out_prefix.size()) != 0 && strncmp(node->name, nemotron_h_block_out_prefix.c_str(), nemotron_h_block_out_prefix.size()) != 0 &&
strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0) { strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0 &&
strncmp(node->name, delta_net_prefix.c_str(), delta_net_prefix.size()) != 0) {
// disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
// by means of matching node names. See // by means of matching node names. See
// https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
@@ -4544,6 +4546,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_UNARY_OP_CEIL: case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_ROUND: case GGML_UNARY_OP_ROUND:
case GGML_UNARY_OP_TRUNC: case GGML_UNARY_OP_TRUNC:
// TODO: should become:
//return ggml_is_contiguous_rows(op->src[0]);
return ggml_is_contiguous(op->src[0]); return ggml_is_contiguous(op->src[0]);
default: default:
return false; return false;
ggml/src/ggml-metal/ggml-metal-common.cpp
+1
View File
@@ -273,6 +273,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
case GGML_OP_DIAG: case GGML_OP_DIAG:
case GGML_OP_MUL: case GGML_OP_MUL:
case GGML_OP_ADD: case GGML_OP_ADD:
case GGML_OP_SUB:
case GGML_OP_DIV: case GGML_OP_DIV:
case GGML_OP_GLU: case GGML_OP_GLU:
case GGML_OP_SCALE: case GGML_OP_SCALE:
src/models/models.h
-6
View File
@@ -489,9 +489,6 @@ private:
ggml_tensor * build_layer_attn_linear( ggml_tensor * build_layer_attn_linear(
llm_graph_input_rs * inp, llm_graph_input_rs * inp,
ggml_tensor * cur, ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il); int il);
ggml_tensor * build_layer_ffn( ggml_tensor * build_layer_ffn(
@@ -506,9 +503,6 @@ private:
ggml_tensor * g, ggml_tensor * g,
ggml_tensor * beta, ggml_tensor * beta,
ggml_tensor * state, ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il); int il);
// returns pair of output and new state // returns pair of output and new state
src/models/qwen3next.cpp
+249 -285
View File
@@ -16,17 +16,6 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
ggml_tensor * inp_pos = build_inp_pos(); ggml_tensor * inp_pos = build_inp_pos();
ggml_tensor * inp_out_ids = build_inp_out_ids(); ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * 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 * identity = ggml_diag(ctx0, ggml_fill_inplace(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) { for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL; ggml_tensor * inpSA = inpL;
@@ -36,7 +25,7 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
// Determine layer type and build appropriate attention mechanism // Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) { if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net) // 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 { } else {
// Full attention layer // Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il); cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
@@ -99,11 +88,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
ggml_tensor * k, ggml_tensor * k,
ggml_tensor * v, ggml_tensor * v,
ggml_tensor * g, ggml_tensor * g,
ggml_tensor * beta, ggml_tensor * b,
ggml_tensor * state, ggml_tensor * s,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) { int il) {
const int64_t S_k = q->ne[0]; const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1]; const int64_t H_k = q->ne[1];
@@ -113,134 +99,123 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
const int64_t S_v = v->ne[0]; const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1]; const int64_t H_v = v->ne[1];
GGML_ASSERT(v->ne[2] == n_tokens); GGML_ASSERT(S_k == S_v);
GGML_ASSERT(k->ne[2] == n_tokens); GGML_ASSERT(H_v % H_k == 0);
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(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(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(H_k == H_v); // we did a repeat to make sure this is the case 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 eps_norm = hparams.f_norm_rms_eps; const float scale = 1.0f / sqrtf(S_k);
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); q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il); cb(q, "q_in", il);
cb(k, "k_in", il); cb(k, "k_in", il);
cb(v, "v_in", il); cb(v, "v_in", il);
cb(beta, "beta_in", il); cb(b, "b_in", il);
cb(g, "g_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); q = 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_v, n_tokens, H_v, n_seqs); k = 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); v = 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); 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]
beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); const int CS = CHUNK_SIZE;
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
cb(q, "q_perm", il); const int pad = (CS - n_tokens % CS) % CS;
cb(k, "k_perm", il); const int n_chunks = (n_tokens + pad) / CS;
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); q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0); k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0); v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, pad, 0, 0, 0); g = ggml_pad(ctx0, g, 0, pad, 0, 0);
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); b = ggml_pad(ctx0, b, 0, pad, 0, 0);
cb(q, "q_pad", il); ggml_tensor * v_b = ggml_mul(ctx0, v, b);
cb(k, "k_pad", il); ggml_tensor * k_b = ggml_mul(ctx0, k, b);
cb(v, "v_pad", il);
cb(beta, "beta_pad", il);
cb(g, "g_pad", il);
ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); cb(v_b, "v_b", il);
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); cb(k_b, "k_b", il);
cb(v_beta, "v_beta", il); q = ggml_reshape_4d(ctx0, q, S_k, CS, n_chunks, H_k * n_seqs);
cb(k_beta, "k_beta", il); 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);
q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * 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); // [CS, 1, n_chunks, H_v * n_seqs]
beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); ggml_tensor * g_cs = ggml_cumsum(ctx0, g);
cb(g_cs, "g_cs", il);
ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); ggml_tensor * g_cs_i = g_cs;
cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, 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); g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, 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 = // [CS, CS, n_chunks, H_v * n_seqs]
ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); ggml_tensor * decay_mask;
decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
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_exp(ctx0, decay_mask);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); cb(decay_mask, "decay_mask", il);
ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); // [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);
ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); // [CS, CS, n_chunks, H_k * n_seqs]
ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); ggml_tensor * attn;
cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); ggml_tensor * identity;
ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); 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); 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, lin_solve, identity);
attn = ggml_add(ctx0, attn, identity); cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs]
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); // [S_v, CS, n_chunks, H_v * n_seqs]
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);
ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); // [CS, 1, n_chunks, H_v * n_seqs]
ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);
ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * k_cumdecay = // [CS, S_k, n_chunks, H_k * n_seqs]
ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp);
cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) cb(kbg, "k_beta_g_exp", il);
ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q); // [S_k, CS, n_chunks, H_k * n_seqs]
attn_kq = ggml_mul(ctx0, attn_kq, decay_mask); ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
attn_kq = ggml_mul(ctx0, attn_kq, diag_mask); cb(k_cd, "k_cumdecay", il);
cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
// [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 // vectorized calculation of key_gdiff
// improved from the chunked version: // improved from the chunked version:
@@ -250,109 +225,98 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along chunk_size dimension (ne0) // get last element in g_cumsum along CS dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...] // 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], // [1, 1, n_chunks, H_v * n_seqs]
g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3], ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3],
(g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum)); 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); g_last = ggml_cont(ctx0, g_last);
cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
// [1, 1, n_chunks, H_v * n_seqs]
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last); 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) cb(g_last_exp, "g_last_exp", il);
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last)); // [CS, 1, n_chunks, H_v * n_seqs]
cb(g_diff, "g_diff", il); // shape: (chunk_size, 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 = ggml_exp(ctx0, g_diff);
ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, ggml_tensor * g_diff_exp_t = ggml_transpose(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); // [S_k, CS, n_chunks, H_v * n_seqs]
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
cb(kg, "key_gdiff", il);
ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); // [CS, S_k, n_chunks, H_v * n_seqs]
cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, 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);
// state to be updated per chunk // [CS, S_v, n_chunks, H_v * n_seqs]
ggml_tensor * new_state = state; // ggml_dup(ctx0, state); ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v));
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++) { for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
// shape: (S_k, chunk_size, 1, H_k * n_seqs) ggml_tensor * ch_k_cd = get_slice_2d(ctx0, k_cd, chunk); // [S_k, CS, 1, H_k * n_seqs]
ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul 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]
// shape: (S_v, chunk_size, 1, H_v * n_seqs) // [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s_t);
cb(v_t_p, "v_prime", il);
// shape: (chunk_size, 1, n_chunks, H_v * n_seqs) // [CS, S_v, 1, H_v * n_seqs]
ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p);
cb(v_t_new, "v_t_new", il);
// shape: (chunk_size, 1, H_v * n_seqs) // [S_v, CS, 1, H_v * n_seqs]
ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq);
cb(v_attn, "v_attn", il);
// attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) // [S_v, CS, 1, H_v * n_seqs]
// replaced by precomputed attn_kq ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s_t, ch_q_g_exp);
ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk); cb(attn_inter, "attn_inter", il);
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); // [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_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]);
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 // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); // TODO: head broadcast might not work here - probably will need a transpose
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew // 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)); ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk);
new_state = ggml_add(ctx0, s_t = ggml_mul(ctx0, s_t, ch_g_last_exp);
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)), s_t = ggml_add(ctx0, s_t, kgv);
ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); 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 // truncate padded tokens
ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, ggml_tensor * o = ggml_view_4d(ctx0, v,
S_v, n_tokens, H_v, n_seqs, S_v, n_tokens, H_v, n_seqs,
ggml_row_size(core_attn_out->type, S_v), ggml_row_size(v->type, S_v),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks), ggml_row_size(v->type, S_v * CS * n_chunks),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0); ggml_row_size(v->type, S_v * CS * 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) o = ggml_permute (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3); s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs]
output_tokens = ggml_cont(ctx0, output_tokens);
return {output_tokens, new_state}; return {o, s};
} }
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_autoregressive( std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_autoregressive(
@@ -360,8 +324,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_aut
ggml_tensor * k, ggml_tensor * k,
ggml_tensor * v, ggml_tensor * v,
ggml_tensor * g, ggml_tensor * g,
ggml_tensor * beta, ggml_tensor * b, // beta
ggml_tensor * state, ggml_tensor * s, // state
int il) { int il) {
const int64_t S_k = q->ne[0]; const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1]; const int64_t H_k = q->ne[1];
@@ -371,75 +335,72 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_aut
const int64_t S_v = v->ne[0]; const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1]; const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing GGML_ASSERT(n_tokens == 1);
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens); GGML_ASSERT(S_k == S_v);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); GGML_ASSERT(H_v % H_k == 0);
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(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(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(H_k == H_v); // we did a repeat to make sure this is the case 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 eps_norm = hparams.f_norm_rms_eps; const float scale = 1.0f / sqrtf(S_k);
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); q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
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(q, "q_in", il);
cb(k, "k_in", il); cb(k, "k_in", il);
cb(v, "v_in", il); cb(v, "v_in", il);
cb(beta, "beta_in", il); cb(b, "b_in", il);
cb(g, "g_in", il); cb(g, "g_in", il);
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs);
b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs); // [S_v, S_v, H_v, n_seqs]
ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs); g = ggml_exp(ctx0, g);
s = ggml_mul(ctx0, s, g);
// Apply exponential to g_t ggml_tensor * s_t = ggml_cont(ctx0, ggml_transpose(ctx0, s));
g_t = ggml_exp(ctx0, g_t);
// Apply the gated delta rule for the single timestep // [1, S_v, H_v, n_seqs]
// last_recurrent_state = last_recurrent_state * g_t ggml_tensor * sk;
state = ggml_mul(ctx0, state, g_t); sk = ggml_mul (ctx0, s_t, k);
sk = ggml_sum_rows(ctx0, sk);
// kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2) // [S_v, 1, H_v, n_seqs]
ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs); ggml_tensor * d;
ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed); d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk));
// we need to sum over dim=-2, so we transpose, sum, then transpose again d = ggml_mul(ctx0, d, b);
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) // [1, S_v, H_v, n_seqs]
ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs); ggml_tensor * d_t;
// delta = (v_t - kv_mem) * beta_t d_t = ggml_transpose(ctx0, d);
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 // [S_v, S_v, H_v, n_seqs]
ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta); ggml_tensor * kd;
state = ggml_add(ctx0, state, k_t_delta); k = ggml_repeat(ctx0, k, s);
kd = ggml_mul (ctx0, k, d_t);
// Compute the attention output s_t = ggml_add(ctx0, s_t, kd);
// 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(s_t, "dnet_add_ar_state", il);
cb(core_attn_out, "output_tokens", il);
cb(state, "new_state", il);
return {core_attn_out, state}; 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 * llm_build_qwen3next::build_norm_gated(
@@ -473,38 +434,28 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
// The split should be along dimension 0 (the feature dimension) // The split should be along dimension 0 (the feature dimension)
ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0); Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);
cb(Qcur, "Qcur_view", il);
ggml_tensor * gate = ggml_tensor * gate =
ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full)); Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));
cb(Qcur, "Qcur", il);
cb(gate, "gate", il); cb(gate, "gate", il);
// Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention
Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
cb(Qcur, "Qcur_reshaped", il);
// Apply Q normalization
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
// Apply K normalization
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 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 = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
cb(Kcur, "Kcur_normed", il); cb(Kcur, "Kcur_normed", il);
// Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads)
gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
cb(gate, "gate_reshaped", il);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
// Apply RoPE
Qcur = ggml_rope_ext( Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr, ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -519,7 +470,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
// Attention computation
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
cur = build_attn(inp, cur = build_attn(inp,
@@ -527,10 +477,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_pregate", il); cb(cur, "attn_pregate", il);
ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont
cb(gate_sigmoid, "gate_sigmoid", il); gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
cur = ggml_mul(ctx0, cur, gate_sigmoid); gate = ggml_sigmoid(ctx0, gate);
cb(gate, "gate_sigmoid", il);
gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
cur = ggml_mul(ctx0, cur, gate);
cb(cur, "attn_gated", il); cb(cur, "attn_gated", il);
cur = build_lora_mm(model.layers[il].wo, cur); cur = build_lora_mm(model.layers[il].wo, cur);
@@ -560,7 +515,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
cb(z, "z", il); cb(z, "z", il);
return { qkv_mixed, z }; return { qkv_mixed, z };
} else { } else {
// legacy (slower) path // legacy (slower) path
ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input); ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input);
@@ -624,9 +578,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
llm_graph_input_rs * inp, llm_graph_input_rs * inp,
ggml_tensor * cur, ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) { int il) {
const auto * mctx_cur = inp->mctx; const auto * mctx_cur = inp->mctx;
@@ -671,7 +622,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped)); split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));
cb(a, "a", il); cb(a, "a", il);
ggml_tensor * beta = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs); // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont
b = ggml_cont(ctx0, b);
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] // 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); ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
@@ -679,6 +635,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
cb(alpha_softplus, "a_softplus", il); cb(alpha_softplus, "a_softplus", il);
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il); cb(gate, "gate", il);
@@ -686,8 +643,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_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 // Build the convolution states tensor
ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
cb(conv_states, "conv_states", il); cb(conv_states, "conv_states", il);
@@ -696,11 +651,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;
const int64_t conv_kernel_size = conv_kernel->ne[0]; 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; 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); cb(conv_states, "conv_states_reshaped", il);
qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); qkv_mixed = ggml_transpose(ctx0, qkv_mixed);
cb(qkv_mixed, "qkv_mixed_permuted", il); cb(qkv_mixed, "qkv_mixed_transposed", il);
ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
cb(conv_input, "conv_input", il); cb(conv_input, "conv_input", il);
@@ -720,7 +676,10 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il); 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); ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
cb(conv_output_proper, "conv_output_raw", il); cb(conv_output_proper, "conv_output_raw", il);
@@ -734,26 +693,36 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
// Extract the convolved Q, K, V from conv_output // Extract the convolved Q, K, V from conv_output
ggml_tensor * q_conv = ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0); ggml_row_size(conv_qkv_mix->type, head_k_dim),
cb(q_conv, "q_conv", il); nb1_qkv,
ggml_tensor * k_conv = nb1_qkv * n_seq_tokens,
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 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)); 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);
cb(k_conv, "k_conv", il); 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); cb(v_conv, "v_conv", il);
// Unsqueeze them const float eps_norm = hparams.f_norm_rms_eps;
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);
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
cb(state, "state_predelta", il);
//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 head keys and value keys are different, repeat to force tensors into matching shapes
if (num_k_heads != num_v_heads) { if (num_k_heads != num_v_heads) {
@@ -786,7 +755,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
if (n_seq_tokens == 1) { if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il); attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else { } 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 * output = attn_out.first;
ggml_tensor * new_state = attn_out.second; ggml_tensor * new_state = attn_out.second;
@@ -799,15 +768,11 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, 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)))); 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);
// z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] // 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) // 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] // 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); ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
@@ -818,7 +783,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
cb(cur, "linear_attn_out", il); cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions // 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; return cur;
} }
@@ -852,11 +818,9 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
cb(shared_gate, "shared_expert_gate", il); cb(shared_gate, "shared_expert_gate", il);
// Apply sigmoid to the gate
shared_gate = ggml_sigmoid(ctx0, shared_gate); shared_gate = ggml_sigmoid(ctx0, shared_gate);
cb(shared_gate, "shared_expert_gate_sigmoid", il); cb(shared_gate, "shared_expert_gate_sigmoid", il);
// Apply the gate to the shared expert output
ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
cb(ffn_shexp, "ffn_shexp_gated", il); cb(ffn_shexp, "ffn_shexp_gated", il);