ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-01 01:24:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "Simplify delta-net (#1335)"

Browse Source 
This reverts commit e5fc30244c.
This commit is contained in:
Kawrakow
2026-02-28 12:09:02 +00:00
parent e5fc30244c
commit 18b1b23ed5
8 changed files with 362 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
common/common.cpp
View File

@@ -1533,7 +1533,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
}
if (arg == "-fdn" || arg == "--fused-delta-net") {
CHECK_ARG
fprintf(stderr, "=================== %s has been deprecated\n", arg.c_str());
params.fused_delta_net = std::stoi(argv[i]);
return true;
}
if (arg == "-smf16" || arg == "--split-mode-f16") {
@@ -2276,6 +2276,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
options.push_back({ "*", "-grt, --graph-reduce-type", "Type for data exchange between GPUs (default: %s)", "f32"});
options.push_back({ "*", "-smgs, --split-mode-graph-scheduling,", "Force Split Mode Graph Scheduling (default: %d)", params.split_mode_graph_scheduling});
options.push_back({ "*", "-sas, --scheduler_async,", "Async evaluation of compute graphs: %d)", params.scheduler_async});
options.push_back({ "*", "-fdn, --fused-delta-net N", "Use fused delta-net when batch size is <= N with recurrent models: %d)", params.fused_delta_net});
options.push_back({ "*", "-vq, --validate-quants", "validate quantized data while loading the model (default: %d)", params.validate_quants});
options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with\n"
"in conversation mode, this will be used as system prompt\n"
@@ -3354,6 +3355,7 @@ struct llama_context_params common_context_params_to_llama(const gpt_params & pa
cparams.split_mode_graph_scheduling = params.split_mode_graph_scheduling;
//cparams.split_mode_f16 = params.split_mode_f16;
cparams.scheduler_async = params.scheduler_async;
cparams.fused_delta_net = params.fused_delta_net;
cparams.min_experts = params.min_experts;
cparams.thresh_experts = params.thresh_experts;
cparams.only_active_experts = params.only_active_exps;
@@ -4364,6 +4366,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
//fprintf(stream, "split_mode_f16: %s # default: true\n", params.split_mode_f16 ? "true" : "false");
fprintf(stream, "reduce_type: %s # default f16\n", params.reduce_type.c_str());
fprintf(stream, "scheduler_async: %s # default: false\n", params.scheduler_async ? "true" : "false");
fprintf(stream, "fused_delta_net: %d # default: 0\n", params.fused_delta_net );
fprintf(stream, "ser: %d,%g # defaulr: -1,0\n", params.min_experts, params.thresh_experts);
fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);

1
common/common.h
View File

@@ -359,6 +359,7 @@ struct gpt_params {
bool split_mode_graph_scheduling = false; // if true, force split mode graph scheduling
//bool split_mode_f16 = true; // if true, intermediate results will be cast to f16 before copying to other GPUs to perform reduce ops
bool scheduler_async = false; // if true, in split mode graph the scheduler will use multiple threads to evaluate the graph
int fused_delta_net = 65536; // use fused delta-net if number of tokens in the batch is less than this value
bool has_mtp = false; // enable MTP if supported by the model
std::string cache_type_k = "f16"; // KV cache data type for the K

33
examples/llama-bench/llama-bench.cpp
View File

@@ -271,6 +271,7 @@ struct cmd_params {
bool muge = false;
bool rcache = false;
bool sas = false;
int fdn = 65536; // fdn = fused delta net
bool print_overrides = false;
output_formats output_format;
output_formats output_format_stderr;
@@ -316,6 +317,7 @@ static const cmd_params cmd_params_defaults = {
/* muge */ false,
/* rcache */ false,
/* sas */ false,
/* fdn */ 65536,
/* print_overrides */ false,
/* output_format */ MARKDOWN,
/* output_format_stderr */ NONE,
@@ -369,6 +371,7 @@ static void print_usage(int /* argc */, char ** argv) {
printf(" -no-fug, --no-fused-up-gate <0|1> (default: %s)\n", cmd_params_defaults.no_fug? "1" : "0");
printf(" -no-ooae, --no-offload-only-active-experts <0|1> (default: %s)\n", cmd_params_defaults.no_ooae? "1" : "0");
printf(" -sas, --scheduler-async <0|1> (default: %s)\n", cmd_params_defaults.sas ? "1" : "0");
printf(" -fdn, --fused-delta-net <n> (default: %d)\n", cmd_params_defaults.fdn);
printf(" --print-overrides <0|1> (default: %s)\n", cmd_params_defaults.print_overrides ? "1" : "0");
printf("\n");
printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
@@ -810,6 +813,12 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
break;
}
params.sas = std::stoi(argv[i]);
} else if (arg == "-fdn" || arg == "--fused-delta-net") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.fdn = std::stoi(argv[i]);
} else if (arg == "-rcache" || arg == "--rope-cache") {
if (++i >= argc) {
invalid_param = true;
@@ -956,6 +965,7 @@ struct cmd_params_instance {
bool muge = false;
bool rcache = false;
bool sas = false;
int fdn = 0;
const llama_model_tensor_buft_override* buft_overrides;
llama_model_params to_llama_mparams() const {
@@ -991,6 +1001,7 @@ struct cmd_params_instance {
muge == other.muge &&
use_thp == other.use_thp &&
sas == other.sas &&
fdn == other.fdn &&
tensor_split == other.tensor_split;
}
@@ -1017,6 +1028,7 @@ struct cmd_params_instance {
cparams.embeddings = embeddings;
cparams.cuda_params = (void *)cuda_params.data();
cparams.scheduler_async = sas;
cparams.fused_delta_net = fdn;
return cparams;
}
@@ -1083,6 +1095,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .muge = */ params.muge,
/* .rcache = */ params.rcache,
/* .sas = */ params.sas,
/* .fdn = */ params.fdn,
/* .buft_overrides=*/ params.buft_overrides.data(),
};
instances.push_back(instance);
@@ -1126,6 +1139,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .muge = */ params.muge,
/* .rcache = */ params.rcache,
/* .sas = */ params.sas,
/* .fdn = */ params.fdn,
/* .buft_overrides=*/ params.buft_overrides.data(),
};
instances.push_back(instance);
@@ -1169,6 +1183,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .muge = */ params.muge,
/* .rcache = */ params.rcache,
/* .sas = */ params.sas,
/* .fdn = */ params.fdn,
/* .buft_overrides=*/ params.buft_overrides.data(),
};
instances.push_back(instance);
@@ -1212,6 +1227,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .muge = */ params.muge,
/* .rcache = */ params.rcache,
/* .sas = */ params.sas,
/* .fdn = */ params.fdn,
/* .buft_overrides=*/ params.buft_overrides.data(),
};
instances.push_back(instance);
@@ -1266,6 +1282,7 @@ struct test {
bool muge = false;
bool rcache = false;
bool sas = false;
int fdn = 0;
std::string override_tensor;
int n_prompt;
int n_gen;
@@ -1307,6 +1324,7 @@ struct test {
ger = inst.ger;
rcache = inst.rcache;
sas = inst.sas;
fdn = inst.fdn;
no_fug = inst.no_fug;
use_thp = inst.use_thp;
no_ooae = inst.no_ooae;
@@ -1411,7 +1429,7 @@ struct test {
field == "model_size" || field == "model_n_params" ||
field == "n_gpu_layers" || field == "main_gpu" ||
field == "n_prompt" || field == "n_gen" || field == "mla_attn" || field == "attn_max_batch" ||
field == "avg_ns" || field == "stddev_ns") {
field == "avg_ns" || field == "stddev_ns" || field == "fdn") {
return INT;
}
if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
@@ -1462,7 +1480,7 @@ struct test {
std::to_string(mla_attn), std::to_string(attn_max_batch), ser_to_string(ser), std::to_string(reuse),
tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
std::to_string(repack), std::to_string(mqkv), std::to_string(muge), std::to_string(fmoe), std::to_string(ger),
std::to_string(no_fug), std::to_string(use_thp), std::to_string(no_ooae), std::to_string(rcache), std::to_string(sas),
std::to_string(no_fug), std::to_string(use_thp), std::to_string(no_ooae), std::to_string(rcache), std::to_string(sas), std::to_string(fdn),
cuda_params, override_tensor,
std::to_string(n_prompt), std::to_string(n_gen), test_time,
std::to_string(avg_ns()), std::to_string(stdev_ns()),
@@ -1483,7 +1501,7 @@ struct test {
"n_gpu_layers", "split_mode",
"main_gpu", "no_kv_offload", "flash_attn", "mla_attn", "attn_max_batch", "ser", "reuse",
"tensor_split", "use_mmap", "embeddings", "repack", "mqkv", "muge", "fused_moe", "grouped_er",
"no_fused_up_gate", "use_thp", "no_ooae", "rcache", "sas", "cuda_params", "override_tensor",
"no_fused_up_gate", "use_thp", "no_ooae", "rcache", "sas", "fdn", "cuda_params", "override_tensor",
"n_prompt", "n_gen", "test_time",
"avg_ns", "stddev_ns",
"avg_ts", "stddev_ts", "test",
@@ -1673,6 +1691,9 @@ struct markdown_printer : public printer {
if (field == "sas") {
return 3;
}
if (field == "fdn") {
return 4;
}
if (field == "use_thp") {
return 3;
}
@@ -1746,6 +1767,9 @@ struct markdown_printer : public printer {
if (field == "sas") {
return "sas";
}
if (field == "fdn") {
return "fdn";
}
if (field == "use_thp") {
return "thp";
}
@@ -1856,6 +1880,9 @@ struct markdown_printer : public printer {
if (params.sas != cmd_params_defaults.sas) {
fields.emplace_back("sas");
}
if (params.fdn != cmd_params_defaults.fdn) {
fields.emplace_back("fdn");
}
if (params.muge != cmd_params_defaults.muge) {
fields.emplace_back("muge");
}

1
include/llama.h
View File

@@ -456,6 +456,7 @@ extern "C" {
bool split_mode_graph_scheduling; // if true, force split mode graph scheduling
//bool split_mode_f16; // if true, cast intermediate results to f16 before copying to other GPUs
bool scheduler_async; // if true, with split mode "graph" graph evaluation will be done using multiple threads
int fused_delta_net;
bool mtp; // Activate MTP if supported
enum llama_mtp_op_type mtp_op_type;

1
src/llama-cparams.h
View File

@@ -43,6 +43,7 @@ struct llama_cparams {
bool split_mode_graph_scheduling;
//bool split_mode_f16;
bool scheduler_async;
int fused_delta_net;
int min_experts;
float thresh_experts;
bool mtp;

314
src/llama-delta-net.cpp
View File

@@ -74,6 +74,304 @@ delta_net::delta_net(llama_context & _lctx, const llama_batch & _batch) : lctx(_
delta_net::~delta_net() = default;
std::pair<ggml_tensor *, ggml_tensor *> delta_net::build_delta_net_chunking(ggml_context * ctx0,
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 llm_build_cb & cb) {
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_seqs == 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);
if (beta->ne[0] != H_v || beta->ne[2] != n_tokens || beta->ne[3] != n_seqs) {
printf("beta: %ld x %ld x %ld, expected %ld x %ld x %ld\n", beta->ne[0], beta->ne[2], beta->ne[3], H_v, n_tokens, 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(H_k == H_v);
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);
cb(state,"state_in", il);
const int64_t chunk_size = DELTA_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_permute(ctx0, q, 0, 2, 1, 3);
k = ggml_permute(ctx0, k, 0, 2, 1, 3);
v = ggml_permute(ctx0, v, 0, 2, 1, 3);
g = ggml_permute(ctx0, g, 2, 0, 3, 1);
beta = ggml_permute(ctx0, beta, 2, 0, 1, 3);
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);
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, pad, 0, 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_v * 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_v * n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
cb(g_cumsum, "g_cumsum", il);
ggml_tensor * gcs_i =
ggml_repeat_4d(ctx0, g_cumsum, chunk_size, chunk_size, 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);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
cb(decay_mask, "decay_mask_1", il);
decay_mask = ggml_exp(ctx0, decay_mask);
cb(decay_mask, "decay_mask_exp", il);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
cb(decay_mask, "decay_mask_2", il);
ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
cb(kmulkbeta, "kk_beta", il);
ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
cb(k_decay, "k_decay_1", il);
k_decay = ggml_mul(ctx0, k_decay, causal_mask);
cb(k_decay, "k_decay_2", il);
ggml_tensor * attn = ggml_neg(ctx0, k_decay);
cb(attn, "attn_pre_solve", il);
ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
cb(attn_lower, "attn_lower", il);
ggml_tensor * identity_repeat =
ggml_repeat_4d(ctx0, identity, attn_lower->ne[0], attn_lower->ne[1], attn_lower->ne[2], attn_lower->ne[3]);
ggml_tensor * lhs = ggml_neg(ctx0, ggml_sub(ctx0, attn_lower, identity_repeat));
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_mul(ctx0, lin_solve, causal_mask);
cb(attn, "attn_mul", il);
attn = ggml_add(ctx0, attn, identity);
cb(attn, "attn_solved", il);
auto v_beta_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_beta));
cb(v_beta_t, "v_beta_t", il);
v = ggml_mul_mat(ctx0, v_beta_t, attn);
cb(v, "v_beta", il);
ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
cb(g_cumsum_t, "g_cumsum_t", il);
ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
cb(gexp, "gexp", il);
ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
cb(kbeta_gexp, "kbeta_gexp", il);
auto kbeta_gexp_t = ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp));
cb(kbeta_gexp_t, "kbeta_gexp_t", il);
auto attn_kbeta = ggml_mul_mat(ctx0, attn, kbeta_gexp_t);
cb(attn_kbeta, "attn_kbeta", il);
ggml_tensor * k_cumdecay = ggml_cont(ctx0, ggml_transpose(ctx0, attn_kbeta));
cb(k_cumdecay, "k_cumdecay", il);
ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
cb(attn_kq, "attn_kq_pre", il);
attn_kq = ggml_mul(ctx0, decay_mask, attn_kq);
cb(attn_kq, "attn_kq_0", il);
attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
cb(attn_kq, "attn_kq", il);
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);
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
cb(g_last_exp, "g_last_exp", il);
ggml_tensor * g_last_repeat =
ggml_repeat_4d(ctx0, g_last, chunk_size, 1, n_chunks, H_v * n_seqs);
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last_repeat));
cb(g_diff, "g_diff", il);
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
cb(g_diff_exp, "g_diff_exp", il);
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);
ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
cb(key_gdiff_t, "key_gdiff_t", il);
cb(state, "new_state", il);
auto get_slice_2d = [ctx0](ggml_tensor * t, int64_t c) -> ggml_tensor * {
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);
};
ggml_tensor * core_attn_out = nullptr;
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
ggml_tensor * q_chunk = get_slice_2d(q, chunk);
ggml_tensor * v_chunk = get_slice_2d(v, chunk);
ggml_tensor * gexp_chunk = get_slice_2d(gexp, chunk);
ggml_tensor * k_cumdecay_chunk = get_slice_2d(k_cumdecay, chunk);
ggml_tensor * attn_chunk = get_slice_2d(attn_kq, chunk);
cb(attn_chunk, "attn_chunk", il);
ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
cb(state_t, "state_t", il);
ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
cb(v_prime, "v_prime_chunk", il);
ggml_tensor * v_new = ggml_sub(ctx0, v_prime, v_chunk);
ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
cb(v_new, "v_new_chunk", il);
ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
cb(q_g_exp, "q_g_exp", il);
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
cb(attn_inter, "attn_inter_chunk", il);
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_sub(ctx0, attn_inter, v_attn);
cb(core_attn_out_chunk, "core_attn_out_chunk", il);
core_attn_out = core_attn_out == nullptr
? core_attn_out_chunk
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
ggml_tensor * k_gdiff_t = get_slice_2d(key_gdiff_t, chunk);
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
cb(kgdmulvnew, "kgdmulvnew", il);
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(g_last_exp, chunk));
cb(gexp_last_chunk, "gexp_last_chunk", il);
auto s_mul = ggml_mul(ctx0, state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs));
cb(s_mul, "s_mul", il);
state = ggml_sub(ctx0, s_mul, ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
}
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 * DELTA_CHUNK_SIZE * n_chunks),
ggml_row_size(core_attn_out->type, S_v * DELTA_CHUNK_SIZE * n_chunks * H_v), 0);
cb(output_tokens, "output_tokens", il);
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
output_tokens = ggml_cont(ctx0, output_tokens);
cb(output_tokens, "output_tokens", il);
return {output_tokens, state};
}
std::pair<ggml_tensor *, ggml_tensor *> delta_net::build_delta_net_autoregressive(ggml_context * ctx0,
ggml_tensor * q, ggml_tensor * k, ggml_tensor * v,
ggml_tensor * g, ggml_tensor * beta, ggml_tensor * state,
int il, const llm_build_cb & cb) {
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(n_seqs == 1);
GGML_ASSERT(H_k == H_v);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v && state->ne[2] == H_v && state->ne[3] == n_seqs);
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);
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);
g_t = ggml_exp(ctx0, g_t);
cb(g_t, "g_t", il);
state = ggml_mul(ctx0, state, g_t);
cb(state, "state", il);
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);
cb(kv_mem, "kv_mem", il);
kv_mem = ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem));
cb(kv_mem, "kv_mem_t_cont", il);
kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, kv_mem));
ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem);
cb(v_diff, "v_diff", il);
ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
cb(delta, "delta", il);
ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
cb(k_t_delta, "k_t_delta", il);
state = ggml_add(ctx0, state, k_t_delta);
ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs);
ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
cb(state_q, "state_q", il);
state_q = ggml_cont(ctx0, ggml_transpose(ctx0, state_q));
cb(state_q, "state_q_t_cont", il);
ggml_tensor * core_attn_out = ggml_transpose(ctx0, ggml_sum_rows(ctx0, state_q));
cb(core_attn_out, "output_tokens", il);
cb(state, "new_state", il);
return {core_attn_out, state};
}
std::pair<ggml_tensor *, ggml_tensor *> delta_net::build_fused_delta_net(ggml_context * ctx0,
ggml_tensor * q, ggml_tensor * k, ggml_tensor * v,
ggml_tensor * g, ggml_tensor * beta, ggml_tensor * state,
@@ -246,7 +544,9 @@ ggml_tensor * delta_net::build_layer_attn_linear_core(ggml_context * ctx0, ggml_
const int64_t n_seqs = 1;
const int64_t n_seq_tokens = n_tok;
auto [qkv_mixed, z] = build_qkvz(ctx0, cur, il, cb);
auto qkvz = build_qkvz(ctx0, cur, il, cb);
ggml_tensor * qkv_mixed = qkvz.first;
ggml_tensor * z = qkvz.second;
ggml_tensor *alpha, *beta;
if (model.layers[il].ssm_beta_alpha) {
@@ -379,8 +679,14 @@ ggml_tensor * delta_net::build_layer_attn_linear_core(ggml_context * ctx0, ggml_
GGML_ASSERT(identity != nullptr);
GGML_ASSERT(diag_mask != nullptr);
auto [output, new_state] = build_fused_delta_net(ctx0, q_conv, k_conv, v_conv, gate, beta, state, il, cb);
std::pair<ggml_tensor *, ggml_tensor *> attn_out;
// The fused delta-net implementation is only faster than chunked for n_tok <= 8, so use it only in that case
attn_out = n_tok <= lctx.cparams.fused_delta_net ? build_fused_delta_net(ctx0, q_conv, k_conv, v_conv, gate, beta, state, il, cb) :
n_tok == 1 ? build_delta_net_autoregressive(ctx0, q_conv, k_conv, v_conv, gate, beta, state, il, cb)
: build_delta_net_chunking(ctx0, q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il, cb);
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
cb(new_state, "new_state", il);
@@ -393,7 +699,8 @@ ggml_tensor * delta_net::build_layer_attn_linear_core(ggml_context * ctx0, ggml_
ggml_tensor * new_ssm_flat = ggml_reshape_2d(ctx0, new_state, ssm_state_dim, 1);
ggml_tensor * new_state_flat = ggml_concat(ctx0, new_conv_flat, new_ssm_flat, 0);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_state_flat, state_dst));
ggml_tensor * state_update = new_state_flat;
ggml_build_forward_expand(gf, ggml_cpy(ctx0, state_update, state_dst));
ggml_tensor * attn_out_2d = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_tok);
ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_tok);
@@ -410,6 +717,7 @@ ggml_tensor * delta_net::build_layer_attn_linear_core(ggml_context * ctx0, ggml_
cb(out, "linear_attn_out", il);
return out;
//return ggml_reshape_2d(ctx0, out, hparams.n_embd, n_tok);
}

11
src/llama-delta-net.h
View File

@@ -8,6 +8,17 @@ struct delta_net {
delta_net(llama_context & lctx, const llama_batch & batch);
~delta_net();
static std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(ggml_context * ctx0,
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 llm_build_cb & cb);
static std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(ggml_context * ctx0,
ggml_tensor * q, ggml_tensor * k, ggml_tensor * v,
ggml_tensor * g, ggml_tensor * beta, ggml_tensor * state,
int il, const llm_build_cb & cb);
static std::pair<ggml_tensor *, ggml_tensor *> build_fused_delta_net(ggml_context * ctx0,
ggml_tensor * q, ggml_tensor * k, ggml_tensor * v,
ggml_tensor * g, ggml_tensor * beta, ggml_tensor * state,

4
src/llama.cpp
View File

@@ -1512,7 +1512,6 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner);
LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state);
LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank);
LLAMA_LOG_INFO("%s: ssm_n_group = %u\n", __func__, hparams.ssm_n_group);
}
LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type));
@@ -4395,6 +4394,7 @@ struct llama_context_params llama_context_default_params() {
/*.split_mode_graph_scheduling =*/ false,
// /*.split_mode_f16 =*/ true,
/*.scheduler_async =*/ false,
/*.fused_delta_net =*/ 65536,
/*.mtp =*/ false,
/*.mtp_op_type =*/ MTP_OP_NONE,
/*.abort_callback =*/ nullptr,
@@ -4766,6 +4766,7 @@ struct llama_context * llama_init_from_model(
cparams.split_mode_graph_scheduling = params.split_mode_graph_scheduling;
//cparams.split_mode_f16 = params.split_mode_f16;
cparams.scheduler_async = params.scheduler_async;
cparams.fused_delta_net = params.fused_delta_net;
cparams.min_experts = params.min_experts;
cparams.thresh_experts = params.thresh_experts;
cparams.cuda_params = params.cuda_params;
@@ -4872,6 +4873,7 @@ struct llama_context * llama_init_from_model(
//LLAMA_LOG_INFO("%s: split_mode_f16= %d\n", __func__, cparams.split_mode_f16);
LLAMA_LOG_INFO("%s: reduce_type = %s\n", __func__, ggml_type_name(cparams.reduce_type));
LLAMA_LOG_INFO("%s: sched_async = %d\n", __func__, cparams.scheduler_async);
LLAMA_LOG_INFO("%s: fused_delta = %d\n", __func__, cparams.fused_delta_net);
LLAMA_LOG_INFO("%s: ser = %d, %g\n", __func__, cparams.min_experts, cparams.thresh_experts);
LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base);
LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale);