Skip to content

Commit

Permalink
llama : add function for model-based max number of graph nodes (ggerg…
Browse files Browse the repository at this point in the history 
…anov#8622)

* llama : model-based max number of graph nodes

ggml-ci

* llama : disable 405B max_nodes path due to lack of complaints

ggml-ci
  • Loading branch information
@ggerganov @arthw
ggerganov authored and arthw committed Jul 27, 2024
1 parent 1c3e99a commit 67501cf
Showing 1 changed file with 53 additions and 43 deletions.
96 changes: 53 additions & 43 deletions src/llama.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -101,7 +101,6 @@
#endif

// bump if necessary
#define LLAMA_MAX_NODES 8192
#define LLAMA_MAX_LAYERS 512
#define LLAMA_MAX_EXPERTS 160 // DeepSeekV2

Expand Down Expand Up @@ -3568,6 +3567,15 @@ namespace GGUFMeta {

using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;

// TODO: update when needed or think of some clever automatic way to do this
static size_t llama_model_max_nodes(const llama_model & /*model*/) {
//if (model.arch == LLM_ARCH_LLAMA && model.hparams.n_layer > ??) { // llama-3 405B
// return 32768;
//}

return 8192;
}

struct llama_model_loader {
int n_kv = 0;
int n_tensors = 0;
Expand Down Expand Up @@ -8397,7 +8405,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_k_shift() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

GGML_ASSERT(kv_self.size == n_ctx);

Expand Down Expand Up @@ -8428,7 +8436,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_s_copy() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

GGML_ASSERT(kv_self.recurrent);

Expand All @@ -8451,7 +8459,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_defrag(const std::vector<uint32_t> & ids) {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

for (uint32_t i = 0; i < ids.size(); ++i) {
const uint32_t id = ids[i];
Expand Down Expand Up @@ -8692,7 +8700,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_llama() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -8835,7 +8843,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_baichuan() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -8950,7 +8958,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_xverse() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -9053,7 +9061,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_falcon() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -9173,7 +9181,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_grok() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -9330,7 +9338,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_dbrx() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -9456,7 +9464,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_starcoder() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -9560,7 +9568,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_refact() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -9654,7 +9662,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_bert() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -9848,7 +9856,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_bloom() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -9949,7 +9957,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_mpt() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -10239,7 +10247,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_qwen() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -10351,7 +10359,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_qwen2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -10463,7 +10471,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_qwen2moe() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -10609,7 +10617,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_phi2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -10730,7 +10738,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_phi3() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -10962,7 +10970,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_gpt2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -11067,7 +11075,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_codeshell() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -11178,7 +11186,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_orion() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -11296,7 +11304,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_internlm2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -11417,7 +11425,7 @@ struct llm_build_context {
// https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738
// based on the original build_llama() function
struct ggml_cgraph * build_minicpm() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -11561,7 +11569,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_gemma() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head_k = hparams.n_embd_head_k;

Expand Down Expand Up @@ -11669,7 +11677,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_gemma2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head_k = hparams.n_embd_head_k;

Expand Down Expand Up @@ -11804,7 +11812,7 @@ struct llm_build_context {


struct ggml_cgraph * build_starcoder2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -11923,7 +11931,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_mamba() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t d_model = n_embd;
const int64_t d_conv = hparams.ssm_d_conv;
Expand Down Expand Up @@ -12072,7 +12080,7 @@ struct llm_build_context {

struct ggml_cgraph * build_command_r() {

struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -12226,7 +12234,7 @@ struct llm_build_context {
// * removed bias
// * removed MoE
struct ggml_cgraph * build_olmo() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -12350,7 +12358,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_openelm() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -12475,7 +12483,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_gptneox() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -12617,7 +12625,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_arctic() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -12749,7 +12757,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_deepseek2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -12977,7 +12985,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_bitnet() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down Expand Up @@ -13117,7 +13125,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_t5() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
Expand Down Expand Up @@ -13434,7 +13442,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_jais() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -13526,7 +13534,7 @@ struct llm_build_context {
}

struct ggml_cgraph * build_chatglm() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);

const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
Expand Down Expand Up @@ -14871,9 +14879,9 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
// each move requires 6*n_layer tensors (see build_defrag)
// - source view, destination view, copy operation
// - x2 for keys and values
//const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer);
//const uint32_t max_moves = llama_model_max_nodes(model)/(6*n_layer);
// TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516
const uint32_t max_moves = (LLAMA_MAX_NODES - 2*n_layer)/(6*n_layer);
const uint32_t max_moves = (llama_model_max_nodes(lctx.model) - 2*n_layer)/(6*n_layer);

// determine which KV cells to move where
//
Expand Down Expand Up @@ -16764,8 +16772,10 @@ struct llama_context * llama_new_context_with_model(
}
}

const size_t max_nodes = llama_model_max_nodes(*model);

// buffer used to store the computation graph and the tensor meta data
ctx->buf_compute_meta.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead_custom(LLAMA_MAX_NODES, false));
ctx->buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));

// enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
bool pipeline_parallel =
Expand All @@ -16778,7 +16788,7 @@ struct llama_context * llama_new_context_with_model(
// currently this is only implemented in the CUDA backend
pipeline_parallel = false;
#endif
ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), LLAMA_MAX_NODES, pipeline_parallel);
ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), max_nodes, pipeline_parallel);

if (pipeline_parallel) {
LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched));
Expand Down

0 comments on commit 67501cf

Please sign in to comment.