vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge pull request #1 from ggerganov/refactor-mpi

Browse source 
mpi : trying to move more MPI stuff into ggml-mpi (#2099)
This commit is contained in:
Evan Miller 2023-07-09 15:23:04 -04:00 • committed by GitHub
commit 1c3a15c5d4
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
11 changed files with 300 additions and 179 deletions
Download patch file Download diff file Expand all files Collapse all files

2
examples/embd-input/embd-input-lib.cpp
View file

@ -34,7 +34,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
}
fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
llama_init_backend(params.numa);
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;

4
examples/embedding/embedding.cpp
View file

@ -35,7 +35,7 @@ int main(int argc, char ** argv) {
params.prompt = gpt_random_prompt(rng);
}
llama_init_backend(params.numa);
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
@ -93,5 +93,7 @@ int main(int argc, char ** argv) {
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;
}

4
examples/main/main.cpp
View file

@ -105,7 +105,7 @@ int main(int argc, char ** argv) {
params.prompt = gpt_random_prompt(rng);
}
llama_init_backend(params.numa);
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
@ -671,7 +671,7 @@ int main(int argc, char ** argv) {
llama_free(ctx);
llama_free_model(model);
llama_finalize_backend();
llama_backend_free();
return 0;
}

4
examples/perplexity/perplexity.cpp
View file

@ -147,7 +147,7 @@ int main(int argc, char ** argv) {
params.prompt = gpt_random_prompt(rng);
}
llama_init_backend(params.numa);
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
@ -172,7 +172,7 @@ int main(int argc, char ** argv) {
llama_free(ctx);
llama_free_model(model);
llama_finalize_backend();
llama_backend_free();
return 0;
}

4
examples/quantize/quantize.cpp
View file

@ -180,7 +180,7 @@ int main(int argc, char ** argv) {
usage(argv[0]);
}
llama_init_backend(false);
llama_backend_init(false);
// parse command line arguments
const std::string fname_inp = argv[arg_idx];
@ -257,5 +257,7 @@ int main(int argc, char ** argv) {
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
}
llama_backend_free();
return 0;
}

4
examples/server/server.cpp
View file

@ -1079,7 +1079,7 @@ int main(int argc, char **argv)
params.model_alias = params.model;
}
llama_init_backend(params.numa);
llama_backend_init(params.numa);
LOG_INFO("build info", {{"build", BUILD_NUMBER},
{"commit", BUILD_COMMIT}});
@ -1309,5 +1309,7 @@ int main(int argc, char **argv)
return 1;
}
llama_backend_free();
return 0;
}

4
examples/simple/simple.cpp
View file

@ -66,7 +66,7 @@ int main(int argc, char ** argv)
// Init LLM :
//---------------------------------
llama_init_backend(params.numa);
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
@ -173,7 +173,7 @@ int main(int argc, char ** argv)
llama_free( ctx );
llama_free_model( model );
llama_finalize_backend();
llama_backend_free();
return 0;
}

241
ggml-mpi.c
View file

@ -2,80 +2,211 @@
#include "ggml.h"
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define UNUSED GGML_UNUSED
struct ggml_mpi_tensor_info {
struct ggml_mpi_context {
int rank;
int size;
};
// ggml_compute_forward_send
void ggml_mpi_backend_init(void) {
MPI_Init(NULL, NULL);
}
static void ggml_mpi_compute_forward_send(
struct ggml_tensor * src,
const struct ggml_tensor * orig) {
UNUSED(orig);
GGML_ASSERT(src->type == GGML_TYPE_F32);
void ggml_mpi_backend_free(void) {
MPI_Finalize();
}
int my_rank;
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
struct ggml_mpi_context * ggml_mpi_init(void) {
struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
int dst_rank = ((struct ggml_mpi_tensor_info *)src->extra)->rank;
// fprintf(stderr, "(%d) Sending to (%d)\n", my_rank, (int)dst->extra);
int retval = MPI_Send(src->data, ggml_nelements(src), MPI_FLOAT, dst_rank, 0, MPI_COMM_WORLD);
// fprintf(stderr, "(%d) Sent to (%d)\n", my_rank, (int)dst->extra);
MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
return ctx;
}
void ggml_mpi_free(struct ggml_mpi_context * ctx) {
free(ctx);
}
int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
return ctx->rank;
}
void ggml_mpi_eval_init(
struct ggml_mpi_context * ctx_mpi,
int * n_tokens,
int * n_past,
int * n_threads) {
UNUSED(ctx_mpi);
// synchronize the worker node parameters with the root node
MPI_Barrier(MPI_COMM_WORLD);
MPI_Bcast(n_tokens, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(n_past, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
}
int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
if (t == NULL) {
fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
return -1;
}
for (int i = 0; i < gf->n_nodes; i++) {
if (gf->nodes[i] == t) {
return i;
}
}
fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
return -1;
}
// TODO: there are many improvements that can be done to this implementation
void ggml_mpi_graph_compute(
struct ggml_mpi_context * ctx_mpi,
struct ggml_context * ctx,
struct ggml_cgraph * gf,
int n_layers) {
const int mpi_rank = ctx_mpi->rank;
const int mpi_size = ctx_mpi->size;
struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
if (inp_tokens == NULL) {
fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
return;
}
struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
if (inp0 == NULL) {
fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
return;
}
GGML_ASSERT(inp0 == gf->nodes[0]);
// distribute the compute graph into slices across the MPI nodes
//
// the main node (0) processes the last layers + the remainder of the compute graph
// and is responsible to pass the input tokens to the first node (1)
//
// node 1: [( 0) * n_per_node, ( 1) * n_per_node)
// node 2: [( 1) * n_per_node, ( 2) * n_per_node)
// ...
// node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
// node 0: [(n-1) * n_per_node, n_nodes)
//
if (mpi_rank > 0) {
if (mpi_rank == 1) { // the first node receives the input tokens from the main node
MPI_Status status; UNUSED(status);
const int mpi_rank_src = mpi_rank - 1;
const int retval = MPI_Recv(inp_tokens->data, ggml_nelements(inp_tokens), MPI_INT, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
GGML_ASSERT(retval == MPI_SUCCESS);
} else { // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
MPI_Status status; UNUSED(status);
const int mpi_rank_src = mpi_rank - 1;
//printf("%s: node %d: waiting for %d elements from %d\n", __func__, mpi_rank, (int) ggml_nelements(inp0), mpi_rank_src);
const int retval = MPI_Recv(inp0->data, ggml_nelements(inp0), MPI_FLOAT, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
GGML_ASSERT(retval == MPI_SUCCESS);
}
} else if (mpi_size > 1) {
// node 0 sends the input tokens to node 1
{
const int mpi_rank_dst = mpi_rank + 1;
const int retval = MPI_Send(inp_tokens->data, ggml_nelements(inp_tokens), MPI_INT, mpi_rank_dst, 0, MPI_COMM_WORLD);
GGML_ASSERT(retval == MPI_SUCCESS);
}
// ggml_compute_forward_recv
// recv the output data from the last node
{
MPI_Status status; UNUSED(status);
static void ggml_mpi_compute_forward_recv(
struct ggml_tensor * dst,
const struct ggml_tensor * orig,
const struct ggml_tensor * parent) {
UNUSED(parent);
UNUSED(orig);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
MPI_Status status;
const int mpi_rank_src = mpi_size - 1;
int my_rank;
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
int src_rank = ((struct ggml_mpi_tensor_info *)dst->extra)->rank;
// fprintf(stderr, "(%d) Receiving from (%d)\n", my_rank, src_extra);
int retval = MPI_Recv(dst->data, ggml_nelements(dst), MPI_FLOAT, src_rank, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
// fprintf(stderr, "(%d) Received from (%d)\n", my_rank, src_extra);
//fprintf(stderr, "%s: node %d: waiting for %d elements from %d\n", __func__, mpi_rank, (int) ggml_nelements(inp0), mpi_rank_src);
const int retval = MPI_Recv(inp0->data, ggml_nelements(inp0), MPI_FLOAT, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
GGML_ASSERT(retval == MPI_SUCCESS);
}
struct ggml_tensor * ggml_mpi_send_tensor(
struct ggml_context * ctx,
struct ggml_tensor *src,
int dst_rank) {
struct ggml_tensor * result = ggml_map_custom1_inplace_f32(ctx, src, ggml_mpi_compute_forward_send);
// TODO how/when to free this struct?
struct ggml_mpi_tensor_info *info = calloc(1, sizeof(struct ggml_mpi_tensor_info));
info->rank = dst_rank;
result->extra = info;
return result;
}
struct ggml_tensor * ggml_mpi_recv_tensor(
struct ggml_context * ctx,
struct ggml_tensor *parent,
struct ggml_tensor *dst,
int src_rank) {
struct ggml_tensor * result = ggml_map_custom2_inplace_f32(ctx, dst, parent, ggml_mpi_compute_forward_recv);
{
const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
// TODO how/when to free this struct?
struct ggml_mpi_tensor_info *info = calloc(1, sizeof(struct ggml_mpi_tensor_info));
info->rank = src_rank;
result->extra = info;
const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
return result;
const int il0 = (mpi_idx + 0) * n_per_node;
const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
char name_l0[GGML_MAX_NAME];
char name_l1[GGML_MAX_NAME];
snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
const int idx_l0 = ggml_graph_get_node_idx(gf, name_l0);
const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
if (idx_l0 < 0 || idx_l1 < 0) {
fprintf(stderr, "%s: layer input nodes not found\n", __func__);
return;
}
// attach the input data to all nodes that need it
// TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
for (int i = idx_l0; i < idx_l1; i++) {
if (gf->nodes[i]->src0 == gf->nodes[idx_l0]) {
gf->nodes[i]->src0 = inp0;
}
if (gf->nodes[i]->src1 == gf->nodes[idx_l0]) {
gf->nodes[i]->src1 = inp0;
}
}
// TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
for (int i = 1; i < idx_l1 - idx_l0; i++) {
gf->nodes[i] = gf->nodes[idx_l0 + i];
gf->grads[i] = gf->grads[idx_l0 + i];
}
// the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
if (mpi_idx != 0) {
gf->nodes[0]->op = GGML_OP_NONE;
}
gf->n_nodes = idx_l1 - idx_l0;
//fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
}
ggml_graph_compute(ctx, gf);
//fprintf(stderr, "%s: node %d: done\n", __func__, mpi_rank);
// send the output data to the next node
if (mpi_rank > 0) {
struct ggml_tensor * output = gf->nodes[gf->n_nodes - 1];
const int mpi_rank_dst = (mpi_rank + 1) % mpi_size;
//fprintf(stderr, "%s: node %d: sending %d elements to node %d\n", __func__, mpi_rank, ggml_nelements(output), mpi_rank_dst);
const int retval = MPI_Send(output->data, ggml_nelements(output), MPI_FLOAT, mpi_rank_dst, 0, MPI_COMM_WORLD);
GGML_ASSERT(retval == MPI_SUCCESS);
}
}

29
ggml-mpi.h
View file

@ -2,20 +2,33 @@
struct ggml_context;
struct ggml_tensor;
struct ggml_cgraph;
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_tensor * ggml_mpi_send_tensor(
struct ggml_mpi_context;
void ggml_mpi_backend_init(void);
void ggml_mpi_backend_free(void);
struct ggml_mpi_context * ggml_mpi_init(void);
void ggml_mpi_free(struct ggml_mpi_context * ctx);
int ggml_mpi_rank(struct ggml_mpi_context * ctx);
void ggml_mpi_eval_init(
struct ggml_mpi_context * ctx_mpi,
int * n_tokens,
int * n_past,
int * n_threads);
void ggml_mpi_graph_compute(
struct ggml_mpi_context * ctx_mpi,
struct ggml_context * ctx,
struct ggml_tensor *src,
int dst_rank);
struct ggml_tensor * ggml_mpi_recv_tensor(
struct ggml_context * ctx,
struct ggml_tensor *parent,
struct ggml_tensor *dst,
int src_rank);
struct ggml_cgraph * gf,
int n_layers);
#ifdef __cplusplus
}

119
llama.cpp
View file

@ -52,10 +52,6 @@
#include <sstream>
#include <numeric>
#ifdef GGML_USE_MPI
#include <mpi.h>
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
@ -359,8 +355,9 @@ struct llama_context {
ggml_metal_context * ctx_metal = NULL;
#endif
int mpi_rank;
int mpi_size;
#ifdef GGML_USE_MPI
ggml_mpi_context * ctx_mpi = NULL;
#endif
int buf_last = 0;
size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 };
@ -880,7 +877,7 @@ bool llama_mlock_supported() {
return llama_mlock::SUPPORTED;
}
void llama_init_backend(bool numa) {
void llama_backend_init(bool numa) {
ggml_time_init();
// needed to initialize f16 tables
@ -893,14 +890,15 @@ void llama_init_backend(bool numa) {
if (numa) {
ggml_numa_init();
}
#ifdef GGML_USE_MPI
MPI_Init(NULL, NULL);
ggml_mpi_backend_init();
#endif
}
void llama_finalize_backend() {
void llama_backend_free() {
#ifdef GGML_USE_MPI
MPI_Finalize();
ggml_mpi_backend_free();
#endif
}
@ -1303,13 +1301,17 @@ static bool llama_eval_internal(
llama_context & lctx,
const llama_token * tokens,
const float * embd,
const int n_tokens,
const int n_past,
int n_tokens,
int n_past,
int n_threads,
const char * cgraph_fname) {
LLAMA_ASSERT((!tokens && embd) || (tokens && !embd));
#ifdef GGML_USE_MPI
ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
#endif
const int64_t t_start_us = ggml_time_us();
const int N = n_tokens;
@ -1349,21 +1351,17 @@ static bool llama_eval_internal(
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
if (lctx.mpi_rank > 0) {
#ifdef GGML_USE_MPI
inpL = ggml_mpi_recv_tensor(ctx0, NULL,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N),
lctx.mpi_rank-1);
ggml_set_name(inpL, "mpi_recv");
#else
GGML_ASSERT(false);
#endif
} else if (tokens) {
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
ggml_set_name(embd, "embd");
memcpy(embd->data, tokens, N*ggml_element_size(embd));
inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
if (tokens) {
struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens));
ggml_set_name(inp_tokens, "inp_tokens");
inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
} else {
#ifdef GGML_USE_MPI
GGML_ASSERT(false && "not implemented");
#endif
inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N);
memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL));
}
@ -1392,9 +1390,9 @@ static bool llama_eval_internal(
}
#endif // GGML_USE_CUBLAS
// EMM TODO distribute work more evenly - maybe rank=0 gets the smallest amount?
int slice_size = (n_layer + (lctx.mpi_size - 1)) / lctx.mpi_size;
for (int il = lctx.mpi_rank * slice_size; il < n_layer && il < (lctx.mpi_rank + 1) * slice_size; ++il) {
for (int il = 0; il < n_layer; ++il) {
ggml_format_name(inpL, "layer_inp_%d", il);
offload_func_t offload_func = llama_nop;
#ifdef GGML_USE_CUBLAS
@ -1601,7 +1599,6 @@ static bool llama_eval_internal(
// input for next layer
inpL = cur;
}
lctx.use_buf(ctx0, 0);
@ -1609,28 +1606,9 @@ static bool llama_eval_internal(
// used at the end to optionally extract the embeddings
struct ggml_tensor * embeddings = NULL;
if (lctx.mpi_size > 1) {
#ifdef GGML_USE_MPI
cur = ggml_mpi_send_tensor(ctx0, cur, (lctx.mpi_rank+1)%lctx.mpi_size);
ggml_set_name(cur, "mpi_send");
#else
GGML_ASSERT(false);
#endif
}
if (lctx.mpi_rank == 0) {
if (lctx.mpi_size > 1) {
#ifdef GGML_USE_MPI
cur = ggml_mpi_recv_tensor(ctx0, cur,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N),
lctx.mpi_size-1);
ggml_set_name(cur, "mpi_recv");
#else
GGML_ASSERT(false);
#endif
}
// norm
{
cur = ggml_rms_norm(ctx0, cur);
cur = ggml_rms_norm(ctx0, inpL);
offload_func_nr(cur);
ggml_set_name(cur, "rms_norm_2");
@ -1642,11 +1620,9 @@ static bool llama_eval_internal(
embeddings = cur;
}
// lm_head
cur = ggml_mul_mat(ctx0, model.output, cur);
ggml_set_name(cur, "result_output");
}
lctx.use_buf(ctx0, -1);
@ -1680,6 +1656,10 @@ static bool llama_eval_internal(
ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
}
#elif GGML_USE_MPI
ggml_mpi_graph_compute(lctx.ctx_mpi, ctx0, &gf, n_layer);
cur = gf.nodes[gf.n_nodes - 1];
#else
ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
#endif
@ -1705,7 +1685,11 @@ static bool llama_eval_internal(
// update kv token count
lctx.kv_self.n = n_past + N;
if (lctx.mpi_rank == 0) {
#ifdef GGML_USE_MPI
if (ggml_mpi_rank(lctx.ctx_mpi) == 0) {
#else
{
#endif
// extract logits
{
auto & logits_out = lctx.logits;
@ -2676,14 +2660,6 @@ struct llama_context * llama_new_context_with_model(
ctx->rng = std::mt19937(params.seed);
ctx->logits_all = params.logits_all;
#ifdef GGML_USE_MPI
MPI_Comm_size(MPI_COMM_WORLD, &ctx->mpi_size);
MPI_Comm_rank(MPI_COMM_WORLD, &ctx->mpi_rank);
#else
ctx->mpi_size = 1;
ctx->mpi_rank = 0;
#endif
ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
// reserve memory for context buffers
@ -2756,15 +2732,17 @@ struct llama_context * llama_new_context_with_model(
}
#endif
if (ctx->mpi_rank > 0) {
// Enter a blocking eval loop with dummy input, letting rank=0 drive the process
const std::vector<llama_token> tmp = { llama_token_bos(), };
while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0));
#ifdef GGML_USE_MPI
MPI_Finalize();
#endif
ctx->ctx_mpi = ggml_mpi_init();
if (ggml_mpi_rank(ctx->ctx_mpi) > 0) {
// Enter a blocking eval loop with dummy input, letting rank=0 drive the process
const std::vector<llama_token> tmp(ctx->model.hparams.n_ctx, llama_token_bos());
while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0)) {};
llama_backend_free();
exit(1);
}
#endif
return ctx;
}
@ -3443,13 +3421,6 @@ int llama_eval(
int n_tokens,
int n_past,
int n_threads) {
#ifdef GGML_USE_MPI
// Synchronize the worker node parameters with the root node
MPI_Barrier(MPI_COMM_WORLD);
MPI_Bcast(&n_past, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(&n_tokens, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(&n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
#endif
if (!llama_eval_internal(*ctx, tokens, nullptr, n_tokens, n_past, n_threads, nullptr)) {
fprintf(stderr, "%s: failed to eval\n", __func__);
return 1;

4
llama.h
View file

@ -158,9 +158,9 @@ extern "C" {
// Initialize the llama + ggml backend
// If numa is true, use NUMA optimizations
// Call once at the start of the program
LLAMA_API void llama_init_backend(bool numa);
LLAMA_API void llama_backend_init(bool numa);
// Call once at the end of the program - currently only used for MPI
LLAMA_API void llama_finalize_backend();
LLAMA_API void llama_backend_free();
LLAMA_API int64_t llama_time_us();