baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

update llama.cpp to f64d44a

Browse source
This commit is contained in:
Jeffrey Morgan 2023-08-12 22:47:15 -04:00
parent ed969d2a06
commit 22885aeaee
19 changed files with 2197 additions and 1073 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
llm/ggml-alloc.c
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *
@ -420,6 +420,14 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
if (parent == NULL) { if (parent == NULL) {
break; break;
} }
// if the node's data is external, then we cannot re-use it
if ((char *) parent->data < (char *) alloc->data ||
(char *) parent->data >= ((char *) alloc->data + alloc->size)) {
AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
continue;
}
struct hash_node * p_hn = hash_get(ht, parent); struct hash_node * p_hn = hash_get(ht, parent);
if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) { if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
if (ggml_is_view(parent)) { if (ggml_is_view(parent)) {

2
llm/ggml-alloc.h
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

2026
llm/ggml-cuda.cu
View file

File diff suppressed because it is too large Load diff

2
llm/ggml-cuda.h
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

2
llm/ggml-metal.h
View file

@ -1,7 +1,7 @@
//go:build darwin //go:build darwin
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

47
llm/ggml-metal.m
View file

@ -1,7 +1,7 @@
//go:build darwin //go:build darwin
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *
@ -35,6 +35,11 @@
#import <Metal/Metal.h> #import <Metal/Metal.h>
#import <MetalPerformanceShaders/MetalPerformanceShaders.h> #import <MetalPerformanceShaders/MetalPerformanceShaders.h>
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#ifdef GGML_METAL_NDEBUG #ifdef GGML_METAL_NDEBUG
#define metal_printf(...) #define metal_printf(...)
#else #else
@ -43,6 +48,8 @@
#define UNUSED(x) (void)(x) #define UNUSED(x) (void)(x)
#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
struct ggml_metal_buffer { struct ggml_metal_buffer {
const char * name; const char * name;
@ -64,7 +71,7 @@ struct ggml_metal_context {
int n_buffers; int n_buffers;
struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS]; struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
int concur_list[GGML_MAX_NODES]; int concur_list[GGML_MAX_CONCUR];
int concur_list_len; int concur_list_len;
// custom kernels // custom kernels
@ -398,9 +405,9 @@ void ggml_metal_graph_find_concurrency(
struct ggml_metal_context * ctx, struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) { struct ggml_cgraph * gf) {
int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
int nodes_unused[GGML_MAX_NODES]; int nodes_unused[GGML_MAX_CONCUR];
for (int i = 0; i < GGML_MAX_NODES; i++) {ctx->concur_list[i] = 0;} for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
for (int i = 0; i < gf->n_nodes; i++) { nodes_unused[i] = 1; } for (int i = 0; i < gf->n_nodes; i++) { nodes_unused[i] = 1; }
ctx->concur_list_len = 0; ctx->concur_list_len = 0;
@ -415,20 +422,32 @@ void ggml_metal_graph_find_concurrency(
if (nodes_unused[i]) { if (nodes_unused[i]) {
// if the requirements for gf->nodes[i] are satisfied // if the requirements for gf->nodes[i] are satisfied
int exe_flag = 1; int exe_flag = 1;
// scan all srcs // scan all srcs
for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) { for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind]; struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
if (src_cur) { if (src_cur) {
// if is leaf nodes it's satisfied. // if is leaf nodes it's satisfied.
if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {continue;} // TODO: ggml_is_leaf()
if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
continue;
}
// otherwise this src should be the output from previous nodes. // otherwise this src should be the output from previous nodes.
int is_found = 0; int is_found = 0;
// scan 2*search_depth back because we inserted barrier. // scan 2*search_depth back because we inserted barrier.
for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) { //for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
if (gf->nodes[ctx->concur_list[j]] == src_cur) {is_found = 1; break;} for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
is_found = 1;
break;
}
}
if (is_found == 0) {
exe_flag = 0;
break;
} }
if (is_found == 0) {exe_flag = 0; break;}
} }
} }
if (exe_flag) { if (exe_flag) {
@ -444,9 +463,9 @@ void ggml_metal_graph_find_concurrency(
if (((int64_t)gf->nodes[j]->data) >= data_start + length || \ if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) { ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
continue; continue;
} else {
exe_flag = 0;
} }
exe_flag = 0;
} }
} }
} }
@ -463,11 +482,13 @@ void ggml_metal_graph_find_concurrency(
ctx->concur_list[level_pos + concurrency] = -1; ctx->concur_list[level_pos + concurrency] = -1;
ctx->concur_list_len++; ctx->concur_list_len++;
// jump all sorted nodes at nodes_bak // jump all sorted nodes at nodes_bak
while (!nodes_unused[n_start]) {n_start++;} while (!nodes_unused[n_start]) {
n_start++;
}
level_pos += concurrency + 1; level_pos += concurrency + 1;
} }
if (ctx->concur_list_len > GGML_MAX_NODES) { if (ctx->concur_list_len > GGML_MAX_CONCUR) {
fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__); fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
} }
} }
@ -481,7 +502,7 @@ void ggml_metal_graph_compute(
// else fallback to serial dispatch // else fallback to serial dispatch
MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor; MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_NODES; const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
const int n_nodes = has_concur ? ctx->concur_list_len : gf->n_nodes; const int n_nodes = has_concur ? ctx->concur_list_len : gf->n_nodes;
edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial; edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;

2
llm/ggml-metal.metal
View file

@ -1,7 +1,7 @@
//go:build darwin //go:build darwin
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

2
llm/ggml-mpi.c
View file

@ -1,7 +1,7 @@
//go:build mpi //go:build mpi
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

2
llm/ggml-mpi.h
View file

@ -1,7 +1,7 @@
//go:build mpi //go:build mpi
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

2
llm/ggml-opencl.cpp
View file

@ -1,7 +1,7 @@
//go:build opencl //go:build opencl
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

2
llm/ggml-opencl.h
View file

@ -1,7 +1,7 @@
//go:build opencl //go:build opencl
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

482
llm/ggml.c
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *
@ -3837,7 +3837,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"CROSS_ENTROPY_LOSS_BACK", "CROSS_ENTROPY_LOSS_BACK",
}; };
static_assert(GGML_OP_COUNT == 59, "GGML_OP_COUNT != 59"); static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none", "none",
@ -3909,7 +3909,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"cross_entropy_loss_back(x,y)", "cross_entropy_loss_back(x,y)",
}; };
static_assert(GGML_OP_COUNT == 59, "GGML_OP_COUNT != 59"); static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@ -4136,7 +4136,7 @@ size_t ggml_nbytes(const struct ggml_tensor * tensor) {
// //
// is enough, but just in case, adding the second part // is enough, but just in case, adding the second part
return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]); return GGML_PAD(MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]), GGML_MEM_ALIGN);
} }
size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) { size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
@ -4279,7 +4279,7 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
} }
static inline bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) { bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return return
@ -4660,6 +4660,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
} }
static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) { static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
assert(params_size <= GGML_MAX_OP_PARAMS); assert(params_size <= GGML_MAX_OP_PARAMS);
memcpy(tensor->op_params, params, params_size); memcpy(tensor->op_params, params, params_size);
} }
@ -6465,7 +6466,7 @@ struct ggml_tensor * ggml_permute(
result->src[0] = a; result->src[0] = a;
int32_t params[] = { axis0, axis1, axis2, axis3 }; int32_t params[] = { axis0, axis1, axis2, axis3 };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
return result; return result;
} }
@ -6591,7 +6592,7 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl(
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
int32_t params[] = { n_past, inplace ? 1 : 0 }; int32_t params[] = { n_past, inplace ? 1 : 0 };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_DIAG_MASK_INF; result->op = GGML_OP_DIAG_MASK_INF;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -6631,7 +6632,7 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl(
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
int32_t params[] = { n_past, inplace ? 1 : 0 }; int32_t params[] = { n_past, inplace ? 1 : 0 };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_DIAG_MASK_ZERO; result->op = GGML_OP_DIAG_MASK_ZERO;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -6749,7 +6750,7 @@ static struct ggml_tensor * ggml_rope_impl(
int32_t params[6] = { n_past, n_dims, mode, n_ctx }; int32_t params[6] = { n_past, n_dims, mode, n_ctx };
memcpy(params + 4, &freq_base, sizeof(float)); memcpy(params + 4, &freq_base, sizeof(float));
memcpy(params + 5, &freq_scale, sizeof(float)); memcpy(params + 5, &freq_scale, sizeof(float));
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_ROPE; result->op = GGML_OP_ROPE;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -6823,7 +6824,7 @@ struct ggml_tensor * ggml_rope_back(
struct ggml_tensor * result = ggml_dup_tensor(ctx, a); struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
int32_t params[] = { n_past, n_dims, mode, n_ctx }; int32_t params[] = { n_past, n_dims, mode, n_ctx };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_ROPE_BACK; result->op = GGML_OP_ROPE_BACK;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -6854,7 +6855,7 @@ struct ggml_tensor * ggml_alibi(
int32_t op_params[3] = { n_past, n_head }; int32_t op_params[3] = { n_past, n_head };
memcpy(op_params + 2, &bias_max, sizeof(float)); memcpy(op_params + 2, &bias_max, sizeof(float));
ggml_set_op_params(result, &op_params, sizeof(op_params)); ggml_set_op_params(result, op_params, sizeof(op_params));
result->op = GGML_OP_ALIBI; result->op = GGML_OP_ALIBI;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -6881,7 +6882,7 @@ struct ggml_tensor * ggml_clamp(
struct ggml_tensor * result = ggml_view_tensor(ctx, a); struct ggml_tensor * result = ggml_view_tensor(ctx, a);
float params[] = { min, max }; float params[] = { min, max };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_CLAMP; result->op = GGML_OP_CLAMP;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -6919,7 +6920,7 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
int32_t params[] = { s0, p0, d0 }; int32_t params[] = { s0, p0, d0 };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_CONV_1D; result->op = GGML_OP_CONV_1D;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -6958,7 +6959,7 @@ struct ggml_tensor* ggml_conv_2d(
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
int32_t params[] = { s0, s1, p0, p1, d0, d1 }; int32_t params[] = { s0, s1, p0, p1, d0, d1 };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_CONV_2D; result->op = GGML_OP_CONV_2D;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -7011,7 +7012,7 @@ struct ggml_tensor* ggml_pool_1d(
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
int32_t params[] = { op, k0, s0, p0 }; int32_t params[] = { op, k0, s0, p0 };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_POOL_1D; result->op = GGML_OP_POOL_1D;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -7048,7 +7049,7 @@ struct ggml_tensor* ggml_pool_2d(
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne); struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
int32_t params[] = { op, k0, k1, s0, s1, p0, p1 }; int32_t params[] = { op, k0, k1, s0, s1, p0, p1 };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_POOL_2D; result->op = GGML_OP_POOL_2D;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -7216,7 +7217,7 @@ struct ggml_tensor * ggml_win_part(
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
int32_t params[] = { npx, npy, w }; int32_t params[] = { npx, npy, w };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_WIN_PART; result->op = GGML_OP_WIN_PART;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -7246,7 +7247,7 @@ struct ggml_tensor * ggml_win_unpart(
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne); struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
int32_t params[] = { w }; int32_t params[] = { w };
ggml_set_op_params(result, &params, sizeof(params)); ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_WIN_UNPART; result->op = GGML_OP_WIN_UNPART;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -7375,7 +7376,7 @@ struct ggml_tensor * ggml_map_binary_inplace_f32(
return ggml_map_binary_impl_f32(ctx, a, b, fun, true); return ggml_map_binary_impl_f32(ctx, a, b, fun, true);
} }
// ggml_map_custom1 // ggml_map_custom1_f32
static struct ggml_tensor * ggml_map_custom1_impl_f32( static struct ggml_tensor * ggml_map_custom1_impl_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
@ -7392,7 +7393,7 @@ static struct ggml_tensor * ggml_map_custom1_impl_f32(
ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_CUSTOM1; result->op = GGML_OP_MAP_CUSTOM1_F32;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a; result->src[0] = a;
@ -7413,7 +7414,7 @@ struct ggml_tensor * ggml_map_custom1_inplace_f32(
return ggml_map_custom1_impl_f32(ctx, a, fun, true); return ggml_map_custom1_impl_f32(ctx, a, fun, true);
} }
// ggml_map_custom2 // ggml_map_custom2_f32
static struct ggml_tensor * ggml_map_custom2_impl_f32( static struct ggml_tensor * ggml_map_custom2_impl_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
@ -7431,7 +7432,7 @@ static struct ggml_tensor * ggml_map_custom2_impl_f32(
ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_CUSTOM2; result->op = GGML_OP_MAP_CUSTOM2_F32;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a; result->src[0] = a;
result->src[1] = b; result->src[1] = b;
@ -7455,7 +7456,7 @@ struct ggml_tensor * ggml_map_custom2_inplace_f32(
return ggml_map_custom2_impl_f32(ctx, a, b, fun, true); return ggml_map_custom2_impl_f32(ctx, a, b, fun, true);
} }
// ggml_map_custom3 // ggml_map_custom3_f32
static struct ggml_tensor * ggml_map_custom3_impl_f32( static struct ggml_tensor * ggml_map_custom3_impl_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
@ -7474,7 +7475,7 @@ static struct ggml_tensor * ggml_map_custom3_impl_f32(
ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_CUSTOM3; result->op = GGML_OP_MAP_CUSTOM3_F32;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a; result->src[0] = a;
result->src[1] = b; result->src[1] = b;
@ -7501,6 +7502,190 @@ struct ggml_tensor * ggml_map_custom3_inplace_f32(
return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true); return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true);
} }
// ggml_map_custom1
struct ggml_map_custom1_op_params {
ggml_custom1_op_t fun;
int n_tasks;
void * userdata;
};
static struct ggml_tensor * ggml_map_custom1_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_custom1_op_t fun,
int n_tasks,
void * userdata,
bool inplace) {
GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
bool is_node = false;
if (!inplace && a->grad) {
is_node = true;
}
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
struct ggml_map_custom1_op_params params = {
/*.fun =*/ fun,
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, (const void *) &params, sizeof(params));
result->op = GGML_OP_MAP_CUSTOM1;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
return result;
}
struct ggml_tensor * ggml_map_custom1(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_custom1_op_t fun,
int n_tasks,
void * userdata) {
return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, false);
}
struct ggml_tensor * ggml_map_custom1_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_custom1_op_t fun,
int n_tasks,
void * userdata) {
return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, true);
}
// ggml_map_custom2
struct ggml_map_custom2_op_params {
ggml_custom2_op_t fun;
int n_tasks;
void * userdata;
};
static struct ggml_tensor * ggml_map_custom2_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_custom2_op_t fun,
int n_tasks,
void * userdata,
bool inplace) {
GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
bool is_node = false;
if (!inplace && (a->grad || b->grad)) {
is_node = true;
}
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
struct ggml_map_custom2_op_params params = {
/*.fun =*/ fun,
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, (const void *) &params, sizeof(params));
result->op = GGML_OP_MAP_CUSTOM2;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = b;
return result;
}
struct ggml_tensor * ggml_map_custom2(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_custom2_op_t fun,
int n_tasks,
void * userdata) {
return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, false);
}
struct ggml_tensor * ggml_map_custom2_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_custom2_op_t fun,
int n_tasks,
void * userdata) {
return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, true);
}
// ggml_map_custom3
struct ggml_map_custom3_op_params {
ggml_custom3_op_t fun;
int n_tasks;
void * userdata;
};
static struct ggml_tensor * ggml_map_custom3_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
const ggml_custom3_op_t fun,
int n_tasks,
void * userdata,
bool inplace) {
GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
bool is_node = false;
if (!inplace && (a->grad || b->grad || c->grad)) {
is_node = true;
}
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
struct ggml_map_custom3_op_params params = {
/*.fun =*/ fun,
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, (const void *) &params, sizeof(params));
result->op = GGML_OP_MAP_CUSTOM3;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = b;
result->src[2] = c;
return result;
}
struct ggml_tensor * ggml_map_custom3(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
const ggml_custom3_op_t fun,
int n_tasks,
void * userdata) {
return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, false);
}
struct ggml_tensor * ggml_map_custom3_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
const ggml_custom3_op_t fun,
int n_tasks,
void * userdata) {
return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true);
}
// ggml_cross_entropy_loss // ggml_cross_entropy_loss
struct ggml_tensor * ggml_cross_entropy_loss( struct ggml_tensor * ggml_cross_entropy_loss(
@ -10572,32 +10757,63 @@ static void ggml_compute_forward_mul_mat(
return; return;
} }
// parallelize by src0 rows
const int64_t dr = (ne01 + nth - 1)/nth;
const int64_t ir10 = dr*ith;
const int64_t ir11 = MIN(ir10 + dr, ne01);
// src1 rows
const int64_t nr1 = ne11*ne12*ne13;
const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type]; const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
for (int64_t ir1 = 0; ir1 < nr1; ++ir1) { const int64_t nr0 = ne01; // src0 rows
const int64_t nr1 = ne11*ne12*ne13; // src1 rows
//printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
// distribute the thread work across the inner or outer loop based on which one is larger
const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
const int64_t ith0 = ith % nth0;
const int64_t ith1 = ith / nth0;
const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
const int64_t ir010 = dr0*ith0;
const int64_t ir011 = MIN(ir010 + dr0, nr0);
const int64_t ir110 = dr1*ith1;
const int64_t ir111 = MIN(ir110 + dr1, nr1);
//printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
// threads with no work simply yield (not sure if it helps)
if (ir010 >= ir011 || ir110 >= ir111) {
sched_yield();
return;
}
assert(ne12 % ne02 == 0);
assert(ne13 % ne03 == 0);
// broadcast factors
const int64_t r2 = ne12/ne02;
const int64_t r3 = ne13/ne03;
// block-tiling attempt
const int64_t blck_0 = 16;
const int64_t blck_1 = 16;
// attempt to reduce false-sharing (does not seem to make a difference)
float tmp[16];
for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
const int64_t i13 = (ir1/(ne12*ne11)); const int64_t i13 = (ir1/(ne12*ne11));
const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11; const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11); const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
const int64_t ir0 = (ir1/ne11)%(ne02*ne03); // broadcast src0 into src1
const int64_t i03 = (ir0/(ne02)); const int64_t i03 = i13/r3;
// Hack for "Falcon multi-query-attention key stutter" / alternative to ggml_repeat2. const int64_t i02 = i12/r2;
// See https://github.com/ggerganov/llama.cpp/issues/1602#issuecomment-1606087470:
// GG: this is likely the correct way to broadcast, though need some more thought
// therefore leaving the comments to remind us for now
const int64_t i02 = (i12 / (ne12 / ne02));
// Original from PR/224 (and also essential/correct for non-broadcast matmuls in Falcon)
// const int64_t i02 = (ir0 - i03*ne02);
const int64_t i1 = i11; const int64_t i1 = i11;
const int64_t i2 = i12; const int64_t i2 = i12;
@ -10616,28 +10832,21 @@ static void ggml_compute_forward_mul_mat(
float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3)); float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
for (int64_t ir = ir10; ir < ir11; ++ir) { //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
vec_dot(ne00, &dst_col[ir], src0_row + ir*nb01, src1_col); // vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
}
}
//int64_t t1 = ggml_time_us();
//static int64_t acc = 0;
//acc += t1 - t0;
//if (t1 - t0 > 10) {
// printf("\n");
// printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
// printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
// printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
// printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
//} //}
}
for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
}
memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
}
}
}
}
// ggml_compute_forward_out_prod // ggml_compute_forward_out_prod
static void ggml_compute_forward_out_prod_f32( static void ggml_compute_forward_out_prod_f32(
const struct ggml_compute_params * params, const struct ggml_compute_params * params,
const struct ggml_tensor * src0, const struct ggml_tensor * src0,
@ -14253,24 +14462,6 @@ static void ggml_compute_forward_map_custom1_f32(
fun(dst, a); fun(dst, a);
} }
static void ggml_compute_forward_map_custom1(
const struct ggml_compute_params * params,
const struct ggml_tensor * a,
struct ggml_tensor * dst,
const ggml_custom1_op_f32_t fun) {
switch (a->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_map_custom1_f32(params, a, dst, fun);
} break;
default:
{
GGML_ASSERT(false);
} break;
}
}
// ggml_compute_forward_map_custom2 // ggml_compute_forward_map_custom2
static void ggml_compute_forward_map_custom2_f32( static void ggml_compute_forward_map_custom2_f32(
@ -14289,24 +14480,6 @@ static void ggml_compute_forward_map_custom2_f32(
} }
static void ggml_compute_forward_map_custom2(
const struct ggml_compute_params * params,
const struct ggml_tensor * a,
const struct ggml_tensor * b,
struct ggml_tensor * dst,
const ggml_custom2_op_f32_t fun) {
switch (a->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_map_custom2_f32(params, a, b, dst, fun);
} break;
default:
{
GGML_ASSERT(false);
} break;
}
}
// ggml_compute_forward_map_custom3 // ggml_compute_forward_map_custom3
static void ggml_compute_forward_map_custom3_f32( static void ggml_compute_forward_map_custom3_f32(
@ -14325,24 +14498,52 @@ static void ggml_compute_forward_map_custom3_f32(
fun(dst, a, b, c); fun(dst, a, b, c);
} }
// ggml_compute_forward_map_custom1
static void ggml_compute_forward_map_custom1(
const struct ggml_compute_params * params,
const struct ggml_tensor * a,
struct ggml_tensor * dst) {
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
return;
}
struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) dst->op_params;
p->fun(dst, a, params->ith, params->nth, p->userdata);
}
// ggml_compute_forward_map_custom2
static void ggml_compute_forward_map_custom2(
const struct ggml_compute_params * params,
const struct ggml_tensor * a,
const struct ggml_tensor * b,
struct ggml_tensor * dst) {
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
return;
}
struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) dst->op_params;
p->fun(dst, a, b, params->ith, params->nth, p->userdata);
}
// ggml_compute_forward_map_custom3
static void ggml_compute_forward_map_custom3( static void ggml_compute_forward_map_custom3(
const struct ggml_compute_params * params, const struct ggml_compute_params * params,
const struct ggml_tensor * a, const struct ggml_tensor * a,
const struct ggml_tensor * b, const struct ggml_tensor * b,
const struct ggml_tensor * c, const struct ggml_tensor * c,
struct ggml_tensor * dst, struct ggml_tensor * dst) {
const ggml_custom3_op_f32_t fun) { if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
switch (a->type) { return;
case GGML_TYPE_F32:
{
ggml_compute_forward_map_custom3_f32(params, a, b, c, dst, fun);
} break;
default:
{
GGML_ASSERT(false);
} break;
} }
struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) dst->op_params;
p->fun(dst, a, b, c, params->ith, params->nth, p->userdata);
} }
// ggml_compute_forward_cross_entropy_loss // ggml_compute_forward_cross_entropy_loss
@ -14864,25 +15065,40 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
ggml_compute_forward_map_binary(params, tensor->src[0], tensor->src[1], tensor, fun); ggml_compute_forward_map_binary(params, tensor->src[0], tensor->src[1], tensor, fun);
} }
break; break;
case GGML_OP_MAP_CUSTOM1: case GGML_OP_MAP_CUSTOM1_F32:
{ {
ggml_custom1_op_f32_t fun; ggml_custom1_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun)); memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom1(params, tensor->src[0], tensor, fun); ggml_compute_forward_map_custom1_f32(params, tensor->src[0], tensor, fun);
}
break;
case GGML_OP_MAP_CUSTOM2_F32:
{
ggml_custom2_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom2_f32(params, tensor->src[0], tensor->src[1], tensor, fun);
}
break;
case GGML_OP_MAP_CUSTOM3_F32:
{
ggml_custom3_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom3_f32(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
}
break;
case GGML_OP_MAP_CUSTOM1:
{
ggml_compute_forward_map_custom1(params, tensor->src[0], tensor);
} }
break; break;
case GGML_OP_MAP_CUSTOM2: case GGML_OP_MAP_CUSTOM2:
{ {
ggml_custom2_op_f32_t fun; ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor);
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor, fun);
} }
break; break;
case GGML_OP_MAP_CUSTOM3: case GGML_OP_MAP_CUSTOM3:
{ {
ggml_custom3_op_f32_t fun; ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
} }
break; break;
case GGML_OP_CROSS_ENTROPY_LOSS: case GGML_OP_CROSS_ENTROPY_LOSS:
@ -15690,6 +15906,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
} break; } break;
case GGML_OP_MAP_UNARY: case GGML_OP_MAP_UNARY:
case GGML_OP_MAP_BINARY: case GGML_OP_MAP_BINARY:
case GGML_OP_MAP_CUSTOM1_F32:
case GGML_OP_MAP_CUSTOM2_F32:
case GGML_OP_MAP_CUSTOM3_F32:
case GGML_OP_MAP_CUSTOM1: case GGML_OP_MAP_CUSTOM1:
case GGML_OP_MAP_CUSTOM2: case GGML_OP_MAP_CUSTOM2:
case GGML_OP_MAP_CUSTOM3: case GGML_OP_MAP_CUSTOM3:
@ -16475,12 +16694,39 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
case GGML_OP_WIN_UNPART: case GGML_OP_WIN_UNPART:
case GGML_OP_MAP_UNARY: case GGML_OP_MAP_UNARY:
case GGML_OP_MAP_BINARY: case GGML_OP_MAP_BINARY:
case GGML_OP_MAP_CUSTOM1: case GGML_OP_MAP_CUSTOM1_F32:
case GGML_OP_MAP_CUSTOM2: case GGML_OP_MAP_CUSTOM2_F32:
case GGML_OP_MAP_CUSTOM3: case GGML_OP_MAP_CUSTOM3_F32:
{ {
n_tasks = 1; n_tasks = 1;
} break; } break;
case GGML_OP_MAP_CUSTOM1:
{
struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params;
if (p->n_tasks == GGML_N_TASKS_MAX) {
n_tasks = n_threads;
} else {
n_tasks = MIN(p->n_tasks, n_threads);
}
} break;
case GGML_OP_MAP_CUSTOM2:
{
struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params;
if (p->n_tasks == GGML_N_TASKS_MAX) {
n_tasks = n_threads;
} else {
n_tasks = MIN(p->n_tasks, n_threads);
}
} break;
case GGML_OP_MAP_CUSTOM3:
{
struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params;
if (p->n_tasks == GGML_N_TASKS_MAX) {
n_tasks = n_threads;
} else {
n_tasks = MIN(p->n_tasks, n_threads);
}
} break;
case GGML_OP_CROSS_ENTROPY_LOSS: case GGML_OP_CROSS_ENTROPY_LOSS:
{ {
n_tasks = n_threads; n_tasks = n_threads;

141
llm/ggml.h
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *
@ -209,6 +209,15 @@
# define GGML_API # define GGML_API
#endif #endif
// TODO: support for clang
#ifdef __GNUC__
# define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
#elif defined(_MSC_VER)
# define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
#else
# define GGML_DEPRECATED(func, hint) func
#endif
#include <stdint.h> #include <stdint.h>
#include <stddef.h> #include <stddef.h>
#include <stdbool.h> #include <stdbool.h>
@ -400,6 +409,10 @@ extern "C" {
GGML_OP_MAP_UNARY, GGML_OP_MAP_UNARY,
GGML_OP_MAP_BINARY, GGML_OP_MAP_BINARY,
GGML_OP_MAP_CUSTOM1_F32,
GGML_OP_MAP_CUSTOM2_F32,
GGML_OP_MAP_CUSTOM3_F32,
GGML_OP_MAP_CUSTOM1, GGML_OP_MAP_CUSTOM1,
GGML_OP_MAP_CUSTOM2, GGML_OP_MAP_CUSTOM2,
GGML_OP_MAP_CUSTOM3, GGML_OP_MAP_CUSTOM3,
@ -596,6 +609,8 @@ extern "C" {
GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor); GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor); GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor);
GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
// use this to compute the memory overhead of a tensor // use this to compute the memory overhead of a tensor
GGML_API size_t ggml_tensor_overhead(void); GGML_API size_t ggml_tensor_overhead(void);
@ -1341,15 +1356,6 @@ extern "C" {
int h0, int h0,
int w); int w);
// custom operators
typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
GGML_API struct ggml_tensor * ggml_unary( GGML_API struct ggml_tensor * ggml_unary(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
@ -1360,63 +1366,138 @@ extern "C" {
struct ggml_tensor * a, struct ggml_tensor * a,
enum ggml_unary_op op); enum ggml_unary_op op);
GGML_API struct ggml_tensor * ggml_map_unary_f32( // custom operators
typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
ggml_unary_op_f32_t fun); ggml_unary_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
ggml_unary_op_f32_t fun); ggml_unary_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_API struct ggml_tensor * ggml_map_binary_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
ggml_binary_op_f32_t fun); ggml_binary_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
ggml_binary_op_f32_t fun); ggml_binary_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_API struct ggml_tensor * ggml_map_custom1_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
ggml_custom1_op_f32_t fun); ggml_custom1_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
ggml_custom1_op_f32_t fun); ggml_custom1_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_API struct ggml_tensor * ggml_map_custom2_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
ggml_custom2_op_f32_t fun); ggml_custom2_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
ggml_custom2_op_f32_t fun); ggml_custom2_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_API struct ggml_tensor * ggml_map_custom3_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
struct ggml_tensor * c, struct ggml_tensor * c,
ggml_custom3_op_f32_t fun); ggml_custom3_op_f32_t fun),
"use ggml_map_custom3 instead");
GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32( GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
struct ggml_tensor * c, struct ggml_tensor * c,
ggml_custom3_op_f32_t fun); ggml_custom3_op_f32_t fun),
"use ggml_map_custom3_inplace instead");
// custom operators v2
typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
#define GGML_N_TASKS_MAX -1
GGML_API struct ggml_tensor * ggml_map_custom1(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom2(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom3(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_t fun,
int n_tasks,
void * userdata);
// loss function // loss function

2
llm/k_quants.c
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

2
llm/k_quants.h
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *

44
llm/llama-util.h
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *
@ -175,6 +175,46 @@ struct llama_file {
} }
}; };
// llama_context_data
struct llama_data_context {
virtual void write(const void * src, size_t size) = 0;
virtual size_t get_size_written() = 0;
virtual ~llama_data_context() = default;
};
struct llama_data_buffer_context : llama_data_context {
uint8_t* ptr;
size_t size_written = 0;
llama_data_buffer_context(uint8_t * p) : ptr(p) {}
void write(const void * src, size_t size) override {
memcpy(ptr, src, size);
ptr += size;
size_written += size;
}
size_t get_size_written() override {
return size_written;
}
};
struct llama_data_file_context : llama_data_context {
llama_file* file;
size_t size_written = 0;
llama_data_file_context(llama_file * f) : file(f) {}
void write(const void * src, size_t size) override {
file->write_raw(src, size);
size_written += size;
}
size_t get_size_written() override {
return size_written;
}
};
#if defined(_WIN32) #if defined(_WIN32)
static std::string llama_format_win_err(DWORD err) { static std::string llama_format_win_err(DWORD err) {
LPSTR buf; LPSTR buf;
@ -205,7 +245,7 @@ struct llama_mmap {
// prefetch/readahead impairs performance on NUMA systems // prefetch/readahead impairs performance on NUMA systems
if (numa) { prefetch = 0; } if (numa) { prefetch = 0; }
#ifdef __linux__ #ifdef __linux__
if (prefetch) { flags |= MAP_POPULATE; } if (prefetch >= file->size) { flags |= MAP_POPULATE; }
#endif #endif
addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0); addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
if (addr == MAP_FAILED) { if (addr == MAP_FAILED) {

350
llm/llama.cpp
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *
@ -82,6 +82,13 @@
#pragma warning(disable: 4244 4267) // possible loss of data #pragma warning(disable: 4244 4267) // possible loss of data
#endif #endif
static void llama_log_internal(llama_log_level level, const char* format, ...);
static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data);
#define LLAMA_LOG_INFO(...) llama_log_internal(LLAMA_LOG_LEVEL_INFO , __VA_ARGS__)
#define LLAMA_LOG_WARN(...) llama_log_internal(LLAMA_LOG_LEVEL_WARN , __VA_ARGS__)
#define LLAMA_LOG_ERROR(...) llama_log_internal(LLAMA_LOG_LEVEL_ERROR, __VA_ARGS__)
#if !defined(GGML_USE_CUBLAS) && !defined(GGML_USE_METAL) #if !defined(GGML_USE_CUBLAS) && !defined(GGML_USE_METAL)
#include "ggml-alloc.h" #include "ggml-alloc.h"
#define LLAMA_USE_ALLOCATOR #define LLAMA_USE_ALLOCATOR
@ -175,7 +182,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
} }
// amount of VRAM needed per batch size to hold temporary results // amount of VRAM needed per batch size to hold temporary results
// the values for 3b and 65b are not derived from testing but instead chosen conservatively // the values for 3b are not derived from testing but instead chosen conservatively
static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE() static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
{ {
static std::map<e_model, size_t> k_sizes = { static std::map<e_model, size_t> k_sizes = {
@ -183,14 +190,14 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
{ MODEL_7B, 512ull * kB }, { MODEL_7B, 512ull * kB },
{ MODEL_13B, 640ull * kB }, { MODEL_13B, 640ull * kB },
{ MODEL_30B, 768ull * kB }, { MODEL_30B, 768ull * kB },
{ MODEL_65B, 1536ull * kB }, { MODEL_65B, 1280ull * kB },
{ MODEL_70B, 1536ull * kB }, // TODO (likely can be reduced) { MODEL_70B, 1280ull * kB },
}; };
return k_sizes; return k_sizes;
} }
// amount of VRAM needed per batch size and context to hold temporary results // amount of VRAM needed per batch size and context to hold temporary results
// the values for 3b and 65b are not derived from testing but instead chosen conservatively // the values for 3b are not derived from testing but instead chosen conservatively
static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT() static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
{ {
static std::map<e_model, size_t> k_sizes = { static std::map<e_model, size_t> k_sizes = {
@ -198,8 +205,8 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
{ MODEL_7B, 128ull }, { MODEL_7B, 128ull },
{ MODEL_13B, 160ull }, { MODEL_13B, 160ull },
{ MODEL_30B, 208ull }, { MODEL_30B, 208ull },
{ MODEL_65B, 416ull }, { MODEL_65B, 256ull },
{ MODEL_70B, 416ull }, // TODO (likely can be reduced) { MODEL_70B, 256ull },
}; };
return k_sizes; return k_sizes;
} }
@ -464,6 +471,14 @@ struct llama_context {
} }
}; };
struct llama_state {
// We save the log callback globally
llama_log_callback log_callback = llama_log_callback_default;
void * log_callback_user_data = nullptr;
};
// global state
static llama_state g_state;
template <typename T> template <typename T>
static T checked_mul(T a, T b) { static T checked_mul(T a, T b) {
T ret = a * b; T ret = a * b;
@ -530,7 +545,7 @@ struct llama_file_loader {
llama_file_loader(const char * fname, llama_load_tensors_map & tensors_map) llama_file_loader(const char * fname, llama_load_tensors_map & tensors_map)
: file(fname, "rb") { : file(fname, "rb") {
fprintf(stderr, "llama.cpp: loading model from %s\n", fname); LLAMA_LOG_INFO("llama.cpp: loading model from %s\n", fname);
read_magic(); read_magic();
read_hparams(); read_hparams();
read_vocab(); read_vocab();
@ -645,7 +660,7 @@ struct llama_file_saver {
llama_file_loader * any_file_loader; llama_file_loader * any_file_loader;
llama_file_saver(const char * fname, llama_file_loader * any_file_loader, enum llama_ftype new_ftype) llama_file_saver(const char * fname, llama_file_loader * any_file_loader, enum llama_ftype new_ftype)
: file(fname, "wb"), any_file_loader(any_file_loader) { : file(fname, "wb"), any_file_loader(any_file_loader) {
fprintf(stderr, "llama.cpp: saving model to %s\n", fname); LLAMA_LOG_INFO("llama.cpp: saving model to %s\n", fname);
write_magic(); write_magic();
write_hparams(new_ftype); write_hparams(new_ftype);
write_vocab(); write_vocab();
@ -666,7 +681,7 @@ struct llama_file_saver {
} }
void write_vocab() { void write_vocab() {
if (any_file_loader->file_version == LLAMA_FILE_VERSION_GGML) { if (any_file_loader->file_version == LLAMA_FILE_VERSION_GGML) {
fprintf(stderr, "llama.cpp: WARNING: input is an old file that doesn't have scores; will add dummy scores\n"); LLAMA_LOG_WARN("llama.cpp: WARNING: input is an old file that doesn't have scores; will add dummy scores\n");
} }
uint32_t n_vocab = any_file_loader->hparams.n_vocab; uint32_t n_vocab = any_file_loader->hparams.n_vocab;
for (uint32_t i = 0; i < n_vocab; i++) { for (uint32_t i = 0; i < n_vocab; i++) {
@ -773,12 +788,12 @@ struct llama_model_loader {
void load_all_data(llama_progress_callback progress_callback, void * progress_callback_user_data, llama_mlock * lmlock) { void load_all_data(llama_progress_callback progress_callback, void * progress_callback_user_data, llama_mlock * lmlock) {
size_t data_size = 0; size_t data_size = 0;
size_t prefetch_size = 0; size_t prefetch_size = file_loader->file.size;
size_t lock_size = 0; size_t lock_size = 0;
for (const llama_load_tensor & lt : tensors_map.tensors) { for (const llama_load_tensor & lt : tensors_map.tensors) {
data_size += lt.size; data_size += lt.size;
if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) { if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) {
prefetch_size += lt.size; prefetch_size -= lt.size;
} }
} }
@ -857,7 +872,7 @@ struct llama_model_loader {
uint8_t byte = lt.data[i]; uint8_t byte = lt.data[i];
sum = byte + (sum << 6) + (sum << 16) - sum; // sdbm hash sum = byte + (sum << 6) + (sum << 16) - sum; // sdbm hash
} }
fprintf(stderr, "%s checksum: %#08x (%s, size %zu)\n", lt.name.c_str(), sum, LLAMA_LOG_INFO("%s checksum: %#08x (%s, size %zu)\n", lt.name.c_str(), sum,
llama_format_tensor_shape(lt.ne).c_str(), lt.size); llama_format_tensor_shape(lt.ne).c_str(), lt.size);
} }
@ -890,7 +905,7 @@ static bool kv_cache_init(
cache.ctx = ggml_init(params); cache.ctx = ggml_init(params);
if (!cache.ctx) { if (!cache.ctx) {
fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__); LLAMA_LOG_ERROR("%s: failed to allocate memory for kv cache\n", __func__);
return false; return false;
} }
@ -1102,7 +1117,7 @@ static void llama_model_load_internal(
LLAMA_ASSERT(hparams.n_head % n_gqa == 0); LLAMA_ASSERT(hparams.n_head % n_gqa == 0);
hparams.n_head_kv = hparams.n_head / n_gqa; hparams.n_head_kv = hparams.n_head / n_gqa;
if (model.type == e_model::MODEL_65B && n_gqa == 8) { if (model.type == e_model::MODEL_65B && n_gqa == 8) {
fprintf(stderr, "%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa); LLAMA_LOG_WARN("%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa);
model.type = e_model::MODEL_70B; model.type = e_model::MODEL_70B;
hparams.f_ffn_mult = 1.3f; // from the params.json of the 70B model hparams.f_ffn_mult = 1.3f; // from the params.json of the 70B model
} }
@ -1118,22 +1133,22 @@ static void llama_model_load_internal(
//const uint32_t n_ff = 28672; //const uint32_t n_ff = 28672;
{ {
fprintf(stderr, "%s: format = %s\n", __func__, llama_file_version_name(file_version)); LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(file_version));
fprintf(stderr, "%s: n_vocab = %u\n", __func__, hparams.n_vocab); LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab);
fprintf(stderr, "%s: n_ctx = %u\n", __func__, hparams.n_ctx); LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, hparams.n_ctx);
fprintf(stderr, "%s: n_embd = %u\n", __func__, hparams.n_embd); LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd);
fprintf(stderr, "%s: n_mult = %u\n", __func__, hparams.n_mult); LLAMA_LOG_INFO("%s: n_mult = %u\n", __func__, hparams.n_mult);
fprintf(stderr, "%s: n_head = %u\n", __func__, hparams.n_head); LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head);
fprintf(stderr, "%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv);
fprintf(stderr, "%s: n_layer = %u\n", __func__, hparams.n_layer); LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer);
fprintf(stderr, "%s: n_rot = %u\n", __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
fprintf(stderr, "%s: n_gqa = %u\n", __func__, hparams.n_gqa()); LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa());
fprintf(stderr, "%s: rnorm_eps = %.1e\n", __func__, hparams.f_rms_norm_eps); LLAMA_LOG_INFO("%s: rnorm_eps = %.1e\n", __func__, hparams.f_rms_norm_eps);
fprintf(stderr, "%s: n_ff = %u\n", __func__, n_ff); LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, n_ff);
fprintf(stderr, "%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base); LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base);
fprintf(stderr, "%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale); LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale);
fprintf(stderr, "%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype)); LLAMA_LOG_INFO("%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model size = %s\n", __func__, llama_model_type_name(model.type));
} }
if (file_version < LLAMA_FILE_VERSION_GGJT_V2) { if (file_version < LLAMA_FILE_VERSION_GGJT_V2) {
@ -1161,7 +1176,7 @@ static void llama_model_load_internal(
size_t ctx_size; size_t ctx_size;
size_t mmapped_size; size_t mmapped_size;
ml->calc_sizes(&ctx_size, &mmapped_size); ml->calc_sizes(&ctx_size, &mmapped_size);
fprintf(stderr, "%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0); LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0);
// create the ggml context // create the ggml context
{ {
@ -1186,13 +1201,13 @@ static void llama_model_load_internal(
(void) main_gpu; (void) main_gpu;
(void) mul_mat_q; (void) mul_mat_q;
#if defined(GGML_USE_CUBLAS) #if defined(GGML_USE_CUBLAS)
fprintf(stderr, "%s: using CUDA for GPU acceleration\n", __func__); LLAMA_LOG_INFO("%s: using CUDA for GPU acceleration\n", __func__);
ggml_cuda_set_main_device(main_gpu); ggml_cuda_set_main_device(main_gpu);
ggml_cuda_set_mul_mat_q(mul_mat_q); ggml_cuda_set_mul_mat_q(mul_mat_q);
#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU #define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU
#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT
#elif defined(GGML_USE_CLBLAST) #elif defined(GGML_USE_CLBLAST)
fprintf(stderr, "%s: using OpenCL for GPU acceleration\n", __func__); LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__);
#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU #define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU
#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU
#else #else
@ -1297,14 +1312,14 @@ static void llama_model_load_internal(
const size_t mem_required_state = const size_t mem_required_state =
scale*hparams.kv_size(); scale*hparams.kv_size();
fprintf(stderr, "%s: mem required = %7.2f MB (+ %7.2f MB per state)\n", __func__, LLAMA_LOG_INFO("%s: mem required = %7.2f MB (+ %7.2f MB per state)\n", __func__,
mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0); mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
(void) vram_scratch; (void) vram_scratch;
(void) n_batch; (void) n_batch;
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
if (low_vram) { if (low_vram) {
fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__); LLAMA_LOG_INFO("%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
ggml_cuda_set_scratch_size(0); // disable scratch ggml_cuda_set_scratch_size(0); // disable scratch
} else { } else {
const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type); const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type);
@ -1312,7 +1327,7 @@ static void llama_model_load_internal(
vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context); vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context);
ggml_cuda_set_scratch_size(vram_scratch); ggml_cuda_set_scratch_size(vram_scratch);
if (n_gpu_layers > 0) { if (n_gpu_layers > 0) {
fprintf(stderr, "%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n", LLAMA_LOG_INFO("%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n",
__func__, vram_scratch_base / kB, vram_scratch_per_context, __func__, vram_scratch_base / kB, vram_scratch_per_context,
(vram_scratch + MB - 1) / MB); // round up (vram_scratch + MB - 1) / MB); // round up
} }
@ -1322,9 +1337,9 @@ static void llama_model_load_internal(
#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu); LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
if (n_gpu_layers > (int) hparams.n_layer) { if (n_gpu_layers > (int) hparams.n_layer) {
fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__); LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__);
} }
size_t vram_kv_cache = 0; size_t vram_kv_cache = 0;
@ -1333,17 +1348,17 @@ static void llama_model_load_internal(
const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3; const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
if (n_gpu_layers > (int) hparams.n_layer + 1) { if (n_gpu_layers > (int) hparams.n_layer + 1) {
if (low_vram) { if (low_vram) {
fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__); LLAMA_LOG_INFO("%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
} else { } else {
fprintf(stderr, "%s: offloading v cache to GPU\n", __func__); LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__);
vram_kv_cache += hparams.kv_size() / 2; vram_kv_cache += hparams.kv_size() / 2;
} }
} }
if (n_gpu_layers > (int) hparams.n_layer + 2) { if (n_gpu_layers > (int) hparams.n_layer + 2) {
if (low_vram) { if (low_vram) {
fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__); LLAMA_LOG_WARN("%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
} else { } else {
fprintf(stderr, "%s: offloading k cache to GPU\n", __func__); LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__);
vram_kv_cache += hparams.kv_size() / 2; vram_kv_cache += hparams.kv_size() / 2;
} }
} }
@ -1352,9 +1367,9 @@ static void llama_model_load_internal(
const int max_offloadable_layers = hparams.n_layer + 1; const int max_offloadable_layers = hparams.n_layer + 1;
#endif // GGML_USE_CUBLAS #endif // GGML_USE_CUBLAS
fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n", LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n",
__func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
fprintf(stderr, "%s: total VRAM used: %zu MB\n", LLAMA_LOG_INFO("%s: total VRAM used: %zu MB\n",
__func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
#else #else
(void) n_gpu_layers; (void) n_gpu_layers;
@ -1413,7 +1428,7 @@ static bool llama_model_load(
use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data); use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
return true; return true;
} catch (const std::exception & err) { } catch (const std::exception & err) {
fprintf(stderr, "error loading model: %s\n", err.what()); LLAMA_LOG_ERROR("error loading model: %s\n", err.what());
return false; return false;
} }
} }
@ -1777,7 +1792,7 @@ static struct ggml_cgraph * llama_build_graph(
} }
#if 0 #if 0
printf("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__, LLAMA_LOG_INFO("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__,
ggml_used_mem(ctx0)/1024.0/1024.0, ggml_used_mem(ctx0)/1024.0/1024.0,
lctx.get_buf_max_mem(0)/1024.0/1024.0, lctx.get_buf_max_mem(0)/1024.0/1024.0,
lctx.get_buf_max_mem(1)/1024.0/1024.0, lctx.get_buf_max_mem(1)/1024.0/1024.0,
@ -1838,7 +1853,7 @@ static bool llama_eval_internal(
ggml_allocr_alloc_graph(lctx.alloc, gf); ggml_allocr_alloc_graph(lctx.alloc, gf);
#endif #endif
// fprintf(stderr, "graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
// for big prompts, if BLAS is enabled, it is better to use only one thread // for big prompts, if BLAS is enabled, it is better to use only one thread
// otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
@ -2025,7 +2040,7 @@ struct llama_tokenizer {
left_sym.n += right_sym.n; left_sym.n += right_sym.n;
right_sym.n = 0; right_sym.n = 0;
//printf("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size); //LLAMA_LOG_INFO("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size);
// remove the right sym from the chain // remove the right sym from the chain
left_sym.next = right_sym.next; left_sym.next = right_sym.next;
@ -3033,7 +3048,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
tensor.data = read_data.addr; tensor.data = read_data.addr;
model_loader->load_data_for(tensor); model_loader->load_data_for(tensor);
printf("[%4zu/%4zu] %36s - %16s, type = %6s, ", LLAMA_LOG_INFO("[%4zu/%4zu] %36s - %16s, type = %6s, ",
++idx, model_loader->tensors_map.tensors.size(), ++idx, model_loader->tensors_map.tensors.size(),
tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(), tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(),
ggml_type_name(tensor.type)); ggml_type_name(tensor.type));
@ -3055,7 +3070,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
new_type = tensor.type; new_type = tensor.type;
new_data = tensor.data; new_data = tensor.data;
new_size = tensor.size; new_size = tensor.size;
printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0); LLAMA_LOG_INFO("size = %8.3f MB\n", tensor.size/1024.0/1024.0);
} else { } else {
new_type = quantized_type; new_type = quantized_type;
#ifdef GGML_USE_K_QUANTS #ifdef GGML_USE_K_QUANTS
@ -3090,17 +3105,17 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
int nx = tensor.ne.at(0); int nx = tensor.ne.at(0);
int ny = tensor.ne.at(1); int ny = tensor.ne.at(1);
if (nx % QK_K != 0 || ny % QK_K != 0) { if (nx % QK_K != 0 || ny % QK_K != 0) {
fprintf(stderr, "\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K); LLAMA_LOG_INFO("\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
convert_incompatible_tensor = true; convert_incompatible_tensor = true;
} }
} }
if (convert_incompatible_tensor) { if (convert_incompatible_tensor) {
if (tensor.name == "output.weight") { if (tensor.name == "output.weight") {
new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing. new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
fprintf(stderr, "F16 will be used for this tensor instead.\n"); LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
} else if (tensor.name == "tok_embeddings.weight") { } else if (tensor.name == "tok_embeddings.weight") {
new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing. new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
fprintf(stderr, "Q4_0 will be used for this tensor instead.\n"); LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
} else { } else {
throw std::runtime_error("Unsupported tensor size encountered\n"); throw std::runtime_error("Unsupported tensor size encountered\n");
} }
@ -3120,7 +3135,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
f32_data = (float *) f32_conv_buf.addr; f32_data = (float *) f32_conv_buf.addr;
} }
printf("quantizing to %s .. ", ggml_type_name(new_type)); LLAMA_LOG_INFO("quantizing to %s .. ", ggml_type_name(new_type));
fflush(stdout); fflush(stdout);
work.resize(nelements * 4); // upper bound on size work.resize(nelements * 4); // upper bound on size
@ -3170,7 +3185,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
} }
} }
printf("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0); LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0);
int64_t tot_count = 0; int64_t tot_count = 0;
for (size_t i = 0; i < hist_cur.size(); i++) { for (size_t i = 0; i < hist_cur.size(); i++) {
hist_all[i] += hist_cur[i]; hist_all[i] += hist_cur[i];
@ -3179,18 +3194,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (tot_count > 0) { if (tot_count > 0) {
for (size_t i = 0; i < hist_cur.size(); i++) { for (size_t i = 0; i < hist_cur.size(); i++) {
printf("%5.3f ", hist_cur[i] / float(nelements)); LLAMA_LOG_INFO("%5.3f ", hist_cur[i] / float(nelements));
} }
} }
printf("\n"); LLAMA_LOG_INFO("\n");
} }
total_size_org += tensor.size; total_size_org += tensor.size;
total_size_new += new_size; total_size_new += new_size;
file_saver.write_tensor(tensor, new_type, new_data, new_size); file_saver.write_tensor(tensor, new_type, new_data, new_size);
} }
printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
printf("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); LLAMA_LOG_INFO("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
{ {
int64_t sum_all = 0; int64_t sum_all = 0;
@ -3199,11 +3214,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
} }
if (sum_all > 0) { if (sum_all > 0) {
printf("%s: hist: ", __func__); LLAMA_LOG_INFO("%s: hist: ", __func__);
for (size_t i = 0; i < hist_all.size(); i++) { for (size_t i = 0; i < hist_all.size(); i++) {
printf("%5.3f ", hist_all[i] / float(sum_all)); LLAMA_LOG_INFO("%5.3f ", hist_all[i] / float(sum_all));
} }
printf("\n"); LLAMA_LOG_INFO("\n");
} }
} }
} }
@ -3227,8 +3242,8 @@ struct llama_model * llama_load_model_from_file(
params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale,params.low_vram, params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale,params.low_vram,
memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback, memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback,
params.progress_callback_user_data)) { params.progress_callback_user_data)) {
LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
delete model; delete model;
fprintf(stderr, "%s: failed to load model\n", __func__);
return nullptr; return nullptr;
} }
@ -3261,10 +3276,9 @@ struct llama_context * llama_new_context_with_model(
unsigned percentage = (unsigned) (100 * progress); unsigned percentage = (unsigned) (100 * progress);
while (percentage > *cur_percentage_p) { while (percentage > *cur_percentage_p) {
*cur_percentage_p = percentage; *cur_percentage_p = percentage;
fprintf(stderr, "."); LLAMA_LOG_INFO(".");
fflush(stderr);
if (percentage >= 100) { if (percentage >= 100) {
fprintf(stderr, "\n"); LLAMA_LOG_INFO("\n");
} }
} }
}; };
@ -3278,14 +3292,14 @@ struct llama_context * llama_new_context_with_model(
// reserve memory for context buffers // reserve memory for context buffers
if (!params.vocab_only) { if (!params.vocab_only) {
if (!kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) { if (!kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__); LLAMA_LOG_ERROR("%s: kv_cache_init() failed for self-attention cache\n", __func__);
llama_free(ctx); llama_free(ctx);
return nullptr; return nullptr;
} }
{ {
const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v); const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v);
fprintf(stderr, "%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0); LLAMA_LOG_INFO("%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
} }
const auto & hparams = ctx->model.hparams; const auto & hparams = ctx->model.hparams;
@ -3319,14 +3333,14 @@ struct llama_context * llama_new_context_with_model(
// measure memory requirements for the graph // measure memory requirements for the graph
size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment; size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment;
fprintf(stderr, "%s: compute buffer total size = %7.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0); LLAMA_LOG_INFO("%s: compute buffer total size = %7.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0);
// debug - for comparison with scratch buffer // debug - for comparison with scratch buffer
//size_t prev_req = //size_t prev_req =
// MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type) + // MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type) +
// MEM_REQ_SCRATCH1().at(ctx->model.type) + // MEM_REQ_SCRATCH1().at(ctx->model.type) +
// MEM_REQ_EVAL().at(ctx->model.type); // MEM_REQ_EVAL().at(ctx->model.type);
//fprintf(stderr, "%s: (debug) equivalent with scratch buffer = %7.2f MB\n", __func__, prev_req / 1024.0 / 1024.0); //LLAMA_LOG_INFO("%s: (debug) equivalent with scratch buffer = %7.2f MB\n", __func__, prev_req / 1024.0 / 1024.0);
// recreate allocator with exact memory requirements // recreate allocator with exact memory requirements
ggml_allocr_free(ctx->alloc); ggml_allocr_free(ctx->alloc);
@ -3362,11 +3376,11 @@ struct llama_context * llama_new_context_with_model(
const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx); const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
fprintf(stderr, "%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0); LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
#define LLAMA_METAL_CHECK_BUF(result) \ #define LLAMA_METAL_CHECK_BUF(result) \
if (!(result)) { \ if (!(result)) { \
fprintf(stderr, "%s: failed to add buffer\n", __func__); \ LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
llama_free(ctx); \ llama_free(ctx); \
return NULL; \ return NULL; \
} }
@ -3422,19 +3436,19 @@ int llama_model_quantize(
llama_model_quantize_internal(fname_inp, fname_out, params); llama_model_quantize_internal(fname_inp, fname_out, params);
return 0; return 0;
} catch (const std::exception & err) { } catch (const std::exception & err) {
fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.what()); LLAMA_LOG_ERROR("%s: failed to quantize: %s\n", __func__, err.what());
return 1; return 1;
} }
} }
int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) { int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) {
fprintf(stderr, "%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora); LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
const int64_t t_start_lora_us = ggml_time_us(); const int64_t t_start_lora_us = ggml_time_us();
auto fin = std::ifstream(path_lora, std::ios::binary); auto fin = std::ifstream(path_lora, std::ios::binary);
if (!fin) { if (!fin) {
fprintf(stderr, "%s: failed to open '%s'\n", __func__, path_lora); LLAMA_LOG_ERROR("%s: failed to open '%s'\n", __func__, path_lora);
return 1; return 1;
} }
@ -3443,14 +3457,14 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
uint32_t magic; uint32_t magic;
fin.read((char *) &magic, sizeof(magic)); fin.read((char *) &magic, sizeof(magic));
if (magic != LLAMA_FILE_MAGIC_GGLA) { if (magic != LLAMA_FILE_MAGIC_GGLA) {
fprintf(stderr, "%s: bad file magic\n", __func__); LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
return 1; return 1;
} }
uint32_t format_version; uint32_t format_version;
fin.read((char *) &format_version, sizeof(format_version)); fin.read((char *) &format_version, sizeof(format_version));
if (format_version != 1) { if (format_version != 1) {
fprintf(stderr, "%s: unsupported file version\n", __func__ ); LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
return 1; return 1;
} }
} }
@ -3461,7 +3475,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
fin.read((char *) &lora_alpha, sizeof(lora_alpha)); fin.read((char *) &lora_alpha, sizeof(lora_alpha));
float scaling = (float)lora_alpha / (float)lora_r; float scaling = (float)lora_alpha / (float)lora_r;
fprintf(stderr, "%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling); LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
// create a temporary ggml context to store the lora tensors // create a temporary ggml context to store the lora tensors
@ -3487,7 +3501,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
ggml_context * base_ctx = NULL; ggml_context * base_ctx = NULL;
llama_buffer base_buf; llama_buffer base_buf;
if (path_base_model) { if (path_base_model) {
fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model); LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true)); model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true));
size_t ctx_size; size_t ctx_size;
@ -3544,17 +3558,17 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
const std::string lora_suffix = ".lora"; const std::string lora_suffix = ".lora";
size_t pos = name.rfind(lora_suffix); size_t pos = name.rfind(lora_suffix);
if (pos == std::string::npos) { if (pos == std::string::npos) {
fprintf(stderr, "%s: error: '%s' is not a lora tensor\n", __func__, name.c_str()); LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
return 1; return 1;
} }
std::string lora_type = name.substr(pos + lora_suffix.length()); std::string lora_type = name.substr(pos + lora_suffix.length());
std::string base_name = name; std::string base_name = name;
base_name.erase(pos); base_name.erase(pos);
// fprintf(stderr, "%s: %s => %s (lora type %s) ", __func__, name.c_str(),base_name.c_str(), lora_type.c_str()); // LLAMA_LOG_INFO("%s: %s => %s (lora type %s) \n", __func__, name.c_str(),base_name.c_str(), lora_type.c_str());
if (model_tensors.find(base_name) == model_tensors.end()) { if (model_tensors.find(base_name) == model_tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in lora adapter\n", __func__, name.data()); LLAMA_LOG_ERROR("%s: unknown tensor '%s' in lora adapter\n", __func__, name.data());
return 1; return 1;
} }
@ -3565,7 +3579,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
case 1: wtype = GGML_TYPE_F16; break; case 1: wtype = GGML_TYPE_F16; break;
default: default:
{ {
fprintf(stderr, "%s: invalid tensor data type '%d'\n", LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
__func__, ftype); __func__, ftype);
return false; return false;
} }
@ -3575,7 +3589,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]); lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]);
} }
else { else {
fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims); LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
return 1; return 1;
} }
ggml_set_name(lora_tensor, "lora_tensor"); ggml_set_name(lora_tensor, "lora_tensor");
@ -3613,7 +3627,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
if (model_loader) { if (model_loader) {
// load from base model // load from base model
if (model_loader->tensors_map.name_to_idx.find(base_name) == model_loader->tensors_map.name_to_idx.end()) { if (model_loader->tensors_map.name_to_idx.find(base_name) == model_loader->tensors_map.name_to_idx.end()) {
fprintf(stderr, "%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str()); LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
return 1; return 1;
} }
size_t idx = model_loader->tensors_map.name_to_idx[base_name]; size_t idx = model_loader->tensors_map.name_to_idx[base_name];
@ -3629,7 +3643,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
if (ggml_is_quantized(base_t->type)) { if (ggml_is_quantized(base_t->type)) {
if (!warned) { if (!warned) {
fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, " LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
"use a f16 or f32 base model with --lora-base\n", __func__); "use a f16 or f32 base model with --lora-base\n", __func__);
warned = true; warned = true;
} }
@ -3644,7 +3658,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
ggml_set_name(loraB, "loraB"); ggml_set_name(loraB, "loraB");
if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) { if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");" LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
" are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]); " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
return 1; return 1;
} }
@ -3690,7 +3704,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
n_tensors++; n_tensors++;
if (n_tensors % 4 == 0) { if (n_tensors % 4 == 0) {
fprintf(stderr, "."); LLAMA_LOG_INFO(".");
} }
} }
} }
@ -3702,7 +3716,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
} }
const int64_t t_lora_us = ggml_time_us() - t_start_lora_us; const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
fprintf(stderr, " done (%.2f ms)\n", t_lora_us / 1000.0); LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
return 0; return 0;
} }
@ -3711,7 +3725,7 @@ int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lor
try { try {
return llama_apply_lora_from_file_internal(ctx->model, path_lora, path_base_model, n_threads); return llama_apply_lora_from_file_internal(ctx->model, path_lora, path_base_model, n_threads);
} catch (const std::exception & err) { } catch (const std::exception & err) {
fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what()); LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
return 1; return 1;
} }
} }
@ -3720,7 +3734,7 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha
try { try {
return llama_apply_lora_from_file_internal(*model, path_lora, path_base_model, n_threads); return llama_apply_lora_from_file_internal(*model, path_lora, path_base_model, n_threads);
} catch (const std::exception & err) { } catch (const std::exception & err) {
fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what()); LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
return 1; return 1;
} }
} }
@ -3769,10 +3783,20 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
return s_total; return s_total;
} }
// Copies the state to the specified destination address /** copy state data into either a buffer or file depending on the passed in context
size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { *
uint8_t * out = dst; * file context:
* llama_file file("/path", "wb");
* llama_data_file_context data_ctx(&file);
* llama_copy_state_data(ctx, &data_ctx);
*
* buffer context:
* std::vector<uint8_t> buf(max_size, 0);
* llama_data_buffer_context data_ctx(&buf.data());
* llama_copy_state_data(ctx, &data_ctx);
*
*/
void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) {
// copy rng // copy rng
{ {
std::stringstream rng_ss; std::stringstream rng_ss;
@ -3784,8 +3808,8 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
memset(&rng_buf[0], 0, LLAMA_MAX_RNG_STATE); memset(&rng_buf[0], 0, LLAMA_MAX_RNG_STATE);
memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size()); memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());
memcpy(out, &rng_size, sizeof(rng_size)); out += sizeof(rng_size); data_ctx->write(&rng_size, sizeof(rng_size));
memcpy(out, &rng_buf[0], LLAMA_MAX_RNG_STATE); out += LLAMA_MAX_RNG_STATE; data_ctx->write(&rng_buf[0], LLAMA_MAX_RNG_STATE);
} }
// copy logits // copy logits
@ -3793,25 +3817,29 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
const size_t logits_cap = ctx->logits.capacity(); const size_t logits_cap = ctx->logits.capacity();
const size_t logits_size = ctx->logits.size(); const size_t logits_size = ctx->logits.size();
memcpy(out, &logits_cap, sizeof(logits_cap)); out += sizeof(logits_cap); data_ctx->write(&logits_cap, sizeof(logits_cap));
memcpy(out, &logits_size, sizeof(logits_size)); out += sizeof(logits_size); data_ctx->write(&logits_size, sizeof(logits_size));
if (logits_size) { if (logits_size) {
memcpy(out, ctx->logits.data(), logits_size * sizeof(float)); data_ctx->write(ctx->logits.data(), logits_size * sizeof(float));
} }
out += logits_cap * sizeof(float); // If there is a gap between the size and the capacity, write padding
size_t padding_size = (logits_cap - logits_size) * sizeof(float);
if (padding_size > 0) {
std::vector<uint8_t> padding(padding_size, 0); // Create a buffer filled with zeros
data_ctx->write(padding.data(), padding_size);
}
} }
// copy embeddings // copy embeddings
{ {
const size_t embedding_size = ctx->embedding.size(); const size_t embedding_size = ctx->embedding.size();
memcpy(out, &embedding_size, sizeof(embedding_size)); out += sizeof(embedding_size); data_ctx->write(&embedding_size, sizeof(embedding_size));
if (embedding_size) { if (embedding_size) {
memcpy(out, ctx->embedding.data(), embedding_size * sizeof(float)); data_ctx->write(ctx->embedding.data(), embedding_size * sizeof(float));
out += embedding_size * sizeof(float);
} }
} }
@ -3826,8 +3854,8 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
const size_t kv_size = kv_self.buf.size; const size_t kv_size = kv_self.buf.size;
const int kv_ntok = llama_get_kv_cache_token_count(ctx); const int kv_ntok = llama_get_kv_cache_token_count(ctx);
memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size); data_ctx->write(&kv_size, sizeof(kv_size));
memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok); data_ctx->write(&kv_ntok, sizeof(kv_ntok));
if (kv_size) { if (kv_size) {
const size_t elt_size = ggml_element_size(kv_self.k); const size_t elt_size = ggml_element_size(kv_self.k);
@ -3836,12 +3864,12 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
ggml_cgraph gf{}; ggml_cgraph gf{};
ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer); ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer);
kout3d->data = out; std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
out += ggml_nbytes(kout3d); kout3d->data = kout3d_data.data();
ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer); ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer);
vout3d->data = out; std::vector<uint8_t> vout3d_data(ggml_nbytes(vout3d), 0);
out += ggml_nbytes(vout3d); vout3d->data = vout3d_data.data();
ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k,
n_embd, kv_ntok, n_layer, n_embd, kv_ntok, n_layer,
@ -3856,15 +3884,20 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
ggml_free(cpy_ctx); ggml_free(cpy_ctx);
// our data is now in the kout3d_data and vout3d_data buffers
// write them to file
data_ctx->write(kout3d_data.data(), kout3d_data.size());
data_ctx->write(vout3d_data.data(), vout3d_data.size());
}
} }
} }
const size_t written = out - dst; size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
const size_t max_size = llama_get_state_size(ctx); llama_data_buffer_context data_ctx(dst);
llama_copy_state_data_internal(ctx, &data_ctx);
LLAMA_ASSERT(written <= max_size); return data_ctx.get_size_written();
return written;
} }
// Sets the state reading from the specified source address // Sets the state reading from the specified source address
@ -3983,7 +4016,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
const uint32_t version = file.read_u32(); const uint32_t version = file.read_u32();
if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) { if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) {
fprintf(stderr, "%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version); LLAMA_LOG_ERROR("%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
return false; return false;
} }
@ -3991,7 +4024,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
file.read_raw(&session_hparams, sizeof(llama_hparams)); file.read_raw(&session_hparams, sizeof(llama_hparams));
if (session_hparams != ctx->model.hparams) { if (session_hparams != ctx->model.hparams) {
fprintf(stderr, "%s : model hparams didn't match from session file!\n", __func__); LLAMA_LOG_INFO("%s : model hparams didn't match from session file!\n", __func__);
return false; return false;
} }
} }
@ -4001,7 +4034,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
const uint32_t n_token_count = file.read_u32(); const uint32_t n_token_count = file.read_u32();
if (n_token_count > n_token_capacity) { if (n_token_count > n_token_capacity) {
fprintf(stderr, "%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity); LLAMA_LOG_ERROR("%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
return false; return false;
} }
@ -4015,7 +4048,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
const size_t n_state_size_max = llama_get_state_size(ctx); const size_t n_state_size_max = llama_get_state_size(ctx);
if (n_state_size_cur > n_state_size_max) { if (n_state_size_cur > n_state_size_max) {
fprintf(stderr, "%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur); LLAMA_LOG_ERROR("%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur);
return false; return false;
} }
@ -4032,7 +4065,7 @@ bool llama_load_session_file(struct llama_context * ctx, const char * path_sessi
try { try {
return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
} catch (const std::exception & err) { } catch (const std::exception & err) {
fprintf(stderr, "error loading session file: %s\n", err.what()); LLAMA_LOG_ERROR("error loading session file: %s\n", err.what());
return false; return false;
} }
} }
@ -4049,15 +4082,9 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi
file.write_u32((uint32_t) n_token_count); file.write_u32((uint32_t) n_token_count);
file.write_raw(tokens, sizeof(llama_token) * n_token_count); file.write_raw(tokens, sizeof(llama_token) * n_token_count);
// save the context state // save the context state using stream saving
{ llama_data_file_context data_ctx(&file);
const size_t n_state_size_max = llama_get_state_size(ctx); llama_copy_state_data_internal(ctx, &data_ctx);
std::vector<uint8_t> state_data(n_state_size_max);
const size_t n_state_size_cur = llama_copy_state_data(ctx, state_data.data());
file.write_raw(state_data.data(), n_state_size_cur);
}
return true; return true;
} }
@ -4069,7 +4096,7 @@ int llama_eval(
int n_past, int n_past,
int n_threads) { int n_threads) {
if (!llama_eval_internal(*ctx, tokens, nullptr, n_tokens, n_past, n_threads, nullptr)) { if (!llama_eval_internal(*ctx, tokens, nullptr, n_tokens, n_past, n_threads, nullptr)) {
fprintf(stderr, "%s: failed to eval\n", __func__); LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
return 1; return 1;
} }
@ -4091,7 +4118,7 @@ int llama_eval_embd(
int n_past, int n_past,
int n_threads) { int n_threads) {
if (!llama_eval_internal(*ctx, nullptr, embd, n_tokens, n_past, n_threads, nullptr)) { if (!llama_eval_internal(*ctx, nullptr, embd, n_tokens, n_past, n_threads, nullptr)) {
fprintf(stderr, "%s: failed to eval\n", __func__); LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
return 1; return 1;
} }
@ -4112,7 +4139,7 @@ int llama_eval_export(struct llama_context * ctx, const char * fname) {
const std::vector<llama_token> tmp(n_batch, llama_token_bos()); const std::vector<llama_token> tmp(n_batch, llama_token_bos());
if (!llama_eval_internal(*ctx, tmp.data(), nullptr, tmp.size(), n_ctx, 1, fname)) { if (!llama_eval_internal(*ctx, tmp.data(), nullptr, tmp.size(), n_ctx, 1, fname)) {
fprintf(stderr, "%s: failed to eval\n", __func__); LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
return 1; return 1;
} }
@ -4128,7 +4155,7 @@ int llama_tokenize_with_model(
auto res = llama_tokenize(model->vocab, text, add_bos); auto res = llama_tokenize(model->vocab, text, add_bos);
if (n_max_tokens < (int) res.size()) { if (n_max_tokens < (int) res.size()) {
fprintf(stderr, "%s: too many tokens\n", __func__); LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
return -((int) res.size()); return -((int) res.size());
} }
@ -4245,15 +4272,15 @@ struct llama_timings llama_get_timings(struct llama_context * ctx) {
void llama_print_timings(struct llama_context * ctx) { void llama_print_timings(struct llama_context * ctx) {
const llama_timings timings = llama_get_timings(ctx); const llama_timings timings = llama_get_timings(ctx);
fprintf(stderr, "\n"); LLAMA_LOG_INFO("\n");
fprintf(stderr, "%s: load time = %8.2f ms\n", __func__, timings.t_load_ms); LLAMA_LOG_INFO("%s: load time = %8.2f ms\n", __func__, timings.t_load_ms);
fprintf(stderr, "%s: sample time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", LLAMA_LOG_INFO("%s: sample time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample); __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample);
fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", LLAMA_LOG_INFO("%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval); __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
fprintf(stderr, "%s: eval time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", LLAMA_LOG_INFO("%s: eval time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval); __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
fprintf(stderr, "%s: total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms)); LLAMA_LOG_INFO("%s: total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
} }
void llama_reset_timings(struct llama_context * ctx) { void llama_reset_timings(struct llama_context * ctx) {
@ -4289,3 +4316,44 @@ const char * llama_print_system_info(void) {
const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) { const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
return ctx->model.tensors_by_name; return ctx->model.tensors_by_name;
} }
void llama_log_set(llama_log_callback log_callback, void * user_data) {
g_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
g_state.log_callback_user_data = user_data;
}
#if defined(_MSC_VER) && !defined(vsnprintf)
#define vsnprintf _vsnprintf
#endif
static void llama_log_internal_v(llama_log_level level, const char * format, va_list args) {
va_list args_copy;
va_copy(args_copy, args);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
g_state.log_callback(level, buffer, g_state.log_callback_user_data);
} else {
char* buffer2 = new char[len+1];
vsnprintf(buffer2, len+1, format, args_copy);
buffer2[len] = 0;
g_state.log_callback(level, buffer2, g_state.log_callback_user_data);
delete[] buffer2;
}
va_end(args_copy);
}
static void llama_log_internal(llama_log_level level, const char * format, ...) {
va_list args;
va_start(args, format);
llama_log_internal_v(level, format, args);
va_end(args);
}
static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
fputs(text, stderr);
fflush(stderr);
}

5
llm/llama.go
View file

@ -1,8 +1,9 @@
package llm package llm
/* /*
#cgo CPPFLAGS: -O3 -Wall -Wextra -Wno-unused-function -Wno-unused-variable -DNDEBUG -DGGML_USE_K_QUANTS #cgo CFLAGS: -Ofast -std=c11 -fPIC
#cgo CXXFLAGS: -std=gnu++11 #cgo CPPFLAGS: -Ofast -Wall -Wextra -Wno-unused-function -Wno-unused-variable -DNDEBUG -DGGML_USE_K_QUANTS
#cgo CXXFLAGS: -std=c++11 -fPIC
#cgo darwin CPPFLAGS: -DGGML_USE_ACCELERATE #cgo darwin CPPFLAGS: -DGGML_USE_ACCELERATE
#cgo darwin,arm64 CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG #cgo darwin,arm64 CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG
#cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders #cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders

19
llm/llama.h
View file

@ -1,5 +1,5 @@
/** /**
* llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12 * llama.cpp - git f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e
* *
* MIT License * MIT License
* *
@ -112,6 +112,19 @@ extern "C" {
typedef void (*llama_progress_callback)(float progress, void *ctx); typedef void (*llama_progress_callback)(float progress, void *ctx);
enum llama_log_level {
LLAMA_LOG_LEVEL_ERROR = 2,
LLAMA_LOG_LEVEL_WARN = 3,
LLAMA_LOG_LEVEL_INFO = 4
};
// Signature for logging events
// Note that text includes the new line character at the end for most events.
// If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
// if it exists.
// It might not exist for progress report where '.' is output repeatedly.
typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data);
struct llama_context_params { struct llama_context_params {
uint32_t seed; // RNG seed, -1 for random uint32_t seed; // RNG seed, -1 for random
int32_t n_ctx; // text context int32_t n_ctx; // text context
@ -221,6 +234,10 @@ extern "C" {
int32_t n_eval; int32_t n_eval;
}; };
// Set callback for all future logging events.
// If this is not called, or NULL is supplied, everything is output on stderr.
LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data);
LLAMA_API int llama_max_devices(); LLAMA_API int llama_max_devices();
LLAMA_API struct llama_context_params llama_context_default_params(); LLAMA_API struct llama_context_params llama_context_default_params();