Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 17:25:07 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build.yml
#	scripts/build-info.sh
This commit is contained in:
Concedo
2023-07-25 22:55:55 +08:00
parent 0c26799e77 07aaa0f63f
commit 6a054b80b0
12 changed files with 753 additions and 146 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml.c
+61 -27
View File
@@ -4230,6 +4230,15 @@ bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
static inline bool ggml_is_contiguous_except_dim_1(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return
tensor->nb[0] == GGML_TYPE_SIZE[tensor->type] &&
tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
bool ggml_is_permuted(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
@@ -7022,14 +7031,16 @@ struct ggml_tensor * ggml_flash_attn(
}
//struct ggml_tensor * result = ggml_dup_tensor(ctx, q);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, q->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, q->n_dims, q->ne);
int32_t t = masked ? 1 : 0;
ggml_set_op_params(result, &t, sizeof(t));
result->op = GGML_OP_FLASH_ATTN;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = q;
result->src[1] = k;
result->src[2] = v;
result->src[3] = ggml_new_i32(ctx, masked ? 1 : 0);
return result;
}
@@ -7053,7 +7064,7 @@ struct ggml_tensor * ggml_flash_ff(
}
//struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, a->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne);
result->op = GGML_OP_FLASH_FF;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7119,13 +7130,15 @@ struct ggml_tensor * ggml_flash_attn_back(
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
int32_t masked_i = masked ? 1 : 0;
ggml_set_op_params(result, &masked_i, sizeof(masked_i));
result->op = GGML_OP_FLASH_ATTN_BACK;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = q;
result->src[1] = k;
result->src[2] = v;
result->src[3] = d;
result->src[4] = ggml_new_i32(ctx, masked ? 1 : 0);
return result;
}
@@ -9815,8 +9828,8 @@ static void ggml_compute_forward_gelu_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -9874,8 +9887,8 @@ static void ggml_compute_forward_gelu_quick_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -9933,8 +9946,8 @@ static void ggml_compute_forward_silu_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -9993,9 +10006,9 @@ static void ggml_compute_forward_silu_back_f32(
const struct ggml_tensor * src0,
const struct ggml_tensor * grad,
struct ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(grad));
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(grad));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
GGML_ASSERT(ggml_are_same_shape(src0, grad));
@@ -14760,7 +14773,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
} break;
case GGML_OP_FLASH_ATTN:
{
const int32_t t = ggml_get_i32_1d(tensor->src[3], 0);
const int32_t t = ggml_get_op_params_i32(tensor, 0);
GGML_ASSERT(t == 0 || t == 1);
const bool masked = t != 0;
ggml_compute_forward_flash_attn(params, tensor->src[0], tensor->src[1], tensor->src[2], masked, tensor);
@@ -14771,7 +14784,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
} break;
case GGML_OP_FLASH_ATTN_BACK:
{
int32_t t = ggml_get_i32_1d(tensor->src[4], 0);
int32_t t = ggml_get_op_params_i32(tensor, 0);
GGML_ASSERT(t == 0 || t == 1);
bool masked = t != 0;
ggml_compute_forward_flash_attn_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], masked, tensor);
@@ -15389,7 +15402,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
{
struct ggml_tensor * flash_grad = NULL;
if (src0->grad || src1->grad || tensor->src[2]->grad) {
int32_t t = ggml_get_i32_1d(tensor->src[3], 0);
int32_t t = ggml_get_op_params_i32(tensor, 0);
GGML_ASSERT(t == 0 || t == 1);
bool masked = t != 0;
flash_grad =
@@ -15661,6 +15674,34 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
}
}
static_assert(GGML_GRAPH_HASHTABLE_SIZE > GGML_MAX_NODES * 2, "GGML_GRAPH_HT_SIZE is too small");
static size_t hash(void * p) {
return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
}
static bool hash_insert(void * hash_table[], void * p) {
size_t h = hash(p);
// linear probing
size_t i = h;
while (hash_table[i] != NULL && hash_table[i] != p) {
i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
if (i == h) {
// hash table is full
GGML_ASSERT(false);
}
}
if (hash_table[i] == p) {
return true;
}
// insert
hash_table[i] = p;
return false;
}
static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
if (node->grad == NULL) {
// this usually happens when we generate intermediate nodes from constants in the backward pass
@@ -15671,16 +15712,8 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
}
// check if already visited
for (int i = 0; i < cgraph->n_nodes; i++) {
if (cgraph->nodes[i] == node) {
return;
}
}
for (int i = 0; i < cgraph->n_leafs; i++) {
if (cgraph->leafs[i] == node) {
return;
}
if (hash_insert(cgraph->visited_hash_table, node)) {
return;
}
for (int i = 0; i < GGML_MAX_SRC; ++i) {
@@ -15743,6 +15776,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) {
/*.nodes =*/ { NULL },
/*.grads =*/ { NULL },
/*.leafs =*/ { NULL },
/*.hash_table =*/ { NULL },
/*.perf_runs =*/ 0,
/*.perf_cycles =*/ 0,
/*.perf_time_us =*/ 0,
@@ -15784,7 +15818,7 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg
if (node->is_param) {
GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
ggml_build_forward_impl(&result, node->grad, true);
ggml_build_forward_expand(&result, node->grad);
}
}