static struct ggml_state g_state;
static atomic_int g_state_barrier = 0;
+// barrier via spin lock
+inline static void ggml_critical_section_start() {
+ int processing = atomic_fetch_add(&g_state_barrier, 1);
+
+ while (processing > 0) {
+ // wait for other threads to finish
+ atomic_fetch_sub(&g_state_barrier, 1);
+ sched_yield(); // TODO: reconsider this
+ processing = atomic_fetch_add(&g_state_barrier, 1);
+ }
+}
+
+// TODO: make this somehow automatically executed
+// some sort of "sentry" mechanism
+inline static void ggml_critical_section_end() {
+ atomic_fetch_sub(&g_state_barrier, 1);
+}
+
////////////////////////////////////////////////////////////////////////////////
void ggml_print_object(const struct ggml_object * obj) {
struct ggml_context * ggml_init(struct ggml_init_params params) {
// make this function thread safe
- {
- int processing = atomic_fetch_add(&g_state_barrier, 1);
- while (processing > 0) {
- // wait for other threads to finish
- atomic_fetch_sub(&g_state_barrier, 1);
- sched_yield();
- processing = atomic_fetch_add(&g_state_barrier, 1);
- }
- }
+ ggml_critical_section_start();
static bool is_first_call = true;
+
if (is_first_call) {
- const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
-
- ggml_fp16_t ii;
- for (int i = 0; i < (1 << 16); ++i) {
- uint16_t ui = i;
- memcpy(&ii, &ui, sizeof(ii));
- const float f = GGML_FP16_TO_FP32(ii);
- table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
- table_exp_f16[i] = GGML_FP32_TO_FP16(exp(f));
+ // initialize GELU and EXP tables
+ {
+ const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
+
+ ggml_fp16_t ii;
+ for (int i = 0; i < (1 << 16); ++i) {
+ uint16_t ui = i;
+ memcpy(&ii, &ui, sizeof(ii));
+ const float f = GGML_FP16_TO_FP32(ii);
+ table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
+ table_exp_f16[i] = GGML_FP32_TO_FP16(exp(f));
+ }
+
+ const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
+
+ GGML_PRINT_DEBUG("%s: GELU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
}
- const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
+ // initialize g_state
+ {
+ const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
+
+ g_state = (struct ggml_state) {
+ /*.contexts =*/ { 0 },
+ };
+
+ for (int i = 0; i < GGML_MAX_CONTEXTS; ++i) {
+ g_state.contexts[i].used = false;
+ }
+
+ const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
- GGML_PRINT_DEBUG("%s: GELU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
+ GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
+ }
is_first_call = false;
}
// find non-used context in g_state
struct ggml_context * ctx = NULL;
- static bool first_time = true;
- if (first_time) {
- for (int i = 0; i < GGML_MAX_CONTEXTS; i++) {
- g_state.contexts[i].used = false;
- }
- first_time = false;
- }
-
for (int i = 0; i < GGML_MAX_CONTEXTS; i++) {
if (!g_state.contexts[i].used) {
g_state.contexts[i].used = true;
if (ctx == NULL) {
GGML_PRINT_DEBUG("%s: no unused context found\n", __func__);
- atomic_fetch_sub(&g_state_barrier, 1);
+ ggml_critical_section_end();
return NULL;
}
GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
- atomic_fetch_sub(&g_state_barrier, 1);
+ ggml_critical_section_end();
return ctx;
}
void ggml_free(struct ggml_context * ctx) {
// make this function thread safe
- {
- int processing = atomic_fetch_add(&g_state_barrier, 1);
- while (processing > 0) {
- // wait for other threads to finish
- atomic_fetch_sub(&g_state_barrier, 1);
- sched_yield();
- processing = atomic_fetch_add(&g_state_barrier, 1);
- }
- }
+ ggml_critical_section_start();
+
+ bool found = false;
for (int i = 0; i < GGML_MAX_CONTEXTS; i++) {
if (&g_state.contexts[i].context == ctx) {
free(ctx->mem_buffer);
}
- atomic_fetch_sub(&g_state_barrier, 1);
-
- return;
+ found = true;
+ break;
}
}
- GGML_PRINT_DEBUG("%s: context not found\n", __func__);
+ if (!found) {
+ GGML_PRINT_DEBUG("%s: context not found\n", __func__);
+ }
- atomic_fetch_sub(&g_state_barrier, 1);
+ ggml_critical_section_end();
}
size_t ggml_used_mem(const struct ggml_context * ctx) {
// ggml_compute_forward_dup
-void ggml_compute_forward_dup_f16(
+static void ggml_compute_forward_dup_f16(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_dup_f32(
+static void ggml_compute_forward_dup_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_dup(
+static void ggml_compute_forward_dup(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_add
-void ggml_compute_forward_add_f32(
+static void ggml_compute_forward_add_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_add(
+static void ggml_compute_forward_add(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_sub
-void ggml_compute_forward_sub_f32(
+static void ggml_compute_forward_sub_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_sub(
+static void ggml_compute_forward_sub(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_mul
-void ggml_compute_forward_mul_f32(
+static void ggml_compute_forward_mul_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_mul(
+static void ggml_compute_forward_mul(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_div
-void ggml_compute_forward_div_f32(
+static void ggml_compute_forward_div_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_div(
+static void ggml_compute_forward_div(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_sqr
-void ggml_compute_forward_sqr_f32(
+static void ggml_compute_forward_sqr_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_sqr(
+static void ggml_compute_forward_sqr(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_sqrt
-void ggml_compute_forward_sqrt_f32(
+static void ggml_compute_forward_sqrt_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_sqrt(
+static void ggml_compute_forward_sqrt(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_sum
-void ggml_compute_forward_sum_f32(
+static void ggml_compute_forward_sum_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_sum(
+static void ggml_compute_forward_sum(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_mean
-void ggml_compute_forward_mean_f32(
+static void ggml_compute_forward_mean_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_mean(
+static void ggml_compute_forward_mean(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_repeat
-void ggml_compute_forward_repeat_f32(
+static void ggml_compute_forward_repeat_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_repeat(
+static void ggml_compute_forward_repeat(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_abs
-void ggml_compute_forward_abs_f32(
+static void ggml_compute_forward_abs_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_abs(
+static void ggml_compute_forward_abs(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_sgn
-void ggml_compute_forward_sgn_f32(
+static void ggml_compute_forward_sgn_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_sgn(
+static void ggml_compute_forward_sgn(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_neg
-void ggml_compute_forward_neg_f32(
+static void ggml_compute_forward_neg_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_neg(
+static void ggml_compute_forward_neg(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_step
-void ggml_compute_forward_step_f32(
+static void ggml_compute_forward_step_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_step(
+static void ggml_compute_forward_step(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_relu
-void ggml_compute_forward_relu_f32(
+static void ggml_compute_forward_relu_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_relu(
+static void ggml_compute_forward_relu(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_gelu
-void ggml_compute_forward_gelu_f32(
+static void ggml_compute_forward_gelu_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_gelu(
+static void ggml_compute_forward_gelu(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_norm
-void ggml_compute_forward_norm_f32(
+static void ggml_compute_forward_norm_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_norm(
+static void ggml_compute_forward_norm(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_mul_mat
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
// helper function to determine if it is better to use BLAS or not
// for large matrices, BLAS is faster
-bool ggml_compute_forward_mul_mat_use_blas(
+static bool ggml_compute_forward_mul_mat_use_blas(
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
struct ggml_tensor * dst) {
return false;
}
+#endif
-void ggml_compute_forward_mul_mat_f32(
+static void ggml_compute_forward_mul_mat_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
//}
}
-void ggml_compute_forward_mul_mat_f16_f32(
+static void ggml_compute_forward_mul_mat_f16_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
//}
}
-void ggml_compute_forward_mul_mat(
+static void ggml_compute_forward_mul_mat(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_scale
-void ggml_compute_forward_scale_f32(
+static void ggml_compute_forward_scale_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_scale(
+static void ggml_compute_forward_scale(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_cpy
-void ggml_compute_forward_cpy(
+static void ggml_compute_forward_cpy(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_reshape
-void ggml_compute_forward_reshape(
+static void ggml_compute_forward_reshape(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_view
-void ggml_compute_forward_view(
+static void ggml_compute_forward_view(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0) {
// NOP
// ggml_compute_forward_permute
-void ggml_compute_forward_permute(
+static void ggml_compute_forward_permute(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0) {
// NOP
// ggml_compute_forward_transpose
-void ggml_compute_forward_transpose(
+static void ggml_compute_forward_transpose(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0) {
// NOP
// ggml_compute_forward_get_rows
-void ggml_compute_forward_get_rows_f16(
+static void ggml_compute_forward_get_rows_f16(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_get_rows_f32(
+static void ggml_compute_forward_get_rows_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_get_rows(
+static void ggml_compute_forward_get_rows(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_diag_mask_inf
-void ggml_compute_forward_diag_mask_inf_f32(
+static void ggml_compute_forward_diag_mask_inf_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_diag_mask_inf(
+static void ggml_compute_forward_diag_mask_inf(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_soft_max
-void ggml_compute_forward_soft_max_f32(
+static void ggml_compute_forward_soft_max_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
}
}
-void ggml_compute_forward_soft_max(
+static void ggml_compute_forward_soft_max(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
struct ggml_tensor * dst) {
// ggml_compute_forward_rope
-void ggml_compute_forward_rope_f32(
+static void ggml_compute_forward_rope_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_rope(
+static void ggml_compute_forward_rope(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_conv_1d_1s
-void ggml_compute_forward_conv_1d_1s_f16_f32(
+static void ggml_compute_forward_conv_1d_1s_f16_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_conv_1d_1s_f32(
+static void ggml_compute_forward_conv_1d_1s_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_conv_1d_1s(
+static void ggml_compute_forward_conv_1d_1s(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_conv_1d_2s
-void ggml_compute_forward_conv_1d_2s_f16_f32(
+static void ggml_compute_forward_conv_1d_2s_f16_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_conv_1d_2s_f32(
+static void ggml_compute_forward_conv_1d_2s_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
}
}
-void ggml_compute_forward_conv_1d_2s(
+static void ggml_compute_forward_conv_1d_2s(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
// ggml_compute_forward_flash_attn
-void ggml_compute_forward_flash_attn_f32(
+static void ggml_compute_forward_flash_attn_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * q,
const struct ggml_tensor * k,
}
}
-void ggml_compute_forward_flash_attn_f16(
+static void ggml_compute_forward_flash_attn_f16(
const struct ggml_compute_params * params,
const struct ggml_tensor * q,
const struct ggml_tensor * k,
}
}
-void ggml_compute_forward_flash_attn(
+static void ggml_compute_forward_flash_attn(
const struct ggml_compute_params * params,
const struct ggml_tensor * q,
const struct ggml_tensor * k,
// ggml_compute_forward_flash_ff
-void ggml_compute_forward_flash_ff_f16(
+static void ggml_compute_forward_flash_ff_f16(
const struct ggml_compute_params * params,
const struct ggml_tensor * a, // F16
const struct ggml_tensor * b0, // F16 fc_w
}
}
-void ggml_compute_forward_flash_ff(
+static void ggml_compute_forward_flash_ff(
const struct ggml_compute_params * params,
const struct ggml_tensor * a,
const struct ggml_tensor * b0,
/////////////////////////////////
-void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
+static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
assert(params);
switch (tensor->op) {
////////////////////////////////////////////////////////////////////////////////
-void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, bool inplace) {
+static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, bool inplace) {
struct ggml_tensor * src0 = tensor->src0;
struct ggml_tensor * src1 = tensor->src1;
}
}
-void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
+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
// it can also happen during forward pass, if the user performs computations with constants
}
}
-void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
+static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
if (!expand) {
cgraph->n_nodes = 0;
cgraph->n_leafs = 0;
#define GGML_LOCK_INITIALIZER 0
+typedef pthread_t ggml_thread_t;
+
+#define ggml_thread_create pthread_create
+#define ggml_thread_join pthread_join
+
#else
//typedef pthread_spinlock_t ggml_lock_t;
#define GGML_LOCK_INITIALIZER 0
+typedef pthread_t ggml_thread_t;
+
+#define ggml_thread_create pthread_create
+#define ggml_thread_join pthread_join
+
#endif
struct ggml_compute_state_shared {
};
struct ggml_compute_state {
- pthread_t thrd;
+ ggml_thread_t thrd;
struct ggml_compute_params params;
struct ggml_tensor * node;
struct ggml_compute_state_shared * shared;
};
-// function used by each compute thread
-void * ggml_graph_compute_one(void * data) {
- struct ggml_compute_state * state = (struct ggml_compute_state *) data;
-
- ggml_compute_forward(&state->params, state->node);
-
- return NULL;
-}
-
-thread_ret_t ggml_graph_compute_thread(void * data) {
+static thread_ret_t ggml_graph_compute_thread(void * data) {
struct ggml_compute_state * state = (struct ggml_compute_state *) data;
const int n_threads = state->shared->n_threads;
.node = NULL,
.shared = &state_shared,
};
- int rc = pthread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);
+ int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);
assert(rc == 0);
UNUSED(rc);
}
atomic_store(&state_shared.has_work, true);
for (int j = 0; j < n_threads - 1; j++) {
- int rc = pthread_join(workers[j].thrd, NULL);
+ int rc = ggml_thread_join(workers[j].thrd, NULL);
assert(rc == 0);
UNUSED(rc);
}
}
// check if node is part of the graph
-bool ggml_graph_find(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
+static bool ggml_graph_find(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
if (cgraph == NULL) {
return true;
}
return false;
}
-struct ggml_tensor * ggml_graph_get_parent(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
+static struct ggml_tensor * ggml_graph_get_parent(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * parent = cgraph->nodes[i];
////////////////////////////////////////////////////////////////////////////////
-void ggml_opt_set_params(int np, struct ggml_tensor * const ps[], const float * x) {
+static void ggml_opt_set_params(int np, struct ggml_tensor * const ps[], const float * x) {
int i = 0;
for (int p = 0; p < np; ++p) {
const int ne = ggml_nelements(ps[p]) ;
}
}
-void ggml_opt_get_params(int np, struct ggml_tensor * const ps[], float * x) {
+static void ggml_opt_get_params(int np, struct ggml_tensor * const ps[], float * x) {
int i = 0;
for (int p = 0; p < np; ++p) {
const int ne = ggml_nelements(ps[p]) ;
}
}
-void ggml_opt_get_grad(int np, struct ggml_tensor * const ps[], float * g) {
+static void ggml_opt_get_grad(int np, struct ggml_tensor * const ps[], float * g) {
int i = 0;
for (int p = 0; p < np; ++p) {
const int ne = ggml_nelements(ps[p]) ;
// ref: https://arxiv.org/pdf/1412.6980.pdf
//
-enum ggml_opt_result ggml_opt_adam(
+static enum ggml_opt_result ggml_opt_adam(
struct ggml_context * ctx,
struct ggml_opt_params params,
struct ggml_tensor * f,
return GGML_LINESEARCH_FAIL;
}
-enum ggml_opt_result ggml_opt_lbfgs(
+static enum ggml_opt_result ggml_opt_lbfgs(
struct ggml_context * ctx,
struct ggml_opt_params params,
struct ggml_tensor * f,
overview / pkg / ggml / sources / whisper.cpp / commitdiff