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

naming : normalize the name of callback-related identifiers

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-09-16 09:11:42 +03:00
parent c4965a64f7
commit cc1c017191
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
14 changed files with 202 additions and 215 deletions
Download patch file Download diff file Expand all files Collapse all files

26
examples/cvector-generator/cvector-generator.cpp
View file

@ -50,7 +50,7 @@ static void print_usage(int, char ** argv) {
// cb_eval is reused for each pair of positive - negative prompt
struct callback_data {
struct callback_context {
ggml_context * ctx_ggml = nullptr; // holds v_pos, v_neg, v_diff_filtered
int n_layers = 0;
@ -155,7 +155,7 @@ struct callback_data {
return diff_filtered;
}
// we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors
// we don't implement destructor, because we want to reuse callback_context. we just want to free the tensors
void reset() {
for (auto ptr : v_pos) free(ptr->data);
for (auto ptr : v_neg) free(ptr->data);
@ -320,7 +320,7 @@ static std::vector<std::string> ctrlvec_load_prompt_file(std::string path, bool
//////////////////////////////////////////////////
static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
auto * cb_data = (callback_data *) user_data;
auto * cb_ctx = (callback_context *) user_data;
static const char * l_out_name = "l_out";
const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0;
@ -328,12 +328,12 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
return is_l_out;
}
if (!is_l_out || t->ne[1] != cb_data->n_tokens) {
if (!is_l_out || t->ne[1] != cb_ctx->n_tokens) {
return true;
}
// save the tensor to current context
cb_data->save_tensor_for_layer(t);
cb_ctx->save_tensor_for_layer(t);
return true;
}
@ -400,12 +400,12 @@ int main(int argc, char ** argv) {
}
callback_data cb_data;
callback_context cb_ctx;
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
params.cb_eval = cb_eval;
params.cb_eval_user_data = &cb_data;
params.cb_eval_ctx = &cb_ctx;
params.warmup = false;
print_build_info();
@ -445,8 +445,8 @@ int main(int argc, char ** argv) {
for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
bool success = false;
tokenized_prompt t = tokenized_prompts[i];
cb_data.n_layers = n_layers;
cb_data.n_tokens = t.max_seq_len;
cb_ctx.n_layers = n_layers;
cb_ctx.n_tokens = t.max_seq_len;
printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n",
(int) i+1, (int) ctx_train.positive_entries.size(),
@ -454,22 +454,22 @@ int main(int argc, char ** argv) {
tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(),
(int) t.max_seq_len);
cb_data.is_eval_pos = true;
cb_ctx.is_eval_pos = true;
success = get_hidden_layers(ctx, t.tokens_pos);
if (!success) break;
cb_data.is_eval_pos = false;
cb_ctx.is_eval_pos = false;
success = get_hidden_layers(ctx, t.tokens_neg);
if (!success) break;
// calculate diff and remove all zero rows
auto v_diff_filtered = cb_data.calc_diff();
auto v_diff_filtered = cb_ctx.calc_diff();
// save & concat the filtered v_diff to ctx_train
ctx_train.concat_diff_tmp(v_diff_filtered);
// reset for next iteration
cb_data.reset();
cb_ctx.reset();
}
// done with the model, we can now free it to make gain some memory

26
examples/eval-callback/eval-callback.cpp
View file

@ -12,7 +12,7 @@
* This the arbitrary data which will be passed to each callback.
* Later on we can for example add operation or tensor name filter from the CLI arg, or a file descriptor to dump the tensor.
*/
struct callback_data {
struct callback_context {
std::vector<uint8_t> data;
};
@ -27,7 +27,7 @@ static std::string ggml_ne_string(const ggml_tensor * t) {
return str;
}
static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
static void ggml_print_tensor(const uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
GGML_ASSERT(n > 0);
float sum = 0;
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
@ -52,15 +52,15 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
float v;
if (type == GGML_TYPE_F16) {
v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
v = ggml_fp16_to_fp32(*(const ggml_fp16_t *) &data[i]);
} else if (type == GGML_TYPE_F32) {
v = *(float *) &data[i];
v = *(const float *) &data[i];
} else if (type == GGML_TYPE_I32) {
v = (float) *(int32_t *) &data[i];
v = (float) *(const int32_t *) &data[i];
} else if (type == GGML_TYPE_I16) {
v = (float) *(int16_t *) &data[i];
v = (float) *(const int16_t *) &data[i];
} else if (type == GGML_TYPE_I8) {
v = (float) *(int8_t *) &data[i];
v = (float) *(const int8_t *) &data[i];
} else {
GGML_ABORT("fatal error");
}
@ -88,7 +88,7 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
* @return true to receive data or continue the graph, false otherwise
*/
static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
auto * cb_data = (callback_data *) user_data;
auto * cb_ctx = (callback_context *) user_data;
const struct ggml_tensor * src0 = t->src[0];
const struct ggml_tensor * src1 = t->src[1];
@ -114,12 +114,12 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
if (!is_host) {
auto n_bytes = ggml_nbytes(t);
cb_data->data.resize(n_bytes);
ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
cb_ctx->data.resize(n_bytes);
ggml_backend_tensor_get(t, cb_ctx->data.data(), 0, n_bytes);
}
if (!ggml_is_quantized(t->type)) {
uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
uint8_t * data = is_host ? (uint8_t *) t->data : cb_ctx->data.data();
ggml_print_tensor(data, t->type, t->ne, t->nb, 3);
}
@ -140,7 +140,7 @@ static bool run(llama_context * ctx, const gpt_params & params) {
}
int main(int argc, char ** argv) {
callback_data cb_data;
callback_context cb_ctx;
gpt_params params;
@ -156,7 +156,7 @@ int main(int argc, char ** argv) {
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
params.cb_eval = ggml_debug;
params.cb_eval_user_data = &cb_data;
params.cb_eval_ctx = &cb_ctx;
params.warmup = false;
// init

2
examples/imatrix/imatrix.cpp
View file

@ -602,7 +602,7 @@ int main(int argc, char ** argv) {
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
params.cb_eval = ik_collect_imatrix;
params.cb_eval_user_data = NULL;
params.cb_eval_ctx = NULL;
params.warmup = false;
// init