check C++ code with -Wmissing-declarations (#3184)
This commit is contained in:
Cebtenzzre
GitHub
parent
69eb67e282
commit
3aefaab9e5
22 changed files with 247 additions and 243 deletions
|
|
@ -9,12 +9,12 @@
|
|||
#endif
|
||||
|
||||
#ifdef LLAMA_DEFAULT_RMS_EPS
|
||||
static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
|
||||
constexpr float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
|
||||
#else
|
||||
static const float rms_norm_eps = 5e-6f;
|
||||
constexpr float rms_norm_eps = 5e-6f;
|
||||
#endif
|
||||
|
||||
float frand() {
|
||||
static float frand() {
|
||||
return (float)rand()/(float)RAND_MAX;
|
||||
}
|
||||
|
||||
|
|
@ -25,19 +25,21 @@ struct random_normal_distribution {
|
|||
float max;
|
||||
};
|
||||
|
||||
void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
|
||||
static void init_random_normal_distribution(
|
||||
struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max
|
||||
) {
|
||||
rnd->gen = std::mt19937(seed);
|
||||
rnd->nd = std::normal_distribution<float>{mean, std};
|
||||
rnd->min = min;
|
||||
rnd->max = max;
|
||||
}
|
||||
|
||||
float frand_normal(struct random_normal_distribution * rnd) {
|
||||
static float frand_normal(struct random_normal_distribution * rnd) {
|
||||
const float r = rnd->nd(rnd->gen);
|
||||
return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r);
|
||||
}
|
||||
|
||||
void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
|
||||
static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
|
||||
struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
|
||||
|
||||
if (plan.work_size > 0) {
|
||||
|
|
@ -48,13 +50,9 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
|
|||
ggml_graph_compute(graph, &plan);
|
||||
}
|
||||
|
||||
struct ggml_tensor * randomize_tensor(
|
||||
struct ggml_tensor * tensor,
|
||||
int ndims,
|
||||
const int64_t ne[],
|
||||
float fmin,
|
||||
float fmax) {
|
||||
|
||||
static struct ggml_tensor * randomize_tensor(
|
||||
struct ggml_tensor * tensor, int ndims, const int64_t ne[], float fmin, float fmax
|
||||
) {
|
||||
switch (ndims) {
|
||||
case 1:
|
||||
for (int i0 = 0; i0 < ne[0]; i0++) {
|
||||
|
|
@ -95,11 +93,9 @@ struct ggml_tensor * randomize_tensor(
|
|||
return tensor;
|
||||
}
|
||||
|
||||
struct ggml_tensor * randomize_tensor_normal(
|
||||
struct ggml_tensor * tensor,
|
||||
int ndims,
|
||||
const int64_t ne[],
|
||||
struct random_normal_distribution * rnd) {
|
||||
static struct ggml_tensor * randomize_tensor_normal(
|
||||
struct ggml_tensor * tensor, int ndims, const int64_t ne[], struct random_normal_distribution * rnd
|
||||
) {
|
||||
float scale = 1.0; // xavier
|
||||
switch (ndims) {
|
||||
case 1:
|
||||
|
|
@ -159,7 +155,7 @@ struct llama_hparams {
|
|||
}
|
||||
};
|
||||
|
||||
uint32_t get_n_ff(const struct llama_hparams* hparams) {
|
||||
static uint32_t get_n_ff(const struct llama_hparams* hparams) {
|
||||
const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
|
||||
return n_ff;
|
||||
}
|
||||
|
|
@ -260,7 +256,7 @@ struct llama_model_lora {
|
|||
std::vector<llama_layer_lora> layers;
|
||||
};
|
||||
|
||||
void init_model(struct llama_model * model) {
|
||||
static void init_model(struct llama_model * model) {
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const uint32_t n_embd = hparams.n_embd;
|
||||
|
|
@ -297,7 +293,7 @@ void init_model(struct llama_model * model) {
|
|||
}
|
||||
|
||||
|
||||
void init_model_lora(struct llama_model_lora * model) {
|
||||
static void init_model_lora(struct llama_model_lora * model) {
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const uint32_t n_embd = hparams.n_embd;
|
||||
|
|
@ -340,7 +336,7 @@ void init_model_lora(struct llama_model_lora * model) {
|
|||
}
|
||||
}
|
||||
|
||||
void set_param_model(struct llama_model * model) {
|
||||
static void set_param_model(struct llama_model * model) {
|
||||
const auto& hparams = model->hparams;
|
||||
|
||||
const uint32_t n_layer = hparams.n_layer;
|
||||
|
|
@ -366,7 +362,7 @@ void set_param_model(struct llama_model * model) {
|
|||
}
|
||||
}
|
||||
|
||||
void set_param_model_lora(struct llama_model_lora * model) {
|
||||
static void set_param_model_lora(struct llama_model_lora * model) {
|
||||
const auto& hparams = model->hparams;
|
||||
|
||||
const uint32_t n_layer = hparams.n_layer;
|
||||
|
|
@ -397,7 +393,7 @@ void set_param_model_lora(struct llama_model_lora * model) {
|
|||
}
|
||||
}
|
||||
|
||||
void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
|
||||
static void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const uint32_t n_layer = hparams.n_layer;
|
||||
|
|
@ -426,7 +422,9 @@ void randomize_model(struct llama_model * model, int seed, float mean, float std
|
|||
}
|
||||
|
||||
|
||||
void randomize_model_lora(struct llama_model_lora * model, int seed, float mean, float std, float min, float max) {
|
||||
static void randomize_model_lora(
|
||||
struct llama_model_lora * model, int seed, float mean, float std, float min, float max
|
||||
) {
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const uint32_t n_layer = hparams.n_layer;
|
||||
|
|
@ -459,7 +457,7 @@ void randomize_model_lora(struct llama_model_lora * model, int seed, float mean,
|
|||
}
|
||||
}
|
||||
|
||||
bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
|
||||
static bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const uint32_t n_ctx = hparams.n_ctx;
|
||||
|
|
@ -495,7 +493,7 @@ bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int
|
|||
return true;
|
||||
}
|
||||
|
||||
bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
|
||||
static bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const uint32_t n_ctx = hparams.n_ctx;
|
||||
|
|
@ -531,15 +529,15 @@ bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora *
|
|||
return true;
|
||||
}
|
||||
|
||||
struct ggml_tensor * forward(
|
||||
struct llama_model * model,
|
||||
struct llama_kv_cache * cache,
|
||||
struct ggml_context * ctx0,
|
||||
struct ggml_cgraph * gf,
|
||||
struct ggml_tensor * tokens_input,
|
||||
const int n_tokens,
|
||||
const int n_past) {
|
||||
|
||||
static struct ggml_tensor * forward(
|
||||
struct llama_model * model,
|
||||
struct llama_kv_cache * cache,
|
||||
struct ggml_context * ctx0,
|
||||
struct ggml_cgraph * gf,
|
||||
struct ggml_tensor * tokens_input,
|
||||
const int n_tokens,
|
||||
const int n_past
|
||||
) {
|
||||
const int N = n_tokens;
|
||||
|
||||
struct llama_kv_cache& kv_self = *cache;
|
||||
|
|
@ -756,25 +754,25 @@ struct ggml_tensor * forward(
|
|||
return inpL;
|
||||
}
|
||||
|
||||
void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
|
||||
static void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
|
||||
GGML_ASSERT(tensor->n_dims == 1);
|
||||
GGML_ASSERT(tensor->ne[0] == ne0);
|
||||
}
|
||||
|
||||
void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
|
||||
static void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
|
||||
GGML_ASSERT(tensor->n_dims == 2);
|
||||
GGML_ASSERT(tensor->ne[0] == ne0);
|
||||
GGML_ASSERT(tensor->ne[1] == ne1);
|
||||
}
|
||||
|
||||
void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
|
||||
static void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
|
||||
GGML_ASSERT(tensor->n_dims == 3);
|
||||
GGML_ASSERT(tensor->ne[0] == ne0);
|
||||
GGML_ASSERT(tensor->ne[1] == ne1);
|
||||
GGML_ASSERT(tensor->ne[2] == ne2);
|
||||
}
|
||||
|
||||
void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
|
||||
static void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
|
||||
GGML_ASSERT(tensor->n_dims == 4);
|
||||
GGML_ASSERT(tensor->ne[0] == ne0);
|
||||
GGML_ASSERT(tensor->ne[1] == ne1);
|
||||
|
|
@ -782,16 +780,16 @@ void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int6
|
|||
GGML_ASSERT(tensor->ne[3] == ne3);
|
||||
}
|
||||
|
||||
struct ggml_tensor * forward_batch(
|
||||
struct llama_model * model,
|
||||
struct llama_kv_cache * cache,
|
||||
struct ggml_context * ctx0,
|
||||
struct ggml_cgraph * gf,
|
||||
struct ggml_tensor * tokens_input,
|
||||
const int n_tokens,
|
||||
const int n_past,
|
||||
const int n_batch) {
|
||||
|
||||
static struct ggml_tensor * forward_batch(
|
||||
struct llama_model * model,
|
||||
struct llama_kv_cache * cache,
|
||||
struct ggml_context * ctx0,
|
||||
struct ggml_cgraph * gf,
|
||||
struct ggml_tensor * tokens_input,
|
||||
const int n_tokens,
|
||||
const int n_past,
|
||||
const int n_batch
|
||||
) {
|
||||
const int N = n_tokens;
|
||||
|
||||
struct llama_kv_cache& kv_self = *cache;
|
||||
|
|
@ -1073,16 +1071,15 @@ struct ggml_tensor * forward_batch(
|
|||
return inpL;
|
||||
}
|
||||
|
||||
|
||||
struct ggml_tensor * forward_lora(
|
||||
struct llama_model_lora * model,
|
||||
struct llama_kv_cache * cache,
|
||||
struct ggml_context * ctx0,
|
||||
struct ggml_cgraph * gf,
|
||||
struct ggml_tensor * tokens_input,
|
||||
const int n_tokens,
|
||||
const int n_past) {
|
||||
|
||||
static struct ggml_tensor * forward_lora(
|
||||
struct llama_model_lora * model,
|
||||
struct llama_kv_cache * cache,
|
||||
struct ggml_context * ctx0,
|
||||
struct ggml_cgraph * gf,
|
||||
struct ggml_tensor * tokens_input,
|
||||
const int n_tokens,
|
||||
const int n_past
|
||||
) {
|
||||
const int N = n_tokens;
|
||||
|
||||
struct llama_kv_cache& kv_self = *cache;
|
||||
|
|
@ -1328,7 +1325,7 @@ struct ggml_tensor * forward_lora(
|
|||
return inpL;
|
||||
}
|
||||
|
||||
void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
|
||||
static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
|
||||
assert(logits->n_dims == 2);
|
||||
assert(probs->n_dims == 2);
|
||||
assert(best_samples->n_dims == 1);
|
||||
|
|
@ -1359,7 +1356,10 @@ void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, str
|
|||
}
|
||||
}
|
||||
|
||||
void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
|
||||
static void sample_softmax_batch(
|
||||
struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
|
||||
struct ggml_tensor * best_samples
|
||||
) {
|
||||
GGML_ASSERT(best_samples->n_dims == 2);
|
||||
GGML_ASSERT(logits->n_dims == 3);
|
||||
GGML_ASSERT(probs->n_dims == 3);
|
||||
|
|
@ -1393,7 +1393,7 @@ void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits
|
|||
}
|
||||
}
|
||||
|
||||
void print_row(struct ggml_tensor * probs, int i) {
|
||||
static void print_row(struct ggml_tensor * probs, int i) {
|
||||
for (int k = 0; k < probs->ne[0]; ++k) {
|
||||
float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
|
||||
printf(" %.2f", p);
|
||||
|
|
@ -1401,7 +1401,7 @@ void print_row(struct ggml_tensor * probs, int i) {
|
|||
printf("\n");
|
||||
}
|
||||
|
||||
void print_matrix(struct ggml_tensor * probs) {
|
||||
static void print_matrix(struct ggml_tensor * probs) {
|
||||
assert(probs->n_dims == 2);
|
||||
for (int i = 0; i < probs->ne[1]; ++i) {
|
||||
for (int k = 0; k < probs->ne[0]; ++k) {
|
||||
|
|
@ -1412,7 +1412,7 @@ void print_matrix(struct ggml_tensor * probs) {
|
|||
}
|
||||
}
|
||||
|
||||
void print_token(int token, int n_vocab) {
|
||||
static void print_token(int token, int n_vocab) {
|
||||
for (int k = 0; k < token; ++k) {
|
||||
printf(" ");
|
||||
}
|
||||
|
|
@ -1423,14 +1423,14 @@ void print_token(int token, int n_vocab) {
|
|||
printf("\n");
|
||||
}
|
||||
|
||||
void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
|
||||
static void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
|
||||
for (int i=0; i<tokens->ne[0]; ++i) {
|
||||
int token = ggml_get_i32_1d(tokens, i);
|
||||
print_token(token, n_vocab);
|
||||
}
|
||||
}
|
||||
|
||||
void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
|
||||
static void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
|
||||
int n_tokens = tokens_input->ne[0];
|
||||
int n_vocab = targets->ne[0];
|
||||
float randomness = 0.0f;
|
||||
|
|
@ -1451,7 +1451,9 @@ void get_example_targets(int example_id, struct ggml_tensor * tokens_input, stru
|
|||
}
|
||||
}
|
||||
|
||||
void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
|
||||
static void get_example_targets_batch(
|
||||
struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
|
||||
) {
|
||||
GGML_ASSERT(tokens_input->n_dims == 2);
|
||||
GGML_ASSERT( targets->n_dims == 3);
|
||||
int n_tokens = tokens_input->ne[0];
|
||||
|
|
@ -1474,7 +1476,7 @@ void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct
|
|||
}
|
||||
}
|
||||
|
||||
void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
|
||||
static void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
|
||||
int n_tokens = tokens_input->ne[0];
|
||||
int n_vocab = targets->ne[0];
|
||||
for (int i=0; i<n_tokens-n_shift; ++i) {
|
||||
|
|
@ -1485,12 +1487,16 @@ void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * tar
|
|||
}
|
||||
}
|
||||
|
||||
struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
|
||||
static struct ggml_tensor * square_error_loss(
|
||||
struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
|
||||
) {
|
||||
// todo: instead of a-b: a[1:]-b[:-1]
|
||||
return ggml_sum(ctx, ggml_sqr(ctx, ggml_sub(ctx, a, b)));
|
||||
}
|
||||
|
||||
struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
|
||||
static struct ggml_tensor * cross_entropy_loss(
|
||||
struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
|
||||
) {
|
||||
const float eps = 1e-3f;
|
||||
return
|
||||
ggml_sum(ctx,
|
||||
|
|
|
|||
|
|
@ -30,7 +30,8 @@ struct ostream_beam_view {
|
|||
llama_context * ctx;
|
||||
llama_beam_view beam_view;
|
||||
};
|
||||
std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) {
|
||||
|
||||
static std::ostream & operator<<(std::ostream & os, const ostream_beam_view & obv) {
|
||||
os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
|
||||
for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
|
||||
os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
|
||||
|
|
@ -46,7 +47,7 @@ struct beam_search_callback_data {
|
|||
|
||||
// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
|
||||
// For example, eob can be flagged due to maximum token length, stop words, etc.
|
||||
bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, const size_t n_tokens) {
|
||||
static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
|
||||
return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
|
||||
}
|
||||
|
||||
|
|
@ -56,7 +57,7 @@ bool is_at_eob(const beam_search_callback_data & callback_data, const llama_toke
|
|||
// * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
|
||||
// This is also called when the stop condition is met.
|
||||
// Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
|
||||
void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
|
||||
static void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
|
||||
auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
|
||||
// Mark beams as EOS as needed.
|
||||
for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
|
||||
|
|
|
|||
|
|
@ -115,7 +115,7 @@ struct TransformerWeights {
|
|||
}
|
||||
};
|
||||
|
||||
void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
|
||||
static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
|
||||
// we calloc instead of malloc to keep valgrind happy
|
||||
w->token_embedding_table = new float[p->vocab_size * p->dim]();
|
||||
printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
|
||||
|
|
@ -158,7 +158,7 @@ void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
|
|||
}
|
||||
}
|
||||
|
||||
int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
|
||||
static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
|
||||
if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
|
||||
if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
|
||||
if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
|
||||
|
|
@ -189,7 +189,7 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shar
|
|||
return 0;
|
||||
}
|
||||
|
||||
void print_sample_weights(TransformerWeights *w){
|
||||
static void print_sample_weights(TransformerWeights *w){
|
||||
printf("----- Quick print of first of the weight vales of all the variables\n");
|
||||
printf("%f\n", w->token_embedding_table[0]);
|
||||
printf("%f\n", w->rms_att_weight[0]);
|
||||
|
|
@ -324,7 +324,7 @@ struct train_params {
|
|||
int mem_compute1_gb;
|
||||
};
|
||||
|
||||
void print_params(struct my_llama_hparams * params) {
|
||||
static void print_params(struct my_llama_hparams * params) {
|
||||
printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
|
||||
printf("%s: n_ctx: %d\n", __func__, params->n_ctx);
|
||||
printf("%s: n_embd: %d\n", __func__, params->n_embd);
|
||||
|
|
@ -335,7 +335,7 @@ void print_params(struct my_llama_hparams * params) {
|
|||
printf("%s: n_rot: %d\n", __func__, params->n_rot);
|
||||
}
|
||||
|
||||
void init_model(struct my_llama_model * model) {
|
||||
static void init_model(struct my_llama_model * model) {
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const uint32_t n_embd = hparams.n_embd;
|
||||
|
|
@ -408,17 +408,17 @@ void init_model(struct my_llama_model * model) {
|
|||
}
|
||||
}
|
||||
|
||||
float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
|
||||
static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
|
||||
float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
|
||||
return *ptr;
|
||||
}
|
||||
|
||||
int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
|
||||
static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
|
||||
int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
|
||||
return *ptr;
|
||||
}
|
||||
|
||||
void print_row(struct ggml_tensor * probs, int i) {
|
||||
static void print_row(struct ggml_tensor * probs, int i) {
|
||||
for (int k = 0; k < probs->ne[0]; ++k) {
|
||||
float p = get_f32_2d(probs, k, i);
|
||||
printf(" %f", p);
|
||||
|
|
@ -426,7 +426,7 @@ void print_row(struct ggml_tensor * probs, int i) {
|
|||
printf("\n");
|
||||
}
|
||||
|
||||
void print_matrix(struct ggml_tensor * probs) {
|
||||
static void print_matrix(struct ggml_tensor * probs) {
|
||||
assert(probs->n_dims == 2);
|
||||
for (int i = 0; i < probs->ne[1]; ++i) {
|
||||
for (int k = 0; k < probs->ne[0]; ++k) {
|
||||
|
|
@ -531,7 +531,7 @@ struct llama_file {
|
|||
}
|
||||
};
|
||||
|
||||
bool is_ggml_file(const char *filename) {
|
||||
static bool is_ggml_file(const char * filename) {
|
||||
llama_file file(filename, "rb");
|
||||
if (file.size < 4) {
|
||||
return false;
|
||||
|
|
@ -540,7 +540,7 @@ bool is_ggml_file(const char *filename) {
|
|||
return magic == GGUF_MAGIC;
|
||||
}
|
||||
|
||||
static std::string llama_escape_whitespaces(const std::string& text) {
|
||||
static std::string llama_escape_whitespaces(const std::string & text) {
|
||||
std::ostringstream out;
|
||||
for (char c : text) {
|
||||
if (c == ' ') out << "\xe2\x96\x81";
|
||||
|
|
@ -549,7 +549,7 @@ static std::string llama_escape_whitespaces(const std::string& text) {
|
|||
return out.str();
|
||||
}
|
||||
|
||||
void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
|
||||
static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
|
||||
if (is_ggml_file(filename)) {
|
||||
struct ggml_context * ctx_data = NULL;
|
||||
|
||||
|
|
@ -637,7 +637,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
|
|||
}
|
||||
}
|
||||
|
||||
void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
|
||||
static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
|
||||
int ct;
|
||||
switch (gg_weights->n_dims){
|
||||
case 1:
|
||||
|
|
@ -673,7 +673,9 @@ void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * kar
|
|||
}
|
||||
}
|
||||
|
||||
void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
|
||||
static void save_as_llama_model(
|
||||
struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
|
||||
) {
|
||||
// convert AK weights into GG weights one by one.
|
||||
// w->token_embedding_table -> model->tok_embeddings
|
||||
// float* -> struct ggml_tensor
|
||||
|
|
@ -785,7 +787,7 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
|
|||
gguf_free(ctx);
|
||||
}
|
||||
|
||||
struct train_params get_default_train_params() {
|
||||
static struct train_params get_default_train_params() {
|
||||
struct train_params params;
|
||||
params.fn_vocab_model = "models/7B/ggml-model-f16.gguf";
|
||||
params.fn_llama2c_output_model = "ak_llama_model.bin";
|
||||
|
|
@ -835,7 +837,7 @@ struct train_params get_default_train_params() {
|
|||
return params;
|
||||
}
|
||||
|
||||
void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
|
||||
static void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
|
||||
fprintf(stderr, "usage: %s [options]\n", argv[0]);
|
||||
fprintf(stderr, "\n");
|
||||
fprintf(stderr, "options:\n");
|
||||
|
|
@ -846,7 +848,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params)
|
|||
fprintf(stderr, "\n");
|
||||
}
|
||||
|
||||
bool params_parse(int argc, char ** argv, struct train_params * params) {
|
||||
static bool params_parse(int argc, char ** argv, struct train_params * params) {
|
||||
bool invalid_param = false;
|
||||
bool reqd_param_found = false;
|
||||
std::string arg;
|
||||
|
|
@ -901,7 +903,7 @@ bool params_parse(int argc, char ** argv, struct train_params * params) {
|
|||
return true;
|
||||
}
|
||||
|
||||
std::string basename(const std::string &path) {
|
||||
static std::string basename(const std::string &path) {
|
||||
size_t pos = path.find_last_of("/\\");
|
||||
if (pos == std::string::npos) {
|
||||
return path;
|
||||
|
|
|
|||
|
|
@ -13,14 +13,14 @@
|
|||
#define MIN(a, b) ((a) < (b) ? (a) : (b))
|
||||
#define MAX(a, b) ((a) > (b) ? (a) : (b))
|
||||
|
||||
template<typename T>
|
||||
template <typename T>
|
||||
static std::string to_string(const T & val) {
|
||||
std::stringstream ss;
|
||||
ss << val;
|
||||
return ss.str();
|
||||
}
|
||||
|
||||
bool gguf_ex_write(const std::string & fname) {
|
||||
static bool gguf_ex_write(const std::string & fname) {
|
||||
struct gguf_context * ctx = gguf_init_empty();
|
||||
|
||||
gguf_set_val_u8 (ctx, "some.parameter.uint8", 0x12);
|
||||
|
|
@ -85,7 +85,7 @@ bool gguf_ex_write(const std::string & fname) {
|
|||
}
|
||||
|
||||
// just read tensor info
|
||||
bool gguf_ex_read_0(const std::string & fname) {
|
||||
static bool gguf_ex_read_0(const std::string & fname) {
|
||||
struct gguf_init_params params = {
|
||||
/*.no_alloc = */ false,
|
||||
/*.ctx = */ NULL,
|
||||
|
|
@ -143,7 +143,7 @@ bool gguf_ex_read_0(const std::string & fname) {
|
|||
}
|
||||
|
||||
// read and create ggml_context containing the tensors and their data
|
||||
bool gguf_ex_read_1(const std::string & fname) {
|
||||
static bool gguf_ex_read_1(const std::string & fname) {
|
||||
struct ggml_context * ctx_data = NULL;
|
||||
|
||||
struct gguf_init_params params = {
|
||||
|
|
|
|||
|
|
@ -41,7 +41,8 @@ static std::ostringstream * g_output_ss;
|
|||
static std::vector<llama_token> * g_output_tokens;
|
||||
static bool is_interacting = false;
|
||||
|
||||
void write_logfile(
|
||||
|
||||
static void write_logfile(
|
||||
const llama_context * ctx, const gpt_params & params, const llama_model * model,
|
||||
const std::vector<llama_token> & input_tokens, const std::string & output,
|
||||
const std::vector<llama_token> & output_tokens
|
||||
|
|
@ -86,7 +87,7 @@ void write_logfile(
|
|||
}
|
||||
|
||||
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
|
||||
void sigint_handler(int signo) {
|
||||
static void sigint_handler(int signo) {
|
||||
if (signo == SIGINT) {
|
||||
if (!is_interacting) {
|
||||
is_interacting = true;
|
||||
|
|
|
|||
|
|
@ -28,9 +28,10 @@ struct results_log_softmax {
|
|||
float prob;
|
||||
};
|
||||
|
||||
void write_logfile(const llama_context * ctx, const gpt_params & params,
|
||||
const llama_model * model, const struct results_perplexity & results) {
|
||||
|
||||
static void write_logfile(
|
||||
const llama_context * ctx, const gpt_params & params, const llama_model * model,
|
||||
const struct results_perplexity & results
|
||||
) {
|
||||
if (params.logdir.empty()) {
|
||||
return;
|
||||
}
|
||||
|
|
@ -76,7 +77,7 @@ void write_logfile(const llama_context * ctx, const gpt_params & params,
|
|||
fclose(logfile);
|
||||
}
|
||||
|
||||
std::vector<float> softmax(const std::vector<float>& logits) {
|
||||
static std::vector<float> softmax(const std::vector<float>& logits) {
|
||||
std::vector<float> probs(logits.size());
|
||||
float max_logit = logits[0];
|
||||
for (float v : logits) max_logit = std::max(max_logit, v);
|
||||
|
|
@ -92,7 +93,7 @@ std::vector<float> softmax(const std::vector<float>& logits) {
|
|||
return probs;
|
||||
}
|
||||
|
||||
results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
|
||||
static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
|
||||
float max_logit = logits[0];
|
||||
for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]);
|
||||
double sum_exp = 0.0;
|
||||
|
|
@ -100,9 +101,10 @@ results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
|
|||
return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
|
||||
}
|
||||
|
||||
void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
|
||||
double & nll, double & nll2, float * logit_history, float * prob_history) {
|
||||
|
||||
static void process_logits(
|
||||
int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
|
||||
double & nll, double & nll2, float * logit_history, float * prob_history
|
||||
) {
|
||||
std::mutex mutex;
|
||||
int counter = 0;
|
||||
auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
|
||||
|
|
@ -130,7 +132,7 @@ void process_logits(int n_vocab, const float * logits, const int * tokens, int n
|
|||
|
||||
}
|
||||
|
||||
results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
|
||||
static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
|
||||
// Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
|
||||
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
|
||||
// Output: `perplexity: 13.5106 [114/114]`
|
||||
|
|
@ -260,8 +262,7 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
|
|||
return {tokens, std::exp(nll / count), logit_history, prob_history};
|
||||
}
|
||||
|
||||
results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
|
||||
|
||||
static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
|
||||
if (params.ppl_stride > 0) {
|
||||
return perplexity_v2(ctx, params);
|
||||
}
|
||||
|
|
@ -400,8 +401,9 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
|
|||
return {tokens, ppl, logit_history, prob_history};
|
||||
}
|
||||
|
||||
std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch,
|
||||
int n_vocab, int n_thread) {
|
||||
static std::vector<float> hellaswag_evaluate_tokens(
|
||||
llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch, int n_vocab, int n_thread
|
||||
) {
|
||||
std::vector<float> result;
|
||||
result.reserve(tokens.size() * n_vocab);
|
||||
size_t n_chunk = (tokens.size() + n_batch - 1)/n_batch;
|
||||
|
|
@ -421,7 +423,7 @@ std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vec
|
|||
return result;
|
||||
}
|
||||
|
||||
void hellaswag_score(llama_context * ctx, const gpt_params & params) {
|
||||
static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
|
||||
// Calculates hellaswag score (acc_norm) from prompt
|
||||
//
|
||||
// Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
|
||||
|
|
|
|||
|
|
@ -34,8 +34,8 @@ struct quantize_stats_params {
|
|||
std::vector<enum ggml_type> include_types;
|
||||
};
|
||||
|
||||
const size_t HISTOGRAM_BUCKETS = 150;
|
||||
const double HISTOGRAM_RANGE = 0.03;
|
||||
constexpr size_t HISTOGRAM_BUCKETS = 150;
|
||||
constexpr double HISTOGRAM_RANGE = 0.03;
|
||||
|
||||
struct error_stats {
|
||||
size_t num_samples;
|
||||
|
|
@ -44,8 +44,7 @@ struct error_stats {
|
|||
uint64_t error_histogram[HISTOGRAM_BUCKETS];
|
||||
};
|
||||
|
||||
|
||||
void quantize_stats_print_usage(int /*argc*/, char ** argv) {
|
||||
static void quantize_stats_print_usage(int /*argc*/, char ** argv) {
|
||||
quantize_stats_params params;
|
||||
fprintf(stderr, "usage: %s [options]\n", argv[0]);
|
||||
fprintf(stderr, "\n");
|
||||
|
|
@ -71,7 +70,7 @@ void quantize_stats_print_usage(int /*argc*/, char ** argv) {
|
|||
}
|
||||
|
||||
// Check if a layer is included/excluded by command line
|
||||
bool layer_included(const quantize_stats_params & params, const std::string & layer) {
|
||||
static bool layer_included(const quantize_stats_params & params, const std::string & layer) {
|
||||
for (const auto& excluded : params.exclude_layers) {
|
||||
if (std::regex_search(layer, std::regex(excluded))) {
|
||||
return false;
|
||||
|
|
@ -86,7 +85,7 @@ bool layer_included(const quantize_stats_params & params, const std::string & la
|
|||
}
|
||||
|
||||
// Update error statistics given vectors with the before/after result of quantization
|
||||
void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
|
||||
static void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
|
||||
for (int64_t i = 0; i < nelements; i++) {
|
||||
double diff = input[i] - output[i];
|
||||
stats.total_error += diff * diff;
|
||||
|
|
@ -96,14 +95,14 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou
|
|||
stats.num_samples += nelements;
|
||||
}
|
||||
|
||||
void combine_error_stats(error_stats & into, const error_stats & from) {
|
||||
static void combine_error_stats(error_stats & into, const error_stats & from) {
|
||||
into.num_samples += from.num_samples;
|
||||
into.total_error += from.total_error;
|
||||
if (from.max_error > into.max_error) into.max_error = from.max_error;
|
||||
for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
|
||||
}
|
||||
|
||||
double find_quantile(const error_stats & stats, double quantile) {
|
||||
static double find_quantile(const error_stats & stats, double quantile) {
|
||||
double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);
|
||||
|
||||
double accum = 0;
|
||||
|
|
@ -116,7 +115,7 @@ double find_quantile(const error_stats & stats, double quantile) {
|
|||
return INFINITY;
|
||||
}
|
||||
|
||||
void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
|
||||
static void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
|
||||
double rmse = sqrt(stats.total_error / (double) stats.num_samples);
|
||||
double median = find_quantile(stats, .5);
|
||||
double pct95 = find_quantile(stats, .95);
|
||||
|
|
@ -143,17 +142,10 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
|
|||
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
|
||||
}
|
||||
|
||||
void test_roundtrip_on_chunk(
|
||||
const ggml_tensor * layer,
|
||||
int64_t offset,
|
||||
int64_t chunk_size,
|
||||
const ggml_type_traits_t & qfns,
|
||||
bool use_reference,
|
||||
float * input_scratch,
|
||||
char * quantized_scratch,
|
||||
float * output_scratch,
|
||||
error_stats & stats) {
|
||||
|
||||
static void test_roundtrip_on_chunk(
|
||||
const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits_t & qfns, bool use_reference,
|
||||
float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
|
||||
) {
|
||||
if (layer->type == GGML_TYPE_F16) {
|
||||
for (int i = 0; i < chunk_size; i++) {
|
||||
input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
|
||||
|
|
@ -174,18 +166,11 @@ void test_roundtrip_on_chunk(
|
|||
|
||||
|
||||
// Run quantization function for a single layer and update error stats
|
||||
void test_roundtrip_on_layer(
|
||||
std::string & name,
|
||||
bool print_layer_stats,
|
||||
const ggml_type_traits_t & qfns,
|
||||
bool use_reference,
|
||||
const ggml_tensor * layer,
|
||||
std::vector<float> & input_scratch,
|
||||
std::vector<char> & quantized_scratch,
|
||||
std::vector<float> & output_scratch,
|
||||
error_stats & total_error,
|
||||
int max_thread = 0) {
|
||||
|
||||
static void test_roundtrip_on_layer(
|
||||
std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference,
|
||||
const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
|
||||
std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
|
||||
) {
|
||||
assert(tensor_is_contiguous(layer));
|
||||
error_stats layer_error {};
|
||||
uint64_t nelements = ggml_nelements(layer);
|
||||
|
|
|
|||
|
|
@ -40,7 +40,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
|
|||
};
|
||||
|
||||
|
||||
bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
|
||||
static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
|
||||
std::string ftype_str;
|
||||
|
||||
for (auto ch : ftype_str_in) {
|
||||
|
|
@ -72,7 +72,7 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std:
|
|||
// usage:
|
||||
// ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
|
||||
//
|
||||
void usage(const char * executable) {
|
||||
static void usage(const char * executable) {
|
||||
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
|
||||
printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
|
||||
printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
|
||||
|
|
|
|||
|
|
@ -1083,8 +1083,9 @@ static json format_final_response(llama_server_context &llama, const std::string
|
|||
return res;
|
||||
}
|
||||
|
||||
static json format_partial_response(llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs)
|
||||
{
|
||||
static json format_partial_response(
|
||||
llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs
|
||||
) {
|
||||
json res = json{
|
||||
{"content", content},
|
||||
{"stop", false},
|
||||
|
|
@ -1215,7 +1216,7 @@ static void log_server_request(const Request &req, const Response &res)
|
|||
});
|
||||
}
|
||||
|
||||
bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) {
|
||||
static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) {
|
||||
return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
|
||||
}
|
||||
|
||||
|
|
@ -1225,7 +1226,7 @@ bool is_at_eob(llama_server_context & server_context, const llama_token * tokens
|
|||
// * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
|
||||
// This is also called when the stop condition is met.
|
||||
// Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
|
||||
void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
|
||||
static void beam_search_callback(void *callback_data, llama_beams_state beams_state) {
|
||||
auto & llama = *static_cast<llama_server_context*>(callback_data);
|
||||
// Mark beams as EOS as needed.
|
||||
for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
|
||||
|
|
@ -1258,7 +1259,8 @@ struct token_translator {
|
|||
std::string operator()(const completion_token_output & cto) const { return (*this)(cto.tok); }
|
||||
};
|
||||
|
||||
void append_to_generated_text_from_generated_token_probs(llama_server_context & llama) {
|
||||
static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama)
|
||||
{
|
||||
auto & gtps = llama.generated_token_probs;
|
||||
auto translator = token_translator{llama.ctx};
|
||||
auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue