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
#	.gitignore
#	CMakeLists.txt
#	Makefile
#	README.md
#	ci/run.sh
#	ggml-opencl.cpp
#	tests/CMakeLists.txt
This commit is contained in:
Concedo
2024-01-27 11:45:07 +08:00
parent 762eeb6204 a1d6df129b
commit ed09a854f0
27 changed files with 1335 additions and 979 deletions
Download Patch File Download Diff File Expand all files Collapse all files
llama.cpp
+115 -64
View File
@@ -8269,10 +8269,57 @@ void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * can
auto comp = [](const llama_token_data & a, const llama_token_data & b) {
return a.logit > b.logit;
};
if (k == (int) candidates->size) {
std::sort(candidates->data, candidates->data + candidates->size, comp);
} else {
if (k <= 128) {
std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp);
} else {
constexpr int nbuckets = 128;
constexpr float bucket_low = -10.0f;
constexpr float bucket_high = 10.0f;
constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
constexpr float bucker_inter = -bucket_low * bucket_scale;
std::vector<int> bucket_idx(candidates->size);
std::vector<int> histo(nbuckets, 0);
for (int i = 0; i < (int)candidates->size; ++i) {
const float val = candidates->data[i].logit;
int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
ib = std::max(0, std::min(nbuckets-1, ib));
bucket_idx[i] = ib;
++histo[ib];
}
int nhave = 0;
int ib = nbuckets - 1;
for ( ; ib >= 0; --ib) {
nhave += histo[ib];
if (nhave >= k) break;
}
std::vector<llama_token_data> tmp_tokens(nhave);
auto ptr = tmp_tokens.data();
std::vector<llama_token_data*> bucket_ptrs;
bucket_ptrs.reserve(nbuckets - ib);
for (int j = nbuckets - 1; j >= ib; --j) {
bucket_ptrs.push_back(ptr);
ptr += histo[j];
}
for (int i = 0; i < (int)candidates->size; ++i) {
int j = bucket_idx[i];
if (j >= ib) {
*bucket_ptrs[nbuckets-1-j]++ = candidates->data[i];
}
}
ptr = tmp_tokens.data();
int ndone = 0;
for (int j = nbuckets-1; j > ib; --j) {
std::sort(ptr, ptr + histo[j], comp);
ptr += histo[j];
ndone += histo[j];
}
std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp);
std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data));
}
candidates->sorted = true;
}
@@ -8467,29 +8514,18 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c
}
}
void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * candidates_p, float min_temp, float max_temp, float exponent_val) {
const int64_t t_start_sample_us = ggml_time_us();
for (size_t i = 0; i < candidates_p->size; ++i) {
candidates_p->data[i].logit /= temp;
// no need to do anything if there is only one (or zero) candidates
if(candidates_p->size <= 1) {
return;
}
if (ctx) {
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
}
}
// Calculate maximum possible entropy
float max_entropy = -logf(1.0f / candidates_p->size);
void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
llama_sample_temp(ctx, candidates_p, temp);
}
void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp, float min_temp = 0, float max_temp = 2.0f, float dynatemp_exponent = 1.0f) {
const int64_t t_start_sample_us = ggml_time_us();
llama_sample_softmax(ctx, candidates_p);
float exponent_val = dynatemp_exponent;
llama_sample_softmax(nullptr, candidates_p);
// Calculate entropy of the softmax probabilities
float entropy = 0.0f;
@@ -8500,28 +8536,20 @@ void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * c
}
}
// Calculate maximum possible entropy
float max_entropy = -logf(1.0f / candidates_p->size);
// Guard against division by zero
if (max_entropy == 0.0f) {
max_entropy = 1.0f; // This ensures that normalized_entropy will be 0 when entropy is 0
}
// Normalize the entropy
// Normalize the entropy (max_entropy cannot be 0 here because we checked candidates_p->size != 1 above)
float normalized_entropy = entropy / max_entropy;
// Map the normalized entropy to the desired temperature range using the power function
float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val);
// //todo: Ensure to hide print statements unless debugging!
// printf("Your text maxtemp value is: %f\n", max_temp);
// // Print the variables
// printf("Entropy: %f\n", entropy);
// printf("Max Possible Entropy: %f\n", max_entropy);
// printf("Normalized Entropy: %f\n", normalized_entropy);
// printf("Exponent: %f\n", exponent_val);
// printf("Dynamic Temperature (dyn_temp): %f\n", dyn_temp);
#ifdef DEBUG
LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp);
LLAMA_LOG_INFO("Entropy: %f\n", entropy);
LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy);
LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy);
LLAMA_LOG_INFO("Exponent: %f\n", exponent_val);
LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp);
#endif
// Apply the dynamically calculated temperature scaling
for (size_t i = 0; i < candidates_p->size; ++i) {
@@ -8540,18 +8568,36 @@ void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * c
candidates_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities
}
// //todo: Ensure to hide print statements unless debugging!
// // Print the updated top 25 probabilities after temperature scaling
// printf("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n");
// for (size_t i = 0; i < 25 && i < candidates_p->size; ++i) {
// printf("Token %zu: %f%%\n", i + 1, candidates_p->data[i].p * 100.0f);
// }
#ifdef DEBUG
// Print the updated top 25 probabilities after temperature scaling
LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n");
for (size_t i = 0; i < 25 && i < candidates_p->size; ++i) {
LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, candidates_p->data[i].p * 100.0f);
}
#endif
if (ctx) {
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
}
}
void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
const int64_t t_start_sample_us = ggml_time_us();
for (size_t i = 0; i < candidates_p->size; ++i) {
candidates_p->data[i].logit /= temp;
}
if (ctx) {
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
}
}
void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
llama_sample_temp(ctx, candidates_p, temp);
}
// The llama.cpp repetition penalty code goes unused in kobold's API
void llama_sample_repetition_penalties(
@@ -9216,6 +9262,23 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
auto use_more_bits = [](int i_layer, int num_layers) -> bool {
return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
};
const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) {
if (n_expert > 1) {
// Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly
// sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
// for getting the current layer as I initially thought, and we need to resort to parsing the
// tensor name.
n_layer /= n_expert;
if (sscanf(name, "blk.%d.", &i_layer) != 1) {
throw std::runtime_error(format("Failed to determine layer for tensor %s", name));
}
if (i_layer < 0 || i_layer >= n_layer) {
throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name, n_layer));
}
}
return std::make_pair(i_layer, n_layer);
};
if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
int nx = tensor->ne[0];
@@ -9277,24 +9340,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
new_type = GGML_TYPE_Q2_K;
}
} else if (name.find("ffn_down") != std::string::npos) {
const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
int i_layer, n_layer;
if (n_expert == 1) {
i_layer = qs.i_ffn_down;
n_layer = qs.n_ffn_down;
} else {
// Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly
// sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
// for getting the current layer as I initially thought, and we need to resort to parsing the
// tensor name.
n_layer = qs.n_ffn_down / n_expert;
if (sscanf(name.c_str(), "blk.%d.ffn_down", &i_layer) != 1) {
throw std::runtime_error(format("Failed to determine layer for tensor %s", name.c_str()));
}
if (i_layer < 0 || i_layer >= n_layer) {
throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name.c_str(), n_layer));
}
}
auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
@@ -9350,13 +9397,17 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
}
else if (name.find("ffn_gate") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_gate, qs.n_ffn_gate)) {
auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
new_type = GGML_TYPE_Q2_K;
}
++qs.i_ffn_gate;
}
else if (name.find("ffn_up") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_up, qs.n_ffn_up)) {
auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
new_type = GGML_TYPE_Q2_K;
}
++qs.i_ffn_up;