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

return earlier if there is only 1 candiate (i.e. max_entropy == 0)

Browse source
This commit is contained in:
l3utterfly 2024-01-22 22:16:05 +09:00
parent babb76a33a
commit 4e97bdb497
1 changed files with 8 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

16
llama.cpp
View file

@ -7786,6 +7786,14 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c
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();
// Calculate maximum possible entropy
float max_entropy = -logf(1.0f / candidates_p->size);
// Guard against division by zero
if (max_entropy == 0.0f) {
return;
}
llama_sample_softmax(ctx, candidates_p);
// Calculate entropy of the softmax probabilities
@ -7797,14 +7805,6 @@ 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
float normalized_entropy = entropy / max_entropy;