nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Log probabilities on token sampling crash (#1519) 

* Log probabilities on token sampling crash

* Typo
This commit is contained in:
Kawrakow 2026-03-26 14:34:41 +01:00 committed by GitHub
parent c06067fa14
commit e46601cb19
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 21 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
src/llama-sampling.cpp
View File

@ -10,6 +10,7 @@
#include <cfloat>
#include <numeric>
#include <unordered_map>
#include <fstream>
static void llama_log_softmax(float * array, size_t size) {
float max_l = *std::max_element(array, array + size);
@ -721,8 +722,27 @@ llama_token llama_sample_token_with_rng_impl(struct llama_sampling * smpl, llama
}
probs.back() += sump;
auto p = sump * rng() / rng.max();
auto r = rng();
auto p = sump * r / rng.max();
auto iter = std::upper_bound(probs.begin(), probs.end(), p);
if (iter == probs.end()) {
LLAMA_LOG_ERROR("=============================== Failed to sample token\n");
std::ofstream out("probabilities.txt");
out << "candidates->size: " << candidates->size << std::endl;
out << "max = " << max << std::endl;
out << "sump = " << sump << std::endl;
out << "r = " << r << std::endl;
out << "probabilities:\n";
for (int j = 0; j < candidates->size; ++j) {
out << j << " " << candidates->data[j].id << " " << candidates->data[j].logit << " " << probs[j] << std::endl;
}
out.flush();
out.close();
LLAMA_LOG_ERROR("Data has been stored in probabilities.txt\n");
LLAMA_LOG_ERROR("Create an issue with full log and attach probabilities.txt to the issue\n");
LLAMA_LOG_ERROR("\n\nCrashing now\n");
GGML_ABORT("Fatal error");
}
GGML_ASSERT(iter != probs.end());
auto idx = std::distance(probs.begin(), iter);
auto id = candidates->data[idx].id;