happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Avoid division by zero if truncation occurs

This commit is contained in:
Ed Addario 2025-08-19 22:34:01 +01:00
parent ee05d6bc0b
commit f22b3097eb
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 4 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
src/llama-quant.cpp
View File

@ -790,28 +790,24 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
} }
// Scale for the rows we didn't sample in this expert: multiply by stride-ish factor // Scale for the rows we didn't sample in this expert: multiply by stride-ish factor
const float scale_rows = rows_per_expert / std::max<int64_t>(1, rs); const float scale_rows = (float)rows_per_expert / std::max(1.0f, (float)rs);
total_err *= scale_rows; total_err *= scale_rows;
} }
return total_err; return total_err;
}; };
// Produce per-tensor candidate lists
std::vector<tensor_info> all; std::vector<tensor_info> all;
all.reserve(tensors.size()); all.reserve(tensors.size());
for (const auto * tw : tensors) { for (const auto * tw : tensors) {
// Temporary workers for dequantization
std::vector<std::thread> workers; std::vector<std::thread> workers;
workers.reserve(std::max(1, nthread)); workers.reserve(std::max(1, nthread));
ggml_tensor * t = tw->tensor; ggml_tensor * t = tw->tensor;
const std::string name = ggml_get_name(t); const std::string name = ggml_get_name(t);
if (!can_quantize(t)) { if (!can_quantize(t)) { continue; }
continue;
}
LLAMA_LOG_INFO("\t%s: - processing tensor %45s \t(%12d elements)\n", __func__, name.c_str(), (int)ggml_nelements(t)); LLAMA_LOG_INFO("\t%s: - processing tensor %45s \t(%12d elements)\n", __func__, name.c_str(), (int)ggml_nelements(t));
if (!ml.use_mmap) { if (!ml.use_mmap) {
@ -820,7 +816,6 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
} }
ml.load_data_for(t); ml.load_data_for(t);
// Prepare f32 weights for error estimates
const int64_t nelem = ggml_nelements(t); const int64_t nelem = ggml_nelements(t);
std::vector<no_init<float>> f32_conv_buf; std::vector<no_init<float>> f32_conv_buf;
float * f32_data = nullptr; float * f32_data = nullptr;
@ -955,13 +950,13 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
if (ti.choice >= (int)ti.candidate.size() - 1) { continue; } if (ti.choice >= (int)ti.candidate.size() - 1) { continue; }
int j = next_distinct_idx(ti); int j = next_distinct_idx(ti);
if (j < 0) { continue; } // no larger-size candidate remains if (j < 0) { continue; }
const auto &cur = ti.candidate[ti.choice]; const auto &cur = ti.candidate[ti.choice];
const auto &nxt = ti.candidate[j]; const auto &nxt = ti.candidate[j];
size_t delta_bytes = nxt.bytes - cur.bytes; size_t delta_bytes = nxt.bytes - cur.bytes;
if (delta_bytes == 0) { continue; } // should not happen after dedup, but be safe if (delta_bytes == 0) { continue; }
double err = (double)cur.error - (double)nxt.error; double err = (double)cur.error - (double)nxt.error;
err = std::max(err, 0.0); // do not penalize due to sampling noise err = std::max(err, 0.0); // do not penalize due to sampling noise