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

Remove storing tensor statistics

This commit is contained in:
Ed Addario 2025-11-23 22:18:12 +00:00
parent 4cfddea9c3
commit 4a0511f3a0
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 8 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
tools/imatrix/imatrix.cpp
View File

@ -757,28 +757,11 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.activations.size(), GGML_MEM_ALIGN);
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.values.size(), GGML_MEM_ALIGN);
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.counts.size(), GGML_MEM_ALIGN);
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * 4, GGML_MEM_ALIGN);
}
// deterministic tensor name order
std::sort(to_store.begin(), to_store.end());
// Compute per-tensor statistics (CosSim, L2 Dist, ECS) to store alongside sums
std::vector<tensor_statistics> tstats;
tstats.reserve(m_stats.size());
bool legacy_tensor = true;
for (const auto & kv : m_stats) {
if (!compute_vector_statistics(tstats, kv.first, kv.second, legacy_tensor)) {
LOG_WRN("%s: tensor %s has no data - skipping\n", __func__, kv.first.c_str());
}
}
if (!tstats.empty()) { compute_tensor_statistics(tstats); }
// index by tensor name
std::unordered_map<std::string, const tensor_statistics *> tstat_index;
tstat_index.reserve(tstats.size());
for (const auto & ts : tstats) { tstat_index[ts.tensor] = &ts; }
struct ggml_init_params params = {
/* .mem_size = */ data_size,
/* .mem_buffer = */ NULL,
@ -835,29 +818,6 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
gguf_add_tensor(ctx_gguf, in_sum);
}
}
// Store per-tensor statistics as a small 1D tensor: [ECS, L2 Dist, CosSim, ZD Score]
{
float l2 = 0.0f;
float cs = 0.0f;
float zd = 0.0f;
float ecs = 0.0f;
auto it_ts = tstat_index.find(name);
if (it_ts != tstat_index.end() && it_ts->second != nullptr) {
l2 = it_ts->second->l2_dist;
cs = it_ts->second->cossim;
zd = it_ts->second->zd_score;
ecs = 100.0f * (1.0f - std::exp(-0.01f * l2) * std::pow(std::fabs(cs), 10.0f));
}
struct ggml_tensor * stats_t = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4);
ggml_format_name(stats_t, "%s.stats", name.c_str());
((float *)stats_t->data)[0] = ecs;
((float *)stats_t->data)[1] = l2;
((float *)stats_t->data)[2] = cs;
((float *)stats_t->data)[3] = zd;
gguf_add_tensor(ctx_gguf, stats_t);
}
}
gguf_write_to_file(ctx_gguf, fname.c_str(), false);
@ -1431,10 +1391,16 @@ static bool show_statistics(const common_params & params) {
"=============================================================================================================="
"=============================================================\n");
// Euclidean-Cosine score
auto ecs = [](const float l2_dist, const float cossim) {
return 100.0f - (100.0f * (1.0f / (1.0f + ((2.0f / 3.0f) * l2_dist * l2_dist))) * ((1 + cossim) * 0.5f));
};
for (const auto & tstat : ts) {
std::string layer;
std::string name;
process_tensor_name(tstat.tensor, layer, name);
const float h_norm = tstat.elements > 1 ? 100.0f * (tstat.entropy / std::log2((float) tstat.elements)) : 0.0f;
int blk;
try {
@ -1443,9 +1409,6 @@ static bool show_statistics(const common_params & params) {
blk = -1; // not a block layer
}
const float h_norm = tstat.elements > 1 ? 100.0f * (tstat.entropy / std::log2((float) tstat.elements)) : 0.0f;
const float ecs = 100.0f * (1.0f - std::exp(-0.01f * tstat.l2_dist) * std::pow(std::fabs(tstat.cossim), 10.0f)); // Euclidean-Cosine score
LOG_INF("%5s\t%-20s\t%11.4f\t%10.4f\t%10.4f\t%8.4f\t%8.4f\t%7d\t%10.2f%%\t%10.4f\t%6.2f%%\t%10.4f\n",
layer.c_str(),
name.c_str(),
@ -1456,7 +1419,7 @@ static bool show_statistics(const common_params & params) {
tstat.std_deviation,
tstat.elements,
h_norm,
legacy ? tstat.entropy : ecs,
legacy ? tstat.entropy : ecs(tstat.l2_dist, tstat.cossim),
100.0f * tstat.zd_score,
tstat.cossim);
@ -1516,13 +1479,12 @@ static bool show_statistics(const common_params & params) {
100.0f * stats.layer_zd / stats.n,
layer_cs);
} else {
const float layer_ecs = 100.0f * (1.0f - std::exp(-0.01f * layer_l2n) * std::pow(std::fabs(layer_cs), 10.0f));
LOG_INF("%5d\t%11.4f\t%6.2f%%\t%11.4f\t%8.4f\n",
layer,
layer_l2n,
100.0f * stats.layer_zd / stats.n,
layer_cs,
layer_ecs);
ecs(layer_l2n, layer_cs));
}
}
LOG_INF("\n");