CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-03 09:22:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

imatrix : use a single count for dense 3d tensors

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-07-19 12:57:57 -04:00
parent 90083283ec
commit 73439beb1b
1 changed files with 36 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
tools/imatrix/imatrix.cpp
View File

@@ -112,13 +112,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
const char * data = is_host ? (const char *) src1->data : m_src1_data.data(); const char * data = is_host ? (const char *) src1->data : m_src1_data.data();
GGML_ASSERT(src1->nb[0] == ggml_element_size(src1)); GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
// TODO: 4d? (is that even used in practice?)
// the extra dimension would need to be stored somewhere to be reflected in the imatrix file
if (ggml_nrows(src1) != src1->ne[1] * src1->ne[2]) {
LOG_ERR("%s: tensor has more than 3 dimensions: %s", __func__, wname.c_str());
GGML_ASSERT(false);
}
// this has been adapted to the new format of storing merged experts in a single 3d tensor // this has been adapted to the new format of storing merged experts in a single 3d tensor
// ref: https://github.com/ggml-org/llama.cpp/pull/6387 // ref: https://github.com/ggml-org/llama.cpp/pull/6387
if (t->op == GGML_OP_MUL_MAT_ID) { if (t->op == GGML_OP_MUL_MAT_ID) {
@@ -134,6 +127,13 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
GGML_ASSERT(ids->ne[1] == src1->ne[2]); GGML_ASSERT(ids->ne[1] == src1->ne[2]);
// TODO: 4d? (is that even used in practice?)
// the extra dimension would need to be stored somewhere to be reflected in the imatrix file
if (ggml_nrows(src1) != src1->ne[1] * src1->ne[2]) {
LOG_ERR("%s: tensor has more than 3 dimensions: %s", __func__, wname.c_str());
GGML_ASSERT(false);
}
m_ids.resize(ggml_nbytes(ids)); m_ids.resize(ggml_nbytes(ids));
ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids)); ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
@@ -199,19 +199,33 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
auto & e = m_stats[wname]; auto & e = m_stats[wname];
const int64_t n_mat = src1->ne[2] * src1->ne[3]; const int64_t n_mat = src1->ne[2] * src1->ne[3];
// use a single count per dense tensor
if ((int64_t) e.counts.size() == n_mat) {
bool all_equal = true;
for (size_t i = 1; i < e.counts.size(); ++i) {
if (e.counts[0] != e.counts[i]) {
all_equal = false;
break;
}
}
if (all_equal) {
e.counts.resize(1);
}
}
if (e.values.empty()) { if (e.values.empty()) {
e.values.resize(src1->ne[0] * n_mat, 0); e.values.resize(src1->ne[0] * n_mat, 0);
e.counts.resize(n_mat, 0); e.counts.resize(1, 0);
} }
else if (e.values.size() != (size_t)(src1->ne[0] * n_mat)) { else if (e.values.size() != (size_t)(src1->ne[0] * n_mat)) {
LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)(src1->ne[0] * n_mat)); LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)(src1->ne[0] * n_mat));
exit(1); //GGML_ABORT("fatal error"); exit(1); //GGML_ABORT("fatal error");
} }
else if (e.counts.size() != (size_t)n_mat) { else if (e.counts.size() != 1) {
LOG_ERR("%s: inconsistent expert count for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.counts.size(), (int)n_mat); LOG_ERR("%s: inconsistent matrix count for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.counts.size(), 1);
exit(1); //GGML_ABORT("fatal error"); exit(1); //GGML_ABORT("fatal error");
} }
LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->ne[2], (int)src1->type); LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->ne[2], (int)src1->type);
for (int64_t i3 = 0; i3 < src1->ne[3]; ++i3) { for (int64_t i3 = 0; i3 < src1->ne[3]; ++i3) {
for (int64_t i2 = 0; i2 < src1->ne[2]; ++i2) { for (int64_t i2 = 0; i2 < src1->ne[2]; ++i2) {
const int64_t mat_id = i3 * src1->ne[2] + i2; const int64_t mat_id = i3 * src1->ne[2] + i2;
@@ -219,7 +233,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
for (int64_t row = 0; row < src1->ne[1]; ++row) { for (int64_t row = 0; row < src1->ne[1]; ++row) {
const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->ne[3]); const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->ne[3]);
e.counts[mat_id]++;
for (int64_t j = 0; j < src1->ne[0]; ++j) { for (int64_t j = 0; j < src1->ne[0]; ++j) {
e.values[mat_start + j] += x[j] * x[j]; e.values[mat_start + j] += x[j] * x[j];
if (!std::isfinite((float)e.values[j])) { if (!std::isfinite((float)e.values[j])) {
@@ -228,17 +241,18 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
} }
} }
} }
const int32_t n_chunk = e.counts[mat_id] / chunk_size; }
if (n_chunk > m_last_chunk) { }
const int32_t chunk_step = n_chunk - m_last_chunk; e.counts[0] += src1->ne[1];
m_last_chunk = n_chunk; const int32_t n_chunk = e.counts[0] / chunk_size;
if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) { if (n_chunk > m_last_chunk) {
save_imatrix(); const int32_t chunk_step = n_chunk - m_last_chunk;
} m_last_chunk = n_chunk;
if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) { if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {
save_imatrix(m_last_chunk); save_imatrix();
} }
} if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) {
save_imatrix(m_last_chunk);
} }
} }
} }