imatrix : fix 3d activation handling for hybrid and recurrent models (#14994)

* imatrix : use a single count for dense 3d tensors

* imatrix : fix 3d activations when model tensor is 2d

* imatrix : fix 3d tensor counts

diff --git a/tools/imatrix/imatrix.cpp b/tools/imatrix/imatrix.cpp
index 9aad3711bae5447ffbbddd7802f8604f50d47295..f5262e5e83da9bacd05eae3c818347e4271e118a 100644 (file)
--- a/tools/imatrix/imatrix.cpp
+++ b/tools/imatrix/imatrix.cpp
@@ -250,13 +250,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     const char * data = is_host ? (const char *) src1->data : m_src1_data.data();
     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
     // ref: https://github.com/ggml-org/llama.cpp/pull/6387
     if (t->op == GGML_OP_MUL_MAT_ID) {
@@ -272,6 +265,12 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
         GGML_ASSERT(ids->ne[1] == src1->ne[2]);
 
+        // 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));
         ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
 
@@ -335,29 +334,40 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
     } else {
         auto & e = m_stats[wname];
-        const int64_t n_mat = src1->ne[2] * src1->ne[3];
-
+        const int64_t n_mat = src0->ne[2] * src0->ne[3];
+
+        // use a single count per dense tensor
+        // (necessary when merging older GGUF-imatrix files with 3d tensors)
+        if (e.counts.size() > 1) {
+            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()) {
             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)) {
             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");
         }
-        else if (e.counts.size() != (size_t)n_mat) {
-            LOG_ERR("%s: inconsistent expert count for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.counts.size(), (int)n_mat);
-            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);
+
         for (int64_t i3 = 0; i3 < src1->ne[3]; ++i3) {
             for (int64_t i2 = 0; i2 < src1->ne[2]; ++i2) {
-                const int64_t mat_id = i3 * src1->ne[2] + i2;
+                // handle 3D+ tensors, but flatten 3D+ activations when model tensor is 2D
+                const int64_t mat_id = (i3 % src0->ne[3]) * src0->ne[2] + (i2 % src0->ne[2]);
                 const int64_t mat_start = mat_id * src1->ne[0];
 
                 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]);
-                    e.counts[mat_id]++;
+                    const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->nb[3]);
                     for (int64_t j = 0; j < src1->ne[0]; ++j) {
                         e.values[mat_start + j] += x[j] * x[j];
                         if (!std::isfinite((float)e.values[j])) {
@@ -366,16 +376,20 @@ 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;
-                    m_last_chunk = n_chunk;
-                    if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {
-                        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);
-                    }
+            }
+        }
+        // only 1 count in practice, except when a tensor is used for both MUL_MAT_ID and MUL_MAT
+        for (size_t i = 0; i < e.counts.size(); ++i) {
+            e.counts[i] += ggml_nrows(src1) / n_mat;
+            const int32_t n_chunk = e.counts[i] / chunk_size;
+            if (n_chunk > m_last_chunk) {
+                const int32_t chunk_step = n_chunk - m_last_chunk;
+                m_last_chunk = n_chunk;
+                if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {
+                    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);
                 }
             }
         }