backend : offload large batches to GPU (#6083)

* backend : offload large batches to GPU * fix hip * code cleanup * fix CUDA split buffers * Update ggml-backend-impl.h Co-authored-by: Johannes Gäßler <[email protected]> * cuda : fix memset without set_device * imatrix : remove sched affix from weight names * sched : add a new split if the current one has too many inputs reduce max inputs per split more cleanup * update backends ggml-ci --------- Co-authored-by: Johannes Gäßler <[email protected]>
ggerganov · Mar 18, 2024 · 2bf8d0f · 2bf8d0f · whoreson · Mar 21, 2024
1 parent 496bc79
commit 2bf8d0f
Show file tree

Hide file tree

Showing 14 changed files with 349 additions and 396 deletions.
diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp
@@ -56,13 +56,31 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
 
+    std::string wname;
+    {
+        // remove any prefix and suffixes from the name
+        // CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+        const char * p = strchr(src0->name, '#');
+        if (p != NULL) {
+            p = p + 1;
+            const char * q = strchr(p, '#');
+            if (q != NULL) {
+                wname = std::string(p, q - p);
+            } else {
+                wname = p;
+            }
+        } else {
+            wname = src0->name;
+        }
+    }
+
     // when ask is true, the scheduler wants to know if we are interested in data from this tensor
     // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
     if (ask) {
         if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
         if (t->op != GGML_OP_MUL_MAT) return false;
         if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
-        if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false;
+        if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
         return true;
     }
 
@@ -94,20 +112,20 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         // this is necessary to guarantee equal number of "ncall" for each tensor
         for (int ex = 0; ex < n_as; ++ex) {
             src0 = t->src[2 + ex];
-            auto& e = m_stats[src0->name];
+            auto& e = m_stats[wname];
             if (e.values.empty()) {
                 e.values.resize(src1->ne[0], 0);
             }
             else if (e.values.size() != (size_t)src1->ne[0]) {
-                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
                 exit(1); //GGML_ASSERT(false);
             }
             // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
             //       using the following line, we can correct for that if needed
             //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
             ++e.ncall;
             if (m_params.verbosity > 1) {
-                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
             }
             for (int row = 0; row < (int)src1->ne[1]; ++row) {
                 const int excur = m_ids[row*n_as + idx];
@@ -129,17 +147,17 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             }
         }
     } else {
-        auto& e = m_stats[src0->name];
+        auto& e = m_stats[wname];
         if (e.values.empty()) {
             e.values.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
             exit(1); //GGML_ASSERT(false);
         }
         ++e.ncall;
         if (m_params.verbosity > 1) {
-            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
         }
         for (int row = 0; row < (int)src1->ne[1]; ++row) {
             const float * x = data + row * src1->ne[0];

diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
@@ -114,10 +114,10 @@ static std::string get_cpu_info() {
 static std::string get_gpu_info() {
     std::string id;
 #ifdef GGML_USE_CUBLAS
-    int count = ggml_cuda_get_device_count();
+    int count = ggml_backend_cuda_get_device_count();
     for (int i = 0; i < count; i++) {
         char buf[128];
-        ggml_cuda_get_device_description(i, buf, sizeof(buf));
+        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
         id += buf;
         if (i < count - 1) {
             id += "/";

diff --git a/ggml-alloc.c b/ggml-alloc.c
@@ -548,7 +548,11 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
-        if (ggml_is_view(node)) {
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
             struct ggml_tensor * view_src = node->view_src;
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
@@ -565,8 +569,8 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
 
             ggml_gallocr_hash_get(galloc, src)->n_children += 1;
 
-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
                 ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
             }
         }

diff --git a/ggml-backend-impl.h b/ggml-backend-impl.h
@@ -103,6 +103,11 @@ extern "C" {
         // check if the backend supports an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
 
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
         // (optional) event synchronization
         ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
         void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);