Simplify and improve CUDA graphs through use of indirect copy pointers (#9017)

* CUDA: Simplify and improve CUDA graphs through use of indirect copy pointers

Previously there was complexity in the CUDA graphs implementation due
frequently changing parameters to copy kernels associated with K and V
cache pointers. This patch simplifies by using indirection to avoid
such parameters frequently changing, avoiding the need for frequent
graph updates.

Fixes #12152

* Addressed comments

* fix HIP builds

* properly sync to stream

* removed ggml_cuda_cpy_fn_ptrs

* move stream sync before free

* guard to only use indirection with graphs

* style fixes

* check for errors

---------

Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Alan Gray 2025-04-03 02:31:15 +01:00 committed by GitHub
parent 2a0dc97e56
commit 3f9da22c2b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 122 additions and 121 deletions

View file

@ -2441,10 +2441,11 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
#ifdef USE_CUDA_GRAPH
static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool use_cuda_graph) {
bool use_cuda_graph) {
// Loop over nodes in GGML graph to obtain info needed for CUDA graph
cuda_ctx->cuda_graph->updated_kernel_arg.clear();
cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
@ -2476,8 +2477,11 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
}
if (node->op == GGML_OP_CPY) {
// store the copy op parameter which changes with each token.
cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
// Store the pointers which are updated for each token, such that these can be sent
// to the device and accessed using indirection from CUDA graph
cuda_ctx->cuda_graph->cpy_dest_ptrs.push_back((char *) node->src[1]->data);
// store a pointer to each copy op CUDA kernel to identify it later
void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
if (!ptr) {
@ -2485,10 +2489,6 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
#endif
} else {
if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
ggml_cuda_cpy_fn_ptrs.push_back(ptr);
}
}
}
@ -2497,6 +2497,12 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
}
}
if (use_cuda_graph) {
cuda_ctx->cuda_graph->use_cpy_indirection = true;
// copy pointers to GPU so they can be accessed via indirection within CUDA graph
ggml_cuda_cpy_dest_ptrs_copy(cuda_ctx->cuda_graph.get(), cuda_ctx->cuda_graph->cpy_dest_ptrs.data(), cuda_ctx->cuda_graph->cpy_dest_ptrs.size(), cuda_ctx->stream());
}
return use_cuda_graph;
}
@ -2551,51 +2557,6 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
return true;
}
static void maintain_cuda_graph(ggml_backend_cuda_context * cuda_ctx, std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool cuda_graph_update_required) {
if (cuda_graph_update_required) {
// Extract nodes from graph
// First call with null argument gets number of nodes in graph
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
// Subsequent call with non-null argument gets nodes
cuda_ctx->cuda_graph->nodes.clear();
cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
cuda_ctx->cuda_graph->params.clear();
cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
if (cuda_ctx->cuda_graph->num_nodes > 0) {
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
// Loop over nodes, and extract kernel parameters from each node
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
cudaGraphNodeType node_type;
CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
if (node_type == cudaGraphNodeTypeKernel) {
cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
if (stat == cudaErrorInvalidDeviceFunction) {
// Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
// We don't need to update blas nodes, so clear error and move on.
(void)cudaGetLastError();
} else {
GGML_ASSERT(stat == cudaSuccess);
}
}
}
}
} else {
// One of the arguments to the copy kernel is updated for each token, hence we need to
// replace that argument with the updated value in the CUDA graph
// on update steps, the live parameters will already be captured
int k = 0;
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
*(void**)cuda_ctx->cuda_graph->params[i].kernelParams[1] = *(void**)updated_kernel_arg_ptr;
CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
}
}
}
}
static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
bool cuda_graph_update_required = false;
@ -2655,8 +2616,7 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
#endif
static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
[[maybe_unused]] std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool & graph_evaluated_or_captured, bool & use_cuda_graph,
bool & cuda_graph_update_required) {
bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
while (!graph_evaluated_or_captured) {
// Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
@ -2706,13 +2666,9 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
}
// Perform update to graph (if required for this token), and change copy parameter (required for every token)
maintain_cuda_graph(cuda_ctx, ggml_cuda_cpy_fn_ptrs, cuda_graph_update_required);
// Update graph executable
update_cuda_graph_executable(cuda_ctx);
if (cuda_graph_update_required) { // Update graph executable
update_cuda_graph_executable(cuda_ctx);
}
// Launch graph
CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
#else
@ -2726,10 +2682,6 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
ggml_cuda_set_device(cuda_ctx->device);
// vector of pointers to CUDA cpy kernels, which are required to identify
// kernel parameters which need updated in the graph for each token
std::vector<void *> ggml_cuda_cpy_fn_ptrs;
#ifdef USE_CUDA_GRAPH
static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
@ -2763,8 +2715,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
if (use_cuda_graph) {
cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph,
ggml_cuda_cpy_fn_ptrs, use_cuda_graph);
use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, use_cuda_graph);
// Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
if (use_cuda_graph && cuda_graph_update_required) {
@ -2785,6 +2736,10 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
}
if (!use_cuda_graph) {
cuda_ctx->cuda_graph->use_cpy_indirection = false;
}
#else
bool use_cuda_graph = false;
bool cuda_graph_update_required = false;
@ -2792,7 +2747,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
bool graph_evaluated_or_captured = false;
evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, ggml_cuda_cpy_fn_ptrs, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
return GGML_STATUS_SUCCESS;
}