Skip to content

cuda : avoid repeated warmup after stable graph replay #29768

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
joelagnel wants to merge 1 commit into
base: master
Choose a base branch
from
Draft

cuda : avoid repeated warmup after stable graph replay #29768

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions ggml/src/ggml-cuda/common.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1276,6 +1276,7 @@ struct ggml_cuda_graph {
std::vector<cudaGraphNode_t> nodes;
bool disable_due_to_gpu_arch = false;
bool warmup_complete = false;
bool replayed_since_capture = false;
uint64_t uid = 0;
int64_t last_used_time = 0;
struct node_properties {
Expand Down
46 changes: 35 additions & 11 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2727,8 +2727,9 @@ static const void * ggml_cuda_graph_get_key(ggml_cgraph * cgraph) {
return cgraph->nodes[0];
}

static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, bool * nodes_changed) {
bool res = false;
*nodes_changed = false;

const void * graph_key = ggml_cuda_graph_get_key(cgraph);
ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
Expand All @@ -2747,6 +2748,7 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
// Check if the graph size has changed
if ((int)graph->node_props.size() != cgraph->n_nodes) {
res = true;
*nodes_changed = true;
graph->node_props.resize(cgraph->n_nodes);
}

Expand All @@ -2763,6 +2765,9 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
}

if (res || memcmp(&graph->node_props[i], &prop, sizeof(prop)) != 0) {
// a new op, or a node that starts or stops running a kernel, changes the CUDA graph topology
*nodes_changed |= graph->node_props[i].node.op != prop.node.op ||
ggml_cuda_is_view_or_noop(&graph->node_props[i].node) != ggml_cuda_is_view_or_noop(&prop.node);
graph->node_props[i] = prop;
res = true;
}
Expand Down Expand Up @@ -4645,10 +4650,15 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
} else {
graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
}
#else
GGML_UNUSED(graph_key);
graph_evaluated_or_captured = true;
#endif // USE_CUDA_GRAPH
}

#ifdef USE_CUDA_GRAPH
ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
if (use_cuda_graph) {
ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
if (graph->instance == nullptr) { // Create executable graph from captured graph.
CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
}
Expand All @@ -4657,11 +4667,10 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
}
// Launch graph
CUDA_CHECK(cudaGraphLaunch(graph->instance, cuda_ctx->stream()));
#else
GGML_UNUSED(graph_key);
graph_evaluated_or_captured = true;
#endif // USE_CUDA_GRAPH
}

graph->replayed_since_capture = use_cuda_graph && !cuda_graph_update_required;
#endif // USE_CUDA_GRAPH
}

#ifdef USE_CUDA_GRAPH
Expand Down Expand Up @@ -4699,7 +4708,8 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
if (graph->is_enabled()) {
const bool graph_compatible = ggml_cuda_graph_check_compability(cgraph);
if (graph_compatible) {
const bool properties_changed = ggml_cuda_graph_update_required(cuda_ctx, cgraph);
bool nodes_changed = false;
const bool properties_changed = ggml_cuda_graph_update_required(cuda_ctx, cgraph, &nodes_changed);

if (!graph->warmup_complete) {
// Warmup: need at least 2 calls with no property change on the 2nd call
Expand All @@ -4715,10 +4725,24 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
} else {
// Post-warmup: normal CUDA graph operation
if (properties_changed) {
// Properties changed - reset warmup, execute directly until stable again
graph->warmup_complete = false;
if (ggml_cuda_graph_debug_enabled()) {
GGML_LOG_DEBUG("%s: CUDA graph warmup reset\n", __func__);
// Legacy pools can synchronize on OOM, so changed shapes need warmup before capture.
bool recapture_safe = false;
#if defined(GGML_USE_VMM)
recapture_safe = true;
for (int device = 0; device < ggml_cuda_info().device_count; ++device) {
recapture_safe &= ggml_cuda_info().devices[device].vmm;
}
#endif // defined(GGML_USE_VMM)
// recapture at once if no node changes op or starts or stops running
if (recapture_safe && graph->instance != nullptr && graph->replayed_since_capture && !nodes_changed) {
use_cuda_graph = true;
cuda_graph_update_required = true;
} else {
// Reset warmup when properties keep changing or nodes change.
graph->warmup_complete = false;
if (ggml_cuda_graph_debug_enabled()) {
GGML_LOG_DEBUG("%s: CUDA graph warmup reset\n", __func__);
}
}
} else {
use_cuda_graph = true;
Expand Down