jdelony/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-06-27 23:50:20 -05:00
Code Issues Packages Projects Releases Wiki Activity

clean up logging and timing

Browse Source
This commit is contained in:
Ruben Ortlam 2026-06-17 13:47:53 +02:00
parent 71d9373b82
commit 42874dfd8f
1 changed files with 6 additions and 51 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -2075,12 +2075,6 @@ struct ggml_backend_vk_context {
int64_t last_graph_eviction_sweep = 0;
vk_cached_graph * active_capture_graph = nullptr;
// Timing for graph compute debug logging
int64_t graph_compute_t_start = 0;
int64_t graph_compute_cpu_us = 0;
int n_graph_nodes = 0;
const char * graph_compute_phase = nullptr;
static const bool disable_vk_graphs_due_to_env;
vk_cached_graph * vk_graph(const void * first_node_ptr) {
@ -2126,7 +2120,6 @@ struct ggml_backend_vk_context {
};
const bool ggml_backend_vk_context::disable_vk_graphs_due_to_env = (getenv("GGML_VK_DISABLE_GRAPHS") != nullptr);
static const bool vk_graph_log_enabled = (getenv("GGML_VK_GRAPH_LOG") != nullptr);
static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000; // NOLINT
@ -15262,13 +15255,7 @@ static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
VK_LOG_DEBUG("ggml_backend_vk_synchronize()");
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
const int64_t t_sync_start = vk_graph_log_enabled ? ggml_time_us() : 0;
ggml_vk_synchronize(ctx);
if (vk_graph_log_enabled && ctx->graph_compute_phase) {
const int64_t t_sync = ggml_time_us() - t_sync_start;
std::cerr << "ggml_vulkan: graph (" << ctx->n_graph_nodes << " nodes) — " << ctx->graph_compute_phase << " | cpu " << ctx->graph_compute_cpu_us << " us, sync " << t_sync << " us, total " << (ggml_time_us() - ctx->graph_compute_t_start) << " us" << std::endl;
}
ctx->graph_compute_phase = nullptr;
ggml_vk_graph_cleanup(ctx);
}
@ -15923,7 +15910,6 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
bool use_vk_graph = false;
bool vk_graph_capture = false;
vk_cached_graph * cached = nullptr;
const int64_t t_start = vk_graph_log_enabled ? ggml_time_us() : 0;
if (cgraph->n_nodes > 0 && !ctx->disable_vk_graphs_due_to_env
#ifdef GGML_VULKAN_CHECK_RESULTS
@ -15939,20 +15925,17 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
if (!cached->warmup_complete) {
if (!properties_changed) {
// Properties stable for the second time — capture
cached->warmup_complete = true;
use_vk_graph = true;
vk_graph_capture = true;
} else {
if (vk_graph_log_enabled) { std::cerr << "ggml_vulkan: graph " << graph_key << " (" << cgraph->n_nodes << " nodes) — warmup" << std::endl; }
VK_LOG_DEBUG("ggml_vulkan: graph " << graph_key << " (" << cgraph->n_nodes << " nodes) - warmup");
}
} else {
if (properties_changed) {
// Graph changed — reset warmup
cached->destroy(ctx->device->device);
if (vk_graph_log_enabled) { std::cerr << "ggml_vulkan: graph " << graph_key << " (" << cgraph->n_nodes << " nodes) — invalidated" << std::endl; }
VK_LOG_DEBUG("ggml_vulkan: graph " << graph_key << " (" << cgraph->n_nodes << " nodes) - invalidated");
} else if (!cached->submissions.empty()) {
// Replay cached graph
use_vk_graph = true;
vk_graph_capture = false;
}
@ -15974,22 +15957,11 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
ctx->do_add_rms_partials_offset_calculation = false;
if (ggml_vk_graph_replay(ctx, cached)) {
if (vk_graph_log_enabled) {
ctx->graph_compute_cpu_us = ggml_time_us() - t_start;
ctx->graph_compute_t_start = t_start;
ctx->n_graph_nodes = cgraph->n_nodes;
ctx->graph_compute_phase = "replayed";
}
VK_LOG_DEBUG("ggml_vulkan: graph (" << cgraph->n_nodes << " nodes, " << cached->submissions.size() << " submissions) - replayed");
if (!ctx->device->support_async) {
const int64_t t_sync_start = vk_graph_log_enabled ? ggml_time_us() : 0;
ggml_vk_synchronize(ctx);
ggml_vk_graph_cleanup(ctx);
if (vk_graph_log_enabled) {
const int64_t t_sync = ggml_time_us() - t_sync_start;
std::cerr << "ggml_vulkan: graph (" << cgraph->n_nodes << " nodes, " << cached->submissions.size() << " submissions) — replayed | cpu " << ctx->graph_compute_cpu_us << " us, sync " << t_sync << " us, total " << (ggml_time_us() - t_start) << " us" << std::endl;
ctx->graph_compute_phase = nullptr;
}
}
if (vk_instance.debug_utils_support) {
@ -15997,8 +15969,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
}
return GGML_STATUS_SUCCESS;
}
// Replay failed (prealloc buffer changed) — fall through to normal recording
if (vk_graph_log_enabled) { std::cerr << "ggml_vulkan: graph " << cgraph->nodes[0] << " (" << cgraph->n_nodes << " nodes) — replay failed (prealloc changed)" << std::endl; }
VK_LOG_DEBUG("ggml_vulkan: graph " << cgraph->nodes[0] << " (" << cgraph->n_nodes << " nodes) - replay failed (prealloc changed)");
use_vk_graph = false;
}
@ -16400,33 +16371,17 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
if (cap->disabled) {
cap->destroy(ctx->device->device);
if (vk_graph_log_enabled) { std::cerr << "ggml_vulkan: graph (" << cgraph->n_nodes << " nodes) — capture disabled (staging copies)" << std::endl; }
VK_LOG_DEBUG("ggml_vulkan: graph (" << cgraph->n_nodes << " nodes) - capture disabled (staging copies)");
} else {
cap->prealloc_x_at_capture = ctx->prealloc_x;
cap->prealloc_y_at_capture = ctx->prealloc_y;
cap->prealloc_split_k_at_capture = ctx->prealloc_split_k;
if (vk_graph_log_enabled) {
ctx->graph_compute_cpu_us = ggml_time_us() - t_start;
ctx->graph_compute_t_start = t_start;
ctx->n_graph_nodes = cgraph->n_nodes;
ctx->graph_compute_phase = "captured";
}
VK_LOG_DEBUG("ggml_vulkan: graph (" << cgraph->n_nodes << " nodes) - captured");
}
} else if (vk_graph_log_enabled && cgraph->n_nodes > 0) {
ctx->graph_compute_cpu_us = ggml_time_us() - t_start;
ctx->graph_compute_t_start = t_start;
ctx->n_graph_nodes = cgraph->n_nodes;
ctx->graph_compute_phase = "recorded";
}
if (!ctx->device->support_async) {
const int64_t t_sync_start = vk_graph_log_enabled ? ggml_time_us() : 0;
ggml_vk_synchronize(ctx);
if (vk_graph_log_enabled && ctx->graph_compute_phase) {
const int64_t t_sync = ggml_time_us() - t_sync_start;
std::cerr << "ggml_vulkan: graph (" << ctx->n_graph_nodes << " nodes) — " << ctx->graph_compute_phase << " | cpu " << ctx->graph_compute_cpu_us << " us, sync " << t_sync << " us, total " << (ggml_time_us() - t_start) << " us" << std::endl;
}
ctx->graph_compute_phase = nullptr;
}
if (vk_instance.debug_utils_support) {