Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-19 17:24:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

llama-context : report graph inputs and input tensors during sched reserve

Browse Source 
- fix the tg (token generation) graph bs label to use n_seqs instead of a hardcoded 1
- report the number of graph inputs from llm_graph_result::inputs for both the pp and tg graphs
- report the number of input tensors (nodes and their src tensors flagged with GGML_TENSOR_FLAG_INPUT)
- log a warning when an input tensor has an op other than GGML_OP_NONE
- log a trace line for each input tensor and the nodes (name and op) that use it

Assisted-by: llama.cpp:DeepSeek-v4-Flash-0731
This commit is contained in:
Georgi Gerganov
2026-08-05 11:48:34 +03:00
parent 6011c34ce6
commit d291feda07
1 changed files with 66 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-context.cpp
+66 -10
View File
@@ -19,6 +19,7 @@
#include <limits>
#include <stdexcept>
#include <string>
#include <unordered_map>
//
// llama_context
@@ -579,6 +580,41 @@ void llama_context::resolve_fused_ops(const llama_memory_context_i * mctx, uint3
}
}
// count the number of tensors in the graph that are flagged as inputs
// this includes the graph nodes themselves as well as their src tensors
static int llama_graph_n_input_tensors(ggml_cgraph * gf) {
// map each input tensor to the graph nodes that use it
std::unordered_map<const ggml_tensor *, std::vector<ggml_tensor *>> users;
for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
ggml_tensor * node = ggml_graph_node(gf, i);
if (node->flags & GGML_TENSOR_FLAG_INPUT) {
users[node].push_back(node);
}
for (int j = 0; j < GGML_MAX_SRC; ++j) {
ggml_tensor * src = node->src[j];
if (!src) {
break;
}
if (src->flags & GGML_TENSOR_FLAG_INPUT) {
users[src].push_back(node);
}
}
}
for (const auto & [tensor, nodes] : users) {
if (tensor->op != GGML_OP_NONE) {
LLAMA_LOG_WARN("%s: input tensor '%s' has op %s, expected GGML_OP_NONE\n",
__func__, tensor->name, ggml_op_name(tensor->op));
}
for (const ggml_tensor * node : nodes) {
LLAMA_LOG_DEBUG("%s: input tensor '%s' is used by node '%s' (%s)\n",
__func__, tensor->name, node->name, ggml_op_name(node->op));
}
}
return (int) users.size();
}
void llama_context::sched_reserve() {
if (!sched_need_reserve) {
return;
@@ -621,11 +657,15 @@ void llama_context::sched_reserve() {
resolve_fused_ops(mctx.get(), n_seqs);
// reserve worst-case graph
int n_splits_pp = -1;
int n_nodes_pp = -1;
int n_splits_pp = -1;
int n_nodes_pp = -1;
int n_inputs_pp = -1;
int n_input_tensors_pp = -1;
int n_splits_tg = -1;
int n_nodes_tg = -1;
int n_splits_tg = -1;
int n_nodes_tg = -1;
int n_inputs_tg = -1;
int n_input_tensors_tg = -1;
const uint32_t n_outputs_pp = std::min(n_tokens, cparams.n_outputs_max);
@@ -645,8 +685,10 @@ void llama_context::sched_reserve() {
}
}
n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
n_nodes_pp = ggml_graph_n_nodes(gf);
n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
n_nodes_pp = ggml_graph_n_nodes(gf);
n_inputs_pp = get_gf_res_reserve()->inputs.size();
n_input_tensors_pp = llama_graph_n_input_tensors(gf);
}
// reserve with tg (token generation) graph to get the number of splits and nodes
@@ -656,8 +698,10 @@ void llama_context::sched_reserve() {
throw std::runtime_error("failed to allocate compute tg buffers");
}
n_splits_tg = ggml_backend_sched_get_n_splits(sched.get());
n_nodes_tg = ggml_graph_n_nodes(gf);
n_splits_tg = ggml_backend_sched_get_n_splits(sched.get());
n_nodes_tg = ggml_graph_n_nodes(gf);
n_inputs_tg = get_gf_res_reserve()->inputs.size();
n_input_tensors_tg = llama_graph_n_input_tensors(gf);
}
// reserve again with pp graph to avoid ggml-alloc reallocations during inference
@@ -698,13 +742,25 @@ void llama_context::sched_reserve() {
if (n_nodes_pp == n_nodes_tg) {
LLAMA_LOG_INFO("%s: graph nodes = %d\n", __func__, n_nodes_pp);
} else {
LLAMA_LOG_INFO("%s: graph nodes = %d (with bs=%d), %d (with bs=1)\n", __func__, n_nodes_pp, n_tokens, n_nodes_tg);
LLAMA_LOG_INFO("%s: graph nodes = %d (with bs=%d), %d (with bs=%d)\n", __func__, n_nodes_pp, n_tokens, n_nodes_tg, n_seqs);
}
if (n_splits_pp == n_splits_tg) {
LLAMA_LOG_INFO("%s: graph splits = %d\n", __func__, n_splits_pp);
} else {
LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg);
LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=%d)\n", __func__, n_splits_pp, n_tokens, n_splits_tg, n_seqs);
}
if (n_inputs_pp == n_inputs_tg) {
LLAMA_LOG_INFO("%s: graph inputs = %d\n", __func__, n_inputs_pp);
} else {
LLAMA_LOG_INFO("%s: graph inputs = %d (with bs=%d), %d (with bs=%d)\n", __func__, n_inputs_pp, n_tokens, n_inputs_tg, n_seqs);
}
if (n_input_tensors_pp == n_input_tensors_tg) {
LLAMA_LOG_INFO("%s: graph input tensors = %d\n", __func__, n_input_tensors_pp);
} else {
LLAMA_LOG_INFO("%s: graph input tensors = %d (with bs=%d), %d (with bs=%d)\n", __func__, n_input_tensors_pp, n_tokens, n_input_tensors_tg, n_seqs);
}
const int64_t t_end_us = ggml_time_us();