wylab/llama.cpp
1
0
Fork 1
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-06-09 07:16:44 +02:00
Code Issues Packages Projects Releases Wiki Activity

Add tensor name to JSON output

Browse Source
This commit is contained in:
Piotr Wilkin
2026-06-06 22:33:01 +02:00
parent 2bfe4ff9ca
commit 22634e0eee
4 changed files with 16 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/include/ggml-profiler.h
+1
View File
@@ -31,6 +31,7 @@ typedef struct ggml_profile_record {
const char * extra; // fusion name for fused ops, or NULL
// Output tensor info
char tensor_name[GGML_MAX_NAME]; // output tensor name (e.g. "ffn_out-0"), "" if unnamed
int64_t ne[4]; // output tensor dimensions
int out_type; // output tensor type (ggml_type), -1 if N/A
ggml/src/ggml-backend.cpp
+6
View File
@@ -1577,6 +1577,7 @@ static ggml_profile_record make_copy_record(const char * copy_dir, int backend_i
rec.end_ns = end_ns;
rec.bytes = bytes;
rec.extra = input ? input->name : NULL;
snprintf(rec.tensor_name, sizeof(rec.tensor_name), "%s", input ? input->name : "");
rec.out_type = -1;
rec.sub_op = -1;
rec.n_src = 0;
@@ -2792,6 +2793,11 @@ int ggml_backend_sched_write_profiling_json(ggml_backend_sched_t sched, FILE * f
}
// Output tensor info
if (rec.tensor_name[0] != '\0') {
fprintf(fp, ", \"tensor_name\": \"%s\"", rec.tensor_name);
} else {
fprintf(fp, ", \"tensor_name\": null");
}
fprintf(fp, ", \"ne\": [%lld, %lld, %lld, %lld]", (long long) rec.ne[0], (long long) rec.ne[1],
(long long) rec.ne[2], (long long) rec.ne[3]);
fprintf(fp, ", \"out_type\": %d", rec.out_type);
ggml/src/ggml-profiler.cpp
+4
View File
@@ -53,10 +53,14 @@ void ggml_profile_record_from_tensor(ggml_profile_record * rec, const struct ggm
memcpy(rec->ne, node->ne, sizeof(rec->ne));
rec->out_type = (int) node->type;
memcpy(rec->op_params, node->op_params, sizeof(rec->op_params));
// Copy the tensor name (rather than aliasing node->name) so the record stays
// valid after the graph's meta context is reused on the next build.
snprintf(rec->tensor_name, sizeof(rec->tensor_name), "%s", node->name);
} else {
memset(rec->ne, 0, sizeof(rec->ne));
rec->out_type = -1;
memset(rec->op_params, 0, sizeof(rec->op_params));
rec->tensor_name[0] = '\0';
}
// Sub-op (UNARY/GLU)
tools/profiler/profiler.py
+5 -1
View File
@@ -169,6 +169,7 @@ class ProfileRecord:
bytes: int
extra: Optional[str]
# Output tensor info
tensor_name: Optional[str] = None
ne: list[int] = field(default_factory=lambda: [0, 0, 0, 0])
out_type: int = -1
# Source tensors (variable length, up to GGML_MAX_SRC)
@@ -266,6 +267,7 @@ class ProfileRecord:
"duration_ns": self.duration_ns,
"bytes": self.bytes,
"extra": self.extra,
"tensor_name": self.tensor_name,
"ne": self.ne,
"out_type": self.out_type,
"n_src": len(self.ne_src),
@@ -416,6 +418,7 @@ class ProfileData:
duration_ns=r.get("duration_ns", 0),
bytes=r.get("bytes", 0),
extra=r.get("extra"),
tensor_name=r.get("tensor_name"),
ne=ne_out,
out_type=int(r.get("out_type", -1)),
ne_src=ne_src,
@@ -565,7 +568,8 @@ class ProfileData:
print(f"\n --- Top 5 Longest Kernels ---")
for rec in top_k:
shape = f" {rec.shape_str}" if rec.shape_str else ""
print(f" {rec.type_name:<5} {rec.name:<28} {rec.duration_us:>10.2f} us{shape} "
tname = f" '{rec.tensor_name}'" if rec.tensor_name else ""
print(f" {rec.type_name:<5} {rec.name:<28}{tname} {rec.duration_us:>10.2f} us{shape} "
f"(split={rec.split_id}, backend={rec.backend_id})")
print()