Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-04 02:37:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

metal : fix absolute output indices in fusion patterns

Browse Source 
ggml_can_fuse_subgraph_ext expects the outputs array to contain absolute graph
node indices (it indexes cgraph->nodes[outputs[i]]), but the fusion table query
was passing a relative index (n_ops - 1). As a result the last node of every
pattern was not recognized as an output and was subjected to the elidable
use-count check, which failed for essentially all fusions. This silently
disabled the norm/MUL fusion and caused a ~5% token-generation regression.

Pass the absolute graph index of the last node instead.

Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
This commit is contained in:
Georgi Gerganov
2026-09-03 17:19:47 +03:00
parent 9f61fc9dfd
commit b91f9e0a03
2 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-fuse.cpp
+3 -3
View File
@@ -7,14 +7,12 @@
// ---- helpers -------------------------------------------------------------
// the pattern outputs, defaulting to the last node
// the pattern outputs (absolute graph node indices); the default is the last node
static const int * ggml_metal_fuse_outputs(const struct ggml_metal_fuse * fuse, int * buf) {
if (fuse->outputs) {
return fuse->outputs;
}
buf[0] = fuse->n_ops - 1;
return buf;
}
@@ -224,7 +222,9 @@ const struct ggml_metal_fuse * ggml_metal_fuse_next(
continue;
}
// ggml_can_fuse_subgraph_ext expects outputs as absolute graph node indices
int outputs_buf[GGML_MAX_SRC];
outputs_buf[0] = node_idxs[idx + fuse->n_ops - 1];
const int * outputs = ggml_metal_fuse_outputs(fuse, outputs_buf);
const int n_outputs = fuse->n_outputs ? fuse->n_outputs : 1;
ggml/src/ggml-metal/ggml-metal-fuse.h
+1 -1
View File
@@ -40,7 +40,7 @@ struct ggml_metal_fuse {
enum ggml_metal_fuse_id id;
const enum ggml_op * ops; // op sequence (fixed length)
int n_ops; // number of ops
const int * outputs; // output node indices relative to the subgraph start (nullptr => { n_ops-1 })
const int * outputs; // output node indices (absolute graph indices; nullptr => the last node)
int n_outputs;// number of outputs (0 => default last node)
// extra backend constraints on top of ggml_can_fuse_subgraph
// nodes[j] is the j-th node of the pattern