CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-06 09:46:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Update topk_moe fusion to handle gpt's late softmax (#16656)

Browse Source 
* vulkan: Update topk_moe fusion to handle gpt's late softmax

Based on #16649.

* Add ggml_check_edges

* Add sync logging to show fusion effects

* handle clamp added in #16655

* Update ggml/src/ggml-impl.h

Co-authored-by: Diego Devesa <slarengh@gmail.com>
This commit is contained in:
Jeff Bolz
2025-10-29 08:44:29 -05:00
committed by GitHub
parent bcf5bda6f5
commit 10fcc41290
3 changed files with 275 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
ggml/src/ggml-impl.h
View File

@@ -682,6 +682,7 @@ static inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
#endif
#ifdef __cplusplus
#include <array>
#include <initializer_list>
#include <vector>
@@ -697,6 +698,21 @@ inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
return ggml_can_fuse_subgraph(cgraph, start_idx, ops.size(), ops.begin(), outputs.begin(), outputs.size());
}
// Return true if the edges in the graph match expectations.
inline bool ggml_check_edges(const struct ggml_cgraph * cgraph,
int start_idx,
std::initializer_list<std::array<int, 3>> edges) {
for (const auto & edge : edges) {
int dst_node = edge[0];
int src_idx = edge[1];
int src_node = edge[2];
if (cgraph->nodes[start_idx + dst_node]->src[src_idx] != cgraph->nodes[start_idx + src_node]) {
return false;
}
}
return true;
}
// expose GGUF internals for test code
GGML_API size_t gguf_type_size(enum gguf_type type);
GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);