CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-12 10:47:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml: add ggml_can_fuse_subgraph (#16662)

Browse Source 
* ggml: add ggml_can_fuse_subgraph

* ggml-cuda: use ggml_can_fuse_subgraph for topk-moe

* format

* 1. remove inputs from signature as they are transient nodes
2. add check for views: view_src should be part of the subgraph

* - combine check into one loop
- check all view_src parents
- other minor review comments

* remove redudant if test

* - rename and other minor review comments

* add assert about count < 32
This commit is contained in:
Aman Gupta
2025-10-21 16:43:14 +08:00
committed by GitHub
parent 6ea37f5739
commit 4926419c4d
3 changed files with 113 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -2821,15 +2821,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
std::initializer_list<enum ggml_op> topk_moe_ops = ggml_cuda_topk_moe_ops(false);
std::initializer_list<enum ggml_op> topk_moe_ops_with_norm = ggml_cuda_topk_moe_ops(true);
if (ops.size() == topk_moe_ops_with_norm.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops_with_norm.begin())) {
if (node_idx + topk_moe_ops_with_norm.size() > (size_t)cgraph->n_nodes) {
return false;
}
for (size_t i = 0; i < topk_moe_ops_with_norm.size(); i++) {
if (cgraph->nodes[node_idx + i]->op != topk_moe_ops_with_norm.begin()[i]) return false;
}
if (ops.size() == topk_moe_ops_with_norm.size() &&
ggml_can_fuse_subgraph(cgraph, node_idx, topk_moe_ops_with_norm, { node_idx + 3, node_idx + 8 })) {
ggml_tensor * softmax = cgraph->nodes[node_idx];
ggml_tensor * weights = cgraph->nodes[node_idx+8];
@@ -2838,16 +2831,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
}
}
if (ops.size() == topk_moe_ops.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops.begin())) {
if (node_idx + topk_moe_ops.size() > (size_t)cgraph->n_nodes) {
return false;
}
for (size_t i = 0; i < topk_moe_ops.size(); i++) {
if (cgraph->nodes[node_idx + i]->op != topk_moe_ops.begin()[i]) return false;
}
if (ops.size() == topk_moe_ops.size() &&
ggml_can_fuse_subgraph(cgraph, node_idx, topk_moe_ops, { node_idx + 3, node_idx + 4 })) {
ggml_tensor * softmax = cgraph->nodes[node_idx];
ggml_tensor * weights = cgraph->nodes[node_idx+4];
if (ggml_cuda_should_use_topk_moe(softmax, weights)) {