CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-01 09:01:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Implement topk_moe fused shader, ported from CUDA (#16641)

Browse Source 
This is similar to the CUDA shader from #16130, but doesn't use shared memory
and handles different subgroup sizes.
This commit is contained in:
Jeff Bolz
2025-10-18 05:22:57 -05:00
committed by GitHub
parent 38355c6c8e
commit e56abd2098
4 changed files with 412 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
ggml/src/ggml-impl.h
View File

@@ -565,14 +565,23 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
static inline int32_t ggml_node_get_use_count(const struct ggml_cgraph * cgraph, int node_idx) {
const struct ggml_tensor * node = cgraph->nodes[node_idx];
size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
return 0;
}
return cgraph->use_counts[hash_pos];
}
// return true if the node's results are only used by N other nodes
// and can be fused into their calculations.
static inline bool ggml_node_has_n_uses(const struct ggml_cgraph * cgraph, int node_idx, int32_t n_uses) {
const struct ggml_tensor * node = cgraph->nodes[node_idx];
// check the use count against how many we're replacing
size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos) || cgraph->use_counts[hash_pos] != n_uses) {
if (ggml_node_get_use_count(cgraph, node_idx) != n_uses) {
return false;
}