mozempk/ggml-org-llama-cpp-mirror
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-10-11 07:20:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

add meta backend scale split

Browse Source
This commit is contained in:
Ruben Ortlam
2026-10-09 21:13:44 +02:00
committed by Oliver Simons
parent ffbebf8b07
commit d8d8e70a9a
1 changed files with 14 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+14
View File
@@ -387,6 +387,20 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
const bool is_dsv4 = ud->model->arch == LLM_ARCH_DEEPSEEK4 || const bool is_dsv4 = ud->model->arch == LLM_ARCH_DEEPSEEK4 ||
(ud->model->arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0); (ud->model->arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0);
// a per-output-channel scale indexes the weight's ne[1], so it splits on ne[0] to match
if (tensor_name.size() >= 6 && tensor_name.compare(tensor_name.size() - 6, 6, ".scale") == 0) {
const std::string weight_name = tensor_name.substr(0, tensor_name.size() - 6) + ".weight";
const ggml_tensor * w = ud->model->get_tensor(weight_name.c_str());
if (w != nullptr) {
ggml_backend_meta_split_state ws = llama_meta_device_get_split_state(w, userdata);
if (ws.axis == GGML_BACKEND_SPLIT_AXIS_1 && tensor->ne[0] == w->ne[1]) {
ws.axis = GGML_BACKEND_SPLIT_AXIS_0;
return ws;
}
}
return {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
}
static const std::regex pattern_q_weight ("blk\\.\\d*\\.attn_q.weight"); static const std::regex pattern_q_weight ("blk\\.\\d*\\.attn_q.weight");
static const std::regex pattern_kv_weight ("blk\\.\\d*\\.attn_(k|v).weight"); static const std::regex pattern_kv_weight ("blk\\.\\d*\\.attn_(k|v).weight");
static const std::regex pattern_v_exps_weight ("blk\\.\\d*\\.attn_v_exps.weight"); // K2 Horizon MoVA static const std::regex pattern_v_exps_weight ("blk\\.\\d*\\.attn_v_exps.weight"); // K2 Horizon MoVA