llama: refer to segment documentation [no ci] (#29074)

This commit is contained in:
Johannes Gäßler authored and GitHub committed 2026-10-01 21:52:40 +02:00
1 parent f1cee9941b
commit 78e2964c23
2 files changed
+2 -1

No files matched your search

+1 -1
View File
@@ -604,8 +604,8 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED);
};
// if a model has fused tensors they need to be separated into "segments", see the comments on ggml_backend_meta_split_state struct
auto get_split_segments = [&](int axis, uint32_t il) -> std::vector<std::pair<int64_t, uint32_t>> {
// TODO: clarify why this is necessary specifically for these models
// TODO: deduplicate condition [TAG_SPLIT_QGATE_QWEN]
if (ud->model->arch == LLM_ARCH_QWEN3NEXT || ud->model->arch == LLM_ARCH_QWEN35 || ud->model->arch == LLM_ARCH_QWEN35MOE ||
ud->model->arch == LLM_ARCH_QWEN4EXP) {