LLM inference in C/C++
C++ ggml
llama: refer to segment documentation [no ci] (#29074)
Johannes Gäßler78e2964c23dc
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.hindex cc3f8cd36e..2a09574366 100644--- a/ggml/include/ggml-backend.h+++ b/ggml/include/ggml-backend.h@@ -383,6 +383,7 @@ extern "C" {// - most tensors have n_segments == 1 and a contiguous slice of the tensor data// - some tensors have an inhomogenenous data layout along the split axis,// those tensors are divided into segments which are each individually split across devices+ // (this usually happens when multiple tensors are fused into a single one)// - ne has one entry per segment and device and that segment repeats nr times,// in total when accounting for repetitions the segments add up to ggml_tensor::ne for that axis,// the outer/inner loops are over segments/devices like [seg0_dev0_r0, seg0_dev1_r0, seg0_dev0_r1, seg0_dev1_r1, seg1_dev0_r0, seg1_dev1_r0],diff --git a/src/llama-model.cpp b/src/llama-model.cppindex 404509888e..742bcd5d58 100644--- a/src/llama-model.cpp+++ b/src/llama-model.cpp@@ -604,8 +604,8 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const strreturn 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 structauto 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) {