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

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],

src/llama-model.cpp

@@ -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) {