Use mudnn::Unary::IDENTITY op to accelerate D2D memory copy

[yeahdongcn]

This is a manual merge of ggml-org/llama.cpp#13647.

Testing Done

❯ ollama run gemma3n:e4b
total duration:       7.828684474s
load duration:        58.973429ms
prompt eval count:    11 token(s)
prompt eval duration: 2.363928229s
prompt eval rate:     4.65 tokens/s
eval count:           53 token(s)
eval duration:        5.405308512s
eval rate:            9.81 tokens/s

The previous eval rate was around 7 tokens/s — this update improves it by approximately 40%.

[Copilot]

Pull Request Overview

This PR integrates MUSA’s mudnn::Unary IDENTITY operation to accelerate device-to-device memory copies for FP16/FP32 tensors, offering ~40% performance improvements.

• Adds mudnnMemcpyAsync API declaration and implementation using mudnn::Unary::IDENTITY
• Updates cpy.cu to route contiguous F32/F16 copies through MUSA when enabled
• Extends CMake configurations to include new sources and link against the mudnn library

Reviewed Changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
ml/backend/ggml/ggml/src/ggml-musa/mudnn.cuh	Declares mudnnMemcpyAsync
ml/backend/ggml/ggml/src/ggml-musa/mudnn.cu	Implements mudnnMemcpyAsync with MUSA DNN identity op
ml/backend/ggml/ggml/src/ggml-musa/CMakeLists.txt	Adds mudnn headers/sources and links mudnn library
ml/backend/ggml/ggml/src/ggml-cuda/cpy.cu	Routes contiguous FP16/FP32 copies through mudnnMemcpyAsync
CMakeLists.txt	Includes mudnn in runtime dependency regex and link steps

Comments suppressed due to low confidence (3)

ml/backend/ggml/ggml/src/ggml-musa/mudnn.cu:88

• There are no unit tests covering mudnnMemcpyAsync; adding tests for both FLOAT and HALF copy paths would help ensure correctness and prevent regressions.

musaError_t mudnnMemcpyAsync(ggml_backend_cuda_context& ctx, const ggml_tensor* dst, const ggml_tensor* src) {

ml/backend/ggml/ggml/src/ggml-musa/mudnn.cu:1

• The file uses std::vector and std::unordered_map but does not include or <unordered_map>, leading to compilation failures.

#include <mutex>

ml/backend/ggml/ggml/src/ggml-musa/CMakeLists.txt:101

• Static builds are not linking against the mudnn library, which will cause unresolved symbol errors when mudnn code is compiled; consider linking mudnn for static configuration or disabling mudnn support in static mode.

        target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)