wasi-nn: Add a new target for llama.cpp as a wasi-nn backend (#3709)

Minimum support:
- [x] accept (WasmEdge) customized model parameters. metadata.
- [x] Target [wasmedge-ggml examples](https://github.com/second-state/WasmEdge-WASINN-examples/tree/master/wasmedge-ggml)
  - [x] basic
  - [x] chatml
  - [x] gemma
  - [x] llama
  - [x] qwen

---

In the future, to support if required:
- [ ] Target [wasmedge-ggml examples](https://github.com/second-state/WasmEdge-WASINN-examples/tree/master/wasmedge-ggml)
  - [ ] command-r. (>70G memory requirement)
  - [ ] embedding. (embedding mode)
  - [ ] grammar. (use the grammar option to constrain the model to generate the JSON output)
  - [ ] llama-stream. (new APIS `compute_single`, `get_output_single`, `fini_single`)
  - [ ] llava. (image representation)
  - [ ] llava-base64-stream. (image representation)
  - [ ] multimodel. (image representation)
- [ ] Target [llamaedge](https://github.com/LlamaEdge/LlamaEdge)

core/iwasm/libraries/wasi-nn/test/Dockerfile.wasi-nn-smoke

@@ -63,21 +63,35 @@ WORKDIR /workspaces/wasmedge-wasinn-examples
 RUN git clone --depth 1 https://github.com/second-state/WasmEdge-WASINN-examples.git .
 COPY core/iwasm/libraries/wasi-nn/test/bump_wasi_nn_to_0_6_0.patch .
 RUN git apply ./bump_wasi_nn_to_0_6_0.patch
-# recompile with wasi-nn 0.6.0
-RUN cd openvino-mobilenet-image/rust && cargo build --target=wasm32-wasi
-RUN cd openvino-mobilenet-raw/rust && cargo build --target=wasm32-wasi
-RUN cd openvino-road-segmentation-adas/openvino-road-seg-adas && cargo build --target=wasm32-wasi
-RUN cd tflite-birds_v1-image/rust && cargo build --target=wasm32-wasi
 
-# preparation
-RUN cd openvino-mobilenet-image \
+# recompile with wasi-nn 0.6.0
+WORKDIR /workspaces/wasmedge-wasinn-examples/openvino-mobilenet-image/
+RUN pushd rust \
+  && cargo build --target=wasm32-wasi \
+  && popd \
   && ./download_mobilenet.sh . \
   && ls -l mobilenet.xml mobilenet.bin
 
-RUN cd openvino-mobilenet-raw \
+WORKDIR /workspaces/wasmedge-wasinn-examples/openvino-mobilenet-raw/
+RUN pushd rust \
+  && cargo build --target=wasm32-wasi \
+  && popd \
   && ./download_mobilenet.sh . \
   && ls -l mobilenet.xml mobilenet.bin tensor-1x224x224x3-f32.bgr
 
+WORKDIR /workspaces/wasmedge-wasinn-examples/openvino-road-segmentation-adas/
+RUN pushd openvino-road-seg-adas \
+  && cargo build --target=wasm32-wasi
+
+WORKDIR /workspaces/wasmedge-wasinn-examples/tflite-birds_v1-image/
+RUN pushd rust \
+  && cargo build --target=wasm32-wasi
+
+# mount models when running
+WORKDIR /workspaces/wasmedge-wasinn-examples/wasmedge-ggml/qwen
+RUN wget --progress=dot:giga https://www.modelscope.cn/models/qwen/Qwen1.5-0.5B-Chat-GGUF/resolve/master/qwen1_5-0_5b-chat-q2_k.gguf
+RUN cargo build --target=wasm32-wasi
+
 #
 # iwasm. build from source
 WORKDIR /workspaces/wamr
@@ -88,15 +102,16 @@ WORKDIR /workspaces/wamr/product-mini/platforms/linux
 RUN OpenVINO_DIR=/usr/lib/openvino-2023.2.0 \
     cmake -S . -B build \
     -DWAMR_BUILD_WASI_NN=1 -DWAMR_BUILD_WASI_EPHEMERAL_NN=1 \
-    -DWAMR_BUILD_WASI_NN_OPENVINO=1 -DWAMR_BUILD_WASI_NN_TFLITE=1 \
-  && cmake --build build
-
-ENV PATH=/workspaces/wamr/product-mini/platforms/linux/build:${PATH}
-ENV LD_LIBRARY_PATH=/workspaces/wamr/product-mini/platforms/linux/build
+    -DWAMR_BUILD_WASI_NN_OPENVINO=1 \
+    -DWAMR_BUILD_WASI_NN_TFLITE=1 \
+    -DWAMR_BUILD_WASI_NN_LLAMACPP=1 \
+  && cmake --build build \
+  && cmake --install build
+ 
+ENV LD_LIBRARY_PATH=/usr/local/lib
 
 # add smoke test script
 COPY core/iwasm/libraries/wasi-nn/test/run_smoke_test.py /
 
-#
 WORKDIR /workspaces/wasmedge-wasinn-examples
 CMD ["python3", "/run_smoke_test.py"]

core/iwasm/libraries/wasi-nn/test/run_smoke_test.py

@@ -260,6 +260,63 @@ def execute_openvino_road_segmentation_adas(
     print("------------------------------------------------------------")
 
 
+def execute_wasmedge_ggml_qwen(iwasm_bin: str, wasmedge_bin: str, cwd: Path):
+    iwasm_args = ["--dir=."]
+    wasm_file = ["./target/wasm32-wasi/debug/wasmedge-ggml-qwen.wasm"]
+    wasm_args = ["./qwen1_5-0_5b-chat-q2_k.gguf"]
+
+    cmd = [iwasm_bin]
+    cmd.extend(iwasm_args)
+    cmd.extend(wasm_file)
+    cmd.extend(wasm_args)
+
+    # print(f'Execute: {" ".join(cmd)}')
+
+    prompt = "what is the capital of Pakistan"
+
+    with subprocess.Popen(
+        cmd,
+        stdin=subprocess.PIPE,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        cwd=cwd,
+    ) as p:
+        # USER
+        p.stdout.readline()
+
+        p.stdin.write(b"hi\n")
+        p.stdin.flush()
+        # ASSITANT
+        p.stdout.readline()
+        # xxx
+        p.stdout.readline()
+        # USER
+        p.stdout.readline()
+
+        p.stdin.write(prompt.encode())
+        p.stdin.write(b"\n")
+        p.stdin.flush()
+        # ASSITANT
+        p.stdout.readline()
+        # xxx
+        answer = p.stdout.readline().decode("utf-8")
+        # USER
+        p.stdout.readline()
+
+        p.terminate()
+
+    if "Karachi" in answer:
+        print(f"- wasmedge_ggml_qwen. PASS")
+        return
+
+    print(f"- wasmedge_ggml_qwen. FAILED")
+    print("------------------------------------------------------------")
+    pprint(answer)
+    print("<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<")
+    pprint("Karachi")
+    print("------------------------------------------------------------")
+
+
 def execute_wasmedge_wasinn_examples(iwasm_bin: str, wasmedge_bin: str):
     assert Path.cwd().name == "wasmedge-wasinn-examples"
     assert shutil.which(iwasm_bin)
@@ -282,6 +339,9 @@ def execute_wasmedge_wasinn_examples(iwasm_bin: str, wasmedge_bin: str):
         iwasm_bin, wasmedge_bin, openvino_road_segmentation_adas_dir
     )
 
+    wasmedge_ggml_qwem_dir = Path.cwd().joinpath("./wasmedge-ggml/qwen")
+    execute_wasmedge_ggml_qwen(iwasm_bin, wasmedge_bin, wasmedge_ggml_qwem_dir)
+
 
 if __name__ == "__main__":
     execute_wasmedge_wasinn_examples("iwasm", "wasmedge")