[Example] ggml: add Qwen2-VL example

[dm4]

• Add Qwen2-VL example after [WASI-NN] ggml: support Qwen2VL and bump llama.cpp b4381 WasmEdge/WasmEdge#3930

[juntao]

Hello, I am a code review agent on flows.network. Here are my reviews of changed source code files in this PR.

---

.github/workflows/llama.yml

Potential issues

N/A

Summary of changes

N/A

wasmedge-ggml/qwen2vl/Cargo.toml

Potential issues

1. Version Pinning: The dependencies (serde_json and wasmedge-wasi-nn) are pinned to specific versions, which can lead to potential issues with newer updates not being tested or integrated. Consider using version ranges instead.
2. Missing License Information: There is no license specified in the package metadata, which could lead to ambiguity about how the software can be used and distributed.
3. Potential Edition Mismatch: The edition = "2021" may not be compatible with all dependencies or the target environment, ensure that this edition is supported by all listed dependencies.

Summary of changes

• Added a new Cargo.toml file: Initializes a Rust package named "wasmedge-ggml-qwen2vl".
• Set the Rust edition to 2021.
• Included dependencies: Added serde_json version 1.0 and wasmedge-wasi-nn version 0.7.1 for JSON processing and neural network functionalities, respectively.

wasmedge-ggml/qwen2vl/src/main.rs

Potential issues

1. Error Handling: The get_options_from_env function uses unwrap() when parsing JSON strings from environment variables, which can cause a panic if the string is not valid JSON. Use Result::map_err() or similar to handle errors gracefully.

2. Buffer Overflow Risk: In get_data_from_context, the code assumes that the output size will never exceed MAX_OUTPUT_BUFFER_SIZE. However, if the actual output exceeds this, it can lead to a buffer overflow. Ensure the buffer size is dynamically adjusted based on the actual output size returned by context.get_output().

3. Redundant Metadata Retrieval: The metadata is retrieved twice in non-interactive mode (get_metadata_from_context(&context) is called twice). This redundant call should be removed for efficiency and to avoid potential inconsistencies if the metadata changes between calls.

Summary of changes

• Key Changes Summary:
• Added Qwen2-VL Example: Introduced a complete Rust program to interact with the Qwen2-VL model using WasmEdge-WASINN, handling image and text inputs.
• Environment Variable Parsing: Implemented a function get_options_from_env to read required (mmproj, image) and optional parameters (enable_log, ctx_size, n_gpu_layers) from environment variables.
• Interactive Mode Support: Enabled both interactive and non-interactive modes for the application, with non-interactive mode handling a prompt provided as a command-line argument.

wasmedge-ggml/qwen2vl/wasmedge-ggml-qwen2vl.wasm

Potential issues

N/A

Summary of changes

N/A