[Example] ggml: add Qwen2-VL example

.github/workflows/llama.yml

@@ -25,8 +25,8 @@ jobs:
     strategy:
       matrix:
         runner: [ubuntu-20.04, macos-m1]
-        wasmedge: ["0.13.5", "0.14.0"]
-        plugin: [wasi_nn-ggml]
+        wasmedge: ["0.14.1"]
+        plugin: [wasi_nn-ggml-b4381]
         job:
           - name: "Tiny Llama"
             run: |
@@ -187,18 +187,6 @@ jobs:
                 default \
                 'hello world'
 
-          - name: Embedding Example (Llama-2)
-            run: |
-              test -f ~/.wasmedge/env && source ~/.wasmedge/env
-              cd wasmedge-ggml/embedding
-              curl -LO https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF/resolve/main/llama-2-7b-chat.Q5_K_M.gguf
-              cargo build --target wasm32-wasi --release
-              time wasmedge --dir .:. \
-                --nn-preload default:GGML:AUTO:llama-2-7b-chat.Q5_K_M.gguf \
-                target/wasm32-wasi/release/wasmedge-ggml-llama-embedding.wasm \
-                default \
-                'hello world'
-
           - name: RPC Example
             run: |
               test -f ~/.wasmedge/env && source ~/.wasmedge/env
@@ -301,6 +289,23 @@ jobs:
                 default \
                 $'[INST] <<SYS>>\nYou are a helpful, respectful and honest assistant. Always output JSON format string.\n<</SYS>>\nGive me a JSON array of Apple products.[/INST]'
 
+          - name: Qwen2-VL
+            run: |
+              test -f ~/.wasmedge/env && source ~/.wasmedge/env
+              cd wasmedge-ggml/qwen2vl
+              curl -LO https://huggingface.co/second-state/Qwen2-VL-2B-Instruct-GGUF/resolve/main/Qwen2-VL-2B-Instruct-vision-encoder.gguf
+              curl -LO https://huggingface.co/second-state/Qwen2-VL-2B-Instruct-GGUF/resolve/main/Qwen2-VL-2B-Instruct-Q5_K_M.gguf
+              curl -LO https://llava-vl.github.io/static/images/monalisa.jpg
+              cargo build --target wasm32-wasi --release
+              time wasmedge --dir .:. \
+                --env n_gpu_layers="$NGL" \
+                --nn-preload default:GGML:AUTO:Qwen2-VL-2B-Instruct-Q5_K_M.gguf \
+                --env mmproj=Qwen2-VL-2B-Instruct-vision-encoder.gguf \
+                --env image=monalisa.jpg \
+                target/wasm32-wasi/release/wasmedge-ggml-qwen2vl.wasm \
+                default \
+                $'<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><image><|vision_end|>what is in this picture?<|im_end|>\n<|im_start|>assistant\n'
+
           - name: Build llama-stream
             run: |
               cd wasmedge-ggml/llama-stream

wasmedge-ggml/qwen2vl/Cargo.toml

@@ -0,0 +1,8 @@
+[package]
+name = "wasmedge-ggml-qwen2vl"
+version = "0.1.0"
+edition = "2021"
+
+[dependencies]
+serde_json = "1.0"
+wasmedge-wasi-nn = "0.7.1"

wasmedge-ggml/qwen2vl/README.md

@@ -0,0 +1,47 @@
+# Qwen-2VL Example For WASI-NN with GGML Backend
+
+> [!NOTE]
+> Please refer to the [wasmedge-ggml/README.md](../README.md) for the general introduction and the setup of the WASI-NN plugin with GGML backend. This document will focus on the specific example of the Qwen2-VL model.
+
+## Get Qwen2-VL Model
+
+In this example, we are going to use the pre-converted [Qwen2-VL-2B](https://huggingface.co/second-state/Qwen2-VL-2B-Instruct-GGUF/tree/main) model.
+
+Download the model:
+
+```bash
+curl -LO https://huggingface.co/second-state/Qwen2-VL-2B-Instruct-GGUF/resolve/main/Qwen2-VL-2B-Instruct-vision-encoder.gguf
+curl -LO https://huggingface.co/second-state/Qwen2-VL-2B-Instruct-GGUF/resolve/main/Qwen2-VL-2B-Instruct-Q5_K_M.gguf
+```
+
+## Prepare the Image
+
+Download the image you want to perform inference on:
+
+```bash
+curl -LO https://llava-vl.github.io/static/images/monalisa.jpg
+```
+
+## Parameters
+
+> [!NOTE]
+> Please check the parameters section of [wasmedge-ggml/README.md](https://github.com/second-state/WasmEdge-WASINN-examples/tree/master/wasmedge-ggml#parameters) first.
+
+In Qwen2-VL inference, we recommend to use the `ctx-size` at least `4096` for better results.
+
+```rust
+options.insert("ctx-size", Value::from(4096));
+```
+
+## Execute
+
+Execute the WASM with the `wasmedge` using the named model feature to preload a large model:
+
+```bash
+wasmedge --dir .:. \
+  --nn-preload default:GGML:AUTO:Qwen2-VL-2B-Instruct-Q5_K_M.gguf \
+  --env mmproj=Qwen2-VL-2B-Instruct-vision-encoder.gguf \
+  --env image=monalisa.jpg \
+  --env ctx_size=4096 \
+  wasmedge-ggml-qwen2vl.wasm default
+```

wasmedge-ggml/qwen2vl/src/main.rs

@@ -0,0 +1,197 @@
+use serde_json::Value;
+use std::collections::HashMap;
+use std::env;
+use std::io;
+use wasmedge_wasi_nn::{
+    self, BackendError, Error, ExecutionTarget, GraphBuilder, GraphEncoding, GraphExecutionContext,
+    TensorType,
+};
+
+fn read_input() -> String {
+    loop {
+        let mut answer = String::new();
+        io::stdin()
+            .read_line(&mut answer)
+            .expect("Failed to read line");
+        if !answer.is_empty() && answer != "\n" && answer != "\r\n" {
+            return answer.trim().to_string();
+        }
+    }
+}
+
+fn get_options_from_env() -> HashMap<&'static str, Value> {
+    let mut options = HashMap::new();
+
+    // Required parameters for llava
+    if let Ok(val) = env::var("mmproj") {
+        options.insert("mmproj", Value::from(val.as_str()));
+    } else {
+        eprintln!("Failed to get mmproj model.");
+        std::process::exit(1);
+    }
+    if let Ok(val) = env::var("image") {
+        options.insert("image", Value::from(val.as_str()));
+    } else {
+        eprintln!("Failed to get the target image.");
+        std::process::exit(1);
+    }
+
+    // Optional parameters
+    if let Ok(val) = env::var("enable_log") {
+        options.insert("enable-log", serde_json::from_str(val.as_str()).unwrap());
+    } else {
+        options.insert("enable-log", Value::from(false));
+    }
+    if let Ok(val) = env::var("ctx_size") {
+        options.insert("ctx-size", serde_json::from_str(val.as_str()).unwrap());
+    } else {
+        options.insert("ctx-size", Value::from(4096));
+    }
+    if let Ok(val) = env::var("n_gpu_layers") {
+        options.insert("n-gpu-layers", serde_json::from_str(val.as_str()).unwrap());
+    } else {
+        options.insert("n-gpu-layers", Value::from(0));
+    }
+    options
+}
+
+fn set_data_to_context(context: &mut GraphExecutionContext, data: Vec<u8>) -> Result<(), Error> {
+    context.set_input(0, TensorType::U8, &[1], &data)
+}
+
+fn get_data_from_context(context: &GraphExecutionContext, index: usize) -> String {
+    // Preserve for 4096 tokens with average token length 6
+    const MAX_OUTPUT_BUFFER_SIZE: usize = 4096 * 6;
+    let mut output_buffer = vec![0u8; MAX_OUTPUT_BUFFER_SIZE];
+    let mut output_size = context
+        .get_output(index, &mut output_buffer)
+        .expect("Failed to get output");
+    output_size = std::cmp::min(MAX_OUTPUT_BUFFER_SIZE, output_size);
+
+    String::from_utf8_lossy(&output_buffer[..output_size]).to_string()
+}
+
+fn get_output_from_context(context: &GraphExecutionContext) -> String {
+    get_data_from_context(context, 0)
+}
+
+fn get_metadata_from_context(context: &GraphExecutionContext) -> Value {
+    serde_json::from_str(&get_data_from_context(context, 1)).expect("Failed to get metadata")
+}
+
+fn main() {
+    let args: Vec<String> = env::args().collect();
+    let model_name: &str = &args[1];
+
+    // Set options for the graph. Check our README for more details:
+    // https://github.com/second-state/WasmEdge-WASINN-examples/tree/master/wasmedge-ggml#parameters
+    let options = get_options_from_env();
+    // You could also set the options manually like this:
+
+    // Create graph and initialize context.
+    let graph = GraphBuilder::new(GraphEncoding::Ggml, ExecutionTarget::AUTO)
+        .config(serde_json::to_string(&options).expect("Failed to serialize options"))
+        .build_from_cache(model_name)
+        .expect("Failed to build graph");
+    let mut context = graph
+        .init_execution_context()
+        .expect("Failed to init context");
+
+    // If there is a third argument, use it as the prompt and enter non-interactive mode.
+    // This is mainly for the CI workflow.
+    if args.len() >= 3 {
+        let prompt = &args[2];
+        // Set the prompt.
+        println!("Prompt:\n{}", prompt);
+        let tensor_data = prompt.as_bytes().to_vec();
+        context
+            .set_input(0, TensorType::U8, &[1], &tensor_data)
+            .expect("Failed to set input");
+        println!("Response:");
+
+        // Get the number of input tokens and llama.cpp versions.
+        let input_metadata = get_metadata_from_context(&context);
+        println!("[INFO] llama_commit: {}", input_metadata["llama_commit"]);
+        println!(
+            "[INFO] llama_build_number: {}",
+            input_metadata["llama_build_number"]
+        );
+        println!(
+            "[INFO] Number of input tokens: {}",
+            input_metadata["input_tokens"]
+        );
+
+        // Get the output.
+        context.compute().expect("Failed to compute");
+        let output = get_output_from_context(&context);
+        println!("{}", output.trim());
+
+        // Retrieve the output metadata.
+        let metadata = get_metadata_from_context(&context);
+        println!(
+            "[INFO] Number of input tokens: {}",
+            metadata["input_tokens"]
+        );
+        println!(
+            "[INFO] Number of output tokens: {}",
+            metadata["output_tokens"]
+        );
+        std::process::exit(0);
+    }
+
+    let mut saved_prompt = String::new();
+    let system_prompt = String::from("You are a helpful assistant.");
+    let image_placeholder = "<image>";
+
+    loop {
+        println!("USER:");
+        let input = read_input();
+
+        // Qwen2VL prompt format: <|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n<|vision_start|>{image_placeholder}<|vision_end|>{user_prompt}<|im_end|>\n<|im_start|>assistant\n";
+        if saved_prompt.is_empty() {
+            saved_prompt = format!(
+                "<|im_start|>system\n{}<|im_end|>\n<|im_start|>user\n<|vision_start|>{}<|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n",
+                system_prompt, image_placeholder, input
+            );
+        } else {
+            saved_prompt = format!(
+                "{}<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n",
+                saved_prompt, input
+            );
+        }
+
+        // Set prompt to the input tensor.
+        set_data_to_context(&mut context, saved_prompt.as_bytes().to_vec())
+            .expect("Failed to set input");
+
+        // Execute the inference.
+        let mut reset_prompt = false;
+        match context.compute() {
+            Ok(_) => (),
+            Err(Error::BackendError(BackendError::ContextFull)) => {
+                println!("\n[INFO] Context full, we'll reset the context and continue.");
+                reset_prompt = true;
+            }
+            Err(Error::BackendError(BackendError::PromptTooLong)) => {
+                println!("\n[INFO] Prompt too long, we'll reset the context and continue.");
+                reset_prompt = true;
+            }
+            Err(err) => {
+                println!("\n[ERROR] {}", err);
+                std::process::exit(1);
+            }
+        }
+
+        // Retrieve the output.
+        let mut output = get_output_from_context(&context);
+        println!("ASSISTANT:\n{}", output.trim());
+
+        // Update the saved prompt.
+        if reset_prompt {
+            saved_prompt.clear();
+        } else {
+            output = output.trim().to_string();
+            saved_prompt = format!("{}{}<|im_end|>\n", saved_prompt, output);
+        }
+    }
+}