add Qwen3VL model

tests/config_tests.rs

@@ -1,6 +1,5 @@
 use aha::models::{
-    minicpm4::config::MiniCPM4Config, qwen2_5vl::config::Qwen2_5VLConfig,
-    voxcpm::config::VoxCPMConfig,
+    minicpm4::config::MiniCPM4Config, qwen2_5vl::config::Qwen2_5VLConfig, qwen3vl::config::Qwen3VLConfig, voxcpm::config::VoxCPMConfig
 };
 use anyhow::Result;
 
@@ -34,3 +33,13 @@ fn voxcpm_config() -> Result<()> {
     println!("{:?}", config);
     Ok(())
 }
+
+#[test]
+fn qwen3vl_config() -> Result<()> {
+    // cargo test -F cuda qwen3vl_config -- --nocapture
+    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen3-VL-4B-Instruct/";
+    let config_path = model_path.to_string() + "/config.json";
+    let config: Qwen3VLConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
+    println!("{:?}", config);
+    Ok(())
+}

tests/messy_test.rs

@@ -1,13 +1,25 @@
-use aha::utils::audio_utils::load_audio_with_resample;
+use aha::utils::{tensor_utils::bitor_tensor};
 use anyhow::Result;
+use candle_core::Tensor;
 
 #[test]
 fn messy_test() -> Result<()> {
-    let device = candle_core::Device::Cpu;
-    let wav_path = "./assets/audio/voice_01.wav";
-    let audio_tensor = load_audio_with_resample(wav_path, device, Some(16000))?;
-
-    println!("audio_tensor: {}", audio_tensor);
+    let device = &candle_core::Device::Cpu;
+    let image_mask = Tensor::new(vec![0u32, 0, 0, 1, 0, 1], device)?;
+    let video_mask = Tensor::new(vec![0u32, 1, 0, 1, 0, 1], device)?;
+    let visual_mask = bitor_tensor(&image_mask, &video_mask)?;
+    println!("visual_mask: {}", visual_mask);
+    // let x = Tensor::arange_step(0.0_f32, 5., 0.5, &device)?;
+    // let x_int = x.to_dtype(candle_core::DType::U32)?;
+    // println!("x: {}", x);
+    // println!("x_int: {}", x_int);
+    // let x_affine = x_int.affine(1.0, 1.0)?;
+    // println!("x_affine: {}", x_affine);
+    // let x_clamp = x_affine.clamp(0u32, 3u32)?;
+    // println!("x_clamp: {}", x_clamp);
+    // let wav_path = "./assets/audio/voice_01.wav";
+    // let audio_tensor = load_audio_with_resample(wav_path, device, Some(16000))?;
+    // println!("audio_tensor: {}", audio_tensor);
     // let string = "你好啊".to_string();
     // let vec_str: Vec<String>= string.chars().map(|c| c.to_string()).collect();
     // println!("vec_str: {:?}", vec_str);

tests/test_qwen3vl.rs

@@ -0,0 +1,94 @@
+use std::{pin::pin, time::Instant};
+
+use aha::models::{GenerateModel, qwen3vl::generate::Qwen3VLGenerateModel};
+use anyhow::Result;
+use openai_dive::v1::resources::chat::ChatCompletionParameters;
+use rocket::futures::StreamExt;
+
+#[test]
+fn qwen3vl_generate() -> Result<()> {
+    // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda qwen3vl_generate -- --nocapture
+
+    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen3-VL-2B-Instruct/";
+
+    let message = r#"
+    {
+        "model": "qwen3vl",
+        "messages": [
+            {
+                "role": "user",
+                "content": [ 
+                    {
+                        "type": "video",
+                        "video_url": 
+                        {
+                            "url": "./assets/video/video_test.mp4"
+                        }
+                    },               
+                    {
+                        "type": "text", 
+                        "text": "视频中发生了什么."
+                    }
+                ]
+            }
+        ]
+    }
+    "#;
+    let mes: ChatCompletionParameters = serde_json::from_str(message)?;
+    let i_start = Instant::now();
+    let mut qwen3vl = Qwen3VLGenerateModel::init(model_path, None, None)?;
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in load model is: {:?}", i_duration);
+
+    let i_start = Instant::now();
+    let res = qwen3vl.generate(mes)?;
+    let i_duration = i_start.elapsed();
+    println!("generate: \n {:?}", res);
+    println!("Time elapsed in generate is: {:?}", i_duration);
+    Ok(())
+}
+
+#[tokio::test]
+async fn qwen3vl_stream() -> Result<()> {
+    // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda qwen3vl_stream -- --nocapture
+
+    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen3-VL-2B-Instruct/";
+
+    let message = r#"
+    {
+        "model": "qwen3vl",
+        "messages": [
+            {
+                "role": "user",
+                "content": [ 
+                    {
+                        "type": "image",
+                        "image_url": 
+                        {
+                            "url": "file://./assets/img/voxcpm.png"
+                        }
+                    },              
+                    {
+                        "type": "text", 
+                        "text": "描述这张图片"
+                    }
+                ]
+            }
+        ]
+    }
+    "#;
+    let mes: ChatCompletionParameters = serde_json::from_str(message)?;
+    let i_start = Instant::now();
+    let mut qwen3vl = Qwen3VLGenerateModel::init(model_path, None, None)?;
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in load model is: {:?}", i_duration);
+
+    let i_start = Instant::now();
+    let mut stream = pin!(qwen3vl.generate_stream(mes)?);
+    while let Some(item) = stream.next().await {
+        println!("generate: \n {:?}", item);
+    }
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in generate is: {:?}", i_duration);
+    Ok(())
+}

tests/weight_test.rs

@@ -45,3 +45,19 @@ fn voxcpm_weight() -> Result<()> {
     );
     Ok(())
 }
+
+#[test]
+fn qwen3vl_weight() -> Result<()> {
+    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen3-VL-4B-Instruct/";
+    let model_list = find_type_files(model_path, "safetensors")?;
+    
+    let device = Device::Cpu;
+    for m in &model_list {
+        let weights = safetensors::load(m, &device)?;
+        for (key, tensor) in weights.iter() {
+            println!("=== {} === {:?}", key, tensor.shape());
+        }
+    }
+    println!("model_list: {:?}", model_list);
+    Ok(())
+}