add hunyuan_ocr

tests/config_tests.rs

@@ -1,7 +1,7 @@
 use aha::models::{
-    deepseek_ocr::config::DeepseekOCRConfig, minicpm4::config::MiniCPM4Config,
-    qwen2_5vl::config::Qwen2_5VLConfig, qwen3vl::config::Qwen3VLConfig,
-    voxcpm::config::VoxCPMConfig,
+    deepseek_ocr::config::DeepseekOCRConfig, hunyuan_ocr::config::HunYuanVLConfig,
+    minicpm4::config::MiniCPM4Config, qwen2_5vl::config::Qwen2_5VLConfig,
+    qwen3vl::config::Qwen3VLConfig, voxcpm::config::VoxCPMConfig,
 };
 use anyhow::Result;
 
@@ -48,10 +48,20 @@ fn qwen3vl_config() -> Result<()> {
 
 #[test]
 fn deepseek_ocr_config() -> Result<()> {
-    // cargo test -F cuda qwen3vl_config -r -- --nocapture
+    // cargo test -F cuda deepseek_ocr_config -r -- --nocapture
     let model_path = "/home/jhq/huggingface_model/deepseek-ai/DeepSeek-OCR/";
     let config_path = model_path.to_string() + "/config.json";
     let config: DeepseekOCRConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
     println!("{:?}", config);
     Ok(())
 }
+
+#[test]
+fn hunyuan_ocr_config() -> Result<()> {
+    // cargo test -F cuda hunyuan_ocr_config -r -- --nocapture
+    let model_path = "/home/jhq/huggingface_model/Tencent-Hunyuan/HunyuanOCR/";
+    let config_path = model_path.to_string() + "/config.json";
+    let config: HunYuanVLConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
+    println!("{:?}", config);
+    Ok(())
+}

tests/messy_test.rs

@@ -1,4 +1,6 @@
-use aha::utils::tensor_utils::interpolate_bicubic;
+use std::time::Instant;
+
+use aha::utils::tensor_utils::interpolate_bilinear;
 use anyhow::Result;
 use candle_core::Tensor;
 
@@ -6,11 +8,24 @@ use candle_core::Tensor;
 fn messy_test() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda messy_test -r -- --nocapture
     let device = &candle_core::Device::Cpu;
-    // let t = Tensor::randn(0.0f32, 1.0, (1, 768, 64, 64), device)?;
-    let t = Tensor::arange(0.0f32, 10.0, device)?.broadcast_as((1, 1, 10, 10))?;
+    let t = Tensor::arange(0.0f32, 40.0, device)?.broadcast_as((1, 1, 40, 40))?;
     println!("t: {}", t);
-    let t_resized = interpolate_bicubic(&t, (5, 5), Some(true), Some(false))?;
-    println!("t_resized: {}", t_resized);
+    let i_start = Instant::now();
+    let t_inter = interpolate_bilinear(&t, (20, 20), Some(false))?;
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in interpolate_bilinear is: {:?}", i_duration);
+    println!("t_inter: {}", t_inter);
+    // let x: Vec<u32> = (0..5).flat_map(|_| 0u32..10).collect();
+    // let id: Vec<u32> = (0..5).flat_map(|h| vec![h; 10]).collect();
+    // println!("x: {:?}", id);
+    // let t = Tensor::randn(0.0f32, 1.0, (1, 768, 64, 64), device)?;
+    // let t = Tensor::arange(0u32, 10, device)?.broadcast_as((1, 10))?;
+    // let eq = t.broadcast_eq(&Tensor::new(5u32, device)?)?;
+    // println!("eq: {}", eq);
+    // let t = Tensor::arange(0.0f32, 10.0, device)?.broadcast_as((1, 1, 10, 10))?;
+    // println!("t: {}", t);
+    // let t_resized = interpolate_bicubic(&t, (5, 5), Some(true), Some(false))?;
+    // println!("t_resized: {}", t_resized);
     // let t1 = Tensor::rand(0.0, 1.0, (1, 5, 5, 10), device)?;
     // let t2 = Tensor::rand(0.0, 1.0, (5, 8, 10), device)?;
     // let t2 = t2.t()?;

tests/test_hunyuan_ocr.rs

@@ -0,0 +1,88 @@
+use std::{pin::pin, time::Instant};
+
+use aha::models::{GenerateModel, hunyuan_ocr::generate::HunyuanOCRGenerateModel};
+use aha_openai_dive::v1::resources::chat::ChatCompletionParameters;
+use anyhow::Result;
+use rocket::futures::StreamExt;
+
+#[test]
+fn hunyuan_ocr_generate() -> Result<()> {
+    // RUST_BACKTRACE=1 cargo test -F cuda hunyuan_ocr_generate -r -- --nocapture
+    let message = r#"
+    {
+        "model": "hunyuan-ocr",
+        "messages": [
+            {
+                "role": "user",
+                "content": [ 
+                    {
+                        "type": "image",
+                        "image_url": 
+                        {
+                            "url": "file://./assets/img/ocr_test1.png"
+                        }
+                    },              
+                    {
+                        "type": "text", 
+                        "text": "检测并识别图片中的文字，将文本坐标格式化输出。"
+                    }
+                ]
+            }
+        ]
+    }
+    "#;
+    let model_path = "/home/jhq/huggingface_model/Tencent-Hunyuan/HunyuanOCR/";
+    let mes: ChatCompletionParameters = serde_json::from_str(message)?;
+    let i_start = Instant::now();
+    let mut model = HunyuanOCRGenerateModel::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 = model.generate(mes)?;
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in generate is: {:?}", i_duration);
+    println!("generate: \n {:?}", res);
+    Ok(())
+}
+
+#[tokio::test]
+async fn hunyuan_ocr_stream() -> Result<()> {
+    // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda hunyuan_ocr_stream -r -- --nocapture
+
+    let message = r#"
+    {
+        "model": "hunyuan-ocr",
+        "messages": [
+            {
+                "role": "user",
+                "content": [ 
+                    {
+                        "type": "image",
+                        "image_url": 
+                        {
+                            "url": "file://./assets/img/ocr_test1.png"
+                        }
+                    },              
+                    {
+                        "type": "text", 
+                        "text": "检测并识别图片中的文字，将文本坐标格式化输出。"
+                    }
+                ]
+            }
+        ]
+    }
+    "#;
+    let model_path = "/home/jhq/huggingface_model/Tencent-Hunyuan/HunyuanOCR/";
+    let mes: ChatCompletionParameters = serde_json::from_str(message)?;
+    let i_start = Instant::now();
+    let mut model = HunyuanOCRGenerateModel::init(model_path, None, None)?;
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in load model is: {:?}", i_duration);
+    let mut stream = pin!(model.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/test_qwen3vl.rs

@@ -19,15 +19,15 @@ fn qwen3vl_generate() -> Result<()> {
                 "role": "user",
                 "content": [
                     {
-                        "type": "video",
-                        "video_url": 
+                        "type": "image",
+                        "image_url": 
                         {
-                            "url": "https://www.w3schools.com/html/movie.mp4"
+                            "url": "file://./assets/img/ocr_test1.png"
                         }
-                    },                          
+                    },             
                     {
                         "type": "text", 
-                        "text": "视频中发生了什么"
+                        "text": "请分析图片并提取所有可见文本内容，按从左到右、从上到下的布局，返回纯文本"
                     }
                 ]
             }

tests/weight_test.rs

@@ -80,3 +80,22 @@ fn deepseekocr_weight() -> Result<()> {
     println!("model_list: {:?}", model_list);
     Ok(())
 }
+
+#[test]
+fn hunyuanocr_weight() -> Result<()> {
+    let model_path = "/home/jhq/huggingface_model/Tencent-Hunyuan/HunyuanOCR/";
+    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() {
+            if key.contains(".image_") {
+                println!("=== {} === {:?}", key, tensor.shape());
+            }
+            // println!("=== {} === {:?}", key, tensor.shape());
+        }
+    }
+    println!("model_list: {:?}", model_list);
+    Ok(())
+}