add minicpm with a bug

tests/config_tests.rs

@@ -1,12 +1,23 @@
-use aha::models::qwen2_5vl::config::Config;
+use aha::models::{minicpm4::config::MiniCPM4Config, qwen2_5vl::config::Qwen2_5VLConfig};
 use anyhow::Result;
 
 #[test]
-fn qwen2_5vl_config() -> Result<()> {
-    // cargo test qwen2_5vl_config -- --nocapture
+fn qwen2_5_vl_config() -> Result<()> {
+    // cargo test -F cuda,flash-attn qwen2_5vl_config -- --nocapture
     let model_path = "/home/jhq/huggingface_model/Qwen/Qwen2.5-VL-3B-Instruct/";
     let config_path = model_path.to_string() + "/config.json";
-    let config: Config = serde_json::from_slice(&std::fs::read(config_path)?)?;
+    let config: Qwen2_5VLConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
     println!("{:?}", config);
     Ok(())
 }
+
+
+#[test]
+fn minicpm4_config() -> Result<()> {
+    // cargo test -F cuda,flash-attn minicpm4_config -- --nocapture
+    let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
+    let config_path = model_path.to_string() + "/config.json";
+    let config: MiniCPM4Config = serde_json::from_slice(&std::fs::read(config_path)?)?;
+    println!("{:?}", config);
+    Ok(())
+}

tests/test_minicpm4.rs

@@ -0,0 +1,97 @@
+use std::time::Instant;
+
+use anyhow::Result;
+use candle_core::{DType, Device};
+use openai_dive::v1::resources::chat::ChatCompletionParameters;
+
+#[test]
+fn qwen2_5vl_generate() -> Result<()> {
+    // test with cpu :(太慢了, : RUST_BACKTRACE=1 cargo test qwen2_5vl_generate -- --nocapture
+    // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda qwen2_5vl_generate -- --nocapture
+    // test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn qwen2_5vl_generate -- --nocapture
+    let device = Device::cuda_if_available(0)?;
+    let dtype = DType::BF16;
+
+    let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
+
+    let message = r#"
+    {
+        "model": "minicpm4",
+        "messages": [
+            {
+                "role": "user",
+                "content": [          
+                    {
+                        "type": "text", 
+                        "text": "你是谁"
+                    }
+                ]
+            }
+        ]
+    }
+    "#;
+    let mes: ChatCompletionParameters = serde_json::from_str(message)?;
+    let i_start = Instant::now();
+    // let mut model = Qwen2_5VLGenerateModel::init(model_path, &device, dtype)?;
+    let mut model = ModelType::init(ModelType::Qwen2_5VL, 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 result = model.generate(mes)?;
+    println!("generate: \n {:?}", result);
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in generate is: {:?}", i_duration);
+
+    Ok(())
+}
+
+#[tokio::test]
+async fn qwen2_5vl_stream() -> Result<()> {
+    // test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn qwen2_5vl_generate -- --nocapture
+    let device = Device::cuda_if_available(0)?;
+    let dtype = DType::BF16;
+
+    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen2.5-VL-3B-Instruct/";
+
+    let message = r#"
+    {
+        "model": "qwen2.5vl",
+        "messages": [
+            {
+                "role": "user",
+                "content": [ 
+                    {
+                        "type": "image",
+                        "image_url": 
+                        {
+                            "url": "file://./assets/img/ocr_test.png"
+                        }
+                    },               
+                    {
+                        "type": "text", 
+                        "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"
+                    }
+                ]
+            }
+        ]
+    }
+    "#;
+    let mes: ChatCompletionParameters = serde_json::from_str(message)?;
+    let i_start = Instant::now();
+    // let mut model = Qwen2_5VLGenerateModel::init(model_path, &device, dtype)?;
+    let mut model = ModelType::init(ModelType::Qwen2_5VL, 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!(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/weight_test.rs

@@ -0,0 +1,18 @@
+use aha::utils::utils::find_safetensors_files;
+use anyhow::Result;
+use candle_core::{safetensors, Device};
+#[test]
+fn minicpm4_weight() -> Result<()> {
+    let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
+    let model_list = find_safetensors_files(&model_path)?;
+    let device = Device::Cpu;
+    for m in model_list {
+        let weights = safetensors::load(m, &device)?;
+        for (key, tensor) in weights.iter() {
+            println!("=== {} ===", key);
+            println!("Shape: {:?}", tensor.shape());
+            println!("DType: {:?}", tensor.dtype());
+        }
+    }
+    Ok(())
+}