tckm.yan/aha
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add voxcpm with some bug

Browse Source
This commit is contained in:
jhqxxx
2025-10-03 22:25:58 +08:00
parent ae36194a4b
commit f33eaaee0d
30 changed files with 2411 additions and 165 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/config_tests.rs
+11 -1
View File
@@ -1,4 +1,4 @@
use aha::models::{minicpm4::config::MiniCPM4Config, qwen2_5vl::config::Qwen2_5VLConfig};
use aha::models::{minicpm4::config::MiniCPM4Config, qwen2_5vl::config::Qwen2_5VLConfig, voxcpm::config::VoxCPMConfig};
use anyhow::Result;
#[test]
@@ -20,4 +20,14 @@ fn minicpm4_config() -> Result<()> {
let config: MiniCPM4Config = serde_json::from_slice(&std::fs::read(config_path)?)?;
println!("{:?}", config);
Ok(())
}
#[test]
fn voxcpm_config() -> Result<()> {
// cargo test -F cuda,flash-attn minicpm4_config -- --nocapture
let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
let config_path = model_path.to_string() + "/config.json";
let config: VoxCPMConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
println!("{:?}", config);
Ok(())
}
tests/messy_test.rs
+20
View File
@@ -0,0 +1,20 @@
use aha::utils::audio_utils::{load_audio_with_resample};
use anyhow::Result;
use candle_core::Tensor;
#[test]
fn messy_test() -> Result<()> {
let device = candle_core::Device::Cpu;
let wav_path = "./assets/audio/example.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);
// let t = Tensor::rand(-1.0, 1.0, (2, 2), &device)?;
// println!("t: {}", t);
// let re_t = t.recip()?;
// println!("re_t: {}", re_t);
Ok(())
}
tests/test_minicpm4.rs
+19 -39
View File
@@ -1,39 +1,34 @@
use std::time::Instant;
use std::{pin::pin, time::Instant};
use aha::models::{minicpm4::generate::MiniCPMGenerateModel, GenerateModel};
use anyhow::Result;
use candle_core::{DType, Device};
use openai_dive::v1::resources::chat::ChatCompletionParameters;
use rocket::futures::StreamExt;
#[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;
fn minicpm_generate() -> Result<()> {
// test with cpu :(太慢了, : RUST_BACKTRACE=1 cargo test minicpm_generate -- --nocapture
// test with cuda: RUST_BACKTRACE=1 cargo test -F cuda minicpm_generate -- --nocapture
// test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn minicpm_generate -- --nocapture
let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
let message = r#"
{
"temperature": 0.3,
"top_p": 0.8,
"model": "minicpm4",
"messages": [
{
"role": "user",
"content": [
{
"type": "text",
"text": "你是谁"
}
]
"content": "贾宝玉和孙悟空有什么关系"
}
]
}
"#;
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 mut model = MiniCPMGenerateModel::init(model_path, None, None)?;
let i_duration = i_start.elapsed();
println!("Time elapsed in load model is: {:?}", i_duration);
@@ -47,40 +42,25 @@ fn qwen2_5vl_generate() -> Result<()> {
}
#[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/";
async fn minicpm_stream() -> Result<()> {
// test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn minicpm_stream -- --nocapture
let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
let message = r#"
{
"model": "qwen2.5vl",
"model": "minicpm4",
"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"
}
]
"content": "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 mut model = MiniCPMGenerateModel::init(model_path, None, None)?;
let i_duration = i_start.elapsed();
println!("Time elapsed in load model is: {:?}", i_duration);
tests/test_qwen2_5vl.rs
+7 -10
View File
@@ -1,7 +1,6 @@
use std::{pin::pin, time::Instant};
use aha::{
ModelType,
models::{GenerateModel, qwen2_5vl::generate::Qwen2_5VLGenerateModel},
};
use anyhow::Result;
@@ -14,8 +13,8 @@ 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 device = Device::cuda_if_available(0)?;
// let dtype = DType::BF16;
let model_path = "/home/jhq/huggingface_model/Qwen/Qwen2.5-VL-3B-Instruct/";
@@ -30,7 +29,7 @@ fn qwen2_5vl_generate() -> Result<()> {
"type": "image",
"image_url":
{
"url": "file://./assets/img/ocr_test.png"
"url": "file://./assets/img/ocr_test1.png"
}
},
{
@@ -44,8 +43,7 @@ fn qwen2_5vl_generate() -> Result<()> {
"#;
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 mut model = Qwen2_5VLGenerateModel::init(model_path, None, None)?;
let i_duration = i_start.elapsed();
println!("Time elapsed in load model is: {:?}", i_duration);
@@ -61,8 +59,8 @@ fn qwen2_5vl_generate() -> Result<()> {
#[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 device = Device::cuda_if_available(0)?;
// let dtype = DType::BF16;
let model_path = "/home/jhq/huggingface_model/Qwen/Qwen2.5-VL-3B-Instruct/";
@@ -91,8 +89,7 @@ async fn qwen2_5vl_stream() -> Result<()> {
"#;
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 mut model = Qwen2_5VLGenerateModel::init(model_path, None, None)?;
let i_duration = i_start.elapsed();
println!("Time elapsed in load model is: {:?}", i_duration);
tests/test_voxcpm.rs
+57
View File
@@ -0,0 +1,57 @@
use std::collections::HashMap;
use anyhow::{Ok, Result};
use aha::{models::voxcpm::{audio_vae::AudioVAE, config::VoxCPMConfig, model::VoxCPMModel, tokenizer::SingleChineseTokenizer}, utils::utils::{find_type_files, get_device}};
use candle_core::pickle::read_all_with_key;
use candle_nn::VarBuilder;
#[test]
fn voxcpm_generate() -> Result<()> {
let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
let model_list = find_type_files(&model_path, "pth")?;
println!(" pth model_list: {:?}", model_list);
let dev = get_device(None);
let mut dict_to_hashmap = HashMap::new();
let mut dtype = candle_core::DType::F32;
for m in model_list {
let dict = read_all_with_key(m, Some("state_dict"))?;
dtype = dict[0].1.dtype();
for (k, v) in dict {
// println!("key: {}, tensor shape: {:?}", k, v);
dict_to_hashmap.insert(k, v);
}
}
let vb = VarBuilder::from_tensors(dict_to_hashmap, dtype, &dev);
let audio_vae = AudioVAE::new(vb, 128, vec![2, 5, 8, 8], Some(64), 1536, vec![8, 8, 5, 2], 16000)?;
println!("audio vae load down");
let model_list = find_type_files(&model_path, "bin")?;
println!(" bin model_list: {:?}", model_list);
dict_to_hashmap = HashMap::new();
for m in model_list {
let dict = read_all_with_key(m, Some("state_dict"))?;
dtype = dict[0].1.dtype();
for (k, v) in dict {
// println!("key: {}, tensor shape: {:?}", k, v);
dict_to_hashmap.insert(k, v);
}
}
let vb_vox = VarBuilder::from_tensors(dict_to_hashmap, dtype, &dev);
let config_path = model_path.to_string() + "/config.json";
let config: VoxCPMConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
let tokenizer = SingleChineseTokenizer::new(model_path)?;
let mut voxcpm = VoxCPMModel::new(vb_vox, config, tokenizer, audio_vae)?;
let generate = voxcpm.generate("你好啊，这是初始测试语句".to_string(), None, None, 2, 30, 10, 2.0, false, 3, 6.0)?;
// let audio_path = "./assets/audio/example.wav";
Ok(())
}
#[test]
fn voxcpm_tokenizer() -> Result<()> {
let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
let tokenizer = SingleChineseTokenizer::new(model_path)?;
let ids = tokenizer.encode("你好啊，你吃饭了吗".to_string())?;
println!("ids: {:?}", ids);
Ok(())
}
tests/weight_test.rs
+29 -3
View File
@@ -1,10 +1,14 @@
use aha::utils::utils::find_safetensors_files;
use std::collections::HashMap;
use aha::utils::utils::{find_type_files, get_device};
use anyhow::Result;
use candle_core::{safetensors, Device};
use candle_core::{pickle::{read_all_with_key, read_pth_tensor_info, PthTensors}, safetensors, Device, Tensor};
use candle_nn::VarBuilder;
#[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 model_list = find_type_files(&model_path, "safetensors")?;
let device = Device::Cpu;
for m in model_list {
let weights = safetensors::load(m, &device)?;
@@ -15,4 +19,26 @@ fn minicpm4_weight() -> Result<()> {
}
}
Ok(())
}
#[test]
fn voxcpm_weight() -> Result<()> {
let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
let model_list = find_type_files(&model_path, "pth")?;
println!("model_list: {:?}", model_list);
let dev = get_device(None);
let mut dict_to_hashmap = HashMap::new();
let mut dtype = candle_core::DType::F16;
for m in model_list {
let dict = read_all_with_key(m, Some("state_dict"))?;
dtype = dict[0].1.dtype();
for (k, v) in dict {
println!("key: {}, tensor shape: {:?}", k, v);
dict_to_hashmap.insert(k, v);
}
}
let vb = VarBuilder::from_tensors(dict_to_hashmap, dtype, &dev);
let contain_key = vb.contains_tensor("encoder.block.4.block.2.block.3.weight_g");
println!("contain encoder.block.4.block.2.block.3.weight_g: {}", contain_key);
Ok(())
}