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

add LFM2.5-VL-1.6B|LFM2-VL-1.6B

Browse Source
This commit is contained in:
jhqxxx
2026-03-30 00:44:41 +08:00
parent b881dfcd8d
commit ea78da7834
27 changed files with 847 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/config_tests.rs
+12 -3
View File
@@ -1,5 +1,13 @@
use aha::models::{
deepseek_ocr::config::DeepseekOCRConfig, hunyuan_ocr::config::HunYuanVLConfig, lfm2::config::Lfm2Config, lfm2vl::config::{Lfm2ProcessorConfig, Lfm2VLConfig}, minicpm4::config::MiniCPM4Config, paddleocr_vl::config::PaddleOCRVLConfig, qwen2_5vl::config::Qwen2_5VLConfig, qwen3vl::config::Qwen3VLConfig, voxcpm::config::VoxCPMConfig
deepseek_ocr::config::DeepseekOCRConfig,
hunyuan_ocr::config::HunYuanVLConfig,
lfm2::config::Lfm2Config,
lfm2vl::config::{Lfm2ProcessorConfig, Lfm2VLConfig},
minicpm4::config::MiniCPM4Config,
paddleocr_vl::config::PaddleOCRVLConfig,
qwen2_5vl::config::Qwen2_5VLConfig,
qwen3vl::config::Qwen3VLConfig,
voxcpm::config::VoxCPMConfig,
};
use anyhow::Result;
@@ -104,7 +112,8 @@ fn lfm2vl_config() -> Result<()> {
let config: Lfm2VLConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
println!("{:?}", config);
let processor_config_path = model_path.to_string() + "/processor_config.json";
let processor_config: Lfm2ProcessorConfig = serde_json::from_slice(&std::fs::read(processor_config_path)?)?;
let processor_config: Lfm2ProcessorConfig =
serde_json::from_slice(&std::fs::read(processor_config_path)?)?;
println!("{:?}", processor_config);
Ok(())
}
}
tests/test_health_models.rs
+8 -8
View File
@@ -8,8 +8,8 @@ fn test_model_type_classification() {
// Since get_model_type and get_model_id are private to api.rs,
// we document the expected behavior here for reference:
//
// LLM models: MiniCPM4_0_5B, Qwen2_5vl3B, Qwen2_5vl7B, Qwen3_0_6B,
// Qwen3vl2B, Qwen3vl4B, Qwen3vl8B, Qwen3vl32B
// LLM models: MiniCPM4_0_5B, Qwen2_5VL3B, Qwen2_5VL7B, Qwen3_0_6B,
// Qwen3VL2B, Qwen3VL4B, Qwen3VL8B, Qwen3VL32B
// OCR models: DeepSeekOCR, HunyuanOCR, PaddleOCRVL
// ASR models: Qwen3ASR0_6B, Qwen3ASR1_7B, GlmASRNano2512, FunASRNano2512
// Image models: RMBG2_0, VoxCPM, VoxCPM1_5
@@ -17,13 +17,13 @@ fn test_model_type_classification() {
// This test documents the expected model type classification
let llm_models = [
WhichModel::MiniCPM4_0_5B,
WhichModel::Qwen2_5vl3B,
WhichModel::Qwen2_5vl7B,
WhichModel::Qwen2_5VL3B,
WhichModel::Qwen2_5VL7B,
WhichModel::Qwen3_0_6B,
WhichModel::Qwen3vl2B,
WhichModel::Qwen3vl4B,
WhichModel::Qwen3vl8B,
WhichModel::Qwen3vl32B,
WhichModel::Qwen3VL2B,
WhichModel::Qwen3VL4B,
WhichModel::Qwen3VL8B,
WhichModel::Qwen3VL32B,
];
let ocr_models = [
tests/test_lfm2.rs
+3 -4
View File
@@ -9,19 +9,18 @@ use std::{pin::pin, time::Instant};
#[test]
fn lfm2_generate() -> Result<()> {
// test with cuda: RUST_BACKTRACE=1 cargo test -F cuda --test test_lfm2 lfm2_generate -r -- --nocapture
// test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn qwen3_0_6b_generate -r -- --nocapture
let save_dir =
aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
// let model_path = format!("{}/LiquidAI/LFM2-1.2B/", save_dir);
let model_path = format!("{}/LiquidAI/LFM2.5-1.2B-Instruct/", save_dir);
let model_path = format!("{}/LiquidAI/LFM2-1.2B/", save_dir);
// let model_path = format!("{}/LiquidAI/LFM2.5-1.2B-Instruct/", save_dir);
let message = r#"
{
"model": "lfm2",
"messages": [
{
"role": "user",
"content": "你如何看待AI"
"content": "你是谁，你如何看待AI"
}
]
}
tests/test_lfm2vl.rs
+68 -13
View File
@@ -1,13 +1,71 @@
use std::time::Instant;
use std::{pin::pin, time::Instant};
use aha::{chat::ChatCompletionParameters, models::lfm2vl::generate::Lfm2VLGenerateModel};
use aha::{
chat::ChatCompletionParameters,
models::{GenerateModel, lfm2vl::generate::Lfm2VLGenerateModel},
};
use anyhow::Result;
use rocket::futures::StreamExt;
#[test]
fn lfm2vl_generate() -> Result<()> {
// test with cuda: RUST_BACKTRACE=1 cargo test -F cuda --test test_lfm2vl lfm2vl_generate -r -- --nocapture
let save_dir =
aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
// let model_path = format!("{}/LiquidAI/LFM2.5-VL-1.6B/", save_dir);
let model_path = format!("{}/LiquidAI/LFM2-VL-1.6B/", save_dir);
let message = r#"
{
"model": "lfm2vl",
"messages": [
{
"role": "user",
"content": [
{
"type": "image",
"image_url":
{
"url": "file://./assets/img/ocr_test1.png"
}
},
{
"type": "text",
"text": "图片里面是什么"
}
]
}
]
}
"#;
let mes: ChatCompletionParameters = serde_json::from_str(message)?;
let i_start = Instant::now();
let mut model = Lfm2VLGenerateModel::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 result = model.generate(mes)?;
let i_duration = i_start.elapsed();
println!("generate: \n {:?}", result);
if let Some(usage) = &result.usage {
let num_token = usage.total_tokens;
let duration_secs = i_duration.as_secs_f64();
let tps = num_token as f64 / duration_secs;
println!("Tokens per second (TPS): {:.2}", tps);
}
println!("Time elapsed in generate is: {:?}", i_duration);
Ok(())
}
#[tokio::test]
async fn lfm2vl_stream() -> Result<()> {
// test with cuda: RUST_BACKTRACE=1 cargo test -F cuda --test test_lfm2vl lfm2vl_stream -r -- --nocapture
// test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn qwen3_0_6b_generate -r -- --nocapture
let save_dir =
aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
// let model_path = format!("{}/LiquidAI/LFM2-1.2B/", save_dir);
let model_path = format!("{}/LiquidAI/LFM2.5-VL-1.6B/", save_dir);
let message = r#"
{
@@ -39,16 +97,13 @@ fn lfm2vl_generate() -> Result<()> {
println!("Time elapsed in load model is: {:?}", i_duration);
let i_start = Instant::now();
let result = model.generate(mes)?;
// let result = model.generate(mes)?;
let mut stream = pin!(model.generate_stream(mes)?);
let i_duration = i_start.elapsed();
// println!("generate: \n {:?}", result);
// if let Some(usage) = &result.usage {
// let num_token = usage.total_tokens;
// let duration_secs = i_duration.as_secs_f64();
// let tps = num_token as f64 / duration_secs;
// println!("Tokens per second (TPS): {:.2}", tps);
// }
// println!("Time elapsed in generate is: {:?}", i_duration);
while let Some(token) = stream.next().await {
println!("generate: \n {:?}", token);
}
println!("Time elapsed in generate is: {:?}", i_duration);
Ok(())
}
}
tests/weight_test.rs
+23
View File
@@ -287,3 +287,26 @@ fn lfm2_weight() -> Result<()> {
println!("model_list: {:?}", model_list);
Ok(())
}
#[test]
fn lfm2vl_weight() -> Result<()> {
// cargo test -F cuda --test weight_test lfm2vl_weight -r -- --nocapture
let save_dir =
aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
// let model_path = format!("{}/LiquidAI/LFM2.5-VL-1.6B/", save_dir);
let model_path = format!("{}/LiquidAI/LFM2-VL-1.6B/", save_dir);
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("lm_head") {
// println!("=== {} === {:?}", key, tensor.shape());
// }
println!("=== {} === {:?}", key, tensor.shape());
}
}
println!("model_list: {:?}", model_list);
Ok(())
}