add qwen3-embedding and all-minilm-l6-v2

src/cli/mod.rs

@@ -238,46 +238,28 @@ pub(crate) fn run_run(args: RunArgs) -> anyhow::Result<()> {
         None => get_default_weight_path(model),
     };
     match model {
+        WhichModel::AllMiniLML6V2 => {
+            all_minilm_l6_v2::AllMiniLML6V2Exec::run(&input, output.as_deref(), &weight_path)?;
+        }
         WhichModel::MiniCPM4_0_5B => {
             minicpm4::MiniCPM4Exec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::LFM2_1_2B => {
+        WhichModel::LFM2_1_2B | WhichModel::LFM2_5_1_2BInstruct => {
             lfm2::Lfm2Exec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::LFM2_5_1_2BInstruct => {
-            lfm2::Lfm2Exec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::LFM2_5VL1_6B => {
+        WhichModel::LFM2_5VL1_6B | WhichModel::LFM2VL1_6B => {
             lfm2vl::Lfm2VLExec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::LFM2VL1_6B => {
-            lfm2vl::Lfm2VLExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen2_5VL3B => {
+        WhichModel::Qwen2_5VL3B | WhichModel::Qwen2_5VL7B => {
             qwen2_5vl::Qwen2_5VLExec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::Qwen2_5VL7B => {
-            qwen2_5vl::Qwen2_5VLExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3_0_6B => {
+        WhichModel::Qwen3_0_6B | WhichModel::Qwen3_1_7B | WhichModel::Qwen3_4B => {
             qwen3::Qwen3Exec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::Qwen3_1_7B => {
-            qwen3::Qwen3Exec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3_4B => {
-            qwen3::Qwen3Exec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3_5_0_8B => {
-            qwen3_5::Qwen3_5Exec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3_5_2B => {
-            qwen3_5::Qwen3_5Exec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3_5_4B => {
-            qwen3_5::Qwen3_5Exec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3_5_9B => {
+        WhichModel::Qwen3_5_0_8B
+        | WhichModel::Qwen3_5_2B
+        | WhichModel::Qwen3_5_4B
+        | WhichModel::Qwen3_5_9B => {
             qwen3_5::Qwen3_5Exec::run(&input, output.as_deref(), &weight_path)?;
         }
         WhichModel::Qwen3_5Gguf => {
@@ -288,46 +270,33 @@ pub(crate) fn run_run(args: RunArgs) -> anyhow::Result<()> {
                 path_common.mmproj_path,
             )?;
         }
-        WhichModel::Qwen3ASR0_6B => {
+        WhichModel::Qwen3ASR0_6B | WhichModel::Qwen3ASR1_7B => {
             qwen3_asr::Qwen3ASRExec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::Qwen3ASR1_7B => {
-            qwen3_asr::Qwen3ASRExec::run(&input, output.as_deref(), &weight_path)?;
+        WhichModel::Qwen3Embedding0_6B
+        | WhichModel::Qwen3Embedding4B
+        | WhichModel::Qwen3Embedding8B => {
+            qwen3_embedding::Qwen3EmbeddingExec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::Qwen3VL2B => {
+        WhichModel::Qwen3VL2B
+        | WhichModel::Qwen3VL4B
+        | WhichModel::Qwen3VL8B
+        | WhichModel::Qwen3VL32B => {
             qwen3vl::Qwen3VLExec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::Qwen3VL4B => {
-            qwen3vl::Qwen3VLExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3VL8B => {
-            qwen3vl::Qwen3VLExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::Qwen3VL32B => {
-            qwen3vl::Qwen3VLExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::DeepSeekOCR => {
-            deepseek_ocr::DeepSeekORExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::DeepSeekOCR2 => {
+        WhichModel::DeepSeekOCR | WhichModel::DeepSeekOCR2 => {
             deepseek_ocr::DeepSeekORExec::run(&input, output.as_deref(), &weight_path)?;
         }
         WhichModel::HunyuanOCR => {
             hunyuan_ocr::HunyuanORExec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::PaddleOCRVL => {
-            paddleocr_vl::PaddleOVLExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::PaddleOCRVL1_5 => {
+        WhichModel::PaddleOCRVL | WhichModel::PaddleOCRVL1_5 => {
             paddleocr_vl::PaddleOVLExec::run(&input, output.as_deref(), &weight_path)?;
         }
         WhichModel::RMBG2_0 => {
             rmbg2_0::RMBG2_0Exec::run(&input, output.as_deref(), &weight_path)?;
         }
-        WhichModel::VoxCPM => {
-            voxcpm::VoxCPMExec::run(&input, output.as_deref(), &weight_path)?;
-        }
-        WhichModel::VoxCPM1_5 => {
+        WhichModel::VoxCPM | WhichModel::VoxCPM1_5 => {
             voxcpm::VoxCPMExec::run(&input, output.as_deref(), &weight_path)?;
         }
         WhichModel::GlmASRNano2512 => {

src/exec/all_minilm_l6_v2.rs

@@ -0,0 +1,41 @@
+use std::time::Instant;
+
+use crate::{
+    exec::ExecModel,
+    models::{all_minilm_l6_v2::AllMiniLML6V2Embedding, common::embedding::TextEmbedding},
+    utils::get_file_path,
+};
+use anyhow::Result;
+
+pub struct AllMiniLML6V2Exec;
+
+impl ExecModel for AllMiniLML6V2Exec {
+    fn run(input: &[String], output: Option<&str>, weight_path: &str) -> Result<()> {
+        let input_text = &input[0];
+        let target_text = if input_text.starts_with("file://") {
+            let path = get_file_path(input_text)?;
+            std::fs::read_to_string(path)?
+        } else {
+            input_text.clone()
+        };
+
+        let i_start = Instant::now();
+        let mut model = AllMiniLML6V2Embedding::init(weight_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.embed_texts(&[target_text])?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/mod.rs

@@ -3,6 +3,7 @@
 //! This module provides model-specific exec implementations for the `run` subcommand.
 //! Each model has its own exec module that handles input/output parsing and model invocation.
 
+pub mod all_minilm_l6_v2;
 pub mod deepseek_ocr;
 pub mod fun_asr_nano;
 pub mod glm_asr_nano;
@@ -16,6 +17,7 @@ pub mod qwen2_5vl;
 pub mod qwen3;
 pub mod qwen3_5;
 pub mod qwen3_asr;
+pub mod qwen3_embedding;
 pub mod qwen3vl;
 pub mod rmbg2_0;
 pub mod voxcpm;

src/exec/qwen3_embedding.rs

@@ -0,0 +1,42 @@
+use std::time::Instant;
+
+use crate::{
+    exec::ExecModel,
+    models::{common::embedding::TextEmbedding, qwen3_embedding::Qwen3Embedding},
+    utils::get_file_path,
+};
+use anyhow::Result;
+
+pub struct Qwen3EmbeddingExec;
+
+impl ExecModel for Qwen3EmbeddingExec {
+    fn run(input: &[String], output: Option<&str>, weight_path: &str) -> Result<()> {
+        let input_text = &input[0];
+        let target_text = if input_text.starts_with("file://") {
+            // let path = &input[7..];
+            let path = get_file_path(input_text)?;
+            std::fs::read_to_string(path)?
+        } else {
+            input_text.clone()
+        };
+
+        let i_start = Instant::now();
+        let mut model = Qwen3Embedding::init(weight_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.embed_texts(&[target_text])?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/main.rs

@@ -6,6 +6,8 @@ use crate::cli::{
 };
 
 mod cli;
+#[allow(unused)]
+mod params;
 mod server;
 
 #[tokio::main]

src/models/all_minilm_l6_v2/mod.rs

@@ -0,0 +1,78 @@
+use crate::{
+    models::common::embedding::{NormalizeType, TextEmbedding},
+    tokenizer::TokenizerModel,
+    utils::{find_type_files, get_device, get_dtype},
+};
+use anyhow::{Result, anyhow};
+use candle_core::{DType, Device, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::models::bert::{BertModel, Config as BertConfig};
+
+pub struct AllMiniLML6V2Embedding {
+    tokenizer: TokenizerModel,
+    model: BertModel,
+    device: Device,
+    normalize: NormalizeType,
+}
+
+impl AllMiniLML6V2Embedding {
+    pub fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self> {
+        let tokenizer = TokenizerModel::init(path)?;
+        let config_path = path.to_string() + "/config.json";
+        let cfg: BertConfig = serde_json::from_slice(&std::fs::read(config_path)?)?;
+        let device = get_device(device);
+        let dtype = get_dtype(dtype, "float32");
+        let model_list = find_type_files(path, "safetensors")?;
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&model_list, dtype, &device)? };
+        let model = BertModel::load(vb, &cfg)?;
+        Ok(Self {
+            tokenizer,
+            model,
+            device,
+            normalize: NormalizeType::L2,
+        })
+    }
+
+    fn prepare_token_ids(&self, text: &str) -> Result<Vec<u32>> {
+        let mut token_ids = self.tokenizer.text_encode_vec(text.to_string(), true)?;
+        token_ids = token_ids
+            .into_iter()
+            .filter(|&x| x != 0)
+            .collect::<Vec<u32>>();
+        if token_ids.is_empty() {
+            return Err(anyhow!("embedding tokenized input cannot be empty"));
+        }
+        Ok(token_ids)
+    }
+    fn embed_one(&mut self, text: &str) -> Result<Vec<f32>> {
+        let token_ids = self.prepare_token_ids(text)?;
+        let seq_len = token_ids.len();
+        let input_ids = Tensor::from_slice(&token_ids, (1, seq_len), &self.device)?;
+        let token_type_ids = Tensor::zeros((1, seq_len), DType::U32, &self.device)?;
+        let attention_mask = Tensor::ones((1, seq_len), DType::U32, &self.device)?;
+        let hidden = self
+            .model
+            .forward(&input_ids, &token_type_ids, Some(&attention_mask))?
+            .to_dtype(DType::F32)?;
+        let hidden = hidden.mean(1)?;
+        let embed = self
+            .normalize
+            .normalize(&hidden, hidden.rank() - 1)?
+            .squeeze(0)?;
+        let embed = embed.to_vec1::<f32>()?;
+        Ok(embed)
+    }
+}
+
+impl TextEmbedding for AllMiniLML6V2Embedding {
+    fn embed_texts(&mut self, input: &[String]) -> Result<Vec<Vec<f32>>> {
+        if input.is_empty() {
+            return Err(anyhow!("embedding input cannot be empty"));
+        }
+        let mut out = Vec::with_capacity(input.len());
+        for text in input {
+            out.push(self.embed_one(text)?);
+        }
+        Ok(out)
+    }
+}

src/models/campplus/mod.rs

@@ -2,9 +2,8 @@ use anyhow::Result;
 use candle_core::{D, Tensor};
 use candle_nn::{BatchNorm, Conv1d, Conv2d, Module, ModuleT, VarBuilder, ops::sigmoid};
 
-use crate::{
-    models::common::modules::{get_batch_norm, get_conv1d, get_conv2d},
-    utils::tensor_utils::{pool1d, statistics_pooling},
+use crate::models::common::modules::{
+    get_batch_norm, get_conv1d, get_conv2d, pool1d, statistics_pooling,
 };
 
 pub struct Shortcut {

src/models/common/embedding.rs

@@ -0,0 +1,29 @@
+use anyhow::Result;
+use candle_core::Tensor;
+
+use crate::models::common::modules::{
+    l1_normalize, l2_normalize, max_abs_normalize, min_max_normalize, z_score_normalize,
+};
+pub trait TextEmbedding {
+    fn embed_texts(&mut self, input: &[String]) -> Result<Vec<Vec<f32>>>;
+}
+
+pub enum NormalizeType {
+    L1,
+    L2,
+    ZScore,
+    MinMax,
+    MaxAbs,
+}
+
+impl NormalizeType {
+    pub fn normalize(&self, t: &Tensor, dim: usize) -> Result<Tensor> {
+        match self {
+            NormalizeType::L1 => l1_normalize(t, dim),
+            NormalizeType::L2 => l2_normalize(t, dim),
+            NormalizeType::ZScore => z_score_normalize(t, dim),
+            NormalizeType::MinMax => min_max_normalize(t, dim),
+            NormalizeType::MaxAbs => max_abs_normalize(t, dim),
+        }
+    }
+}

src/models/common/mod.rs

@@ -1,5 +1,6 @@
 use anyhow::Result;
 use candle_core::Tensor;
+pub mod embedding;
 pub mod generate;
 pub mod gguf;
 pub mod model_mapping;

src/models/common/model_mapping.rs

@@ -2,6 +2,8 @@ use clap::ValueEnum;
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, clap::ValueEnum)]
 pub enum WhichModel {
+    #[value(name = "sentence-transformers/all-MiniLM-L6-v2")]
+    AllMiniLML6V2,
     #[value(name = "LiquidAI/LFM2-1.2B")]
     LFM2_1_2B,
     #[value(name = "LiquidAI/LFM2.5-1.2B-Instruct")]
@@ -36,6 +38,12 @@ pub enum WhichModel {
     Qwen3ASR0_6B,
     #[value(name = "Qwen/Qwen3-ASR-1.7B")]
     Qwen3ASR1_7B,
+    #[value(name = "Qwen/Qwen3-Embedding-0.6B")]
+    Qwen3Embedding0_6B,
+    #[value(name = "Qwen/Qwen3-Embedding-4B")]
+    Qwen3Embedding4B,
+    #[value(name = "Qwen/Qwen3-Embedding-8B")]
+    Qwen3Embedding8B,
     #[value(name = "Qwen/Qwen3-VL-2B-Instruct")]
     Qwen3VL2B,
     #[value(name = "Qwen/Qwen3-VL-4B-Instruct")]
@@ -153,6 +161,10 @@ impl WhichModel {
             WhichModel::RMBG2_0 => "image",
             // TTS models
             WhichModel::VoxCPM | WhichModel::VoxCPM1_5 => "tts",
+            WhichModel::Qwen3Embedding0_6B
+            | WhichModel::Qwen3Embedding4B
+            | WhichModel::Qwen3Embedding8B
+            | WhichModel::AllMiniLML6V2 => "embedding",
         }
     }
 }

src/models/common/modules.rs

@@ -1,5 +1,5 @@
-use anyhow::Result;
-use candle_core::{D, IndexOp, Tensor};
+use anyhow::{Result, anyhow};
+use candle_core::{D, DType, IndexOp, Tensor};
 use candle_nn::{
     Activation, BatchNorm, BatchNormConfig, Conv1d, Conv1dConfig, Conv2d, Conv2dConfig,
     ConvTranspose1d, ConvTranspose1dConfig, Embedding, LayerNorm, LayerNormConfig, Linear, Module,
@@ -9,7 +9,7 @@ use candle_nn::{
 
 use crate::{
     position_embed::rope::{RoPE, apply_rotary_pos_emb, apply_rotary_pos_emb_roformer},
-    utils::tensor_utils::{prepare_causal_attention_mask, repeat_kv},
+    utils::tensor_utils::{pad_replicate_last_dim, prepare_causal_attention_mask, repeat_kv},
 };
 
 #[derive(Debug, Clone)]
@@ -1327,3 +1327,134 @@ pub fn conv1d_depthwise(input: &Tensor, weight: &Tensor, bias: Option<&Tensor>)
         }
     }
 }
+
+pub fn log10(t: &Tensor) -> Result<Tensor> {
+    Ok(t.log()?.affine(1.0 / 10.0_f64.ln(), 0.0)?)
+}
+
+pub fn max_abs_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
+    let rank = t.rank();
+    if dim >= rank {
+        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
+    }
+    Ok(t.broadcast_div(&t.abs()?.max_keepdim(dim)?)?)
+}
+
+pub fn min_max_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
+    let rank = t.rank();
+    if dim >= rank {
+        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
+    }
+    let t_min = t.min_keepdim(dim)?;
+    Ok(t.broadcast_sub(&t_min)?
+        .broadcast_div(&t.max_keepdim(dim)?.sub(&t_min)?)?)
+}
+
+pub fn z_score_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
+    let rank = t.rank();
+    if dim >= rank {
+        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
+    }
+    Ok(t.broadcast_sub(&t.mean_keepdim(dim)?)?
+        .broadcast_div(&t.var_keepdim(dim)?.sqrt()?)?)
+}
+
+pub fn l2_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
+    let rank = t.rank();
+    if dim >= rank {
+        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
+    }
+    let l2_norm = t.sqr()?.sum_keepdim(dim)?.affine(1.0, 1e-6)?.sqrt()?;
+    Ok(t.broadcast_div(&l2_norm)?)
+}
+
+pub fn l1_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
+    let rank = t.rank();
+    if dim >= rank {
+        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
+    }
+    let l1_norm = t.abs()?.sum_keepdim(dim)?;
+    Ok(t.broadcast_div(&l1_norm)?)
+}
+
+pub fn pool1d(xs: &Tensor, pool_size: usize, ceil_mode: bool, stype: &str) -> Result<Tensor> {
+    // xs: (bs, c, dim)
+    // ceil_mode: 是否保留不完整窗口，为true时通过pad实现
+    if pool_size == 0 {
+        return Err(anyhow!("pool_size must be greater than 0"));
+    }
+    let (bs, c, dim) = xs.dims3()?;
+    let xs_reshape = if ceil_mode {
+        let remain = dim % pool_size;
+        if remain > 0 {
+            let pad = pool_size - remain;
+            let xs_pad = pad_replicate_last_dim(xs, (0, pad))?;
+            xs_pad.reshape((bs, c, (), pool_size))?
+        } else {
+            xs.reshape((bs, c, (), pool_size))?
+        }
+    } else {
+        let remain = dim % pool_size;
+        if remain > 0 {
+            let xs_del = xs.narrow(D::Minus1, 0, dim - remain)?;
+            xs_del.reshape((bs, c, (), pool_size))?
+        } else {
+            xs.reshape((bs, c, (), pool_size))?
+        }
+    };
+    let xs_pool = match stype {
+        "avg" => xs_reshape.mean(D::Minus1)?,
+        "max" => xs_reshape.max(D::Minus1)?,
+        "min" => xs_reshape.min(D::Minus1)?,
+        _ => {
+            return Err(anyhow!(
+                "unsupported pool type: {}, supported types are: avg, max, min",
+                stype
+            ));
+        }
+    };
+    Ok(xs_pool)
+}
+
+pub fn statistics_pooling(xs: &Tensor, dim: D, keepdim: bool) -> Result<Tensor> {
+    let mean = xs.mean(dim)?;
+    let std = xs.var(dim)?.sqrt()?;
+    let mut stats = Tensor::cat(&[mean, std], D::Minus1)?;
+    if keepdim {
+        stats = stats.unsqueeze(dim)?;
+    }
+    Ok(stats)
+}
+pub fn float_range_normalize(t: &Tensor) -> Result<Tensor> {
+    let peak = t
+        .to_dtype(DType::F32)?
+        .abs()?
+        .max_all()?
+        .to_scalar::<f32>()?;
+    if peak == 0.0 {
+        return Ok(t.clone());
+    }
+    let mut t = t.clone();
+    if peak > 1.0 {
+        t = t.affine(1.0 / peak as f64, 0.0)?;
+    }
+    t = t.clamp(-1.0, 1.0)?;
+    Ok(t)
+}
+
+pub fn cosine_similarity(query_vector: &Tensor, matrix: &Tensor) -> Result<Tensor> {
+    // query_vector: (n, dim)
+    // matrix: (m, dim)
+    let query_norm = l2_normalize(query_vector, query_vector.rank() - 1)?;
+    let matrix_norm = l2_normalize(matrix, matrix.rank() - 1)?;
+    let similarity = query_norm
+        .matmul(&matrix_norm.transpose(D::Minus1, D::Minus2)?)?
+        .squeeze(D::Minus1)?;
+    Ok(similarity)
+}
+
+pub fn quick_gelu(xs: &Tensor) -> Result<Tensor> {
+    let x = xs.affine(1.702, 0.0)?;
+    let x = sigmoid(&x)?;
+    Ok(xs.mul(&x)?)
+}

src/models/deepseek_ocr/model.rs

@@ -16,7 +16,7 @@ use crate::{
             InferenceModel,
             modules::{
                 GateUpDownMLP, NaiveAttention, QKVCatAttention, TwoLinearMLP,
-                eager_attention_forward, get_conv2d, get_layer_norm,
+                eager_attention_forward, get_conv2d, get_layer_norm, quick_gelu,
             },
         },
         deepseek_ocr::config::{DeepseekOCRConfig, DeepseekV2Config},
@@ -27,7 +27,7 @@ use crate::{
         interpolate::{interpolate_bicubic, interpolate_linear_1d},
         tensor_utils::{
             attn_masked_fill, index_select_2d, masked_scatter_dim0, nonzero, onehot,
-            prepare_causal_attention_mask, quick_gelu, topk,
+            prepare_causal_attention_mask, topk,
         },
     },
 };

src/models/feature_extractor/feature_extraction_whisper.rs

@@ -1,9 +1,12 @@
 use anyhow::Result;
 use candle_core::{D, Device, Tensor};
 
-use crate::utils::{
-    audio_utils::{create_hann_window, mel_filter_bank, torch_stft},
-    tensor_utils::{log10, pad_reflect_last_dim},
+use crate::{
+    models::common::modules::log10,
+    utils::{
+        audio_utils::{create_hann_window, mel_filter_bank, torch_stft},
+        tensor_utils::pad_reflect_last_dim,
+    },
 };
 
 pub struct WhisperFeatureExtractor {

src/models/feature_extractor/seamless_m4t_feature_extractor.rs

@@ -1,9 +1,12 @@
 use anyhow::Result;
 use candle_core::{D, Device, Tensor};
 
-use crate::utils::{
-    audio_utils::{create_povey_window, mel_filter_bank, spectrogram},
-    tensor_utils::{PaddingSide, z_score_normalize},
+use crate::{
+    models::common::modules::z_score_normalize,
+    utils::{
+        audio_utils::{create_povey_window, mel_filter_bank, spectrogram},
+        tensor_utils::PaddingSide,
+    },
 };
 
 pub struct SeamlessM4TFeatureExtractor {

src/models/mask_gct/model.rs

@@ -6,10 +6,10 @@ use candle_nn::{
 
 use crate::{
     models::{
-        common::modules::{WNConv1d, conv1d_depthwise, get_conv1d, get_layer_norm},
+        common::modules::{WNConv1d, conv1d_depthwise, get_conv1d, get_layer_norm, l2_normalize},
         mask_gct::config::SemanticCodec,
     },
-    utils::{interpolate::interpolate_nearest_1d, tensor_utils::l2_normalize},
+    utils::interpolate::interpolate_nearest_1d,
 };
 
 pub struct ConvNeXtBlock {

src/models/mod.rs

@@ -1,3 +1,4 @@
+pub mod all_minilm_l6_v2;
 pub mod bigvgan;
 pub mod campplus;
 pub mod common;
@@ -17,16 +18,22 @@ pub mod qwen2_5vl;
 pub mod qwen3;
 pub mod qwen3_5;
 pub mod qwen3_asr;
+pub mod qwen3_embedding;
 pub mod qwen3vl;
 pub mod rmbg2_0;
 pub mod voxcpm;
 pub mod w2v_bert_2_0;
 
 use crate::{
-    models::common::model_mapping::WhichModel,
+    models::{
+        all_minilm_l6_v2::AllMiniLML6V2Embedding,
+        common::{embedding::TextEmbedding, model_mapping::WhichModel},
+        qwen3_embedding::Qwen3Embedding,
+    },
     params::chat::{ChatCompletionChunkResponse, ChatCompletionParameters, ChatCompletionResponse},
 };
 use anyhow::{Result, anyhow};
+use candle_core::{DType, Device};
 use rocket::futures::Stream;
 
 use crate::models::{
@@ -57,6 +64,7 @@ pub trait GenerateModel {
 }
 
 pub enum ModelInstance<'a> {
+    AllMiniLML6V2(AllMiniLML6V2Embedding),
     MiniCPM4(MiniCPMGenerateModel<'a>),
     Lfm2(Lfm2GenerateModel<'a>),
     Lfm2VL(Lfm2VLGenerateModel<'a>),
@@ -64,6 +72,7 @@ pub enum ModelInstance<'a> {
     Qwen3(Qwen3GenerateModel<'a>),
     Qwen3_5(Qwen3_5GenerateModel<'a>),
     Qwen3ASR(Qwen3AsrGenerateModel<'a>),
+    Qwen3Embedding(Qwen3Embedding),
     Qwen3VL(Box<Qwen3VLGenerateModel<'a>>),
     DeepSeekOCR(DeepseekOCRGenerateModel),
     HunyuanOCR(HunyuanOCRGenerateModel<'a>),
@@ -78,12 +87,18 @@ pub enum ModelInstance<'a> {
 impl<'a> GenerateModel for ModelInstance<'a> {
     fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse> {
         match self {
+            ModelInstance::AllMiniLML6V2(_) => {
+                Err(anyhow!("embedding model does not support chat completions"))
+            }
             ModelInstance::MiniCPM4(model) => model.generate(mes),
             ModelInstance::Lfm2(model) => model.generate(mes),
             ModelInstance::Lfm2VL(model) => model.generate(mes),
             ModelInstance::Qwen2_5VL(model) => model.generate(mes),
             ModelInstance::Qwen3(model) => model.generate(mes),
             ModelInstance::Qwen3_5(model) => model.generate(mes),
+            ModelInstance::Qwen3Embedding(_) => {
+                Err(anyhow!("embedding model does not support chat completions"))
+            }
             ModelInstance::Qwen3ASR(model) => model.generate(mes),
             ModelInstance::Qwen3VL(model) => model.generate(mes),
             ModelInstance::DeepSeekOCR(model) => model.generate(mes),
@@ -109,12 +124,18 @@ impl<'a> GenerateModel for ModelInstance<'a> {
         >,
     > {
         match self {
+            ModelInstance::AllMiniLML6V2(_) => {
+                Err(anyhow!("embedding model does not support chat completions"))
+            }
             ModelInstance::MiniCPM4(model) => model.generate_stream(mes),
             ModelInstance::Lfm2(model) => model.generate_stream(mes),
             ModelInstance::Lfm2VL(model) => model.generate_stream(mes),
             ModelInstance::Qwen2_5VL(model) => model.generate_stream(mes),
             ModelInstance::Qwen3(model) => model.generate_stream(mes),
             ModelInstance::Qwen3_5(model) => model.generate_stream(mes),
+            ModelInstance::Qwen3Embedding(_) => Err(anyhow!(
+                "embedding model does not support streaming chat completions"
+            )),
             ModelInstance::Qwen3VL(model) => model.generate_stream(mes),
             ModelInstance::Qwen3ASR(model) => model.generate_stream(mes),
             ModelInstance::DeepSeekOCR(model) => model.generate_stream(mes),
@@ -129,16 +150,27 @@ impl<'a> GenerateModel for ModelInstance<'a> {
     }
 }
 
+impl<'a> ModelInstance<'a> {
+    pub fn embedding(&mut self, input: &[String]) -> Result<Vec<Vec<f32>>> {
+        match self {
+            ModelInstance::Qwen3Embedding(model) => model.embed_texts(input),
+            ModelInstance::AllMiniLML6V2(model) => model.embed_texts(input),
+            _ => Err(anyhow!("current model does not support embeddings")),
+        }
+    }
+}
+
 #[allow(unused)]
 pub fn load_gguf_model<'a>(
     model_type: WhichModel,
     config_path: Option<&str>, // 有些gguf未包含模型其他配置，需额外指定
     gguf_path: &str,
     mmproj_path: Option<&str>,
+    device: Option<&Device>,
 ) -> Result<ModelInstance<'a>> {
     let model = match model_type {
         WhichModel::Qwen3_5Gguf => {
-            let model = Qwen3_5GenerateModel::init_from_gguf(gguf_path, mmproj_path, None)?;
+            let model = Qwen3_5GenerateModel::init_from_gguf(gguf_path, mmproj_path, device)?;
             ModelInstance::Qwen3_5(model)
         }
         _ => {
@@ -149,135 +181,101 @@ pub fn load_gguf_model<'a>(
     Ok(model)
 }
 
-pub fn load_model<'a>(model_type: WhichModel, path: &str) -> Result<ModelInstance<'a>> {
+pub fn load_model<'a>(
+    model_type: WhichModel,
+    path: &str,
+    device: Option<&Device>,
+    dtype: Option<DType>,
+) -> Result<ModelInstance<'a>> {
     let model = match model_type {
+        WhichModel::AllMiniLML6V2 => {
+            let model = AllMiniLML6V2Embedding::init(path, device, dtype)?;
+            ModelInstance::AllMiniLML6V2(model)
+        }
         WhichModel::MiniCPM4_0_5B => {
-            let model = MiniCPMGenerateModel::init(path, None, None)?;
+            let model = MiniCPMGenerateModel::init(path, device, dtype)?;
             ModelInstance::MiniCPM4(model)
         }
-        WhichModel::LFM2_1_2B => {
-            let model = Lfm2GenerateModel::init(path, None, None)?;
+        WhichModel::LFM2_1_2B | WhichModel::LFM2_5_1_2BInstruct => {
+            let model = Lfm2GenerateModel::init(path, device, dtype)?;
             ModelInstance::Lfm2(model)
         }
-        WhichModel::LFM2_5_1_2BInstruct => {
-            let model = Lfm2GenerateModel::init(path, None, None)?;
-            ModelInstance::Lfm2(model)
-        }
-        WhichModel::LFM2_5VL1_6B => {
-            let model = Lfm2VLGenerateModel::init(path, None, None)?;
+        WhichModel::LFM2_5VL1_6B | WhichModel::LFM2VL1_6B => {
+            let model = Lfm2VLGenerateModel::init(path, device, dtype)?;
             ModelInstance::Lfm2VL(model)
         }
-        WhichModel::LFM2VL1_6B => {
-            let model = Lfm2VLGenerateModel::init(path, None, None)?;
-            ModelInstance::Lfm2VL(model)
-        }
-        WhichModel::Qwen2_5VL3B => {
-            let model = Qwen2_5VLGenerateModel::init(path, None, None)?;
+        WhichModel::Qwen2_5VL3B | WhichModel::Qwen2_5VL7B => {
+            let model = Qwen2_5VLGenerateModel::init(path, device, dtype)?;
             ModelInstance::Qwen2_5VL(model)
         }
-        WhichModel::Qwen2_5VL7B => {
-            let model = Qwen2_5VLGenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen2_5VL(model)
-        }
-        WhichModel::Qwen3_0_6B => {
-            let model = Qwen3GenerateModel::init(path, None, None)?;
+        WhichModel::Qwen3_0_6B | WhichModel::Qwen3_1_7B | WhichModel::Qwen3_4B => {
+            let model = Qwen3GenerateModel::init(path, device, dtype)?;
             ModelInstance::Qwen3(model)
         }
-        WhichModel::Qwen3_1_7B => {
-            let model = Qwen3GenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3(model)
-        }
-        WhichModel::Qwen3_4B => {
-            let model = Qwen3GenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3(model)
-        }
-        WhichModel::Qwen3_5_0_8B => {
-            let model = Qwen3_5GenerateModel::init(path, None, None)?;
+        WhichModel::Qwen3_5_0_8B
+        | WhichModel::Qwen3_5_2B
+        | WhichModel::Qwen3_5_4B
+        | WhichModel::Qwen3_5_9B => {
+            let model = Qwen3_5GenerateModel::init(path, device, dtype)?;
             ModelInstance::Qwen3_5(model)
         }
-        WhichModel::Qwen3_5_2B => {
-            let model = Qwen3_5GenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3_5(model)
-        }
-        WhichModel::Qwen3_5_4B => {
-            let model = Qwen3_5GenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3_5(model)
-        }
-        WhichModel::Qwen3_5_9B => {
-            let model = Qwen3_5GenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3_5(model)
-        }
-        WhichModel::Qwen3ASR0_6B => {
-            let model = Qwen3AsrGenerateModel::init(path, None, None)?;
+        WhichModel::Qwen3ASR0_6B | WhichModel::Qwen3ASR1_7B => {
+            let model = Qwen3AsrGenerateModel::init(path, device, dtype)?;
             ModelInstance::Qwen3ASR(model)
         }
-        WhichModel::Qwen3ASR1_7B => {
-            let model = Qwen3AsrGenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3ASR(model)
+        WhichModel::Qwen3Embedding0_6B
+        | WhichModel::Qwen3Embedding4B
+        | WhichModel::Qwen3Embedding8B => {
+            let model = Qwen3Embedding::init(path, device, dtype)?;
+            ModelInstance::Qwen3Embedding(model)
         }
-        WhichModel::Qwen3VL2B => {
-            let model = Qwen3VLGenerateModel::init(path, None, None)?;
+        WhichModel::Qwen3VL2B
+        | WhichModel::Qwen3VL4B
+        | WhichModel::Qwen3VL8B
+        | WhichModel::Qwen3VL32B => {
+            let model = Qwen3VLGenerateModel::init(path, device, dtype)?;
             ModelInstance::Qwen3VL(Box::new(model))
         }
-        WhichModel::Qwen3VL4B => {
-            let model = Qwen3VLGenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3VL(Box::new(model))
-        }
-        WhichModel::Qwen3VL8B => {
-            let model = Qwen3VLGenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3VL(Box::new(model))
-        }
-        WhichModel::Qwen3VL32B => {
-            let model = Qwen3VLGenerateModel::init(path, None, None)?;
-            ModelInstance::Qwen3VL(Box::new(model))
-        }
-        WhichModel::DeepSeekOCR => {
-            let model = DeepseekOCRGenerateModel::init(path, None, None)?;
-            ModelInstance::DeepSeekOCR(model)
-        }
-        WhichModel::DeepSeekOCR2 => {
-            let model = DeepseekOCRGenerateModel::init(path, None, None)?;
+        WhichModel::DeepSeekOCR | WhichModel::DeepSeekOCR2 => {
+            let model = DeepseekOCRGenerateModel::init(path, device, dtype)?;
             ModelInstance::DeepSeekOCR(model)
         }
         WhichModel::HunyuanOCR => {
-            let model = HunyuanOCRGenerateModel::init(path, None, None)?;
+            let model = HunyuanOCRGenerateModel::init(path, device, dtype)?;
             ModelInstance::HunyuanOCR(model)
         }
-        WhichModel::PaddleOCRVL => {
-            let model = PaddleOCRVLGenerateModel::init(path, None, None)?;
-            ModelInstance::PaddleOCRVL(Box::new(model))
-        }
-        WhichModel::PaddleOCRVL1_5 => {
-            let model = PaddleOCRVLGenerateModel::init(path, None, None)?;
+        WhichModel::PaddleOCRVL | WhichModel::PaddleOCRVL1_5 => {
+            let model = PaddleOCRVLGenerateModel::init(path, device, dtype)?;
             ModelInstance::PaddleOCRVL(Box::new(model))
         }
         WhichModel::RMBG2_0 => {
-            let model = RMBG2_0Model::init(path, None, None)?;
+            let model = RMBG2_0Model::init(path, device, dtype)?;
             ModelInstance::RMBG2_0(Box::new(model))
         }
-        WhichModel::VoxCPM => {
-            let model = VoxCPMGenerate::init(path, None, None)?;
-            ModelInstance::VoxCPM(Box::new(model))
-        }
-        WhichModel::VoxCPM1_5 => {
-            let model = VoxCPMGenerate::init(path, None, None)?;
+        WhichModel::VoxCPM | WhichModel::VoxCPM1_5 => {
+            let model = VoxCPMGenerate::init(path, device, dtype)?;
             ModelInstance::VoxCPM(Box::new(model))
         }
         WhichModel::GlmASRNano2512 => {
-            let model = GlmAsrNanoGenerateModel::init(path, None, None)?;
+            let model = GlmAsrNanoGenerateModel::init(path, device, dtype)?;
             ModelInstance::GlmASRNano(model)
         }
         WhichModel::FunASRNano2512 => {
-            let model = FunAsrNanoGenerateModel::init(path, None, None)?;
+            let model = FunAsrNanoGenerateModel::init(path, device, dtype)?;
             ModelInstance::FunASRNano(model)
         }
         WhichModel::GlmOCR => {
-            let model = GlmOcrGenerateModel::init(path, None, None)?;
+            let model = GlmOcrGenerateModel::init(path, device, dtype)?;
             ModelInstance::GlmOCR(model)
         }
         _ => {
             let model_id = model_type.as_string();
-            return Err(anyhow!("model id {model_id} is not safetensor model"));
+            if model_id.to_lowercase().contains("gguf") || model_id.to_lowercase().contains("onnx")
+            {
+                return Err(anyhow!("model id {model_id} is not safetensor model"));
+            } else {
+                return Err(anyhow!("model id {model_id} not impl load_model function"));
+            }
         }
     };
     Ok(model)

src/models/qwen3/model.rs

@@ -102,7 +102,11 @@ pub struct Qwen3Model {
 
 impl Qwen3Model {
     pub fn new(config: &Qwen3Config, vb: VarBuilder, eos_ids: Vec<u32>) -> Result<Self> {
-        let vb = vb.pp("model");
+        let vb = if vb.contains_tensor("model.embed_tokens.weight") {
+            vb.pp("model")
+        } else {
+            vb
+        };
         let vocab_size = config.vocab_size;
         let embed_tokens = embedding(vocab_size, config.hidden_size, vb.pp("embed_tokens"))?;
         let mut layers = vec![];
@@ -133,6 +137,17 @@ impl Qwen3Model {
         input_ids: Option<&Tensor>,
         inputs_embeds: Option<&Tensor>,
         seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        let hidden_state = self.forward_hidden(input_ids, inputs_embeds, seqlen_offset)?;
+        let logits = self.lm_head.forward(&hidden_state)?;
+        Ok(logits)
+    }
+
+    pub fn forward_hidden(
+        &mut self,
+        input_ids: Option<&Tensor>,
+        inputs_embeds: Option<&Tensor>,
+        seqlen_offset: usize,
     ) -> Result<Tensor> {
         if input_ids.is_none() && inputs_embeds.is_none() {
             return Err(anyhow::anyhow!(
@@ -170,9 +185,9 @@ impl Qwen3Model {
         }
         hidden_states = self.norm.forward(&hidden_states)?;
         let hidden_state = hidden_states.narrow(1, seq_len - 1, 1)?;
-        let logits = self.lm_head.forward(&hidden_state)?;
-        Ok(logits)
+        Ok(hidden_state)
     }
+
     pub fn embedding_token_id(&self, input_ids: &Tensor) -> Result<Tensor> {
         Ok(self.embed_tokens.forward(input_ids)?)
     }

src/models/qwen3_5/model.rs

@@ -12,14 +12,16 @@ use crate::{
         common::{
             InferenceModel,
             gguf::{GateUpDownMLPGguf, Gguf, ProjKind, QuantizedLinear},
-            modules::{conv1d_depthwise, eager_attention_forward, get_conv1d, softplus},
+            modules::{
+                conv1d_depthwise, eager_attention_forward, get_conv1d, l2_normalize, softplus,
+            },
         },
         qwen3_5::config::{Qwen3_5Config, Qwen3_5TextConfig},
         qwen3vl::model::Qwen3VLVisionModel,
     },
     position_embed::rope::{Qwen3VLTextRotaryEmbedding, glm_asr_apply_rotary_pos_emb},
     utils::tensor_utils::{
-        get_equal_mask, get_vision_next_indices, l2_normalize, masked_scatter_dim0, nonzero_index,
+        get_equal_mask, get_vision_next_indices, masked_scatter_dim0, nonzero_index,
         prepare_causal_attention_mask, repeat_interleave, split_tensor, zero_index,
     },
 };

src/models/qwen3_asr/processor.rs

@@ -1,4 +1,6 @@
-use crate::params::chat::ChatCompletionParameters;
+use crate::{
+    models::common::modules::float_range_normalize, params::chat::ChatCompletionParameters,
+};
 use anyhow::Result;
 use candle_core::{Device, Tensor};
 
@@ -10,7 +12,6 @@ use crate::{
     utils::{
         audio_utils::{extract_audios, split_audio_into_chunks},
         capitalize_first_letter,
-        tensor_utils::float_range_normalize,
     },
 };
 

src/models/qwen3_embedding/mod.rs

@@ -0,0 +1,66 @@
+use crate::{
+    models::{
+        common::embedding::{NormalizeType, TextEmbedding},
+        qwen3::{config::Qwen3Config, model::Qwen3Model},
+    },
+    tokenizer::TokenizerModel,
+    utils::{find_type_files, get_device, get_dtype},
+};
+use anyhow::{Result, anyhow};
+use candle_core::{DType, Device};
+use candle_nn::VarBuilder;
+
+pub struct Qwen3Embedding {
+    tokenizer: TokenizerModel,
+    model: Qwen3Model,
+    device: Device,
+    normalize: NormalizeType,
+}
+
+impl Qwen3Embedding {
+    pub fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self> {
+        let tokenizer = TokenizerModel::init(path)?;
+        let config_path = path.to_string() + "/config.json";
+        let cfg: Qwen3Config = serde_json::from_slice(&std::fs::read(config_path)?)?;
+        let device = get_device(device);
+        let dtype = get_dtype(dtype, cfg.torch_dtype.as_str());
+        let model_list = find_type_files(path, "safetensors")?;
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&model_list, dtype, &device)? };
+        let model = Qwen3Model::new(&cfg, vb, vec![])?;
+        Ok(Self {
+            tokenizer,
+            model,
+            device,
+            normalize: NormalizeType::L2,
+        })
+    }
+
+    fn embed_one(&mut self, text: &str) -> Result<Vec<f32>> {
+        let input_ids = self.tokenizer.text_encode(text.to_string(), &self.device)?;
+        let hidden = self
+            .model
+            .forward_hidden(Some(&input_ids), None, 0)?
+            .squeeze(0)?
+            .to_dtype(DType::F32)?;
+        let norm = self
+            .normalize
+            .normalize(&hidden, hidden.rank() - 1)?
+            .squeeze(0)?;
+        let norm = norm.to_vec1::<f32>()?;
+        Ok(norm)
+    }
+}
+
+impl TextEmbedding for Qwen3Embedding {
+    fn embed_texts(&mut self, input: &[String]) -> Result<Vec<Vec<f32>>> {
+        if input.is_empty() {
+            return Err(anyhow!("embedding input cannot be empty"));
+        }
+        let mut out = Vec::with_capacity(input.len());
+        for text in input {
+            out.push(self.embed_one(text)?);
+            self.model.clear_kv_cache();
+        }
+        Ok(out)
+    }
+}

src/params/asr.rs

@@ -46,33 +46,3 @@ pub(crate) struct ErrorDetail {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub(crate) code: Option<String>,
 }
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn test_transcription_response_serialization() {
-        let response = TranscriptionResponse {
-            text: "Hello, world!".to_string(),
-        };
-        let json = serde_json::to_string(&response).unwrap();
-        assert_eq!(json, r#"{"text":"Hello, world!"}"#);
-    }
-
-    #[test]
-    fn test_error_response_serialization() {
-        let error = ErrorResponse {
-            error: ErrorDetail {
-                message: "Invalid audio file".to_string(),
-                error_type: "invalid_request_error".to_string(),
-                code: Some("invalid_audio".to_string()),
-            },
-        };
-        let json = serde_json::to_string(&error).unwrap();
-        let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
-        assert_eq!(parsed["error"]["message"], "Invalid audio file");
-        assert_eq!(parsed["error"]["type"], "invalid_request_error");
-        assert_eq!(parsed["error"]["code"], "invalid_audio");
-    }
-}

src/params/embedding.rs

@@ -0,0 +1,22 @@
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+
+#[derive(Debug, Deserialize)]
+pub(crate) struct EmbeddingRequest {
+    pub model: Option<String>,
+    pub input: Value,
+}
+
+#[derive(Debug, Serialize)]
+pub(crate) struct EmbeddingData {
+    pub object: String,
+    pub index: usize,
+    pub embedding: Vec<f32>,
+}
+
+#[derive(Debug, Serialize)]
+pub(crate) struct EmbeddingResponse {
+    pub object: String,
+    pub model: String,
+    pub data: Vec<EmbeddingData>,
+}

src/params/mod.rs

@@ -1,2 +1,8 @@
+#[allow(unused)]
+pub mod asr;
 pub mod chat;
+#[allow(unused)]
+pub mod embedding;
+#[allow(unused)]
+pub mod rerank;
 pub mod shared;

src/params/rerank.rs

@@ -0,0 +1,23 @@
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Deserialize)]
+pub(crate) struct RerankRequest {
+    pub model: Option<String>,
+    pub query: String,
+    pub documents: Vec<String>,
+    pub top_n: Option<usize>,
+}
+
+#[derive(Debug, Serialize)]
+struct RerankResult {
+    index: usize,
+    relevance_score: f32,
+    document: String,
+}
+
+#[derive(Debug, Serialize)]
+struct RerankResponse {
+    object: String,
+    model: String,
+    results: Vec<RerankResult>,
+}

src/params/shared.rs

@@ -179,7 +179,7 @@ pub enum FinishReason {
     EndTurn,
     /// The finish reason is unspecified. [Gemini]
     #[serde(rename = "FINISH_REASON_UNSPECIFIED	")]
-    FinishReasonUnspecified,
+    Unspecified,
     #[serde(rename = "MALFORMED_FUNCTION_CALL")]
     MalformedFunctionCall,
     #[serde(rename = "OTHER")]

src/server/api.rs

@@ -43,7 +43,7 @@ pub fn init(
         if let Some(gguf_path) = gguf {
             let gguf_path = string_to_static_str(gguf_path);
             let mmproj_path = mmproj.map(string_to_static_str);
-            load_gguf_model(model_type, None, gguf_path, mmproj_path)?
+            load_gguf_model(model_type, None, gguf_path, mmproj_path, None)?
         } else {
             return Err(anyhow!("gguf model need gguf model path"));
         }
@@ -51,7 +51,7 @@ pub fn init(
         return Err(anyhow!("onnx comming soon but now not support"));
     } else {
         let model_path = string_to_static_str(path);
-        load_model(model_type, model_path)?
+        load_model(model_type, model_path, None, None)?
     };
 
     MODEL.get_or_init(|| {

src/server/asr.rs

@@ -11,10 +11,8 @@ use rocket::http::Status;
 use rocket::serde::json::Json;
 use rocket::{form::Form, post};
 
+use crate::params::asr::{ErrorDetail, ErrorResponse, TranscriptionRequest, TranscriptionResponse};
 use crate::server::api::MODEL;
-use crate::server::asr_types::{
-    ErrorDetail, ErrorResponse, TranscriptionRequest, TranscriptionResponse,
-};
 
 /// Handle audio transcription requests
 ///

src/server/embedding.rs

@@ -0,0 +1,77 @@
+use rocket::{http::Status, post, serde::json::Json};
+use serde_json::Value;
+
+use crate::{
+    params::embedding::{EmbeddingData, EmbeddingRequest, EmbeddingResponse},
+    server::api::MODEL,
+};
+
+fn parse_embedding_input(input: &Value) -> anyhow::Result<Vec<String>> {
+    match input {
+        Value::String(s) => Ok(vec![s.clone()]),
+        Value::Array(arr) => {
+            let mut out = Vec::with_capacity(arr.len());
+            for v in arr {
+                let s = v.as_str().ok_or_else(|| {
+                    anyhow::anyhow!("embedding input array must contain only strings")
+                })?;
+                out.push(s.to_string());
+            }
+            if out.is_empty() {
+                return Err(anyhow::anyhow!("embedding input cannot be empty"));
+            }
+            Ok(out)
+        }
+        _ => Err(anyhow::anyhow!(
+            "embedding input must be a string or an array of strings"
+        )),
+    }
+}
+
+#[post("/embeddings", data = "<req>")]
+pub(crate) async fn embeddings(req: Json<EmbeddingRequest>) -> (Status, Json<Value>) {
+    let texts = match parse_embedding_input(&req.input) {
+        Ok(v) => v,
+        Err(e) => {
+            return (
+                Status::BadRequest,
+                Json(serde_json::json!({ "error": e.to_string() })),
+            );
+        }
+    };
+    let model_ref = match MODEL.get().cloned() {
+        Some(v) => v,
+        None => {
+            return (
+                Status::ServiceUnavailable,
+                Json(serde_json::json!({ "error": "model not init" })),
+            );
+        }
+    };
+    let mut guard = model_ref.write().await;
+    let embeddings = match guard.instance.embedding(&texts) {
+        Ok(v) => v,
+        Err(e) => {
+            return (
+                Status::BadRequest,
+                Json(serde_json::json!({ "error": e.to_string() })),
+            );
+        }
+    };
+    let model_name = guard.which_model.as_string();
+    let data = embeddings
+        .into_iter()
+        .enumerate()
+        .map(|(index, embedding)| EmbeddingData {
+            object: "embedding".to_string(),
+            index,
+            embedding,
+        })
+        .collect::<Vec<_>>();
+    let response = EmbeddingResponse {
+        object: "list".to_string(),
+        data,
+        model: model_name,
+    };
+    (Status::Ok, Json(serde_json::to_value(response).unwrap()))
+}

src/server/mod.rs

@@ -10,7 +10,7 @@ use std::sync::atomic::{AtomicBool, Ordering};
 // ASR (Automatic Speech Recognition) API module
 pub(crate) mod api;
 pub(crate) mod asr;
-pub(crate) mod asr_types;
+pub(crate) mod embedding;
 pub(crate) mod process;
 
 pub(crate) async fn start_http_server(
@@ -61,6 +61,9 @@ pub(crate) async fn start_http_server(
     builder = builder.mount("/audio", routes![api::speech, asr::transcriptions]);
     // /v1/audio/transcriptions (OpenAI standard ASR transcription endpoint)
     builder = builder.mount("/v1/audio", routes![asr::transcriptions]);
+    // /embeddings and /v1/embeddings (OpenAI-compatible embeddings endpoint)
+    builder = builder.mount("/", routes![embedding::embeddings]);
+    builder = builder.mount("/v1", routes![embedding::embeddings]);
     // Health check and model info endpoints
     builder = builder.mount("/", routes![api::health, api::models]);
     // Shutdown endpoint

src/utils/audio_utils.rs

@@ -4,6 +4,7 @@ use std::path::{Path, PathBuf};
 use std::thread;
 use std::{f64::consts::PI, io::Cursor};
 
+use crate::models::common::modules::log10;
 use crate::params::chat::{
     ChatCompletionParameters, ChatCompletionResponse, ChatMessage, ChatMessageContent,
     ChatMessageContentPart,
@@ -33,7 +34,7 @@ use symphonia::core::probe::Hint;
 
 use crate::utils::get_default_save_dir;
 use crate::utils::tensor_utils::{
-    linspace, log10, pad_reflect_last_dim, pad_replicate_last_dim, split_tensor,
+    linspace, pad_reflect_last_dim, pad_replicate_last_dim, split_tensor,
 };
 
 // 重采样方法枚举

src/utils/tensor_utils.rs

@@ -2,7 +2,6 @@ use std::f32;
 
 use anyhow::{Result, anyhow};
 use candle_core::{D, DType, Device, IndexOp, Tensor, shape::Dim};
-use candle_nn::ops::sigmoid;
 
 pub enum PaddingSide {
     Left,
@@ -463,12 +462,6 @@ pub fn index_select_2d(t: &Tensor, index: &Tensor) -> Result<Tensor> {
     Ok(res)
 }
 
-pub fn quick_gelu(xs: &Tensor) -> Result<Tensor> {
-    let x = xs.affine(1.702, 0.0)?;
-    let x = sigmoid(&x)?;
-    Ok(xs.mul(&x)?)
-}
-
 pub fn topk(weight: &Tensor, topk: usize) -> Result<(Tensor, Tensor)> {
     let topk_idx = weight
         .arg_sort_last_dim(false)?
@@ -558,102 +551,6 @@ pub fn pad_replicate_last_dim(t: &Tensor, pad: (usize, usize)) -> Result<Tensor>
     Ok(pad_tensor)
 }
 
-pub fn log10(t: &Tensor) -> Result<Tensor> {
-    Ok(t.log()?.affine(1.0 / 10.0_f64.ln(), 0.0)?)
-}
-
-pub fn z_score_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
-    let rank = t.rank();
-    if dim >= rank {
-        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
-    }
-    Ok(t.broadcast_sub(&t.mean_keepdim(dim)?)?
-        .broadcast_div(&t.var_keepdim(dim)?.sqrt()?)?)
-}
-
-pub fn l2_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
-    let rank = t.rank();
-    if dim >= rank {
-        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
-    }
-    let l2_norm = t.sqr()?.sum_keepdim(dim)?.affine(1.0, 1e-6)?.sqrt()?;
-    Ok(t.broadcast_div(&l2_norm)?)
-}
-
-pub fn l1_normalize(t: &Tensor, dim: usize) -> Result<Tensor> {
-    let rank = t.rank();
-    if dim >= rank {
-        return Err(anyhow!(format!("input dim {} must < rank {}", dim, rank)));
-    }
-    let l1_norm = t.abs()?.sum_keepdim(dim)?;
-    Ok(t.broadcast_div(&l1_norm)?)
-}
-
-pub fn pool1d(xs: &Tensor, pool_size: usize, ceil_mode: bool, stype: &str) -> Result<Tensor> {
-    // xs: (bs, c, dim)
-    // ceil_mode: 是否保留不完整窗口，为true时通过pad实现
-    if pool_size == 0 {
-        return Err(anyhow!("pool_size must be greater than 0"));
-    }
-    let (bs, c, dim) = xs.dims3()?;
-    let xs_reshape = if ceil_mode {
-        let remain = dim % pool_size;
-        if remain > 0 {
-            let pad = pool_size - remain;
-            let xs_pad = pad_replicate_last_dim(xs, (0, pad))?;
-            xs_pad.reshape((bs, c, (), pool_size))?
-        } else {
-            xs.reshape((bs, c, (), pool_size))?
-        }
-    } else {
-        let remain = dim % pool_size;
-        if remain > 0 {
-            let xs_del = xs.narrow(D::Minus1, 0, dim - remain)?;
-            xs_del.reshape((bs, c, (), pool_size))?
-        } else {
-            xs.reshape((bs, c, (), pool_size))?
-        }
-    };
-    let xs_pool = match stype {
-        "avg" => xs_reshape.mean(D::Minus1)?,
-        "max" => xs_reshape.max(D::Minus1)?,
-        "min" => xs_reshape.min(D::Minus1)?,
-        _ => {
-            return Err(anyhow!(
-                "unsupported pool type: {}, supported types are: avg, max, min",
-                stype
-            ));
-        }
-    };
-    Ok(xs_pool)
-}
-
-pub fn statistics_pooling(xs: &Tensor, dim: D, keepdim: bool) -> Result<Tensor> {
-    let mean = xs.mean(dim)?;
-    let std = xs.var(dim)?.sqrt()?;
-    let mut stats = Tensor::cat(&[mean, std], D::Minus1)?;
-    if keepdim {
-        stats = stats.unsqueeze(dim)?;
-    }
-    Ok(stats)
-}
-pub fn float_range_normalize(t: &Tensor) -> Result<Tensor> {
-    let peak = t
-        .to_dtype(DType::F32)?
-        .abs()?
-        .max_all()?
-        .to_scalar::<f32>()?;
-    if peak == 0.0 {
-        return Ok(t.clone());
-    }
-    let mut t = t.clone();
-    if peak > 1.0 {
-        t = t.affine(1.0 / peak as f64, 0.0)?;
-    }
-    t = t.clamp(-1.0, 1.0)?;
-    Ok(t)
-}
-
 pub fn sequence_mask(length: &Tensor, max_length: Option<u32>) -> Result<Tensor> {
     let max_length = max_length.unwrap_or(length.max_all()?.to_scalar::<u32>()?);
     let x = Tensor::arange(0, max_length, length.device())?.unsqueeze(0)?;
@@ -662,17 +559,6 @@ pub fn sequence_mask(length: &Tensor, max_length: Option<u32>) -> Result<Tensor>
     Ok(mask)
 }
 
-pub fn cosine_similarity(query_vector: &Tensor, matrix: &Tensor) -> Result<Tensor> {
-    // query_vector: (n, dim)
-    // matrix: (m, dim)
-    let query_norm = l2_normalize(query_vector, query_vector.rank() - 1)?;
-    let matrix_norm = l2_normalize(matrix, matrix.rank() - 1)?;
-    let similarity = query_norm
-        .matmul(&matrix_norm.transpose(D::Minus1, D::Minus2)?)?
-        .squeeze(D::Minus1)?;
-    Ok(similarity)
-}
-
 pub fn repeat_interleave(t: &Tensor, repeats: usize, dim: usize) -> Result<Tensor> {
     if repeats == 1 {
         return Ok(t.clone());

tests/test_all_minilm_l6_v2.rs

@@ -0,0 +1,24 @@
+use aha::models::{
+    all_minilm_l6_v2::AllMiniLML6V2Embedding, common::embedding::TextEmbedding,
+};
+use anyhow::Result;
+use std::time::Instant;
+
+#[test]
+fn all_minilm_l6_v2_embedding() -> Result<()> {
+    // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda --test test_all_minilm_l6_v2 all_minilm_l6_v2_embedding -r -- --nocapture
+
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/sentence-transformers/all-MiniLM-L6-v2/", save_dir);
+
+    let i_start = Instant::now();
+    let mut model = AllMiniLML6V2Embedding::init(&model_path, None, None)?;
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in load model is: {:?}", i_duration);
+    let input_texts = ["test ALL_MINILM_L6_V2 embedding".to_string()];
+    let result = model.embed_texts(&input_texts)?;
+    println!("result: {:?}", result);
+
+    Ok(())
+}

tests/test_qwen3_embedding.rs

@@ -0,0 +1,22 @@
+use aha::models::{common::embedding::TextEmbedding, qwen3_embedding::Qwen3Embedding};
+use anyhow::Result;
+use std::time::Instant;
+
+#[test]
+fn qwen3_embedding() -> Result<()> {
+    // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda --test test_qwen3_embedding qwen3_embedding -r -- --nocapture
+
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Qwen/Qwen3-Embedding-0.6B/", save_dir);
+
+    let i_start = Instant::now();
+    let mut model = Qwen3Embedding::init(&model_path, None, None)?;
+    let i_duration = i_start.elapsed();
+    println!("Time elapsed in load model is: {:?}", i_duration);
+    let input_texts = ["test Qwen3-Embedding-0.6B".to_string()];
+    let result = model.embed_texts(&input_texts)?;
+    println!("result: {:?}", result);
+
+    Ok(())
+}

tests/weight_test.rs

@@ -2,7 +2,7 @@ use std::collections::HashMap;
 
 use aha::utils::{find_type_files, get_device};
 use anyhow::Result;
-use candle_core::{Device, pickle::read_all_with_key, safetensors};
+use candle_core::{Device, pickle::read_all_with_key, quantized::gguf_file, safetensors};
 use candle_nn::VarBuilder;
 
 #[test]
@@ -310,3 +310,21 @@ fn lfm2vl_weight() -> Result<()> {
     println!("model_list: {:?}", model_list);
     Ok(())
 }
+
+#[test]
+fn gguf_weight() -> Result<()> {
+    // cargo test -F cuda --test weight_test gguf_weight -r -- --nocapture
+    let gguf_path = "/home/jhq/.aha/Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-f16.gguf";
+    let mut model_file = std::fs::File::open(gguf_path)?;
+    let model = gguf_file::Content::read(&mut model_file)?;
+    for (key, value) in model.tensor_infos {
+        println!("{key}: {:#?}", value);
+    }
+    // for (key, value) in model.metadata {
+    //     if key.contains("tokeni") {
+    //         continue;
+    //     }
+    //     println!("{key}: {:#?}", value);
+    // }
+    Ok(())
+}