update qwen2.5vl

Cargo.toml

@@ -21,7 +21,12 @@ base64 = "0.22.1"
 num = "0.4.3"
 minijinja = "2.12.0"
 tokenizers = "0.22.1"
-openai_dive = "1.3.0"
+openai_dive = { version = "1.3.0", features = ["stream"]}
+uuid = { version = "1.18.1", features = ["v4"]}
+chrono = "0.4.42"
+rocket = "0.5.1"
+tokio = "1.47.1"
+
 [features]
 flash-attn=["candle-flash-attn"]
 cuda=["candle-nn/cuda", "candle-core/cuda", "candle-transformers/cuda"]

src/lib.rs

@@ -17,7 +17,7 @@ impl ModelType {
         model_path: &str,
         device: Option<&Device>,
         dtype: Option<DType>,
-    ) -> Result<Box<dyn GenerateModel>> {
+    ) -> Result<Box<impl GenerateModel>> {
         match model_type {
             ModelType::Qwen2_5VL => {
                 let model = Qwen2_5VLGenerateModel::init(model_path, device, dtype)?;

src/models/mod.rs

@@ -1,11 +1,14 @@
 pub mod qwen2_5vl;
 use anyhow::Result;
 use candle_core::{DType, Device};
-use openai_dive::v1::resources::chat::ChatCompletionParameters;
+use openai_dive::v1::resources::chat::{ChatCompletionChunkResponse, ChatCompletionParameters, ChatCompletionResponse};
+use rocket::futures::Stream;
 
 pub trait GenerateModel {
     fn init(path: &str, device: Option<&Device>, dtype: Option<DType>) -> Result<Self>
     where
         Self: Sized;
-    fn generate(&mut self, mes: ChatCompletionParameters) -> Result<String>;
+    fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse>;
+    fn generate_stream(&mut self, mes: ChatCompletionParameters) -> Result<impl Stream<Item = Result<ChatCompletionChunkResponse, anyhow::Error>>>;
 }
+

src/models/qwen2_5vl/generate.rs

@@ -1,5 +1,9 @@
+// use crate::models::GenerateStream;
 use crate::models::qwen2_5vl::config::Config;
-use crate::utils::utils::{find_safetensors_files, get_device, get_dtype};
+use crate::utils::utils::{
+    build_completion_chunk_response, build_completion_response, find_safetensors_files, get_device,
+    get_dtype, get_logit_processor,
+};
 use crate::{
     chat_template::chat_template::ChatTemplate,
     models::{
@@ -12,7 +16,11 @@ use anyhow::{Result, anyhow};
 use candle_core::{D, DType, Device, IndexOp, Tensor};
 use candle_nn::VarBuilder;
 use candle_transformers::generation::LogitsProcessor;
-use openai_dive::v1::resources::chat::ChatCompletionParameters;
+use openai_dive::v1::resources::chat::{
+    ChatCompletionChunkResponse, ChatCompletionParameters, ChatCompletionResponse,
+};
+use rocket::async_stream::stream;
+use rocket::futures::Stream;
 
 pub struct Qwen2_5VLGenerateModel<'a> {
     chat_template: ChatTemplate<'a>,
@@ -20,7 +28,8 @@ pub struct Qwen2_5VLGenerateModel<'a> {
     pre_processor: Qwen2_5VLProcessor,
     qwen2_5_vl: Qwen2_5VLModel,
     device: Device,
-    dtype: DType,
+    endoftext_id: u32,
+    im_end_id: u32,
 }
 
 impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
@@ -33,29 +42,25 @@ impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
         let cfg_dtype = cfg.torch_dtype.as_str();
         let dtype = get_dtype(dtype, cfg_dtype);
         let pre_processor = Qwen2_5VLProcessor::new(device, dtype)?;
-        
+        let endoftext_id = cfg.bos_token_id as u32;
+        let im_end_id = cfg.eos_token_id as u32;
         let model_list = find_safetensors_files(&path)?;
         let vb = unsafe { VarBuilder::from_mmaped_safetensors(&model_list, dtype, device)? };
         let qwen2_5_vl = Qwen2_5VLModel::new(cfg, vb)?;
+
         Ok(Qwen2_5VLGenerateModel {
             chat_template,
             tokenizer,
             pre_processor,
             qwen2_5_vl,
             device: device.clone(),
-            dtype: dtype,
+            endoftext_id,
+            im_end_id,
         })
     }
-    fn generate(&mut self, mes: ChatCompletionParameters) -> Result<String> {
+
-        let temperature = match mes.temperature {
+    fn generate(&mut self, mes: ChatCompletionParameters) -> Result<ChatCompletionResponse> {
-            Some(temp) => Some(temp as f64),
+        let mut logit_processor = get_logit_processor(mes.temperature, mes.top_p);
-            None => None,
-        };
-        let top_p = match mes.top_p {
-            Some(tp) => Some(tp as f64),
-            None => None,
-        };
-        let mut logit_processor = LogitsProcessor::new(34562, temperature, top_p);
         let mes_render = self.chat_template.apply_chat_template(&mes)?;
         let input = self.pre_processor.process_info(&mes, &mes_render)?;
         let mut input_ids = self
@@ -63,33 +68,11 @@ impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
             .text_encode(input.replace_text.clone(), &self.device)?;
         let mut seq_len = input_ids.dim(1)?;
         let mut seqlen_offset = 0;
-        let end_of_text_id = self.qwen2_5_vl.cfg.bos_token_id as u32;
+        let mut pixel_values = input.pixel_values.as_ref();
-        let im_end_id = self.qwen2_5_vl.cfg.eos_token_id as u32;
+        let image_grid_thw = input.image_grid_thw.as_ref();
-        let mut pixel_values = if input.pixel_values.is_some() {
+        let mut pixel_values_video = input.pixel_values_video.as_ref();
-            Some(&input.pixel_values.unwrap().clone())
+        let video_grid_thw = input.video_grid_thw.as_ref();
-        } else {
+        let second_per_grid_ts = input.second_per_grid_ts.clone();
-            None
-        };
-        let image_grid_thw = if input.image_grid_thw.is_some() {
-            Some(&input.image_grid_thw.unwrap().clone())
-        } else {
-            None
-        };
-        let mut pixel_values_video = if input.pixel_values_video.is_some() {
-            Some(&input.pixel_values_video.unwrap().clone())
-        } else {
-            None
-        };
-        let video_grid_thw = if input.video_grid_thw.is_some() {
-            Some(&input.video_grid_thw.unwrap().clone())
-        } else {
-            None
-        };
-        let second_per_grid_ts = if input.second_per_grid_ts.is_some() {
-            Some(input.second_per_grid_ts.unwrap().clone())
-        } else {
-            None
-        };
 
         let mut mask = Tensor::ones_like(&input_ids)?;
         let mut cache_position = Tensor::ones_like(&input_ids.i(0)?)?
@@ -118,7 +101,7 @@ impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
             let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
             let next_token = logit_processor.sample(&logits)?;
             generate.push(next_token);
-            if next_token == end_of_text_id || next_token == im_end_id {
+            if next_token == self.endoftext_id || next_token == self.im_end_id {
                 break;
             }
             seqlen_offset += seq_len;
@@ -132,6 +115,96 @@ impl<'a> GenerateModel for Qwen2_5VLGenerateModel<'a> {
         }
         let res = self.tokenizer.token_decode(generate)?;
         self.qwen2_5_vl.clear_kv_cache();
-        Ok(res)
+        let response = build_completion_response(res, "qwen2.5vl");
+        Ok(response)
+    }
+    fn generate_stream(
+        &mut self,
+        mes: ChatCompletionParameters,
+    ) -> Result<impl Stream<Item = Result<ChatCompletionChunkResponse, anyhow::Error>>> {
+        let mut logit_processor = get_logit_processor(mes.temperature, mes.top_p);
+        let mes_render = self.chat_template.apply_chat_template(&mes)?;
+        let input = self.pre_processor.process_info(&mes, &mes_render)?;
+        let mut input_ids = self
+            .tokenizer
+            .text_encode(input.replace_text.clone(), &self.device)?;
+        let mut seq_len = input_ids.dim(1)?;
+        let mut seqlen_offset = 0;
+        let pixel_values = input.pixel_values.clone();
+        let image_grid_thw = input.image_grid_thw.clone();
+        let pixel_values_video = input.pixel_values_video.clone();
+        let video_grid_thw = input.video_grid_thw.clone();
+        let second_per_grid_ts = input.second_per_grid_ts.clone();
+
+        let mut mask = Tensor::ones_like(&input_ids)?;
+        let mut cache_position = Tensor::ones_like(&input_ids.i(0)?)?
+            .to_dtype(candle_core::DType::F64)?
+            .cumsum(D::Minus1)?
+            .to_dtype(candle_core::DType::U32)?
+            .broadcast_sub(&Tensor::new(vec![1_u32], input_ids.device())?)?;
+
+        let sample_len = match mes.max_tokens {
+            Some(max) => max,
+            None => 512,
+        };
+        let stream = stream! {
+            let mut error_tokens = Vec::new();
+            let mut pixel_values = pixel_values.as_ref();
+            let image_grid_thw = image_grid_thw.as_ref();
+            let mut pixel_values_video = pixel_values_video.as_ref();
+            let video_grid_thw = video_grid_thw.as_ref();
+            for _ in 0..sample_len {
+                let logits = self.qwen2_5_vl.forward(
+                    &input_ids,
+                    pixel_values,
+                    image_grid_thw,
+                    pixel_values_video,
+                    video_grid_thw,
+                    &mask,
+                    Some(&cache_position),
+                    seqlen_offset,
+                    second_per_grid_ts.clone(),
+                )?;
+                let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
+                let next_token = logit_processor.sample(&logits)?;
+                let mut decode_ids = Vec::new();
+                if error_tokens.len() > 0 {
+                    decode_ids.extend_from_slice(&error_tokens);
+                }
+                decode_ids.push(next_token);
+                let decoded_token = self.tokenizer.token_decode(decode_ids).map_err(|e| anyhow!(format!("stream decode error{}", e)))?;
+                if decoded_token.contains("�") {
+                    error_tokens.push(next_token);
+                    if error_tokens.len() > 3 {
+                        error_tokens.clear();
+                    }
+                    seqlen_offset += seq_len;
+                    seq_len = 1;
+                    input_ids = Tensor::from_vec(vec![next_token], (1, 1), &self.device)?;
+                    let appendd_mask = Tensor::ones((1, 1), mask.dtype(), &self.device)?;
+                    mask = Tensor::cat(&[mask, appendd_mask], 1)?;
+                    cache_position = Tensor::from_vec(vec![seqlen_offset as u32], 1, &self.device)?;
+                    pixel_values = None;
+                    pixel_values_video = None;
+                    continue;
+                }
+                error_tokens.clear();
+                let chunk = build_completion_chunk_response(decoded_token, "qwen2.5vl", None, None);
+                yield Ok(chunk);
+                if next_token == self.endoftext_id || next_token == self.im_end_id {
+                    break;
+                }
+                seqlen_offset += seq_len;
+                seq_len = 1;
+                input_ids = Tensor::from_vec(vec![next_token], (1, 1), &self.device)?;
+                let appendd_mask = Tensor::ones((1, 1), mask.dtype(), &self.device)?;
+                mask = Tensor::cat(&[mask, appendd_mask], 1)?;
+                cache_position = Tensor::from_vec(vec![seqlen_offset as u32], 1, &self.device)?;
+                pixel_values = None;
+                pixel_values_video = None;
+            }
+            self.qwen2_5_vl.clear_kv_cache();
+        };
+        Ok(stream)
     }
 }

src/utils/utils.rs

@@ -1,5 +1,12 @@
 use anyhow::Result;
 use candle_core::{DType, Device};
+use candle_transformers::generation::LogitsProcessor;
+use openai_dive::v1::resources::{
+    chat::{
+        ChatCompletionChoice, ChatCompletionChunkChoice, ChatCompletionChunkResponse, ChatCompletionResponse, ChatMessage, ChatMessageContent, DeltaChatMessage, DeltaFunction, DeltaToolCall, Function, ToolCall
+    },
+    shared::FinishReason,
+};
 
 pub fn get_device(device: Option<&Device>) -> Device {
     match device {
@@ -85,3 +92,168 @@ pub fn ceil_by_factor(num: f32, factor: u32) -> u32 {
     let ceil = (num / factor as f32).ceil() as u32;
     ceil * factor
 }
+
+pub fn build_completion_response(res: String, model_name: &str) -> ChatCompletionResponse {
+    let id = uuid::Uuid::new_v4().to_string();
+    let mut response = ChatCompletionResponse {
+        id: Some(id),
+        choices: vec![],
+        created: chrono::Utc::now().timestamp() as u32,
+        model: model_name.to_string(),
+        service_tier: None,
+        system_fingerprint: None,
+        object: "chat.completion".to_string(),
+        usage: None,
+    };
+    let choice = if res.contains("<tool_call>") {
+        let mes: Vec<&str> = res.split("<tool_call>").collect();
+        let content = mes[0].to_string();
+        let mut tool_vec = Vec::new();
+        for i in 1..mes.len() {
+            let tool_mes = mes[i].replace("</tool_call>", "");
+            let function = match serde_json::from_str::<serde_json::Value>(&tool_mes) {
+                Ok(json_value) => {
+                    let name = json_value
+                        .get("name")
+                        .and_then(|v| v.as_str())
+                        .map(|s| s.to_string())
+                        .unwrap_or_default();
+
+                    let arguments = json_value
+                        .get("arguments")
+                        .map(|v| v.to_string())
+                        .unwrap_or_default();
+
+                    Function { name, arguments }
+                }
+                Err(_) => Function {
+                    name: "".to_string(),
+                    arguments: "".to_string(),
+                },
+            };
+            let tool_call = ToolCall {
+                id: (i - 1).to_string(),
+                r#type: "function".to_string(),
+                function: function,
+            };
+            tool_vec.push(tool_call);
+        }
+        ChatCompletionChoice {
+            index: 0,
+            message: ChatMessage::Assistant {
+                content: Some(ChatMessageContent::Text(content)),
+                reasoning_content: None,
+                refusal: None,
+                name: None,
+                audio: None,
+                tool_calls: Some(tool_vec),
+            },
+            finish_reason: Some(FinishReason::ToolCalls),
+            logprobs: None,
+        }
+    } else {
+        ChatCompletionChoice {
+            index: 0,
+            message: ChatMessage::Assistant {
+                content: Some(ChatMessageContent::Text(res)),
+                reasoning_content: None,
+                refusal: None,
+                name: None,
+                audio: None,
+                tool_calls: None,
+            },
+            finish_reason: Some(FinishReason::StopSequenceReached),
+            logprobs: None,
+        }
+    };
+    response.choices.push(choice);
+    response
+}
+
+pub fn build_completion_chunk_response(
+    res: String,
+    model_name: &str,
+    tool_call_id: Option<String>,
+    tool_call_content: Option<String>,
+) -> ChatCompletionChunkResponse {
+    let id = uuid::Uuid::new_v4().to_string();
+    let mut response = ChatCompletionChunkResponse {
+        id: Some(id),
+        choices: vec![],
+        created: chrono::Utc::now().timestamp() as u32,
+        model: model_name.to_string(),
+        system_fingerprint: None,
+        object: "chat.completion.chunk".to_string(),
+        usage: None,
+    };
+    let choice = if tool_call_id.is_some() {
+        let tool_call_id = tool_call_id.unwrap();
+        let function = if let Some(content) = tool_call_content {
+            match serde_json::from_str::<serde_json::Value>(&content) {
+                Ok(json_value) => {
+                    let name = json_value
+                        .get("name")
+                        .and_then(|v| v.as_str())
+                        .map(|s| s.to_string());
+
+                    let arguments = json_value.get("arguments").map(|v| v.to_string());
+
+                    DeltaFunction { name, arguments }
+                }
+                Err(_) => DeltaFunction {
+                    name: None,
+                    arguments: Some(content),
+                },
+            }
+        } else {
+            DeltaFunction {
+                name: None,
+                arguments: None,
+            }
+        };
+        ChatCompletionChunkChoice {
+            index: Some(0),
+            delta: DeltaChatMessage::Assistant {
+                content: None,
+                reasoning_content: None,
+                refusal: None,
+                name: None,
+                tool_calls: Some(vec![DeltaToolCall {
+                    index: Some(0),
+                    id: Some(tool_call_id),
+                    r#type: Some("function".to_string()),
+                    function,
+                }]),
+            },
+            finish_reason: None,
+            logprobs: None,
+        }
+    } else {
+        ChatCompletionChunkChoice {
+            index: Some(0),
+            delta: DeltaChatMessage::Assistant {
+                content: Some(ChatMessageContent::Text(res)),
+                reasoning_content: None,
+                refusal: None,
+                name: None,
+                tool_calls: None,
+            },
+            finish_reason: None,
+            logprobs: None,
+        }
+    };
+    response.choices.push(choice);
+    response
+}
+
+pub fn get_logit_processor(temperature: Option<f32>, top_p: Option<f32>) -> LogitsProcessor {
+    let temperature = match temperature {
+        Some(temp) => Some(temp as f64),
+        None => None,
+    };
+    let top_p = match top_p {
+        Some(tp) => Some(tp as f64),
+        None => None,
+    };
+    LogitsProcessor::new(34562, temperature, top_p)
+}

tests/test_qwen2_5vl.rs

@@ -1,9 +1,10 @@
-use std::time::Instant;
+use std::{pin::pin, time::Instant};
 
 use aha::{models::{qwen2_5vl::generate::Qwen2_5VLGenerateModel, GenerateModel}, ModelType};
 use anyhow::{Result};
 use candle_core::{DType, Device};
 use openai_dive::v1::resources::chat::ChatCompletionParameters;
+use rocket::futures::StreamExt;
 
 
 #[test]
@@ -48,7 +49,58 @@ fn qwen2_5vl_generate() -> Result<()> {
 
     let i_start = Instant::now();
     let result = model.generate(mes)?;
-    println!("generate: \n{}", result);    
+    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);