[Iiluvatar] fix ci error and update readme (#7453)

docs/get_started/installation/iluvatar_gpu.md

@@ -34,17 +34,24 @@ Note: Because the 4.3.8 SDK in the image is incompatible with KMD, paddle cannot
 ### 3.2 Install paddle
 
 ```bash
-pip3 install paddlepaddle-iluvatar==3.4.0.dev20260326 -i https://www.paddlepaddle.org.cn/packages/nightly/ixuca/
+pip3 install paddlepaddle-iluvatar==3.4.0.dev20260326 -i https://www.paddlepaddle.org.cn/packages/nightly/ixuca/ --extra-index-url https://mirrors.aliyun.com/pypi/simple/
 ```
 
 ### 3.3 Install or build FastDeploy
+
+You can install FastDeploy in either of the following ways:
+
+- pip install
 ```bash
 pip3 install fastdeploy_iluvatar_gpu==2.5.0.dev0 -i https://www.paddlepaddle.org.cn/packages/stable/ixuca/ --extra-index-url https://mirrors.aliyun.com/pypi/simple/
 ```
+
+- Build from source
 You can build FastDeploy from source if you need the ```latest version```.
 ```bash
-git clone https://github.com/PaddlePaddle/FastDeploy
+git clone --recursive https://github.com/PaddlePaddle/FastDeploy.git
 cd FastDeploy
+pip3 install -r requirements_iluvatar.txt -i https://mirrors.aliyun.com/pypi/simple/
 bash build.sh
 ```
 
@@ -451,9 +458,9 @@ curl -X POST "http://0.0.0.0:8180/v1/chat/completions" \
 ### 4.3 PaddleOCR-VL series
 #### 4.3.1 PaddleOCR-VL-0.9B
 
-- (Optional) Install paddleocr
+- (Optional) Build and install paddleocr from source
 
-To install the latest `paddleocr`, you can compile it from source. The image contains a compilation and installation based on source code `39128c2c7fd40be44d8f33498cabd4ec10f1bfcd`.
+To install the latest `paddleocr`, you can compile it from source. The version in the image is `3.3.2`.
 
 ```bash
 git clone -b main https://github.com/PaddlePaddle/PaddleOCR.git
@@ -515,7 +522,7 @@ import os
 from paddleocr import PaddleOCRVL
 
 input_path = "./images"
-pipeline = PaddleOCRVL(vl_rec_backend="fastdeploy-server", vl_rec_server_url="http://127.0.0.1:8180/v1")
+pipeline = PaddleOCRVL(vl_rec_backend="fastdeploy-server", vl_rec_server_url="http://127.0.0.1:8180/v1", device="iluvatar_gpu")
 file_list = os.listdir(input_path)
 for file_name in file_list:
     file_path = os.path.join(input_path, file_name)

docs/zh/get_started/installation/iluvatar_gpu.md

@@ -34,17 +34,24 @@ docker exec -it paddle_infer bash
 ### 3.2 安装paddle
 
 ```bash
-pip3 install paddlepaddle-iluvatar==3.4.0.dev20260326 -i https://www.paddlepaddle.org.cn/packages/nightly/ixuca/
+pip3 install paddlepaddle-iluvatar==3.4.0.dev20260326 -i https://www.paddlepaddle.org.cn/packages/nightly/ixuca/ --extra-index-url https://mirrors.aliyun.com/pypi/simple/
 ```
 
 ### 3.3 安装fastdeploy
+
+以下两种方式安装均可以：
+
+- pip安装
 ```bash
 pip3 install fastdeploy_iluvatar_gpu==2.5.0.dev0 -i https://www.paddlepaddle.org.cn/packages/stable/ixuca/ --extra-index-url https://mirrors.aliyun.com/pypi/simple/
 ```
-可以按如下步骤编译FastDeploy，，得到```最新版本```。
+
+- 源码编译
+可以按如下步骤编译FastDeploy得到```最新版本```。
 ```bash
-git clone https://github.com/PaddlePaddle/FastDeploy
+git clone --recursive https://github.com/PaddlePaddle/FastDeploy.git
 cd FastDeploy
+pip3 install -r requirements_iluvatar.txt -i https://mirrors.aliyun.com/pypi/simple/
 bash build.sh
 ```
 
@@ -451,9 +458,9 @@ curl -X POST "http://0.0.0.0:8180/v1/chat/completions" \
 ### 4.3 PaddleOCR-VL系列
 #### 4.3.1 PaddleOCR-VL-0.9B
 
-- (可选) 安装 paddleocr
+- (可选) 源码编译安装 paddleocr
 
-如果想要安装最新的`paddleocr`，可以源码编译。镜像里是基于`39128c2c7fd40be44d8f33498cabd4ec10f1bfcd`源码编译安装的
+如果想要安装最新的`paddleocr`，可以源码编译。镜像里的版本是`3.3.2`
 
 ```bash
 git clone -b main https://github.com/PaddlePaddle/PaddleOCR.git
@@ -512,7 +519,7 @@ import os
 from paddleocr import PaddleOCRVL
 
 input_path = "./images"
-pipeline = PaddleOCRVL(vl_rec_backend="fastdeploy-server", vl_rec_server_url="http://127.0.0.1:8180/v1")
+pipeline = PaddleOCRVL(vl_rec_backend="fastdeploy-server", vl_rec_server_url="http://127.0.0.1:8180/v1", device="iluvatar_gpu")
 file_list = os.listdir(input_path)
 for file_name in file_list:
     file_path = os.path.join(input_path, file_name)

fastdeploy/model_executor/layers/backends/iluvatar/moe/fuse_moe_cutlass_iluvatar_backend.py

@@ -124,12 +124,18 @@ class IluvatarCutlassMoEMethod(UnquantizedFusedMoEMethod):
         x: paddle.Tensor,
         gate: nn.Layer,
         topk_ids_hookfunc: Callable = None,
+        fc1_latent_proj: nn.Layer = None,
+        fc2_latent_proj: nn.Layer = None,
     ) -> paddle.Tensor:
         """
         Paddle Cutlass compute Fused MoE.
         """
         gate_out = gate(x)
         gate_out = gate_out.cast("float32")
+
+        if fc1_latent_proj is not None:
+            x = fc1_latent_proj(x)
+
         if layer.topk_method == "noaux_tc":
             gate_out, topk_weights, topk_idx = get_moe_scores(
                 gate_out,
@@ -206,6 +212,9 @@ class IluvatarCutlassMoEMethod(UnquantizedFusedMoEMethod):
             routed_scaling_factor=1.0,
         )
 
+        if fc2_latent_proj is not None:
+            fused_moe_out = fc2_latent_proj(fused_moe_out)
+
         return fused_moe_out
 
 

requirements_iluvatar.txt

@@ -43,3 +43,4 @@ msgspec
 safetensors>=0.7.0
 py-cpuinfo
 transformers>=4.55.1,<5.0.0
+paddleocr[doc-parser]==3.3.2