Fix bugs (#4616)

* Rename files and classes

* Fix doc

* Fix bug

Author: Bobholamovic

docs/pipeline_usage/tutorials/ocr_pipelines/PaddleOCR-VL.md

@@ -878,15 +878,11 @@ for res in output:
 
 #### 3.1.1 使用 Docker 镜像
 
-PaddleX 针对不同推理加速框架提供了相应的 Docker 镜像，用于快速启动 VLM 推理服务：
-
-* **vLLM**：`ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddlex-genai-vllm-server`
-* **SGLang**：`ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddlex-genai-sglang-server`
-
-以 vLLM 为例，可使用以下命令启动服务：
+PaddleX 提供了 Docker 镜像，用于快速启动 vLLM 推理服务。可使用以下命令启动服务：
 
 ```bash
 docker run \
+    -it \
     --rm \
     --gpus all \
     --network host \
@@ -899,6 +895,7 @@ docker run \
 
 ```bash
 docker run \
+    -it \
     --rm \
     --gpus all \
     --network host \

paddlex/inference/models/doc_vlm/modeling/__init__.py

@@ -13,6 +13,6 @@
 # limitations under the License.
 
 from .GOT_ocr_2_0 import PPChart2TableInference
-from .ppocrvl import PPOCRVLForConditionalGeneration
+from .paddleocr_vl import PaddleOCRVLForConditionalGeneration
 from .qwen2_5_vl import PPDocBee2Inference
 from .qwen2_vl import PPDocBeeInference, Qwen2VLForConditionalGeneration

paddlex/inference/models/doc_vlm/modeling/ppocrvl/__init__.py renamed to paddlex/inference/models/doc_vlm/modeling/paddleocr_vl/__init__.py

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ._ppocrvl import PPOCRVLForConditionalGeneration
+from ._paddleocr_vl import PaddleOCRVLForConditionalGeneration

paddlex/inference/models/doc_vlm/modeling/ppocrvl/_config.py renamed to paddlex/inference/models/doc_vlm/modeling/paddleocr_vl/_config.py

@@ -30,7 +30,7 @@
 
 
 class PPOCRVisionConfig(PretrainedConfig):
-    model_type = "ppocrvl"
+    model_type = "paddleocr_vl"
     base_config_key = "vision_config"
 
     def __init__(
@@ -67,8 +67,8 @@ def __init__(
         self.tokens_per_second = tokens_per_second
 
 
-class PPOCRVLConfig(PretrainedConfig):
-    model_type = "ppocrvl"
+class PaddleOCRVLConfig(PretrainedConfig):
+    model_type = "paddleocr_vl"
     keys_to_ignore_at_inference = ["past_key_values"]
     sub_configs = {"vision_config": PPOCRVisionConfig}
 

paddlex/inference/models/doc_vlm/modeling/ppocrvl/_distributed/__init__.py renamed to paddlex/inference/models/doc_vlm/modeling/paddleocr_vl/_distributed/__init__.py

@@ -43,7 +43,7 @@
 from ....common.vlm.transformers.model_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
 )
-from ._config import PPOCRVLConfig
+from ._config import PaddleOCRVLConfig
 from ._distributed import (
     AllGatherVarlenOp,
     ColumnParallelLinear,
@@ -76,7 +76,7 @@ def calc_lm_head_logits(
     handling sequence parallelism and tensor parallelism configurations.
 
     Args:
-        config (PPOCRVLConfig): Model configuration.
+        config (PaddleOCRVLConfig): Model configuration.
         hidden_states (Tensor): Hidden states from the transformer layers
         weight (Tensor): Weight matrix for the language model head
         bias (Tensor): Bias vector for the language model head
@@ -263,7 +263,7 @@ def __init__(self, config):
         Initialize RMSNorm layer.
 
         Args:
-            config (PPOCRVLConfig): Model configuration.
+            config (PaddleOCRVLConfig): Model configuration.
         """
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -324,7 +324,7 @@ def __init__(self, config):
         Initialize LayerNorm with configuration.
 
         Args:
-            config (PPOCRVLConfig): Model configuration contains normalization parameters and flags.
+            config (PaddleOCRVLConfig): Model configuration contains normalization parameters and flags.
         """
         super().__init__(config.hidden_size, epsilon=config.rms_norm_eps)
         self.config = config
@@ -334,7 +334,7 @@ def __init__(self, config):
 
 
 class KeyeRotaryEmbedding(nn.Layer):
-    def __init__(self, config: PPOCRVLConfig, device=None):
+    def __init__(self, config: PaddleOCRVLConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         if config is None:
@@ -408,7 +408,7 @@ def __init__(self, config, layer_idx=0):
         Initialize the MLP module with configuration options.
 
         Args:
-            config (PPOCRVLConfig): Model configurations.
+            config (PaddleOCRVLConfig): Model configurations.
             layer_idx (int): Index of current layer (default: 0)
         """
         super().__init__()
@@ -507,7 +507,7 @@ def __init__(self, config, layer_idx=0):
         """Initialize the attention layer.
 
         Args:
-            config (PPOCRVLConfig): Model configuration.
+            config (PaddleOCRVLConfig): Model configuration.
             layer_idx (int, optional): Index in transformer stack. Defaults to 0.
         """
         super().__init__()
@@ -1223,7 +1223,7 @@ def __init__(self, config, return_tuple=True):
         """Initialize the pretraining criterion.
 
         Args:
-            config (PPOCRVLConfig): Model configuration.
+            config (PaddleOCRVLConfig): Model configuration.
             return_tuple (bool): Whether to return loss as tuple (loss, loss_sum). Defaults to True.
         """
         super(ErniePretrainingCriterion, self).__init__()
@@ -1516,7 +1516,7 @@ def __init__(self, config):
         """Initialize the language model head.
 
         Args:
-            config (PPOCRVLConfig): Model configuration containing:
+            config (PaddleOCRVLConfig): Model configuration containing:
                 - vocab_size: Size of vocabulary
                 - hidden_size: Dimension of hidden states
                 - tensor_parallel_degree: Degree of tensor parallelism
@@ -1632,7 +1632,7 @@ def __init__(self, config, layer_idx):
         """Initialize the decoder layer.
 
         Args:
-            config (PPOCRVLConfig): Model configuration.
+            config (PaddleOCRVLConfig): Model configuration.
             layer_idx (int): Index of this layer in the transformer stack
         """
         super().__init__()
@@ -1776,15 +1776,15 @@ def model_parallel_dropout(self):
 class Ernie4_5PretrainedModel(PretrainedModel):
     """Base class for ERNIE pretrained models."""
 
-    config_class = PPOCRVLConfig
+    config_class = PaddleOCRVLConfig
     base_model_prefix = "ernie"
 
     @classmethod
     def _get_tensor_parallel_mappings(cls, config, is_split=True):
         """Generate tensor parallel mappings for model conversion.
 
         Args:
-            config (PPOCRVLConfig): Model configuration.
+            config (PaddleOCRVLConfig): Model configuration.
             is_split (bool): Whether to generate split mappings (True)
                             or merge mappings (False). Defaults to True.
 
@@ -2005,11 +2005,11 @@ def get_tensor_parallel_split_mappings(num_hidden_layers):
 class Ernie4_5Model(Ernie4_5PretrainedModel):
     """The core ERNIE transformer model"""
 
-    def __init__(self, config: PPOCRVLConfig):
+    def __init__(self, config: PaddleOCRVLConfig):
         """Initialize the ERNIE model architecture.
 
         Args:
-            config (PPOCRVLConfig): Model configuration.
+            config (PaddleOCRVLConfig): Model configuration.
         """
         super().__init__(config)
         self.padding_idx = config.pad_token_id