[BugFix] PaddleOCR-VL fix FD_DEBUG type and support v1 loader (#4605)

ming1753 · web-flow · commit 7681375a1944 · 2025-10-28T09:47:47.000+08:00
* [Bug Fix] PaddleOCRVL fix FD_DEBUG type and support HF model

* fix bug

* fix bug

* fix bug
diff --git a/docs/usage/environment_variables.md b/docs/usage/environment_variables.md
@@ -16,7 +16,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
 
     # Enable debug mode (0 or 1)
     "FD_DEBUG":
-    lambda: os.getenv("FD_DEBUG", "0"),
+    lambda: int(os.getenv("FD_DEBUG", "0")),
 
     # FastDeploy log retention days
     "FD_LOG_BACKUP_COUNT":
diff --git a/docs/zh/usage/environment_variables.md b/docs/zh/usage/environment_variables.md
@@ -16,7 +16,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
 
     # 是否启用调试模式，可设置为 0 或 1
     "FD_DEBUG":
-    lambda: os.getenv("FD_DEBUG", "0"),
+    lambda: int(os.getenv("FD_DEBUG", "0")),
 
     # FastDeploy 日志保留天数
     "FD_LOG_BACKUP_COUNT":
diff --git a/fastdeploy/__init__.py b/fastdeploy/__init__.py
@@ -30,7 +30,7 @@
 from fastdeploy.entrypoints.llm import LLM
 from fastdeploy.utils import current_package_version, envs
 
-if envs.FD_DEBUG != "1":
+if envs.FD_DEBUG != 1:
     import logging
 
     pf_logger.logger.setLevel(logging.INFO)
diff --git a/fastdeploy/engine/args_utils.py b/fastdeploy/engine/args_utils.py
@@ -16,6 +16,7 @@
 
 import argparse
 import json
+import os
 from dataclasses import asdict, dataclass
 from dataclasses import fields as dataclass_fields
 from typing import Any, Dict, List, Optional, Union
@@ -54,6 +55,16 @@ def nullable_str(x: str) -> Optional[str]:
     return x if x else None
 
 
+def get_model_architecture(model: str, model_config_name: Optional[str] = "config.json") -> Optional[str]:
+    config_path = os.path.join(model, model_config_name)
+    if os.path.exists(config_path):
+        model_config = json.load(open(config_path, "r", encoding="utf-8"))
+        architecture = model_config["architectures"][0]
+        return architecture
+    else:
+        return model
+
+
 @dataclass
 class EngineArgs:
     # Model configuration parameters
@@ -440,6 +451,9 @@ def __post_init__(self):
         if self.guided_decoding_backend != "off":
             envs.ENABLE_V1_KVCACHE_SCHEDULER = 0
 
+        if "PaddleOCR" in get_model_architecture(self.model, self.model_config_name):
+            envs.FD_ENABLE_MAX_PREFILL = 1
+
     @staticmethod
     def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         """
diff --git a/fastdeploy/engine/sched/resource_manager_v1.py b/fastdeploy/engine/sched/resource_manager_v1.py
@@ -340,7 +340,6 @@ def _get_num_new_tokens(self, request, token_budget):
         if not self.config.model_config.enable_mm:
             return num_new_tokens
 
-        request.with_image = False
         inputs = request.multimodal_inputs
         if inputs.get("patch_idx", None) is not None and inputs.get("patch_map", None) is not None:
             pre_end_idx = request.num_computed_tokens
diff --git a/fastdeploy/envs.py b/fastdeploy/envs.py
@@ -27,7 +27,7 @@
     # Log directory.
     "FD_LOG_DIR": lambda: os.getenv("FD_LOG_DIR", "log"),
     # Whether to use debug mode, can set 0 or 1
-    "FD_DEBUG": lambda: os.getenv("FD_DEBUG", "0"),
+    "FD_DEBUG": lambda: int(os.getenv("FD_DEBUG", "0")),
     # Number of days to keep fastdeploy logs.
     "FD_LOG_BACKUP_COUNT": lambda: os.getenv("FD_LOG_BACKUP_COUNT", "7"),
     # Model download source, can set "AISTUDIO", "MODELSCOPE" or "HUGGINGFACE".
diff --git a/fastdeploy/input/paddleocr_vl_processor/process.py b/fastdeploy/input/paddleocr_vl_processor/process.py
@@ -80,14 +80,15 @@ def __init__(
         self.temporal_conv_size = self.image_processor.temporal_patch_size
 
         # Special tokens and IDs
-        self.image_token = "<|image_pad|>"
+
+        self.image_token = "<|IMAGE_PLACEHOLDER|>"
         self.video_token = "<|video_pad|>"
 
         self.image_token_id = self.tokenizer.convert_tokens_to_ids(self.image_token)
         self.video_token_id = self.tokenizer.convert_tokens_to_ids(self.video_token)
         self.image_patch_id = self.image_token_id
 
-        self.vision_start = "<|vision_start|>"
+        self.vision_start = "<|IMAGE_START|>"
         self.vision_start_id = self.tokenizer.convert_tokens_to_ids(self.vision_start)
 
         self.tokens_per_second = tokens_per_second
@@ -167,9 +168,8 @@ def text2ids(self, text, images=None, videos=None):
             "vit_position_ids": [],
         }
         # Define placeholders and their lengths
-        # IMAGE_PLACEHOLDER = "<|vision_start|><|image_pad|><|vision_end|>"
-        IMAGE_PLACEHOLDER = "<|image_pad|>"
-        VIDEO_PLACEHOLDER = "<|video@placeholder|>"
+        IMAGE_PLACEHOLDER = self.image_token
+        VIDEO_PLACEHOLDER = self.video_token
         IMAGE_PLACEHOLDER_LEN = len(IMAGE_PLACEHOLDER)
         VIDEO_PLACEHOLDER_LEN = len(VIDEO_PLACEHOLDER)
 
diff --git a/fastdeploy/logger/logger.py b/fastdeploy/logger/logger.py
@@ -104,7 +104,7 @@ def _get_legacy_logger(self, name, file_name, without_formater=False, print_to_c
         if not os.path.exists(log_dir):
             os.makedirs(log_dir, exist_ok=True)
 
-        is_debug = int(envs.FD_DEBUG)
+        is_debug = envs.FD_DEBUG
         # logger = logging.getLogger(name)
         # 为了兼容原有接口，使用命名空间进行隔离，避免logger覆盖、混乱等问题
         legacy_name = f"legacy.{name}"
diff --git a/fastdeploy/model_executor/models/paddleocr_vl/config.py b/fastdeploy/model_executor/models/paddleocr_vl/config.py
@@ -17,7 +17,7 @@
 from paddleformers.transformers.configuration_utils import PretrainedConfig
 
 
-class PPOCRVisionConfig(PretrainedConfig):
+class PaddleOCRVisionConfig(PretrainedConfig):
     model_type = "paddleocr_vl"
     base_config_key = "vision_config"
 
@@ -58,7 +58,7 @@ def __init__(
 class PaddleOCRConfig(PretrainedConfig):
     model_type = "paddleocr_vl"
     keys_to_ignore_at_inference = ["past_key_values"]
-    sub_configs = {"vision_config": PPOCRVisionConfig}
+    sub_configs = {"vision_config": PaddleOCRVisionConfig}
 
     base_model_tp_plan = {
         "layers.*.self_attn.q_proj": "colwise",
diff --git a/fastdeploy/model_executor/models/paddleocr_vl/projector.py b/fastdeploy/model_executor/models/paddleocr_vl/projector.py
@@ -15,6 +15,7 @@
 """
 
 import math
+from typing import Optional
 
 import paddle
 import paddle.nn as nn
@@ -57,8 +58,10 @@ def __init__(self, text_config, vision_config, prefix=""):
 
         self.pre_norm = nn.LayerNorm(self.vision_config.hidden_size, epsilon=1e-05)
         self.linear_1 = nn.Linear(self.hidden_size, self.hidden_size)
+        self.linear_1.weight.weight_loader = self.weight_loader
         self.act = GELUActivation()
         self.linear_2 = nn.Linear(self.hidden_size, self.text_config.hidden_size)
+        self.linear_2.weight.weight_loader = self.weight_loader
 
     def forward(self, image_features, image_grid_thw):
         m1, m2 = self.merge_kernel_size
@@ -94,6 +97,20 @@ def forward(self, image_features, image_grid_thw):
         hidden_states = self.linear_2(hidden_states)
         return hidden_states
 
+    def weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = None):
+        loaded_weight = get_tensor(loaded_weight)
+        loaded_weight = loaded_weight.transpose([1, 0])
+        assert param.shape == loaded_weight.shape, (
+            f" Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({param.shape})"
+        )
+        # Ensure loaded weight dtype matches model param dtype
+        if loaded_weight.dtype != param.dtype:
+            if loaded_weight.dtype == paddle.int8 and param.dtype == paddle.float8_e4m3fn:
+                loaded_weight = loaded_weight.view(param.dtype)
+            else:
+                loaded_weight = loaded_weight.cast(param.dtype)
+        param.copy_(loaded_weight, False)
+
     def load_state_dict(self, state_dict):
         params_dict = dict(self.named_parameters())
         for param_name, param in params_dict.items():
diff --git a/fastdeploy/model_executor/models/paddleocr_vl/siglip.py b/fastdeploy/model_executor/models/paddleocr_vl/siglip.py
@@ -14,25 +14,20 @@
 # limitations under the License.
 """
 
+import os
 from typing import List, Optional, Tuple, Union
 
 import numpy as np
 import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
-from paddle.nn.functional.flash_attention import flash_attn_unpadded
 from paddleformers.transformers.activations import ACT2FN
 from paddleformers.transformers.model_utils import PretrainedModel
 
 from fastdeploy.model_executor.layers.utils import get_tensor
 from fastdeploy.model_executor.utils import slice_fn
 
-try:
-    from paddle.nn.functional.flash_attention import flash_attention_v3_varlen
-except:
-    flash_attention_v3_varlen = None
-
-from .config import PPOCRVisionConfig
+from .config import PaddleOCRVisionConfig
 
 
 def rotate_half(x):
@@ -61,22 +56,51 @@ def apply_rotary_pos_emb_vision(x, cos, sin):
     return x_embed.astype(orig_dtype)
 
 
-class QKVLinear(nn.Linear):
-    def __init__(self, config, in_features, out_features, weight_attr=None, bias_attr=None):
-        super().__init__(in_features, out_features, weight_attr, bias_attr)
+class SiglipAttention(nn.Layer):
+    def __init__(self, config):
+        super().__init__()
         self.config = config
-        self.in_features = in_features
-        self.out_features = out_features
         self.embed_dim = config.hidden_size
         self.num_heads = config.num_attention_heads
         self.head_dim = self.embed_dim // self.num_heads
         assert self.head_dim * self.num_heads == self.embed_dim
-        self.weight.weight_loader = self.weight_loader
-        self.bias.weight_loader = self.weight_loader
+        self.scale = self.head_dim**-0.5
 
-    def weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = None):
+        # qkv_linear
+        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias_attr=True)
+        self.qkv_proj.weight.weight_loader = self.qkv_weight_loader
+        self.qkv_proj.bias.weight_loader = self.qkv_weight_loader
+
+        # out_linear
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj.weight.weight_loader = self.out_proj_weight_loader
+
+        enable_fa3 = False
+        flash_attn_version = int(os.environ.get("FLAGS_flash_attn_version", "2"))
+        if flash_attn_version == 3:
+            prop = paddle.device.cuda.get_device_properties()
+            cc = prop.major * 10 + prop.minor
+            is_current_sm_supported = cc >= 90
+            is_paddle_supported = any(num >= 90 for num in paddle.version.cuda_archs())
+            enable_fa3 = is_current_sm_supported and is_paddle_supported
+
+        if enable_fa3:
+            from paddle.nn.functional.flash_attention import flash_attention_v3_varlen
+
+            self.flash_attn_func = flash_attention_v3_varlen
+            self.flash_attn_kwargs = {}
+        else:
+            from paddle.nn.functional.flash_attention import flash_attn_unpadded
+
+            self.flash_attn_func = flash_attn_unpadded
+            self.flash_attn_kwargs = {"scale": self.scale, "training": False}
+
+    def qkv_weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = None):
         # Tensor parallelism splits the weight along the output_dim
         loaded_weight = get_tensor(loaded_weight)
+        if loaded_weight.dim() == 2:
+            loaded_weight = loaded_weight.transpose([1, 0])
+
         if not param._is_initialized():
             param.initialize()
         if loaded_shard_id == "q":
@@ -90,7 +114,7 @@ def weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = N
             param_shard_offset = self.num_heads * self.head_dim * 2
             param_shard_size = self.num_heads * self.head_dim
 
-        param = slice_fn(param, self.out_features, start=param_shard_offset, end=param_shard_offset + param_shard_size)
+        param = slice_fn(param, -1, start=param_shard_offset, end=param_shard_offset + param_shard_size)
         assert param.shape == loaded_weight.shape, (
             f" Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({param.shape})"
         )
@@ -102,30 +126,19 @@ def weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = N
                 loaded_weight = loaded_weight.cast(param.dtype)
         param.copy_(loaded_weight, False)
 
-
-class SiglipAttention(nn.Layer):
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.embed_dim = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.head_dim = self.embed_dim // self.num_heads
-        assert self.head_dim * self.num_heads == self.embed_dim
-        self.scale = self.head_dim**-0.5
-
-        self.qkv_proj = QKVLinear(config, self.embed_dim, self.embed_dim * 3, bias_attr=True)
-        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
-
-        prop = paddle.device.cuda.get_device_properties()
-        cc = prop.major * 10 + prop.minor
-        is_current_sm_supported = cc >= 90
-        is_paddle_supported = any(num >= 90 for num in paddle.version.cuda_archs())
-        if is_current_sm_supported and is_paddle_supported:
-            self.flash_attn_func = flash_attention_v3_varlen
-            self.flash_attn_kwargs = {}
-        else:
-            self.flash_attn_func = flash_attn_unpadded
-            self.flash_attn_kwargs = {"scale": self.scale, "training": False}
+    def out_proj_weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = None):
+        loaded_weight = get_tensor(loaded_weight)
+        loaded_weight = loaded_weight.transpose([1, 0])
+        assert param.shape == loaded_weight.shape, (
+            f" Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({param.shape})"
+        )
+        # Ensure loaded weight dtype matches model param dtype
+        if loaded_weight.dtype != param.dtype:
+            if loaded_weight.dtype == paddle.int8 and param.dtype == paddle.float8_e4m3fn:
+                loaded_weight = loaded_weight.view(param.dtype)
+            else:
+                loaded_weight = loaded_weight.cast(param.dtype)
+        param.copy_(loaded_weight, False)
 
     def forward(
         self,
@@ -170,7 +183,7 @@ def forward(
         )[0]
         # --------
 
-        attn_output = attn_output.reshape(seq_length, -1)
+        attn_output = attn_output.reshape((seq_length, -1))
         attn_output = self.out_proj(attn_output)
 
         return attn_output
@@ -315,11 +328,28 @@ def __init__(self, config):
         super().__init__()
         self.config = config
         if config.hidden_act == "gelu_pytorch_tanh":
-            config.hidden_act = "silu"
+            config.hidden_act = "gelu_new"
+
         self.activation_fn = ACT2FN[config.hidden_act]
 
         self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc1.weight.weight_loader = self.weight_loader
         self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.fc2.weight.weight_loader = self.weight_loader
+
+    def weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = None):
+        loaded_weight = get_tensor(loaded_weight)
+        loaded_weight = loaded_weight.transpose([1, 0])
+        assert param.shape == loaded_weight.shape, (
+            f" Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({param.shape})"
+        )
+        # Ensure loaded weight dtype matches model param dtype
+        if loaded_weight.dtype != param.dtype:
+            if loaded_weight.dtype == paddle.int8 and param.dtype == paddle.float8_e4m3fn:
+                loaded_weight = loaded_weight.view(param.dtype)
+            else:
+                loaded_weight = loaded_weight.cast(param.dtype)
+        param.copy_(loaded_weight, False)
 
     def forward(self, hidden_states: paddle.Tensor) -> paddle.Tensor:
         hidden_states = self.fc1(hidden_states)
@@ -576,7 +606,7 @@ def forward(
 class SiglipMultiheadAttentionPoolingHead(nn.Layer):
     """Multihead Attention Pooling."""
 
-    def __init__(self, config: PPOCRVisionConfig):
+    def __init__(self, config: PaddleOCRVisionConfig):
         super().__init__()
 
         self.probe = self.create_parameter(
@@ -601,7 +631,7 @@ def forward(self, hidden_state, key_padding_mask=None):
 
 
 class SiglipVisionTransformer(nn.Layer):
-    def __init__(self, config: PPOCRVisionConfig):
+    def __init__(self, config: PaddleOCRVisionConfig):
         super().__init__()
         self.config = config
         embed_dim = config.hidden_size
@@ -666,10 +696,10 @@ def forward(
 
 
 class SiglipVisionModel(PretrainedModel):
-    config_class = PPOCRVisionConfig
+    config_class = PaddleOCRVisionConfig
     main_input_name = "pixel_values"
 
-    def __init__(self, config: PPOCRVisionConfig, prefix=""):
+    def __init__(self, config: PaddleOCRVisionConfig, prefix=""):
         super().__init__(config)
         self.prefix_name = prefix
         self.vision_model = SiglipVisionTransformer(config)
diff --git a/fastdeploy/worker/worker_process.py b/fastdeploy/worker/worker_process.py
@@ -815,6 +815,10 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
     update_fd_config_for_mm(fd_config)
     if fd_config.load_config.load_choices == "default_v1" and not v1_loader_support(fd_config):
         fd_config.load_config.load_choices = "default"
+
+    architecture = fd_config.model_config.architectures[0]
+    if "PaddleOCR" in architecture:
+        envs.FD_ENABLE_MAX_PREFILL = 1
     return fd_config
 
 
diff --git a/tests/logger/test_logger.py b/tests/logger/test_logger.py
diff --git a/tests/logger/test_setup_logging.py b/tests/logger/test_setup_logging.py
