[Misc] Clean up InternVL family config registration (#19992)

Signed-off-by: Isotr0py <2037008807@qq.com>

vllm/transformers_utils/config.py

@@ -33,10 +33,8 @@ from vllm.logger import init_logger
 from vllm.transformers_utils.configs import (ChatGLMConfig, Cohere2Config,
                                              DbrxConfig, DeepseekVLV2Config,
                                              EAGLEConfig, ExaoneConfig,
-                                             H2OVLChatConfig,
+                                             JAISConfig, KimiVLConfig,
-                                             InternVLChatConfig, JAISConfig,
+                                             MedusaConfig, MiniMaxText01Config,
-                                             KimiVLConfig, MedusaConfig,
-                                             MiniMaxText01Config,
                                              MiniMaxVL01Config, MllamaConfig,
                                              MLPSpeculatorConfig, MPTConfig,
                                              NemotronConfig, NVLM_D_Config,
@@ -90,8 +88,6 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = {
     "medusa": MedusaConfig,
     "eagle": EAGLEConfig,
     "exaone": ExaoneConfig,
-    "h2ovl_chat": H2OVLChatConfig,
-    "internvl_chat": InternVLChatConfig,
     "minimax_text_01": MiniMaxText01Config,
     "minimax_vl_01": MiniMaxVL01Config,
     "nemotron": NemotronConfig,
@@ -104,6 +100,10 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = {
     **_CONFIG_REGISTRY_OVERRIDE_HF
 }
 
+_CONFIG_ATTRS_MAPPING: dict[str, str] = {
+    "llm_config": "text_config",
+}
+
 
 class ConfigFormat(str, enum.Enum):
     AUTO = "auto"
@@ -286,6 +286,18 @@ def is_encoder_decoder(config: PretrainedConfig) -> bool:
     return getattr(config, "is_encoder_decoder", False)
 
 
+def _maybe_remap_hf_config_attrs(config: PretrainedConfig) -> PretrainedConfig:
+    """Remap config attributes to match the expected names."""
+    for old_attr, new_attr in _CONFIG_ATTRS_MAPPING.items():
+        if hasattr(config, old_attr):
+            if not hasattr(config, new_attr):
+                config.update({new_attr: getattr(config, old_attr)})
+            delattr(config, old_attr)
+            logger.debug("Remapped config attribute '%s' to '%s'", old_attr,
+                         new_attr)
+    return config
+
+
 def get_config(
     model: Union[str, Path],
     trust_remote_code: bool,
@@ -361,6 +373,9 @@ def get_config(
                     revision=revision,
                     code_revision=code_revision,
                     token=_get_hf_token(),
+                    # some old custom model's config needs
+                    # `has_no_defaults_at_init=True` to work.
+                    has_no_defaults_at_init=trust_remote_code,
                     **kwargs,
                 )
             except ValueError as e:
@@ -376,6 +391,7 @@ def get_config(
                     raise RuntimeError(err_msg) from e
                 else:
                     raise e
+        config = _maybe_remap_hf_config_attrs(config)
 
     elif config_format == ConfigFormat.MISTRAL:
         config = load_params_config(model, revision, **kwargs)

vllm/transformers_utils/configs/__init__.py

@@ -11,8 +11,6 @@ from vllm.transformers_utils.configs.exaone import ExaoneConfig
 # tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the
 # `FalconConfig` class from the official HuggingFace transformers library.
 from vllm.transformers_utils.configs.falcon import RWConfig
-from vllm.transformers_utils.configs.h2ovl import H2OVLChatConfig
-from vllm.transformers_utils.configs.internvl import InternVLChatConfig
 from vllm.transformers_utils.configs.jais import JAISConfig
 from vllm.transformers_utils.configs.kimi_vl import KimiVLConfig
 from vllm.transformers_utils.configs.medusa import MedusaConfig
@@ -38,8 +36,6 @@ __all__ = [
     "DeepseekVLV2Config",
     "MPTConfig",
     "RWConfig",
-    "H2OVLChatConfig",
-    "InternVLChatConfig",
     "JAISConfig",
     "MedusaConfig",
     "EAGLEConfig",

vllm/transformers_utils/configs/h2ovl.py

@@ -1,16 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Adapted from
-# https://huggingface.co/h2oai/h2ovl-mississippi-2b/blob/main/configuration_h2ovl_chat.py
-# --------------------------------------------------------
-# H2OVL-Mississippi
-# Copyright (c) 2024 H2O.AI
-# Licensed under Apache 2.0 License [see LICENSE for details]
-# --------------------------------------------------------
-
-from .internvl import InternVLChatConfig
-
-
-class H2OVLChatConfig(InternVLChatConfig):
-    model_type = "h2ovl_chat"

vllm/transformers_utils/configs/internvl.py

@@ -1,54 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Adapted from
-# https://huggingface.co/OpenGVLab/InternVL2-1B/blob/main/configuration_internvl_chat.py
-# --------------------------------------------------------
-# InternVL
-# Copyright (c) 2024 OpenGVLab
-# Licensed under The MIT License [see LICENSE for details]
-# --------------------------------------------------------
-from transformers.configuration_utils import PretrainedConfig
-
-
-class InternVLChatConfig(PretrainedConfig):
-    model_type = 'internvl_chat'
-    is_composition = True
-
-    def __init__(self,
-                 vision_config=None,
-                 llm_config=None,
-                 use_backbone_lora=0,
-                 use_llm_lora=0,
-                 select_layer=-1,
-                 force_image_size=None,
-                 downsample_ratio=0.5,
-                 template=None,
-                 dynamic_image_size=False,
-                 use_thumbnail=False,
-                 ps_version='v1',
-                 min_dynamic_patch=1,
-                 max_dynamic_patch=6,
-                 **kwargs):
-        super().__init__(**kwargs)
-
-        if vision_config is None:
-            vision_config = {}
-
-        if llm_config is None:
-            llm_config = {}
-
-        self.vision_config = PretrainedConfig(**vision_config)
-        self.text_config = PretrainedConfig(**llm_config)
-
-        self.use_backbone_lora = use_backbone_lora
-        self.use_llm_lora = use_llm_lora
-        self.select_layer = select_layer
-        self.force_image_size = force_image_size
-        self.downsample_ratio = downsample_ratio
-        self.template = template
-        self.dynamic_image_size = dynamic_image_size
-        self.use_thumbnail = use_thumbnail
-        self.ps_version = ps_version  # pixel shuffle version
-        self.min_dynamic_patch = min_dynamic_patch
-        self.max_dynamic_patch = max_dynamic_patch

vllm/transformers_utils/configs/nvlm_d.py

@@ -8,8 +8,24 @@
 # Copyright (c) 2024 NVIDIA
 # Licensed under Apache 2.0 License [see LICENSE for details]
 # --------------------------------------------------------
-from .internvl import InternVLChatConfig
+from transformers import Qwen2Config
+from transformers.configuration_utils import PretrainedConfig
 
 
-class NVLM_D_Config(InternVLChatConfig):
+class NVLM_D_Config(PretrainedConfig):
     model_type = 'NVLM_D'
+    is_composition = True
+
+    def __init__(self, vision_config=None, llm_config=None, **kwargs):
+        super().__init__(**kwargs)
+
+        # Handle vision_config initialization
+        if vision_config is None:
+            vision_config = {}
+
+        # Handle llm_config initialization
+        if llm_config is None:
+            llm_config = {}
+
+        self.vision_config = PretrainedConfig(**vision_config)
+        self.text_config = Qwen2Config(**llm_config)