[Model][VLM] Add Kimi-VL model support (#16387)

Signed-off-by: courage17340 <courage17340@163.com>

vllm/transformers_utils/configs/__init__.py

@@ -13,9 +13,11 @@ 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
 from vllm.transformers_utils.configs.mllama import MllamaConfig
 from vllm.transformers_utils.configs.mlp_speculator import MLPSpeculatorConfig
+from vllm.transformers_utils.configs.moonvit import MoonViTConfig
 from vllm.transformers_utils.configs.mpt import MPTConfig
 from vllm.transformers_utils.configs.nemotron import NemotronConfig
 from vllm.transformers_utils.configs.nvlm_d import NVLM_D_Config
@@ -40,6 +42,8 @@ __all__ = [
     "ExaoneConfig",
     "MllamaConfig",
     "MLPSpeculatorConfig",
+    "MoonViTConfig",
+    "KimiVLConfig",
     "NemotronConfig",
     "NVLM_D_Config",
     "Olmo2Config",

vllm/transformers_utils/configs/kimi_vl.py

@@ -0,0 +1,36 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
+from typing import Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+
+from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
+from vllm.transformers_utils.configs.moonvit import MoonViTConfig
+
+
+class KimiVLConfig(PretrainedConfig):
+    model_type = "kimi_vl"
+
+    def __init__(self,
+                 vision_config: Optional[Union[dict, MoonViTConfig]] = None,
+                 text_config: Optional[Union[dict, DeepseekV2Config]] = None,
+                 ignore_index: int = -100,
+                 media_placeholder_token_id: int = 163605,
+                 pad_token_id: int = 0,
+                 **kwargs):
+        if vision_config is None:
+            vision_config = MoonViTConfig()
+        elif isinstance(vision_config, dict):
+            vision_config = MoonViTConfig(**vision_config)
+        self.vision_config = vision_config
+
+        if text_config is None:
+            text_config = DeepseekV2Config()
+        elif isinstance(text_config, dict):
+            text_config = DeepseekV2Config(**text_config)
+        self.text_config = text_config
+
+        self.ignore_index = ignore_index
+        self.media_placeholder_token_id = media_placeholder_token_id
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)

vllm/transformers_utils/configs/moonvit.py

@@ -0,0 +1,32 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
+from transformers.configuration_utils import PretrainedConfig
+
+
+class MoonViTConfig(PretrainedConfig):
+    model_type = "moonvit"
+
+    def __init__(
+            self,
+            patch_size: int = 14,
+            init_pos_emb_height: int = 64,
+            init_pos_emb_width: int = 64,
+            num_attention_heads: int = 16,
+            num_hidden_layers: int = 27,
+            hidden_size: int = 1152,
+            intermediate_size: int = 4304,
+            merge_kernel_size: tuple[int, int] = (2, 2),
+            **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.patch_size = patch_size
+        # Positional embedding config
+        self.init_pos_emb_height = init_pos_emb_height
+        self.init_pos_emb_width = init_pos_emb_width
+        # Transformer config
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        # Patch merger config
+        self.merge_kernel_size = merge_kernel_size