xinference 1.1.0__py3-none-any.whl → 1.2.0__py3-none-any.whl

xinference/model/image/model_spec_modelscope.json

@@ -12,8 +12,24 @@
     ],
     "default_model_config": {
       "quantize": true,
-      "quantize_text_encoder": "text_encoder_2"
-    }
+      "quantize_text_encoder": "text_encoder_2",
+      "torch_dtype": "bfloat16"
+    },
+    "gguf_model_id": "Xorbits/FLUX.1-schnell-gguf",
+    "gguf_quantizations": [
+      "F16",
+      "Q2_K",
+      "Q3_K_S",
+      "Q4_0",
+      "Q4_1",
+      "Q4_K_S",
+      "Q5_0",
+      "Q5_1",
+      "Q5_K_S",
+      "Q6_K",
+      "Q8_0"
+    ],
+    "gguf_model_file_name_template": "flux1-schnell-{quantization}.gguf"
   },
   {
     "model_name": "FLUX.1-dev",
@@ -28,8 +44,24 @@
     ],
     "default_model_config": {
       "quantize": true,
-      "quantize_text_encoder": "text_encoder_2"
-    }
+      "quantize_text_encoder": "text_encoder_2",
+      "torch_dtype": "bfloat16"
+    },
+    "gguf_model_id": "AI-ModelScope/FLUX.1-dev-gguf",
+    "gguf_quantizations": [
+      "F16",
+      "Q2_K",
+      "Q3_K_S",
+      "Q4_0",
+      "Q4_1",
+      "Q4_K_S",
+      "Q5_0",
+      "Q5_1",
+      "Q5_K_S",
+      "Q6_K",
+      "Q8_0"
+    ],
+    "gguf_model_file_name_template": "flux1-dev-{quantization}.gguf"
   },
   {
     "model_name": "sd3-medium",
@@ -47,6 +79,120 @@
       "quantize_text_encoder": "text_encoder_3"
     }
   },
+  {
+    "model_name": "sd3.5-medium",
+    "model_family": "stable_diffusion",
+    "model_hub": "modelscope",
+    "model_id": "AI-ModelScope/stable-diffusion-3.5-medium",
+    "model_revision": "master",
+    "model_ability": [
+      "text2image",
+      "image2image",
+      "inpainting"
+    ],
+    "default_model_config": {
+      "quantize": true,
+      "quantize_text_encoder": "text_encoder_3",
+      "torch_dtype": "bfloat16"
+    },
+    "gguf_model_id": "Xorbits/stable-diffusion-3.5-medium-gguf",
+    "gguf_quantizations": [
+      "F16",
+      "Q3_K_M",
+      "Q3_K_S",
+      "Q4_0",
+      "Q4_1",
+      "Q4_K_M",
+      "Q4_K_S",
+      "Q5_0",
+      "Q5_1",
+      "Q5_K_M",
+      "Q5_K_S",
+      "Q6_K",
+      "Q8_0"
+    ],
+    "gguf_model_file_name_template": "sd3.5_medium-{quantization}.gguf"
+  },
+  {
+    "model_name": "sd3.5-large",
+    "model_family": "stable_diffusion",
+    "model_hub": "modelscope",
+    "model_id": "AI-ModelScope/stable-diffusion-3.5-large",
+    "model_revision": "master",
+    "model_ability": [
+      "text2image",
+      "image2image",
+      "inpainting"
+    ],
+    "default_model_config": {
+      "quantize": true,
+      "quantize_text_encoder": "text_encoder_3",
+      "torch_dtype": "bfloat16",
+      "transformer_nf4": true
+    },
+    "gguf_model_id": "Xorbits/stable-diffusion-3.5-large-gguf",
+    "gguf_quantizations": [
+      "F16",
+      "Q4_0",
+      "Q4_1",
+      "Q5_0",
+      "Q5_1",
+      "Q8_0"
+    ],
+    "gguf_model_file_name_template": "sd3.5_large-{quantization}.gguf"
+  },
+  {
+    "model_name": "sd3.5-large-turbo",
+    "model_family": "stable_diffusion",
+    "model_hub": "modelscope",
+    "model_id": "AI-ModelScope/stable-diffusion-3.5-large-turbo",
+    "model_revision": "master",
+    "model_ability": [
+      "text2image",
+      "image2image",
+      "inpainting"
+    ],
+    "default_model_config": {
+      "quantize": true,
+      "quantize_text_encoder": "text_encoder_3",
+      "torch_dtype": "bfloat16",
+      "transformer_nf4": true
+    },
+    "default_generate_config": {
+      "guidance_scale": 1.0,
+      "num_inference_steps": 4
+    },
+    "gguf_model_id": "Xorbits/stable-diffusion-3.5-large-turbo-gguf",
+    "gguf_quantizations": [
+      "F16",
+      "Q4_0",
+      "Q4_1",
+      "Q5_0",
+      "Q5_1",
+      "Q8_0"
+    ],
+    "gguf_model_file_name_template": "sd3.5_large_turbo-{quantization}.gguf"
+  },
+  {
+    "model_name": "HunyuanDiT-v1.2",
+    "model_family": "stable_diffusion",
+    "model_hub": "modelscope",
+    "model_id": "Xorbits/HunyuanDiT-v1.2-Diffusers",
+    "model_revision": "master",
+    "model_ability": [
+      "text2image"
+    ]
+  },
+  {
+    "model_name": "HunyuanDiT-v1.2-Distilled",
+    "model_family": "stable_diffusion",
+    "model_hub": "modelscope",
+    "model_id": "Xorbits/HunyuanDiT-v1.2-Diffusers-Distilled",
+    "model_revision": "master",
+    "model_ability": [
+      "text2image"
+    ]
+  },
   {
     "model_name": "sd-turbo",
     "model_family": "stable_diffusion",

xinference/model/image/stable_diffusion/core.py

@@ -14,8 +14,10 @@
 
 import contextlib
 import gc
+import importlib
 import inspect
 import itertools
+import json
 import logging
 import os
 import re
@@ -86,6 +88,7 @@ class DiffusionModel(SDAPIDiffusionModelMixin):
         lora_load_kwargs: Optional[Dict] = None,
         lora_fuse_kwargs: Optional[Dict] = None,
         model_spec: Optional["ImageModelFamilyV1"] = None,
+        gguf_model_path: Optional[str] = None,
         **kwargs,
     ):
         self._model_uid = model_uid
@@ -109,6 +112,8 @@ class DiffusionModel(SDAPIDiffusionModelMixin):
         self._model_spec = model_spec
         self._abilities = model_spec.model_ability or []  # type: ignore
         self._kwargs = kwargs
+        # gguf
+        self._gguf_model_path = gguf_model_path
 
     @property
     def model_ability(self):
@@ -184,7 +189,17 @@ class DiffusionModel(SDAPIDiffusionModelMixin):
             self._model.fuse_lora(**self._lora_fuse_kwargs)
             logger.info(f"Successfully loaded the LoRA for model {self._model_uid}.")
 
+    def _get_layer_cls(self, layer: str):
+        with open(os.path.join(self._model_path, "model_index.json")) as f:  # type: ignore
+            model_index = json.load(f)
+            layer_info = model_index[layer]
+            module_name, class_name = layer_info
+            module = importlib.import_module(module_name)
+            return getattr(module, class_name)
+
     def load(self):
+        from transformers import BitsAndBytesConfig, T5EncoderModel
+
         if "text2image" in self._abilities or "image2image" in self._abilities:
             from diffusers import AutoPipelineForText2Image as AutoPipelineModel
         elif "inpainting" in self._abilities:
@@ -200,7 +215,9 @@ class DiffusionModel(SDAPIDiffusionModelMixin):
                 glob(os.path.join(self._model_path, "*/*.safetensors"))
             )
         if isinstance(torch_dtype, str):
-            self._kwargs["torch_dtype"] = getattr(torch, torch_dtype)
+            self._torch_dtype = torch_dtype = self._kwargs["torch_dtype"] = getattr(
+                torch, torch_dtype
+            )
 
         controlnet = self._kwargs.get("controlnet")
         if controlnet is not None:
@@ -212,18 +229,7 @@ class DiffusionModel(SDAPIDiffusionModelMixin):
                 ]
 
         quantize_text_encoder = self._kwargs.pop("quantize_text_encoder", None)
-        if quantize_text_encoder:
-            try:
-                from transformers import BitsAndBytesConfig, T5EncoderModel
-            except ImportError:
-                error_message = "Failed to import module 'transformers'"
-                installation_guide = [
-                    "Please make sure 'transformers' is installed. ",
-                    "You can install it by `pip install transformers`\n",
-                ]
-
-                raise ImportError(f"{error_message}\n\n{''.join(installation_guide)}")
-
+        if quantize_text_encoder and not self._gguf_model_path:
             try:
                 import bitsandbytes  # noqa: F401
             except ImportError:
@@ -249,6 +255,32 @@ class DiffusionModel(SDAPIDiffusionModelMixin):
                 self._kwargs[text_encoder_name] = text_encoder
                 self._kwargs["device_map"] = "balanced"
 
+        if self._gguf_model_path:
+            from diffusers import GGUFQuantizationConfig
+
+            # GGUF transformer
+            self._kwargs["transformer"] = self._get_layer_cls(
+                "transformer"
+            ).from_single_file(
+                self._gguf_model_path,
+                quantization_config=GGUFQuantizationConfig(compute_dtype=torch_dtype),
+                torch_dtype=torch_dtype,
+                config=os.path.join(self._model_path, "transformer"),
+            )
+        elif self._kwargs.get("transformer_nf4"):
+            nf4_config = BitsAndBytesConfig(
+                load_in_4bit=True,
+                bnb_4bit_quant_type="nf4",
+                bnb_4bit_compute_dtype=torch_dtype,
+            )
+            model_nf4 = self._get_layer_cls("transformer").from_pretrained(
+                self._model_path,
+                subfolder="transformer",
+                quantization_config=nf4_config,
+                torch_dtype=torch_dtype,
+            )
+            self._kwargs["transformer"] = model_nf4
+
         logger.debug(
             "Loading model from %s, kwargs: %s", self._model_path, self._kwargs
         )

xinference/model/llm/__init__.py

@@ -134,6 +134,7 @@ def _install():
     from .mlx.core import MLXChatModel, MLXModel, MLXVisionModel
     from .sglang.core import SGLANGChatModel, SGLANGModel
     from .transformers.chatglm import ChatglmPytorchChatModel
+    from .transformers.cogagent import CogAgentChatModel
     from .transformers.cogvlm2 import CogVLM2Model
     from .transformers.cogvlm2_video import CogVLM2VideoModel
     from .transformers.core import PytorchChatModel, PytorchModel
@@ -195,6 +196,7 @@ def _install():
             DeepSeekV2PytorchChatModel,
             OptPytorchModel,
             GlmEdgeVModel,
+            CogAgentChatModel,
         ]
     )
     if OmniLMMModel:  # type: ignore

xinference/model/llm/llm_family.json

@@ -8942,5 +8942,148 @@
       "<|user|>",
       "<|observation|>"
     ]
+  },
+  {
+    "version": 1,
+    "context_length": 32768,
+    "model_name": "QvQ-72B-Preview",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "vision"
+    ],
+    "model_description": "QVQ-72B-Preview is an experimental research model developed by the Qwen team, focusing on enhancing visual reasoning capabilities.",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 72,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "Qwen/QVQ-72B-Preview"
+      },
+      {
+        "model_format": "mlx",
+        "model_size_in_billions": 72,
+        "quantizations": [
+          "3bit",
+          "4bit",
+          "6bit",
+          "8bit",
+          "bf16"
+        ],
+        "model_id": "mlx-community/QVQ-72B-Preview-{quantization}"
+      }
+    ],
+    "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful and harmless assistant. You are Qwen developed by Alibaba. You should think step-by-step.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}",
+    "stop_token_ids": [
+      151645,
+      151643
+    ],
+    "stop": [
+      "<|im_end|>",
+      "<|endoftext|>"
+    ]
+  },
+  {
+    "version": 1,
+    "context_length": 32768,
+    "model_name": "marco-o1",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "tools"
+    ],
+    "model_description": "Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 7,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "AIDC-AI/Marco-o1"
+      },
+      {
+        "model_format": "ggufv2",
+        "model_size_in_billions": 7,
+        "quantizations": [
+          "Q2_K",
+          "Q3_K_L",
+          "Q3_K_M",
+          "Q3_K_S",
+          "Q4_0",
+          "Q4_1",
+          "Q4_K_M",
+          "Q4_K_S",
+          "Q5_0",
+          "Q5_1",
+          "Q5_K_M",
+          "Q5_K_S",
+          "Q6_K",
+          "Q8_0"
+        ],
+        "model_id": "QuantFactory/Marco-o1-GGUF",
+        "model_file_name_template": "Marco-o1.{quantization}.gguf"
+      }
+    ],
+    "chat_template": "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|im_start|>system\n\n你是一个经过良好训练的AI助手，你的名字是Marco-o1.由阿里国际数字商业集团的AI Business创造.\n        \n## 重要！！！！！\n当你回答问题时，你的思考应该在<Thought>内完成，<Output>内输出你的结果。\n<Thought>应该尽可能是英文，但是有2个特例，一个是对原文中的引用，另一个是是数学应该使用markdown格式，<Output>内的输出需要遵循用户输入的语言。\n        <|im_end|>\n' }}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+    "stop_token_ids": [
+      151643,
+      151644,
+      151645
+    ],
+    "stop": [
+      "<|endoftext|>",
+      "<|im_start|>",
+      "<|im_end|>"
+    ]
+  },
+  {
+    "version": 1,
+    "context_length": 4096,
+    "model_name": "cogagent",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "vision"
+    ],
+    "model_description": "The CogAgent-9B-20241220 model is based on GLM-4V-9B, a bilingual open-source VLM base model. Through data collection and optimization, multi-stage training, and strategy improvements, CogAgent-9B-20241220 achieves significant advancements in GUI perception, inference prediction accuracy, action space completeness, and task generalizability. ",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": "9",
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "THUDM/cogagent-9b-20241220"
+      }
+    ],
+    "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+    "stop_token_ids": [
+      151329,
+      151336,
+      151338
+    ],
+    "stop": [
+      "<|endoftext|>",
+      "<|user|>",
+      "<|observation|>"
+    ]
   }
 ]

xinference/model/llm/llm_family.py

@@ -972,46 +972,25 @@ def match_llm(
         return spec
 
     # priority: download_hub > download_from_modelscope() and download_from_csghub()
-    if download_hub == "modelscope":
-        all_families = (
-            BUILTIN_MODELSCOPE_LLM_FAMILIES
-            + BUILTIN_LLM_FAMILIES
-            + user_defined_llm_families
-        )
-    elif download_hub == "openmind_hub":
-        all_families = (
-            BUILTIN_OPENMIND_HUB_LLM_FAMILIES
-            + BUILTIN_LLM_FAMILIES
-            + user_defined_llm_families
-        )
-    elif download_hub == "csghub":
-        all_families = (
-            BUILTIN_CSGHUB_LLM_FAMILIES
-            + BUILTIN_LLM_FAMILIES
-            + user_defined_llm_families
-        )
-    elif download_hub == "huggingface":
-        all_families = BUILTIN_LLM_FAMILIES + user_defined_llm_families
+    # set base model
+    base_families = BUILTIN_LLM_FAMILIES + user_defined_llm_families
+    hub_families_map = {
+        "modelscope": BUILTIN_MODELSCOPE_LLM_FAMILIES,
+        "openmind_hub": BUILTIN_OPENMIND_HUB_LLM_FAMILIES,
+        "csghub": BUILTIN_CSGHUB_LLM_FAMILIES,
+    }
+    if download_hub == "huggingface":
+        all_families = base_families
+    elif download_hub in hub_families_map:
+        all_families = hub_families_map[download_hub] + base_families
     elif download_from_modelscope():
-        all_families = (
-            BUILTIN_MODELSCOPE_LLM_FAMILIES
-            + BUILTIN_LLM_FAMILIES
-            + user_defined_llm_families
-        )
+        all_families = BUILTIN_MODELSCOPE_LLM_FAMILIES + base_families
     elif download_from_openmind_hub():
-        all_families = (
-            BUILTIN_OPENMIND_HUB_LLM_FAMILIES
-            + BUILTIN_LLM_FAMILIES
-            + user_defined_llm_families
-        )
+        all_families = BUILTIN_OPENMIND_HUB_LLM_FAMILIES + base_families
     elif download_from_csghub():
-        all_families = (
-            BUILTIN_CSGHUB_LLM_FAMILIES
-            + BUILTIN_LLM_FAMILIES
-            + user_defined_llm_families
-        )
+        all_families = BUILTIN_CSGHUB_LLM_FAMILIES + base_families
     else:
-        all_families = BUILTIN_LLM_FAMILIES + user_defined_llm_families
+        all_families = base_families
 
     for family in all_families:
         if model_name != family.model_name:

xinference/model/llm/llm_family_modelscope.json

@@ -6673,5 +6673,153 @@
       "<|user|>",
       "<|observation|>"
     ]
+  },
+  {
+    "version": 1,
+    "context_length": 32768,
+    "model_name": "QvQ-72B-Preview",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "vision"
+    ],
+    "model_description": "QVQ-72B-Preview is an experimental research model developed by the Qwen team, focusing on enhancing visual reasoning capabilities.",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 72,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "Qwen/QVQ-72B-Preview",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "mlx",
+        "model_size_in_billions": 72,
+        "quantizations": [
+          "3bit",
+          "4bit",
+          "6bit",
+          "8bit",
+          "bf16"
+        ],
+        "model_id": "mlx-community/QVQ-72B-Preview-{quantization}",
+        "model_hub": "modelscope"
+      }
+    ],
+    "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful and harmless assistant. You are Qwen developed by Alibaba. You should think step-by-step.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}",
+    "stop_token_ids": [
+      151645,
+      151643
+    ],
+    "stop": [
+      "<|im_end|>",
+      "<|endoftext|>"
+    ]
+  },
+   {
+    "version": 1,
+    "context_length": 32768,
+    "model_name": "marco-o1",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "tools"
+    ],
+    "model_description": "Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 7,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "AIDC-AI/Marco-o1",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "ggufv2",
+        "model_size_in_billions": 7,
+        "quantizations": [
+          "Q2_K",
+          "Q3_K_L",
+          "Q3_K_M",
+          "Q3_K_S",
+          "Q4_0",
+          "Q4_1",
+          "Q4_K_M",
+          "Q4_K_S",
+          "Q5_0",
+          "Q5_1",
+          "Q5_K_M",
+          "Q5_K_S",
+          "Q6_K",
+          "Q8_0"
+        ],
+        "model_file_name_template": "Marco-o1.{quantization}.gguf",
+        "model_hub": "modelscope",
+        "model_id": "QuantFactory/Marco-o1-GGUF"
+      }
+    ],
+    "chat_template": "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|im_start|>system\n\n你是一个经过良好训练的AI助手，你的名字是Marco-o1.由阿里国际数字商业集团的AI Business创造.\n        \n## 重要！！！！！\n当你回答问题时，你的思考应该在<Thought>内完成，<Output>内输出你的结果。\n<Thought>应该尽可能是英文，但是有2个特例，一个是对原文中的引用，另一个是是数学应该使用markdown格式，<Output>内的输出需要遵循用户输入的语言。\n        <|im_end|>\n' }}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+    "stop_token_ids": [
+      151643,
+      151644,
+      151645
+    ],
+    "stop": [
+      "<|endoftext|>",
+      "<|im_start|>",
+      "<|im_end|>"
+    ]
+  },
+  {
+    "version": 1,
+    "context_length": 4096,
+    "model_name": "cogagent",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "vision"
+    ],
+    "model_description": "The CogAgent-9B-20241220 model is based on GLM-4V-9B, a bilingual open-source VLM base model. Through data collection and optimization, multi-stage training, and strategy improvements, CogAgent-9B-20241220 achieves significant advancements in GUI perception, inference prediction accuracy, action space completeness, and task generalizability. ",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": "9",
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "ZhipuAI/cogagent-9b-20241220",
+        "model_hub": "modelscope"
+      }
+    ],
+    "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+    "stop_token_ids": [
+      151329,
+      151336,
+      151338
+    ],
+    "stop": [
+      "<|endoftext|>",
+      "<|user|>",
+      "<|observation|>"
+    ]
   }
 ]