xinference 0.11.3__py3-none-any.whl → 0.12.0__py3-none-any.whl

xinference/model/llm/llm_family_modelscope.json

@@ -522,6 +522,142 @@
       ]
     }
   },
+  {
+    "version": 1,
+    "context_length": 131072,
+    "model_name": "glm4-chat",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "tools"
+    ],
+    "model_description": "GLM4 is the open source version of the latest generation of pre-trained models in the GLM-4 series launched by Zhipu AI.",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 9,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_hub": "modelscope",
+        "model_id": "ZhipuAI/glm-4-9b-chat",
+        "model_revision": "master"
+      }
+    ],
+    "prompt_style": {
+      "style_name": "CHATGLM3",
+      "system_prompt": "",
+      "roles": [
+        "user",
+        "assistant"
+      ],
+      "stop_token_ids": [
+        151329,
+        151336,
+        151338
+      ],
+      "stop": [
+        "<|endoftext|>",
+        "<|user|>",
+        "<|observation|>"
+      ]
+    }
+  },
+  {
+    "version": 1,
+    "context_length": 1048576,
+    "model_name": "glm4-chat-1m",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "tools"
+    ],
+    "model_description": "GLM4 is the open source version of the latest generation of pre-trained models in the GLM-4 series launched by Zhipu AI.",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 9,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_hub": "modelscope",
+        "model_id": "ZhipuAI/glm-4-9b-chat-1m",
+        "model_revision": "master"
+      }
+    ],
+    "prompt_style": {
+      "style_name": "CHATGLM3",
+      "system_prompt": "",
+      "roles": [
+        "user",
+        "assistant"
+      ],
+      "stop_token_ids": [
+        151329,
+        151336,
+        151338
+      ],
+      "stop": [
+        "<|endoftext|>",
+        "<|user|>",
+        "<|observation|>"
+      ]
+    }
+  },
+  {
+    "version": 1,
+    "context_length": 8192,
+    "model_name": "glm-4v",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "vision"
+    ],
+    "model_description": "GLM4 is the open source version of the latest generation of pre-trained models in the GLM-4 series launched by Zhipu AI.",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 9,
+        "quantizations": [
+          "none"
+        ],
+        "model_hub": "modelscope",
+        "model_id": "ZhipuAI/glm-4v-9b",
+        "model_revision": "master"
+      }
+    ],
+    "prompt_style": {
+      "style_name": "CHATGLM3",
+      "system_prompt": "",
+      "roles": [
+        "user",
+        "assistant"
+      ],
+      "stop_token_ids": [
+        151329,
+        151336,
+        151338
+      ],
+      "stop": [
+        "<|endoftext|>",
+        "<|user|>",
+        "<|observation|>"
+      ]
+    }
+  },
   {
     "version": 1,
     "context_length": 2048,
@@ -2648,6 +2784,233 @@
       ]
     }
   },
+  {
+    "version": 1,
+    "context_length": 32768,
+    "model_name": "qwen2-instruct",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat",
+      "tools"
+    ],
+    "model_description": "Qwen2 is the new series of Qwen large language models",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": "0_5",
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "qwen/Qwen2-0.5B-Instruct",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": "1_5",
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "qwen/Qwen2-1.5B-Instruct",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 7,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "qwen/Qwen2-7B-Instruct",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 72,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "qwen/Qwen2-72B-Instruct",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "gptq",
+        "model_size_in_billions": "0_5",
+        "quantizations": [
+          "Int4",
+          "Int8"
+        ],
+        "model_id": "qwen/Qwen2-0.5B-Instruct-GPTQ-{quantization}",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "gptq",
+        "model_size_in_billions": "1_5",
+        "quantizations": [
+          "Int4",
+          "Int8"
+        ],
+        "model_id": "qwen/Qwen2-1.5B-Instruct-GPTQ-{quantization}",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "gptq",
+        "model_size_in_billions": 7,
+        "quantizations": [
+          "Int4",
+          "Int8"
+        ],
+        "model_id": "qwen/Qwen2-7B-Instruct-GPTQ-{quantization}",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "gptq",
+        "model_size_in_billions": 72,
+        "quantizations": [
+          "Int4",
+          "Int8"
+        ],
+        "model_id": "qwen/Qwen2-72B-Instruct-GPTQ-{quantization}",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "awq",
+        "model_size_in_billions": "0_5",
+        "quantizations": [
+          "Int4"
+        ],
+        "model_id": "qwen/Qwen2-0.5B-Instruct-AWQ",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "awq",
+        "model_size_in_billions": "1_5",
+        "quantizations": [
+          "Int4"
+        ],
+        "model_id": "qwen/Qwen2-1.5B-Instruct-AWQ",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "awq",
+        "model_size_in_billions": 7,
+        "quantizations": [
+          "Int4"
+        ],
+        "model_id": "qwen/Qwen2-7B-Instruct-AWQ",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "awq",
+        "model_size_in_billions": 72,
+        "quantizations": [
+          "Int4"
+        ],
+        "model_id": "qwen/Qwen2-72B-Instruct-AWQ",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "ggufv2",
+        "model_size_in_billions": "0_5",
+        "quantizations": [
+          "q2_k",
+          "q3_k_m",
+          "q4_0",
+          "q4_k_m",
+          "q5_0",
+          "q5_k_m",
+          "q6_k",
+          "q8_0",
+          "fp16"
+        ],
+        "model_id": "qwen/Qwen2-0.5B-Instruct-GGUF",
+        "model_file_name_template": "qwen2-0_5b-instruct-{quantization}.gguf",
+        "model_hub": "modelscope"
+      }
+    ],
+    "prompt_style": {
+      "style_name": "QWEN",
+      "system_prompt": "You are a helpful assistant.",
+      "roles": [
+        "user",
+        "assistant"
+      ],
+      "intra_message_sep": "\n",
+      "stop_token_ids": [
+        151643,
+        151644,
+        151645
+      ],
+      "stop": [
+        "<|endoftext|>",
+        "<|im_start|>",
+        "<|im_end|>"
+      ]
+    }
+  },
+  {
+    "version": 1,
+    "context_length": 32768,
+    "model_name": "qwen2-moe-instruct",
+    "model_lang": [
+      "en",
+      "zh"
+    ],
+    "model_ability": [
+      "chat"
+    ],
+    "model_description": "Qwen2 is the new series of Qwen large language models. ",
+    "model_specs": [
+      {
+        "model_format": "pytorch",
+        "model_size_in_billions": 14,
+        "quantizations": [
+          "4-bit",
+          "8-bit",
+          "none"
+        ],
+        "model_id": "qwen/Qwen2-57B-A14B-Instruct",
+        "model_hub": "modelscope"
+      },
+      {
+        "model_format": "gptq",
+        "model_size_in_billions": 14,
+        "quantizations": [
+          "Int4"
+        ],
+        "model_id": "qwen/Qwen2-57B-A14B-Instruct-GPTQ-Int4",
+        "model_hub": "modelscope"
+      }
+    ],
+    "prompt_style": {
+      "style_name": "QWEN",
+      "system_prompt": "You are a helpful assistant.",
+      "roles": [
+        "user",
+        "assistant"
+      ],
+      "intra_message_sep": "\n",
+      "stop_token_ids": [
+        151643,
+        151644,
+        151645
+      ],
+      "stop": [
+        "<|endoftext|>",
+        "<|im_start|>",
+        "<|im_end|>"
+      ]
+    }
+  },
   {
     "version": 1,
     "context_length": 4096,
@@ -3236,7 +3599,7 @@
       "chat",
       "vision"
     ],
-    "model_description":"mniLMM is a family of open-source large multimodal models (LMMs) adept at vision & language modeling.",
+    "model_description":"OmniLMM is a family of open-source large multimodal models (LMMs) adept at vision & language modeling.",
     "model_specs":[
       {
         "model_format":"pytorch",
@@ -3468,6 +3831,50 @@
       ]
     }
   },
+  {
+    "version":1,
+    "context_length":2048,
+    "model_name":"MiniCPM-Llama3-V-2_5",
+    "model_lang":[
+      "en",
+      "zh"
+    ],
+    "model_ability":[
+      "chat",
+      "vision"
+    ],
+    "model_description":"MiniCPM-Llama3-V 2.5 is the latest model in the MiniCPM-V series. The model is built on SigLip-400M and Llama3-8B-Instruct with a total of 8B parameters.",
+    "model_specs":[
+      {
+        "model_format":"pytorch",
+        "model_size_in_billions":8,
+        "quantizations":[
+          "none"
+        ],
+        "model_hub": "modelscope",
+        "model_id":"OpenBMB/MiniCPM-Llama3-V-2_5",
+        "model_revision":"master"
+      },
+      {
+        "model_format":"pytorch",
+        "model_size_in_billions":8,
+        "quantizations":[
+          "int4"
+        ],
+        "model_hub": "modelscope",
+        "model_id":"OpenBMB/MiniCPM-Llama3-V-2_5-{quantization}",
+        "model_revision":"master"
+      }
+    ],
+    "prompt_style":{
+      "style_name":"OmniLMM",
+      "system_prompt":"The role of first msg should be user",
+      "roles":[
+        "user",
+        "assistant"
+      ]
+    }
+  },
   {
     "version": 1,
     "context_length": 2048,
@@ -3860,7 +4267,7 @@
             "<|im_end|>"
         ]
     }
-},
+  },
   {
     "version": 1,
     "context_length": 8192,

xinference/model/llm/pytorch/chatglm.py

@@ -82,7 +82,8 @@ class ChatglmPytorchChatModel(PytorchChatModel):
     ) -> bool:
         if llm_spec.model_format != "pytorch":
             return False
-        if "chatglm" not in llm_family.model_name:
+        model_family = llm_family.model_family or llm_family.model_name
+        if "chatglm" not in model_family and "glm4" not in model_family:
             return False
         if "chat" not in llm_family.model_ability:
             return False

xinference/model/llm/pytorch/cogvlm2.py

@@ -30,6 +30,7 @@ from ....types import (
     ChatCompletionMessage,
     Completion,
     CompletionChoice,
+    CompletionChunk,
     CompletionUsage,
 )
 from ..llm_family import LLMFamilyV1, LLMSpecV1
@@ -183,10 +184,7 @@ class CogVLM2Model(PytorchChatModel):
         generate_config: Optional[PytorchGenerateConfig] = None,
     ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
         system_prompt = system_prompt if system_prompt else ""
-        if generate_config and generate_config.get("stream"):
-            raise Exception(
-                f"Chat with model {self.model_family.model_name} does not support stream."
-            )
+        stream = generate_config.get("stream", False) if generate_config else False
 
         sanitized_config = {
             "pad_token_id": 128002,
@@ -234,24 +232,85 @@ class CogVLM2Model(PytorchChatModel):
             if image is not None
             else None,
         }
-        with torch.no_grad():
-            outputs = self._model.generate(**inputs, **sanitized_config)
-            outputs = outputs[:, inputs["input_ids"].shape[1] :]
-            response = self._tokenizer.decode(outputs[0])
-            response = response.split("<|end_of_text|>")[0]
 
-        chunk = Completion(
-            id=str(uuid.uuid1()),
+        if stream:
+            it = self._streaming_chat_response(inputs, sanitized_config)
+            return self._to_chat_completion_chunks(it)
+        else:
+            with torch.no_grad():
+                outputs = self._model.generate(**inputs, **sanitized_config)
+                outputs = outputs[:, inputs["input_ids"].shape[1] :]
+                response = self._tokenizer.decode(outputs[0])
+                response = response.split("<|end_of_text|>")[0]
+
+            chunk = Completion(
+                id=str(uuid.uuid1()),
+                object="text_completion",
+                created=int(time.time()),
+                model=self.model_uid,
+                choices=[
+                    CompletionChoice(
+                        index=0, text=response, finish_reason="stop", logprobs=None
+                    )
+                ],
+                usage=CompletionUsage(
+                    prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+                ),
+            )
+            return self._to_chat_completion(chunk)
+
+    def _streaming_chat_response(
+        self, inputs: Dict, config: Dict
+    ) -> Iterator[CompletionChunk]:
+        from threading import Thread
+
+        from transformers import TextIteratorStreamer
+
+        streamer = TextIteratorStreamer(
+            self._tokenizer, skip_prompt=True, skip_special_tokens=True
+        )
+        generation_kwargs = {
+            "input_ids": inputs["input_ids"],
+            "attention_mask": inputs["attention_mask"],
+            "token_type_ids": inputs["token_type_ids"],
+            "images": inputs["images"],
+            "max_new_tokens": config["max_new_tokens"],
+            "pad_token_id": config["pad_token_id"],
+            "streamer": streamer,
+        }
+
+        thread = Thread(target=self._model.generate, kwargs=generation_kwargs)
+        thread.start()
+
+        completion_id = str(uuid.uuid1())
+        for new_text in streamer:
+            chunk = CompletionChunk(
+                id=completion_id,
+                object="text_completion",
+                created=int(time.time()),
+                model=self.model_uid,
+                choices=[
+                    CompletionChoice(
+                        index=0, text=new_text, finish_reason=None, logprobs=None
+                    )
+                ],
+                usage=CompletionUsage(
+                    prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+                ),
+            )
+            yield chunk
+
+        completion_choice = CompletionChoice(
+            text="", index=0, logprobs=None, finish_reason="stop"
+        )
+        chunk = CompletionChunk(
+            id=completion_id,
             object="text_completion",
             created=int(time.time()),
             model=self.model_uid,
-            choices=[
-                CompletionChoice(
-                    index=0, text=response, finish_reason="stop", logprobs=None
-                )
-            ],
+            choices=[completion_choice],
             usage=CompletionUsage(
                 prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
             ),
         )
-        return self._to_chat_completion(chunk)
+        yield chunk