xinference 1.4.1__py3-none-any.whl → 1.5.0.post1__py3-none-any.whl

xinference/model/llm/transformers/intern_vl.py

@@ -19,14 +19,14 @@ from typing import Dict, Iterator, List, Optional, Union
 import torch
 
 from ....types import ChatCompletion, ChatCompletionChunk
-from ..llm_family import LLMFamilyV1, LLMSpecV1
+from ..llm_family import LLMFamilyV1, LLMSpecV1, register_transformer
 from ..utils import (
     _decode_image,
     generate_chat_completion,
     generate_completion_chunk,
     parse_messages,
 )
-from .core import PytorchChatModel, PytorchGenerateConfig
+from .core import PytorchChatModel, PytorchGenerateConfig, register_non_default_model
 from .utils import cache_clean
 
 logger = logging.getLogger(__name__)
@@ -232,6 +232,10 @@ def _load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=
     return pixel_values, num_patches_list
 
 
+@register_transformer
+@register_non_default_model(
+    "internvl-chat", "internvl2", "Internvl2.5", "Internvl2.5-MPO", "InternVL3"
+)
 class InternVLChatModel(PytorchChatModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -257,6 +261,8 @@ class InternVLChatModel(PytorchChatModel):
     def _split_model(self):
         import math
 
+        from transformers import AutoConfig
+
         device_map = {}
         world_size = torch.cuda.device_count()
         # single gpu
@@ -265,22 +271,26 @@ class InternVLChatModel(PytorchChatModel):
         model_size = f"{self.model_spec.model_size_in_billions}B"
         model_name = self.model_family.model_name.lower().replace("-mpo", "")
         model_name = f"{model_name}-{model_size}"
-        num_layers = {
-            "internvl2-1B": 24,
-            "internvl2-2B": 24,
-            "internvl2-4B": 32,
-            "internvl2-8B": 32,
-            "internvl2-26B": 48,
-            "internvl2-40B": 60,
-            "internvl2-76B": 80,
-            "internvl2.5-1B": 24,
-            "internvl2.5-2B": 24,
-            "internvl2.5-4B": 36,
-            "internvl2.5-8B": 32,
-            "internvl2.5-26B": 48,
-            "internvl2.5-38B": 64,
-            "internvl2.5-78B": 80,
-        }[model_name]
+        if "internvl3" in model_name.lower():
+            config = AutoConfig.from_pretrained(self.model_path, trust_remote_code=True)
+            num_layers = config.llm_config.num_hidden_layers
+        else:
+            num_layers = {
+                "internvl2-1B": 24,
+                "internvl2-2B": 24,
+                "internvl2-4B": 32,
+                "internvl2-8B": 32,
+                "internvl2-26B": 48,
+                "internvl2-40B": 60,
+                "internvl2-76B": 80,
+                "internvl2.5-1B": 24,
+                "internvl2.5-2B": 24,
+                "internvl2.5-4B": 36,
+                "internvl2.5-8B": 32,
+                "internvl2.5-26B": 48,
+                "internvl2.5-38B": 64,
+                "internvl2.5-78B": 80,
+            }[model_name]
 
         # Since the first GPU will be used for ViT, treat it as half a GPU.
         num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))

xinference/model/llm/transformers/minicpmv26.py

@@ -20,7 +20,12 @@ import torch
 from PIL import Image
 
 from ....core.scheduler import InferenceRequest
-from ....types import ChatCompletion, ChatCompletionChunk, CompletionChunk
+from ....types import (
+    ChatCompletion,
+    ChatCompletionChunk,
+    CompletionChunk,
+    PytorchModelConfig,
+)
 from ...utils import select_device
 from ..llm_family import LLMFamilyV1, LLMSpecV1
 from ..utils import (
@@ -52,6 +57,15 @@ class MiniCPMV26Model(PytorchChatModel):
             return True
         return False
 
+    def _sanitize_model_config(
+        self, pytorch_model_config: Optional[PytorchModelConfig]
+    ) -> PytorchModelConfig:
+        pytorch_model_config = super()._sanitize_model_config(pytorch_model_config)
+        assert pytorch_model_config is not None
+        pytorch_model_config.setdefault("min_pixels", 256 * 28 * 28)
+        pytorch_model_config.setdefault("max_pixels", 1280 * 28 * 28)
+        return pytorch_model_config
+
     def _get_model_class(self):
         from transformers import AutoModel
 
@@ -99,8 +113,13 @@ class MiniCPMV26Model(PytorchChatModel):
             self.model_path,
             trust_remote_code=True,
         )
+        min_pixels = self._pytorch_model_config.get("min_pixels")
+        max_pixels = self._pytorch_model_config.get("max_pixels")
         self._processor = AutoProcessor.from_pretrained(
-            self.model_path, trust_remote_code=True
+            self.model_path,
+            trust_remote_code=True,
+            min_pixels=min_pixels,
+            max_pixels=max_pixels,
         )
         self._device = self._model.device
         self._save_tensorizer()

xinference/model/llm/transformers/qwen-omni.py

@@ -0,0 +1,308 @@
+# Copyright 2022-2025 XProbe Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import base64
+import importlib.util
+import io
+import logging
+import sys
+import time
+import uuid
+from typing import Dict, Iterator, List, Optional, Union
+
+from ....model.utils import select_device
+from ....types import (
+    ChatCompletion,
+    ChatCompletionAudio,
+    ChatCompletionChoice,
+    ChatCompletionChunk,
+    ChatCompletionMessage,
+    CompletionChunk,
+    CompletionUsage,
+)
+from ..llm_family import LLMFamilyV1, LLMSpecV1, register_transformer
+from ..utils import generate_completion_chunk
+from .core import PytorchChatModel, PytorchGenerateConfig, register_non_default_model
+from .utils import cache_clean
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_SYSTEM_PROMPT = (
+    "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, "
+    "capable of perceiving auditory and visual inputs, as well as generating text and speech."
+)
+
+
+@register_transformer
+@register_non_default_model("qwen2.5-omni")
+class Qwen2_5OmniChatModel(PytorchChatModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        self._tokenizer = None
+        self._model = None
+        self._device = None
+        self._processor = None
+
+    @classmethod
+    def match(
+        cls, model_family: "LLMFamilyV1", model_spec: "LLMSpecV1", quantization: str
+    ) -> bool:
+        if model_spec.model_format not in ["pytorch", "gptq", "awq"]:
+            return False
+        llm_family = model_family.model_family or model_family.model_name
+        if "qwen2.5-omni".lower() in llm_family.lower():
+            return True
+        return False
+
+    def load(self):
+        from transformers import (
+            Qwen2_5OmniForConditionalGeneration,
+            Qwen2_5OmniProcessor,
+        )
+
+        device = self._pytorch_model_config.get("device", "auto")
+        device = select_device(device)
+        self._device = device
+        # for multiple GPU, set back to auto to make multiple devices work
+        device = "auto" if device == "cuda" else device
+        flash_attn_installed = importlib.util.find_spec("flash_attn") is not None
+        kwargs = (
+            {}
+            if not flash_attn_installed
+            else {"attn_implementation": "flash_attention_2"}
+        )
+        logger.debug("Loading model with extra kwargs: %s", kwargs)
+
+        self._processor = Qwen2_5OmniProcessor.from_pretrained(
+            self.model_path, trust_remote_code=True
+        )
+        self._tokenizer = self._processor.tokenizer
+        self._model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
+            self.model_path,
+            torch_dtype="auto",
+            device_map=device,
+            trust_remote_code=True,
+            **kwargs,
+        )
+
+    @cache_clean
+    def chat(
+        self,
+        messages: List[Dict],
+        generate_config: Optional[PytorchGenerateConfig] = None,
+    ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
+        messages = self._transform_messages(messages)
+
+        generate_config = generate_config if generate_config else {}
+
+        stream = generate_config.get("stream", False) if generate_config else False
+
+        if stream:
+            it = self._generate_stream(messages, generate_config)
+            return self._to_chat_completion_chunks(it)
+        else:
+            c = self._generate(messages, generate_config)
+            return c
+
+    def _transform_messages(
+        self,
+        messages: Union[List[ChatCompletionMessage], List[dict]],
+    ):
+        messages = super()._transform_messages(messages)
+        if messages[0]["role"] != "system":
+            messages.insert(
+                0,
+                {
+                    "role": "system",
+                    "content": [{"type": "text", "text": DEFAULT_SYSTEM_PROMPT}],  # type: ignore
+                },
+            )
+        else:
+            logger.debug("Force to set system prompt")
+            messages[0]["content"] = [{"type": "text", "text": DEFAULT_SYSTEM_PROMPT}]  # type: ignore
+        return messages
+
+    def _generate(
+        self, messages: List, config: PytorchGenerateConfig = {}
+    ) -> ChatCompletion:
+        import soundfile as sf
+        from qwen_omni_utils import process_mm_info
+
+        use_audio_in_video = config.get("use_audio_in_video", True)
+        voice = config.get("voice", "Chelsie")
+
+        text = self._processor.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=True
+        )
+        audios, images, videos = process_mm_info(
+            messages, use_audio_in_video=use_audio_in_video
+        )
+        logger.debug(
+            "Text, audio, image, video: %s, %s, %s, %s", text, audios, images, videos
+        )
+        inputs = self._processor(
+            text=text,
+            images=images,
+            audio=audios,
+            videos=videos,
+            padding=True,
+            return_tensors="pt",
+            use_audio_in_video=use_audio_in_video,
+        )
+        inputs = inputs.to(self._device)
+
+        # Inference: Generation of the output
+        generated_ids, audio = self._model.generate(
+            **inputs,
+            speaker=voice,
+            max_new_tokens=config.get("max_tokens", 512),
+            temperature=config.get("temperature", 1),
+            use_audio_in_video=use_audio_in_video,
+        )
+        generated_ids_trimmed = [
+            out_ids[len(in_ids) :]
+            for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+        ]
+        output_text = self._processor.batch_decode(
+            generated_ids_trimmed,
+            skip_special_tokens=True,
+            clean_up_tokenization_spaces=False,
+        )[0]
+
+        wav_io = io.BytesIO()
+        sf.write(
+            wav_io,
+            audio.reshape(-1).detach().cpu().numpy(),
+            samplerate=24000,
+            format="WAV",
+        )
+        wav_bytes = wav_io.getvalue()
+        audio_content = base64.b64encode(wav_bytes).decode()
+
+        return ChatCompletion(
+            id="chat" + str(uuid.uuid1()),
+            object="chat.completion",
+            created=int(time.time()),
+            model=self.model_uid,
+            choices=[
+                ChatCompletionChoice(
+                    index=0,
+                    message={
+                        "role": "assistant",
+                        "content": output_text,
+                        "audio": ChatCompletionAudio(
+                            id="audio" + str(uuid.uuid1()),
+                            data=audio_content,
+                            expires_at=int(time.time()),
+                            transcript="",
+                        ),
+                    },
+                    finish_reason="stop",
+                )
+            ],
+            usage=CompletionUsage(
+                prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+            ),
+        )
+
+    def _generate_stream(
+        self, messages: List, config: PytorchGenerateConfig = {}
+    ) -> Iterator[CompletionChunk]:
+        from threading import Thread
+
+        from qwen_omni_utils import process_mm_info
+        from transformers import TextIteratorStreamer
+
+        use_audio_in_video = config.get("use_audio_in_video", True)
+        voice = config.get("voice", "Chelsie")
+
+        text = self._processor.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=True
+        )
+        audios, images, videos = process_mm_info(
+            messages, use_audio_in_video=use_audio_in_video
+        )
+        logger.debug(
+            "Text, audio, image, video: %s, %s, %s, %s", text, audios, images, videos
+        )
+        inputs = self._processor(
+            text=text,
+            images=images,
+            audio=audios,
+            videos=videos,
+            padding=True,
+            return_tensors="pt",
+            use_audio_in_video=use_audio_in_video,
+        )
+        inputs = inputs.to(self._device)
+
+        tokenizer = self._tokenizer
+        streamer = TextIteratorStreamer(
+            tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True
+        )
+
+        # TODO(xuye): Cannot find a way to streaming output,
+        # will implement it when it's supported
+
+        gen_kwargs = {
+            "max_new_tokens": config.get("max_tokens", 512),
+            "temperature": config.get("temperature", 1),
+            "streamer": streamer,
+            "speaker": voice,
+            **inputs,
+        }
+        error = None
+
+        def model_generate():
+            try:
+                return self._model.generate(**gen_kwargs)
+            except Exception:
+                nonlocal error
+                error = sys.exc_info()
+                streamer.end()
+                raise
+
+        thread = Thread(target=model_generate)
+        thread.start()
+
+        completion_id = str(uuid.uuid1())
+        for new_text in streamer:
+            yield generate_completion_chunk(
+                chunk_text=new_text,
+                finish_reason=None,
+                chunk_id=completion_id,
+                model_uid=self.model_uid,
+                prompt_tokens=-1,
+                completion_tokens=-1,
+                total_tokens=-1,
+                has_choice=True,
+                has_content=True,
+            )
+
+        if error:
+            _, err, tb = error  # type: ignore
+            raise err.with_traceback(tb)
+
+        yield generate_completion_chunk(
+            chunk_text=None,
+            finish_reason="stop",
+            chunk_id=completion_id,
+            model_uid=self.model_uid,
+            prompt_tokens=-1,
+            completion_tokens=-1,
+            total_tokens=-1,
+            has_choice=True,
+            has_content=False,
+        )

xinference/model/llm/transformers/qwen2_audio.py

@@ -74,7 +74,7 @@ class Qwen2AudioChatModel(PytorchChatModel):
 
     def _transform_messages(
         self,
-        messages: List[ChatCompletionMessage],
+        messages: Union[List[ChatCompletionMessage], List[dict]],
     ):
         import librosa
 

xinference/model/llm/transformers/qwen2_vl.py

@@ -24,15 +24,18 @@ from ....types import (
     ChatCompletionChunk,
     ChatCompletionMessage,
     CompletionChunk,
+    PytorchModelConfig,
 )
-from ..llm_family import LLMFamilyV1, LLMSpecV1
+from ..llm_family import LLMFamilyV1, LLMSpecV1, register_transformer
 from ..utils import generate_chat_completion, generate_completion_chunk
-from .core import PytorchChatModel, PytorchGenerateConfig
+from .core import PytorchChatModel, PytorchGenerateConfig, register_non_default_model
 from .utils import cache_clean
 
 logger = logging.getLogger(__name__)
 
 
+@register_transformer
+@register_non_default_model("qwen2-vl-instruct", "qwen2.5-vl-instruct")
 class Qwen2VLChatModel(PytorchChatModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -41,6 +44,15 @@ class Qwen2VLChatModel(PytorchChatModel):
         self._device = None
         self._processor = None
 
+    def _sanitize_model_config(
+        self, pytorch_model_config: Optional[PytorchModelConfig]
+    ) -> PytorchModelConfig:
+        pytorch_model_config = super()._sanitize_model_config(pytorch_model_config)
+        assert pytorch_model_config is not None
+        pytorch_model_config.setdefault("min_pixels", 256 * 28 * 28)
+        pytorch_model_config.setdefault("max_pixels", 1280 * 28 * 28)
+        return pytorch_model_config
+
     @classmethod
     def match(
         cls, model_family: "LLMFamilyV1", model_spec: "LLMSpecV1", quantization: str
@@ -69,9 +81,13 @@ class Qwen2VLChatModel(PytorchChatModel):
         self._device = device
         # for multiple GPU, set back to auto to make multiple devices work
         device = "auto" if device == "cuda" else device
-
+        min_pixels = self._pytorch_model_config.get("min_pixels")
+        max_pixels = self._pytorch_model_config.get("max_pixels")
         self._processor = AutoProcessor.from_pretrained(
-            self.model_path, trust_remote_code=True
+            self.model_path,
+            trust_remote_code=True,
+            min_pixels=min_pixels,
+            max_pixels=max_pixels,
         )
         self._tokenizer = self._processor.tokenizer
         flash_attn_installed = importlib.util.find_spec("flash_attn") is not None

xinference/model/llm/utils.py

@@ -31,6 +31,7 @@ from typing import (
     List,
     Optional,
     Tuple,
+    Union,
     cast,
 )
 
@@ -762,7 +763,7 @@ class ChatModelMixin:
 
     def _transform_messages(
         self,
-        messages: List[ChatCompletionMessage],
+        messages: Union[List[ChatCompletionMessage], List[dict]],
     ):
         transformed_messages = []
         for msg in messages:
@@ -783,6 +784,15 @@ class ChatModelMixin:
                         new_content.append(
                             {"type": "video", "video": item["video_url"]["url"]}
                         )
+                    elif "audio_url" in item:
+                        new_content.append(
+                            {"type": "audio", "audio": item["audio_url"]["url"]}
+                        )
+                    else:
+                        logger.warning(
+                            "Unknown message type, message: %s, this message may be ignored",
+                            messages,
+                        )
             new_message = {"role": role, "content": new_content}
             transformed_messages.append(new_message)
 

xinference/model/llm/vllm/core.py

@@ -37,6 +37,7 @@ from typing import (
 )
 
 import xoscar as xo
+from typing_extensions import NotRequired
 
 from ....types import (
     ChatCompletion,
@@ -81,6 +82,9 @@ class VLLMModelConfig(TypedDict, total=False):
     scheduling_policy: Optional[str]
     reasoning_content: bool
     model_quantization: Optional[str]
+    mm_processor_kwargs: NotRequired[dict[str, Any]]
+    min_pixels: NotRequired[int]
+    max_pixels: NotRequired[int]
 
 
 class VLLMGenerateConfig(TypedDict, total=False):
@@ -170,6 +174,8 @@ if VLLM_INSTALLED and vllm.__version__ >= "0.3.0":
     VLLM_SUPPORTED_CHAT_MODELS.append("marco-o1")
     VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-r1-distill-qwen")
     VLLM_SUPPORTED_CHAT_MODELS.append("fin-r1")
+    VLLM_SUPPORTED_CHAT_MODELS.append("seallms-v3")
+    VLLM_SUPPORTED_CHAT_MODELS.append("skywork-or1-preview")
 
 if VLLM_INSTALLED and vllm.__version__ >= "0.3.2":
     VLLM_SUPPORTED_CHAT_MODELS.append("gemma-it")
@@ -205,6 +211,7 @@ if VLLM_INSTALLED and vllm.__version__ >= "0.6.1":
     VLLM_SUPPORTED_VISION_MODEL_LIST.append("internvl2")
     VLLM_SUPPORTED_VISION_MODEL_LIST.append("InternVL2.5")
     VLLM_SUPPORTED_VISION_MODEL_LIST.append("InternVL2.5-MPO")
+    VLLM_SUPPORTED_VISION_MODEL_LIST.append("InternVL3")
 
 if VLLM_INSTALLED and vllm.__version__ >= "0.6.2":
     VLLM_SUPPORTED_CHAT_MODELS.append("minicpm3-4b")
@@ -229,6 +236,9 @@ if VLLM_INSTALLED and vllm.__version__ >= "0.8.0":
     VLLM_SUPPORTED_CHAT_MODELS.append("gemma-3-1b-it")
     VLLM_SUPPORTED_VISION_MODEL_LIST.append("gemma-3-it")
 
+if VLLM_INSTALLED and vllm.__version__ >= "0.8.4":
+    VLLM_SUPPORTED_CHAT_MODELS.append("glm4-0414")
+
 
 class VLLMModel(LLM):
     def __init__(
@@ -531,6 +541,31 @@ class VLLMModel(LLM):
         # Add scheduling policy if vLLM version is 0.6.3 or higher
         if vllm.__version__ >= "0.6.3":
             model_config.setdefault("scheduling_policy", "fcfs")
+            # init mm_processor_kwargs params
+            mm_processor_kwargs = model_config.get("mm_processor_kwargs", {})
+            if isinstance(mm_processor_kwargs, str):
+                try:
+                    mm_processor_kwargs = json.loads(mm_processor_kwargs)
+                except json.JSONDecodeError:
+                    logger.warning(
+                        "Failed to parse mm_processor_kwargs as JSON, using default empty dict"
+                    )
+                    mm_processor_kwargs = {}
+                except Exception as e:
+                    logger.warning(
+                        f"Unexpected error parsing mm_processor_kwargs: {e}, using default empty dict"
+                    )
+                    mm_processor_kwargs = {}
+            pixel_params: Dict[str, int] = {}
+            if "min_pixels" in model_config:
+                pixel_params["min_pixels"] = model_config.pop("min_pixels")
+            if "max_pixels" in model_config:
+                pixel_params["max_pixels"] = model_config.pop("max_pixels")
+            if pixel_params or mm_processor_kwargs:
+                model_config["mm_processor_kwargs"] = {
+                    **mm_processor_kwargs,
+                    **pixel_params,
+                }
         return model_config
 
     @staticmethod

xinference/model/llm/vllm/distributed_executor.py

@@ -84,7 +84,7 @@ class WorkerWrapper:
         return await self._worker_actor_ref.execute_method(method, *args, **kwargs)
 
     def kill(self):
-        coro = xo.kill_actor(self._worker_actor_ref)
+        coro = xo.destroy_actor(self._worker_actor_ref)
         return asyncio.run_coroutine_threadsafe(coro, self._loop)
 
 
@@ -108,6 +108,7 @@ class XinferenceDistributedExecutor(DistributedExecutorBase):
         self._pool_addresses = pool_addresses
         self._loop = loop
         self._n_worker = n_worker
+        self._is_shutdown = False
         super().__init__(vllm_config, *args, **kwargs)
 
     def _init_executor(self) -> None:
@@ -247,11 +248,16 @@ class XinferenceDistributedExecutor(DistributedExecutorBase):
         return
 
     def shutdown(self) -> None:
+        if self._is_shutdown:
+            return
+
         try:
+            self._is_shutdown = True
             futs = [worker.kill() for worker in self.workers]
             _ = [fut.result() for fut in futs]
-        except (RuntimeError, ConnectionError):
+        except (RuntimeError, ConnectionError, xo.ActorNotExist):
             # event loop closed already, ignore
+            # or actor already removed
             pass
 
     def __del__(self):

xinference/model/rerank/core.py

@@ -29,7 +29,7 @@ import torch.nn as nn
 from ...constants import XINFERENCE_CACHE_DIR
 from ...device_utils import empty_cache
 from ...types import Document, DocumentObj, Rerank, RerankTokens
-from ..core import CacheableModelSpec, ModelDescription
+from ..core import CacheableModelSpec, ModelDescription, VirtualEnvSettings
 from ..utils import is_model_cached
 
 logger = logging.getLogger(__name__)
@@ -56,6 +56,7 @@ class RerankModelSpec(CacheableModelSpec):
     model_id: str
     model_revision: Optional[str]
     model_hub: str = "huggingface"
+    virtualenv: Optional[VirtualEnvSettings]
 
 
 class RerankModelDescription(ModelDescription):
@@ -69,6 +70,10 @@ class RerankModelDescription(ModelDescription):
         super().__init__(address, devices, model_path=model_path)
         self._model_spec = model_spec
 
+    @property
+    def spec(self):
+        return self._model_spec
+
     def to_dict(self):
         return {
             "model_type": "rerank",