xinference 1.6.0.post1__py3-none-any.whl → 1.6.1__py3-none-any.whl

xinference/_version.py

@@ -8,11 +8,11 @@ import json
 
 version_json = '''
 {
- "date": "2025-05-17T15:09:06+0800",
+ "date": "2025-05-30T19:36:43+0800",
  "dirty": false,
  "error": null,
- "full-revisionid": "1adc5d3e5cffb2752cd3e05ca782c4cfe3c0ce57",
- "version": "1.6.0.post1"
+ "full-revisionid": "72cc5e39040bdc49981b240c2b59af998554a75f",
+ "version": "1.6.1"
 }
 '''  # END VERSION_JSON
 

xinference/client/restful/restful_client.py

@@ -1017,7 +1017,7 @@ class Client:
         model_path: Optional[str]
             Model path, if gguf format, should be the file path, otherwise, should be directory of the model.
         **kwargs:
-            Any other parameters been specified.
+            Any other parameters been specified. e.g. multimodal_projector for multimodal inference with the llama.cpp backend.
 
         Returns
         -------

xinference/conftest.py

@@ -304,10 +304,3 @@ def setup_with_auth():
             os.remove(auth_file)
         except:
             pass
-
-
-@pytest.fixture
-def set_use_xllamacpp():
-    os.environ["USE_XLLAMACPP"] = "1"
-    yield
-    del os.environ["USE_XLLAMACPP"]

xinference/core/media_interface.py

@@ -19,7 +19,7 @@ import os
 import threading
 import time
 import uuid
-from typing import Dict, List, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import gradio as gr
 import PIL.Image
@@ -463,7 +463,7 @@ class MediaInterface:
 
     def image2video_interface(self) -> "gr.Blocks":
         def image_generate_video(
-            image: "PIL.Image",
+            image: "PIL.Image.Image",
             prompt: str,
             negative_prompt: str,
             num_frames: int,
@@ -653,13 +653,14 @@ class MediaInterface:
                 with open(prompt_speech_file, "rb") as f:
                     prompt_speech_bytes = f.read()
 
+            kw: Dict[str, Any] = {}
+            if prompt_speech_bytes:
+                kw["prompt_speech"] = prompt_speech_bytes
+            if prompt_text:
+                kw["prompt_text"] = prompt_text
+
             response = model.speech(
-                input=input_text,
-                voice=voice,
-                speed=speed,
-                response_format="mp3",
-                prompt_speech=prompt_speech_bytes,
-                prompt_text=prompt_text,
+                input=input_text, voice=voice, speed=speed, response_format="mp3", **kw
             )
 
             # Write to a temp .mp3 file and return its path

xinference/core/model.py

@@ -71,12 +71,8 @@ except ImportError:
     OutOfMemoryError = _OutOfMemoryError
 
 
-XINFERENCE_BATCHING_ALLOWED_VISION_MODELS = [
-    "qwen-vl-chat",
-    "cogvlm2",
-    "glm-4v",
-    "MiniCPM-V-2.6",
-]
+# !!!!! DO NOT add model_name to this list, using `register_batching_multimodal_models` below instead.
+XINFERENCE_BATCHING_ALLOWED_VISION_MODELS = []
 
 XINFERENCE_TEXT_TO_IMAGE_BATCHING_ALLOWED_MODELS = ["FLUX.1-dev", "FLUX.1-schnell"]
 XINFERENCE_TEST_OUT_OF_MEMORY_ERROR = bool(
@@ -84,6 +80,16 @@ XINFERENCE_TEST_OUT_OF_MEMORY_ERROR = bool(
 )
 
 
+def register_batching_multimodal_models(*model_names: str):
+    def decorator(cls):
+        for name in model_names:
+            if name not in XINFERENCE_BATCHING_ALLOWED_VISION_MODELS:
+                XINFERENCE_BATCHING_ALLOWED_VISION_MODELS.append(name)
+        return cls
+
+    return decorator
+
+
 def request_limit(fn):
     """
     Used by ModelActor.
@@ -977,6 +983,7 @@ class ModelActor(xo.StatelessActor, CancelMixin):
                 response_format,
                 temperature,
                 timestamp_granularities,
+                **kwargs,
             )
         raise AttributeError(
             f"Model {self._model.model_spec} is not for creating transcriptions."

xinference/core/scheduler.py

@@ -272,15 +272,6 @@ class InferenceRequest:
         )
 
 
-def _get_valid_batch_kv_cache(cache, skipped_indexes: Set[int]):
-    batch_size = cache.key_cache[0].shape[0]
-    batch_slices = [num for num in range(batch_size) if num not in skipped_indexes]
-    for idx in range(len(cache)):
-        cache.key_cache[idx] = cache.key_cache[idx][batch_slices, ::].contiguous()
-        cache.value_cache[idx] = cache.value_cache[idx][batch_slices, ::].contiguous()
-    return cache
-
-
 class SchedulerActor(xo.StatelessActor):
     @classmethod
     def gen_uid(cls, model_uid: str, replica_id: str):
@@ -409,7 +400,7 @@ class SchedulerActor(xo.StatelessActor):
         # Some requests have been completed. Batch size needs to be reduced for kv cache.
         if stopped_batch_indexes and len(self._running_queue) > 0:
             kv_cache = self._running_queue[0].kv_cache
-            reduced_kv_cache = _get_valid_batch_kv_cache(
+            reduced_kv_cache = self._model.build_reduced_kv_cache(
                 kv_cache, stopped_batch_indexes
             )
             for r in self._running_queue:

xinference/core/worker.py

@@ -533,16 +533,6 @@ class WorkerActor(xo.StatelessActor):
                 existing_model_uids.append(rep_uid)
             if idx in self._gpu_to_embedding_model_uids:
                 existing_model_uids.extend(self._gpu_to_embedding_model_uids[idx])
-            # If user has run the vLLM model on the GPU that was forced to be specified,
-            # it is not possible to force this GPU to be allocated again
-            if idx in self._user_specified_gpu_to_model_uids:
-                for rep_uid, _ in self._user_specified_gpu_to_model_uids[idx]:
-                    is_vllm_model = await self.is_model_vllm_backend(rep_uid)
-                    if is_vllm_model:
-                        raise RuntimeError(
-                            f"User specified GPU index {idx} has been occupied with a vLLM model: {rep_uid}, "
-                            f"therefore cannot allocate GPU memory for a new model."
-                        )
 
             if existing_model_uids:
                 logger.warning(

xinference/model/audio/model_spec.json

@@ -218,13 +218,65 @@
       "batch_size_s": 300
     }
   },
+  {
+    "model_name": "paraformer-zh-hotword",
+    "model_family": "funasr",
+    "model_id": "JunHowie/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404",
+    "model_revision": "26d622993683d7b0c517ee5ec9c1c8bdde76e324",
+    "model_ability": ["audio2text"],
+    "multilingual": false,
+    "default_model_config": {
+      "vad_model": "fsmn-vad",
+      "punc_model": "ct-punc"
+    },
+    "default_transcription_config": {
+      "batch_size_s": 300
+    }
+  },
+  {
+    "model_name": "paraformer-zh-long",
+    "model_family": "funasr",
+    "model_id": "JunHowie/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
+    "model_revision": "b6d8cb81645e34056cd3dda41e5624a740587de3",
+    "model_ability": ["audio2text"],
+    "multilingual": false,
+    "default_model_config": {
+      "vad_model": "fsmn-vad",
+      "punc_model": "ct-punc"
+    },
+    "default_transcription_config": {
+      "batch_size_s": 300
+    }
+  },
+  {
+    "model_name": "paraformer-zh-spk",
+    "model_family": "funasr",
+    "model_id": "JunHowie/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn",
+    "model_revision": "36abd64af4392fe02bf76453bc86c081cf1ca6da",
+    "model_ability": ["audio2text"],
+    "multilingual": false,
+    "default_model_config": {
+      "vad_model": "fsmn-vad",
+      "punc_model": "ct-punc"
+    },
+    "default_transcription_config": {
+      "batch_size_s": 300
+    }
+  },   
   {
     "model_name": "ChatTTS",
     "model_family": "ChatTTS",
     "model_id": "2Noise/ChatTTS",
     "model_revision": "1a3c04a8b0651689bd9242fbb55b1f4b5a9aef84",
     "model_ability": ["text2audio"],
-    "multilingual": true
+    "multilingual": true,
+    "virtualenv": {
+      "packages": [
+        "ChatTTS>=0.2.1",
+        "#system_torch#",
+        "#system_numpy#"
+      ]
+    }
   },
   {
     "model_name": "CosyVoice-300M",

xinference/model/audio/model_spec_modelscope.json

@@ -51,6 +51,55 @@
     "model_name": "paraformer-zh",
     "model_family": "funasr",
     "model_hub": "modelscope",
+    "model_id": "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
+    "model_revision": "master",
+    "model_ability": ["audio2text"],
+    "multilingual": false,
+    "default_model_config": {
+      "vad_model": "fsmn-vad",
+      "punc_model": "ct-punc"
+    },
+    "default_transcription_config": {
+      "batch_size_s": 300
+    }
+  },
+  {
+    "model_name": "paraformer-zh-hotword",
+    "model_family": "funasr",
+    "model_hub": "modelscope",
+    "model_id": "iic/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404",
+    "model_revision": "master",
+    "model_ability": ["audio2text"],
+    "multilingual": false,
+    "default_model_config": {
+      "vad_model": "fsmn-vad",
+      "punc_model": "ct-punc"
+    },
+    "default_transcription_config": {
+      "hotword": "",
+      "batch_size_s": 300
+    }
+  },
+    {
+    "model_name": "paraformer-zh-long",
+    "model_family": "funasr",
+    "model_hub": "modelscope",
+    "model_id": "iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
+    "model_revision": "master",
+    "model_ability": ["audio2text"],
+    "multilingual": false,
+    "default_model_config": {
+      "vad_model": "fsmn-vad",
+      "punc_model": "ct-punc"
+    },
+    "default_transcription_config": {
+      "batch_size_s": 300
+    }
+  },
+  {
+    "model_name": "paraformer-zh-spk",
+    "model_family": "funasr",
+    "model_hub": "modelscope",
     "model_id": "iic/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn",
     "model_revision": "master",
     "model_ability": ["audio2text"],
@@ -70,7 +119,14 @@
     "model_id": "AI-ModelScope/ChatTTS",
     "model_revision": "master",
     "model_ability": ["text2audio"],
-    "multilingual": true
+    "multilingual": true,
+    "virtualenv": {
+      "packages": [
+        "ChatTTS>=0.2.1",
+        "#system_torch#",
+        "#system_numpy#"
+      ]
+    }
   },
   {
     "model_name": "CosyVoice-300M",

xinference/model/embedding/core.py

@@ -651,19 +651,27 @@ class EmbeddingModel:
                 img = Image.open(image_data)
                 return img
 
-            objs: list[dict[str, str]] = []
-            for item in sentences:
-                if isinstance(item, dict):
-                    if item.get("text") is not None:
-                        objs.append(item["text"])
-                    elif item.get("image") is not None:
-                        if re.match(r"^data:image/.+;base64,", item["image"]):
-                            image = base64_to_image(item["image"])
-                            objs.append(image)
+            objs: list[str] = []
+            if isinstance(sentences, str):
+                objs.append(sentences)
+            else:
+                for item in sentences:
+                    if isinstance(item, dict):
+                        if item.get("text") is not None:
+                            objs.append(item["text"])
+                        elif item.get("image") is not None:
+                            if re.match(r"^data:image/.+;base64,", item["image"]):
+                                image = base64_to_image(item["image"])
+                                objs.append(image)
+                            else:
+                                objs.append(item["image"])
                         else:
-                            objs.append(item["image"])
+                            raise ValueError("Please check the input data.")
+                    elif isinstance(item, str):
+                        objs.append(item)
                     else:
-                        logger.error("Please check the input data.")
+                        raise ValueError("Please check the input data.")
+
             all_embeddings, all_token_nums = encode(
                 self._model,
                 objs,

xinference/model/image/model_spec.json

@@ -303,7 +303,16 @@
     "model_ability": [
       "text2image",
       "image2image"
-    ]
+    ],
+    "default_model_config": {
+      "variant": "fp16"
+    },
+    "virtualenv": {
+      "packages": [
+        "diffusers>=0.30.0",
+        "#system_numpy#"
+      ]
+    }
   },
   {
     "model_name": "stable-diffusion-inpainting",

xinference/model/image/model_spec_modelscope.json

@@ -307,6 +307,26 @@
       }
     ]
   },
+  {
+    "model_name": "kolors",
+    "model_family": "stable_diffusion",
+    "model_hub": "modelscope",
+    "model_id": "JunHowie/Kolors-diffusers",
+    "model_revision": "master",
+    "model_ability": [
+      "text2image",
+      "image2image"
+    ],
+    "default_model_config": {
+      "variant": "fp16"
+    },
+    "virtualenv": {
+      "packages": [
+        "diffusers>=0.30.0",
+        "#system_numpy#"
+      ]
+    }
+  },
   {
     "model_name": "GOT-OCR2_0",
     "model_family": "ocr",

xinference/model/llm/__init__.py

@@ -73,7 +73,7 @@ def generate_engine_config_by_model_family(model_family):
         model_size_in_billions = spec.model_size_in_billions
         quantizations = spec.quantizations
         for quantization in quantizations:
-            # traverse all supported engines to match the name, format, size in billions and quatization of model
+            # traverse all supported engines to match the name, format, size in billions and quantization of model
             for engine in SUPPORTED_ENGINES:
                 if not check_format_with_engine(
                     model_format, engine
@@ -107,6 +107,10 @@ def generate_engine_config_by_model_family(model_family):
                                     "llm_class": cls,
                                 }
                             )
+                            if hasattr(spec, "multimodal_projectors"):
+                                engine_params[-1][
+                                    "multimodal_projectors"
+                                ] = spec.multimodal_projectors
                         engines[engine] = engine_params
                         break
     LLM_ENGINES[model_name] = engines
@@ -163,36 +167,9 @@ def _install():
     from .lmdeploy.core import LMDeployChatModel, LMDeployModel
     from .mlx.core import MLXChatModel, MLXModel, MLXVisionModel
     from .sglang.core import SGLANGChatModel, SGLANGModel, SGLANGVisionModel
-    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
-    from .transformers.deepseek_v2 import (
-        DeepSeekV2PytorchChatModel,
-        DeepSeekV2PytorchModel,
-    )
-    from .transformers.deepseek_vl import DeepSeekVLChatModel
-    from .transformers.deepseek_vl2 import DeepSeekVL2ChatModel
-    from .transformers.gemma3 import Gemma3ChatModel, Gemma3TextChatModel
-    from .transformers.glm4v import Glm4VModel
-    from .transformers.glm_edge_v import GlmEdgeVModel
-    from .transformers.minicpmv25 import MiniCPMV25Model
-    from .transformers.minicpmv26 import MiniCPMV26Model
-    from .transformers.opt import OptPytorchModel
-    from .transformers.ovis2 import Ovis2ChatModel
-    from .transformers.qwen2_audio import Qwen2AudioChatModel
-    from .transformers.qwen_vl import QwenVLChatModel
     from .vllm.core import VLLMChatModel, VLLMModel, VLLMVisionModel
 
-    try:
-        from .transformers.omnilmm import OmniLMMModel
-    except ImportError as e:
-        # For quite old transformers version,
-        # import will generate error
-        OmniLMMModel = None
-        warnings.warn(f"Cannot import OmniLLMModel due to reason: {e}")
-
     # register llm classes.
     LLAMA_CLASSES.extend(
         [
@@ -203,32 +180,7 @@ def _install():
     VLLM_CLASSES.extend([VLLMModel, VLLMChatModel, VLLMVisionModel])
     MLX_CLASSES.extend([MLXModel, MLXChatModel, MLXVisionModel])
     LMDEPLOY_CLASSES.extend([LMDeployModel, LMDeployChatModel])
-    TRANSFORMERS_CLASSES.extend(
-        [
-            ChatglmPytorchChatModel,
-            PytorchChatModel,
-            QwenVLChatModel,
-            Qwen2AudioChatModel,
-            DeepSeekVLChatModel,
-            DeepSeekVL2ChatModel,
-            PytorchModel,
-            CogVLM2Model,
-            CogVLM2VideoModel,
-            MiniCPMV25Model,
-            MiniCPMV26Model,
-            Glm4VModel,
-            DeepSeekV2PytorchModel,
-            DeepSeekV2PytorchChatModel,
-            OptPytorchModel,
-            GlmEdgeVModel,
-            CogAgentChatModel,
-            Gemma3TextChatModel,
-            Gemma3ChatModel,
-            Ovis2ChatModel,
-        ]
-    )
-    if OmniLMMModel:  # type: ignore
-        TRANSFORMERS_CLASSES.append(OmniLMMModel)
+    TRANSFORMERS_CLASSES.extend([PytorchChatModel, PytorchModel])
 
     # support 4 engines for now
     SUPPORTED_ENGINES["vLLM"] = VLLM_CLASSES

xinference/model/llm/core.py

@@ -160,12 +160,14 @@ class LLMDescription(ModelDescription):
         llm_family: "LLMFamilyV1",
         llm_spec: "LLMSpecV1",
         quantization: Optional[str],
+        multimodal_projector: Optional[str] = None,
         model_path: Optional[str] = None,
     ):
         super().__init__(address, devices, model_path=model_path)
         self._llm_family = llm_family
         self._llm_spec = llm_spec
         self._quantization = quantization
+        self._multimodal_projector = multimodal_projector
 
     @property
     def spec(self):
@@ -185,6 +187,7 @@ class LLMDescription(ModelDescription):
             "model_family": self._llm_family.model_family
             or self._llm_family.model_name,
             "quantization": self._quantization,
+            "multimodal_projector": self._multimodal_projector,
             "model_hub": self._llm_spec.model_hub,
             "revision": self._llm_spec.model_revision,
             "context_length": self._llm_family.context_length,
@@ -204,6 +207,7 @@ class LLMDescription(ModelDescription):
             "model_file_location": model_file_location,
             "cache_status": cache_status,
             "quantization": self._quantization,
+            "multimodal_projector": self._multimodal_projector,
             "model_format": self._llm_spec.model_format,
             "model_size_in_billions": self._llm_spec.model_size_in_billions,
         }
@@ -212,10 +216,19 @@ class LLMDescription(ModelDescription):
 def generate_llm_description(llm_family: "LLMFamilyV1") -> Dict[str, List[Dict]]:
     res = defaultdict(list)
     for spec in llm_family.model_specs:
+        multimodal_projectors = getattr(spec, "multimodal_projectors", None)
         for q in spec.quantizations:
-            res[llm_family.model_name].append(
-                LLMDescription(None, None, llm_family, spec, q).to_version_info()
-            )
+            if multimodal_projectors:
+                for mmproj in multimodal_projectors:
+                    res[llm_family.model_name].append(
+                        LLMDescription(
+                            None, None, llm_family, spec, q, mmproj
+                        ).to_version_info()
+                    )
+            else:
+                res[llm_family.model_name].append(
+                    LLMDescription(None, None, llm_family, spec, q).to_version_info()
+                )
     return res
 
 
@@ -260,8 +273,9 @@ def create_llm_model_instance(
     )
     logger.debug(f"Launching {model_uid} with {llm_cls.__name__}")
 
+    multimodal_projector = kwargs.get("multimodal_projector")
     if not model_path:
-        model_path = cache(llm_family, llm_spec, quantization)
+        model_path = cache(llm_family, llm_spec, quantization, multimodal_projector)
 
     peft_model = peft_model_config.peft_model if peft_model_config else None
     if peft_model is not None:
@@ -288,5 +302,5 @@ def create_llm_model_instance(
             model_uid, llm_family, llm_spec, quantization, model_path, kwargs
         )
     return model, LLMDescription(
-        subpool_addr, devices, llm_family, llm_spec, quantization
+        subpool_addr, devices, llm_family, llm_spec, quantization, multimodal_projector
     )