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

xinference/model/llm/llm_family_openmind_hub.json

@@ -785,40 +785,6 @@
       "</s>"
     ]
   },
-  {
-    "version": 1,
-    "context_length": 8192,
-    "model_name": "cogvlm2",
-    "model_lang": [
-      "en",
-      "zh"
-    ],
-    "model_ability": [
-      "chat",
-      "vision"
-    ],
-    "model_description": "CogVLM2 have achieved good results in many lists compared to the previous generation of CogVLM open source models. Its excellent performance can compete with some non-open source models.",
-    "model_specs": [
-      {
-        "model_format": "pytorch",
-        "model_size_in_billions": 20,
-        "quantizations": [
-          "none"
-        ],
-        "model_id": "AI-Research/cogvlm2-llama3-chinese-chat-19b",
-        "model_hub": "openmind_hub"
-      }
-    ],
-    "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = '<|begin_of_text|>' + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% else %}{{ '<|end_of_text|>' }}{% endif %}",
-    "stop_token_ids": [
-      128001,
-      128009
-    ],
-    "stop": [
-      "<|end_of_text|>",
-      "<|eot_id|>"
-    ]
-  },
   {
     "version": 1,
     "context_length": 8192,

xinference/model/llm/sglang/core.py

@@ -107,7 +107,11 @@ SGLANG_SUPPORTED_CHAT_MODELS = [
     "deepseek-r1-distill-qwen",
     "deepseek-r1-distill-llama",
     "deepseek-v3",
+    "deepseek-v3-0324",
     "deepseek-r1",
+    "deepseek-r1-0528",
+    "deepseek-r1-0528-qwen3",
+    "deepseek-prover-v2",
     "DianJin-R1",
     "qwen3",
     "HuatuoGPT-o1-Qwen2.5",

xinference/model/llm/transformers/__init__.py

@@ -16,12 +16,33 @@
 import importlib
 import os
 import pkgutil
+from typing import Dict
 
-# Get the path of the current package
+
+def import_submodules(package_path: str, package_name: str, globals_dict: Dict) -> None:
+    """
+    Recursively import all classes in submodules and subpackages
+    """
+    for _, module_name, is_pkg in pkgutil.iter_modules([package_path]):
+        full_module_name = f"{package_name}.{module_name}"
+
+        if module_name.startswith(
+            ("_", "test_")
+        ):  # Skip the modules which start with "_" or "test_"
+            continue
+
+        module = importlib.import_module(full_module_name)
+        globals_dict[module_name] = module
+
+        # If it's a pkg, recursive processing
+        if is_pkg:
+            subpackage_path = os.path.join(package_path, module_name)
+            import_submodules(subpackage_path, full_module_name, globals_dict)
+
+
+# Get the path and name of the current package
 __path__ = [os.path.dirname(os.path.abspath(__file__))]
+__package__ = __name__
 
-# Automatically import all modules under the current package
-for _, module_name, is_pkg in pkgutil.iter_modules(__path__):
-    if not module_name.startswith("_"):  # Skip modules starting with underscore
-        module = importlib.import_module(f"{__name__}.{module_name}")
-        globals()[module_name] = module
+# Automatic import of all sub-modules and sub-packages
+import_submodules(__path__[0], __package__, globals())

xinference/model/llm/transformers/chatglm.py

@@ -22,17 +22,19 @@ import torch
 
 from ....core.scheduler import InferenceRequest
 from ....types import ChatCompletion, ChatCompletionChunk, LoRA, PytorchGenerateConfig
-from ..llm_family import LLMFamilyV1, LLMSpecV1
+from ..llm_family import LLMFamilyV1, LLMSpecV1, register_transformer
 from ..utils import (
     GLM4_TOOL_CALL_FAMILY,
     generate_chat_completion,
     generate_completion_chunk,
 )
-from .core import PytorchChatModel, PytorchModelConfig
+from .core import PytorchChatModel, PytorchModelConfig, register_non_default_model
 
 logger = logging.getLogger(__name__)
 
 
+@register_transformer
+@register_non_default_model("glm4-chat", "glm4-chat-1m")
 class ChatglmPytorchChatModel(PytorchChatModel):
     def __init__(
         self,

xinference/model/llm/transformers/core.py

@@ -16,7 +16,7 @@ import json
 import logging
 import os
 from functools import lru_cache
-from typing import Any, Dict, Iterable, Iterator, List, Optional, Tuple, Union
+from typing import Any, Dict, Iterable, Iterator, List, Optional, Set, Tuple, Union
 
 import torch
 
@@ -45,36 +45,17 @@ from ..utils import (
     QWEN_TOOL_CALL_FAMILY,
     ChatModelMixin,
 )
-from .utils import get_context_length, get_max_src_len, pad_prefill_tokens
+from .utils import (
+    _get_pad_param,
+    get_context_length,
+    get_max_src_len,
+    pad_prefill_tokens,
+)
 
 logger = logging.getLogger(__name__)
 
-NON_DEFAULT_MODEL_LIST: List[str] = [
-    "opt",
-    "glm4-chat",
-    "glm4-chat-1m",
-    "qwen-vl-chat",
-    "OmniLMM",
-    "deepseek-vl-chat",
-    "cogvlm2",
-    "cogvlm2-video-llama3-chat",
-    "MiniCPM-Llama3-V-2_5",
-    "MiniCPM-V-2.6",
-    "glm-4v",
-    "qwen2-audio",
-    "qwen2-audio-instruct",
-    "deepseek-v2",
-    "deepseek-v2-chat",
-    "deepseek-v2.5",
-    "deepseek-v2-chat-0628",
-    "glm-edge-v",
-    "QvQ-72B-Preview",
-    "cogagent",
-    "gemma-3-1b-it",
-    "gemma-3-it",
-    "Ovis2",
-    "deepseek-vl2",
-]
+# !!!!! Do not add model_name to this list, use `register_non_default_model` below instead!
+NON_DEFAULT_MODEL_LIST: List[str] = []
 
 
 # Define the decorator to support multiple names registration
@@ -551,6 +532,36 @@ class PytorchModel(LLM):
     def prepare_sanitize_generate_config(self, req: InferenceRequest):
         return self._sanitize_generate_config(req.generate_config)
 
+    def merge_kv_cache(self, past_cache, new_cache):
+        from torch.nn.functional import pad
+        from transformers import DynamicCache
+
+        _, seq_len_idx = self.get_batch_size_and_seq_len_indexes_from_kv()
+        past_seq_len = past_cache[0][0].shape[seq_len_idx]
+        new_seq_len = new_cache[0][0].shape[seq_len_idx]
+        if past_seq_len != new_seq_len:
+            padding_target = new_cache if past_seq_len > new_seq_len else past_cache
+            padding_len = abs(past_seq_len - new_seq_len)
+            pad_param = _get_pad_param(seq_len_idx, padding_len)
+            for idx in range(len(padding_target)):
+                k = padding_target.key_cache[idx]
+                v = padding_target.value_cache[idx]
+                _k = pad(k, pad_param)
+                _v = pad(v, pad_param)
+                padding_target.key_cache[idx] = _k
+                padding_target.value_cache[idx] = _v
+
+        ret_kv = DynamicCache()
+        for idx in range(len(past_cache)):
+            k1, k2 = new_cache.key_cache[idx], past_cache.key_cache[idx]
+            v1, v2 = new_cache.value_cache[idx], past_cache.value_cache[idx]
+            ret_kv.update(
+                torch.cat((k1, k2), 0).contiguous(),
+                torch.cat((v1, v2), 0).contiguous(),
+                idx,
+            )
+        return ret_kv
+
     def prepare_batch_inference(self, req_list: List[InferenceRequest]):
         # check some parameters
         for r in req_list:
@@ -642,6 +653,16 @@ class PytorchModel(LLM):
         )
         self.handle_batch_inference_results(req_list)
 
+    def build_reduced_kv_cache(self, 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 PytorchChatModel(PytorchModel, ChatModelMixin):
     def __init__(

xinference/model/llm/transformers/deepseek_v2.py

@@ -15,59 +15,16 @@ import logging
 
 import torch
 
-from ..llm_family import LLMFamilyV1, LLMSpecV1
-from .core import PytorchChatModel, PytorchModel
+from ..llm_family import LLMFamilyV1, LLMSpecV1, register_transformer
+from .core import PytorchChatModel, register_non_default_model
 
 logger = logging.getLogger(__name__)
 
 
-class DeepSeekV2PytorchModel(PytorchModel):
-    def _load_model(self, **kwargs):
-        try:
-            from transformers import (
-                AutoModelForCausalLM,
-                AutoTokenizer,
-                GenerationConfig,
-            )
-        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)}")
-
-        tokenizer = AutoTokenizer.from_pretrained(
-            self.model_path,
-            trust_remote_code=kwargs["trust_remote_code"],
-        )
-        model = AutoModelForCausalLM.from_pretrained(
-            self.model_path,
-            attn_implementation="eager",
-            torch_dtype=torch.bfloat16,
-            trust_remote_code=True,
-            device_map="auto",
-            **kwargs,
-        )
-        model.generation_config = GenerationConfig.from_pretrained(self.model_path)
-        model.generation_config.pad_token_id = model.generation_config.eos_token_id
-        return model, tokenizer
-
-    @classmethod
-    def match_json(
-        cls, llm_family: "LLMFamilyV1", llm_spec: "LLMSpecV1", quantization: str
-    ) -> bool:
-        if llm_spec.model_format != "pytorch":
-            return False
-        model_family = llm_family.model_family or llm_family.model_name
-        if "deepseek-v2" not in model_family:
-            return False
-        if "generate" not in llm_family.model_ability:
-            return False
-        return True
-
-
+@register_transformer
+@register_non_default_model(
+    "deepseek-v2-chat", "deepseek-v2.5", "deepseek-v2-chat-0628"
+)
 class DeepSeekV2PytorchChatModel(PytorchChatModel):
     def _load_model(self, **kwargs):
         try:

xinference/model/llm/transformers/gemma3.py

@@ -11,29 +11,18 @@
 # 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 logging
-import sys
-import uuid
-from typing import Iterator, List, Optional, Union
-
-from ....model.utils import select_device
-from ....types import (
-    ChatCompletion,
-    ChatCompletionChunk,
-    ChatCompletionMessage,
-    CompletionChunk,
-    PytorchModelConfig,
-)
-from ..llm_family import LLMFamilyV1, LLMSpecV1
-from ..utils import generate_chat_completion, generate_completion_chunk
-from .core import PytorchChatModel, PytorchGenerateConfig
-from .utils import cache_clean
+from typing import Dict, List, Set
+
+from ....core.scheduler import InferenceRequest
+from ..llm_family import LLMFamilyV1, LLMSpecV1, register_transformer
+from .core import PytorchChatModel, register_non_default_model
 
 logger = logging.getLogger(__name__)
 
 
+@register_transformer
+@register_non_default_model("gemma-3-1b-it")
 class Gemma3TextChatModel(PytorchChatModel):
     @classmethod
     def match_json(
@@ -46,163 +35,129 @@ class Gemma3TextChatModel(PytorchChatModel):
             return True
         return False
 
+    def _load_model(self, **kwargs):
+        import torch
+        from transformers import AutoModelForCausalLM, AutoTokenizer
 
-class Gemma3ChatModel(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_json(
-        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 "gemma-3-it".lower() in llm_family.lower():
-            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 load(self):
-        from transformers import AutoProcessor, Gemma3ForConditionalGeneration
-
-        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
-        min_pixels = self._pytorch_model_config.get("min_pixels")
-        max_pixels = self._pytorch_model_config.get("max_pixels")
-        kwargs = self.apply_bnb_quantization()
-        self._processor = AutoProcessor.from_pretrained(
+        tokenizer = AutoTokenizer.from_pretrained(
             self.model_path,
-            min_pixels=min_pixels,
-            max_pixels=max_pixels,
+            trust_remote_code=kwargs["trust_remote_code"],
+            revision=kwargs["revision"],
         )
-        self._tokenizer = self._processor.tokenizer
-        self._model = Gemma3ForConditionalGeneration.from_pretrained(
-            self.model_path, device_map="auto", torch_dtype="bfloat16", **kwargs
-        )
-
-    @cache_clean
-    def chat(
-        self,
-        messages: List[ChatCompletionMessage],  # type: ignore
-        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 _generate(
-        self, messages: List, config: PytorchGenerateConfig = {}
-    ) -> ChatCompletion:
-        inputs = self._processor.apply_chat_template(
-            messages,
-            add_generation_prompt=True,
-            tokenize=True,
-            return_dict=True,
-            return_tensors="pt",
-        ).to(self._device)
-        input_len = inputs["input_ids"].shape[-1]
-
-        generation = self._model.generate(
-            **inputs,
-            do_sample=False,
-            max_new_tokens=config.get("max_tokens", 512),
-            temperature=config.get("temperature", 1),
+        kwargs["torch_dtype"] = torch.bfloat16
+        model = AutoModelForCausalLM.from_pretrained(
+            self.model_path,
+            **kwargs,
         )
-        generation = generation[0][input_len:]
-
-        decoded = self._processor.decode(generation, skip_special_tokens=True)
-        return generate_chat_completion(self.model_uid, decoded)
+        self._device = model.device
+        return model, tokenizer
 
-    def _generate_stream(
-        self, messages: List, config: PytorchGenerateConfig = {}
-    ) -> Iterator[CompletionChunk]:
-        from threading import Thread
+    def _get_full_prompt(self, messages: List[Dict], tools, generate_config: dict):
+        return messages
 
-        from transformers import TextIteratorStreamer
+    def build_prefill_kwargs(self, prompts: List, req_list: List[InferenceRequest]):
+        """
+        Note that it is important to prepare `past_key_values` for gemma3 prefill phase
+        """
+        from transformers import HybridCache
 
-        inputs = self._processor.apply_chat_template(
-            messages,
-            add_generation_prompt=True,
+        inputs = self._tokenizer.apply_chat_template(
+            prompts,
             tokenize=True,
-            return_dict=True,
+            add_generation_prompt=True,
             return_tensors="pt",
+            return_dict=True,
+            padding=True,
         ).to(self._device)
 
-        tokenizer = self._tokenizer
-        streamer = TextIteratorStreamer(
-            tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True
+        for i, r in enumerate(req_list):
+            r.prompt_tokens = inputs["input_ids"][i].tolist()
+
+        batch_size = len(prompts)
+        max_cache_len = self.get_context_len()
+        kv = HybridCache(
+            self._model.config,
+            max_batch_size=batch_size,
+            max_cache_len=max_cache_len,
+            dtype=self._model.dtype,
+            device=self._device,
+        )
+        return {**inputs, "past_key_values": kv}
+
+    def merge_kv_cache(self, past_cache, new_cache):
+        """
+        Note that: DO NOT use the `update` func of `HybridCache`, that is unrelated to KV cache merging.
+        """
+        import torch
+        from transformers import HybridCache
+
+        max_cache_len = new_cache.max_cache_len
+        batch_size = past_cache.max_batch_size + new_cache.max_batch_size
+
+        kv_batch = HybridCache(
+            self._model.config,
+            max_batch_size=batch_size,
+            max_cache_len=max_cache_len,
+            dtype=self._model.dtype,
+            device=self._device,
         )
 
-        gen_kwargs = {"streamer": streamer, **inputs}
-        error = None
-
-        def model_generate():
-            try:
-                return self._model.generate(
-                    **gen_kwargs,
-                    max_new_tokens=config.get("max_tokens", 512),
-                    temperature=config.get("temperature", 1),
-                )
-            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,
+        new_ks = [
+            torch.cat([nk, pk], dim=0).contiguous()
+            for nk, pk in zip(new_cache.key_cache, past_cache.key_cache)
+        ]
+        new_vs = [
+            torch.cat([nv, pv], dim=0).contiguous()
+            for nv, pv in zip(new_cache.value_cache, past_cache.value_cache)
+        ]
+
+        kv_batch.key_cache.clear()
+        kv_batch.value_cache.clear()
+        kv_batch.key_cache.extend(new_ks)
+        kv_batch.value_cache.extend(new_vs)
+
+        return kv_batch
+
+    def build_decode_attention_mask(
+        self, batch_size: int, seq_length: int, reqs: List[InferenceRequest]
+    ):
+        """
+        In Gemma3's inference script, attention_mask is handled internally for decode phase.
+        """
+        return None
+
+    def build_decode_position_ids(
+        self, batch_size: int, seq_length: int, reqs: List[InferenceRequest]
+    ):
+        """
+        In Gemma3's inference script, position_ids is handled internally for decode phase.
+        """
+        return None
+
+    def build_reduced_kv_cache(self, cache, skipped_indexes: Set[int]):
+        from transformers import HybridCache
+
+        batch_slices = [
+            num for num in range(cache.max_batch_size) if num not in skipped_indexes
+        ]
+        batch_size = len(batch_slices)
+
+        kv_batch = HybridCache(
+            self._model.config,
+            max_batch_size=batch_size,
+            max_cache_len=cache.max_cache_len,
+            dtype=self._model.dtype,
+            device=self._device,
         )
+
+        ks = cache.key_cache
+        vs = cache.value_cache
+
+        new_ks = [_k[batch_slices, ::].contiguous() for _k in ks]
+        new_vs = [_v[batch_slices, ::].contiguous() for _v in vs]
+        kv_batch.key_cache.clear()
+        kv_batch.value_cache.clear()
+        kv_batch.key_cache.extend(new_ks)
+        kv_batch.value_cache.extend(new_vs)
+
+        return kv_batch

xinference/{thirdparty/omnilmm/train → model/llm/transformers/multimodal}/__init__.py

@@ -1,4 +1,4 @@
-# Copyright 2022-2023 XProbe Inc.
+# 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.