chore(docs): update generated model docs

Author: qinxuye

doc/source/getting_started/installation.rst

@@ -65,7 +65,7 @@ Currently, supported models include:
 
 .. vllm_start
 
-- ``code-llama``, ``code-llama-instruct``, ``code-llama-python``, ``deepseek``, ``deepseek-chat``, ``deepseek-coder``, ``deepseek-coder-instruct``, ``deepseek-r1-distill-llama``, ``gorilla-openfunctions-v2``, ``HuatuoGPT-o1-LLaMA-3.1``, ``llama-2``, ``llama-2-chat``, ``llama-3``, ``llama-3-instruct``, ``llama-3.1``, ``llama-3.1-instruct``, ``llama-3.3-instruct``, ``tiny-llama``, ``wizardcoder-python-v1.0``, ``wizardmath-v1.0``, ``Yi``, ``Yi-1.5``, ``Yi-1.5-chat``, ``Yi-1.5-chat-16k``, ``Yi-200k``, ``Yi-chat``
+- ``code-llama``, ``code-llama-instruct``, ``code-llama-python``, ``deepseek``, ``deepseek-chat``, ``deepseek-coder``, ``deepseek-coder-instruct``, ``deepseek-r1-distill-llama``, ``gorilla-openfunctions-v2``, ``HuatuoGPT-o1-LLaMA-3.1``, ``llama-2``, ``llama-2-chat``, ``llama-3``, ``llama-3-instruct``, ``llama-3.1``, ``llama-3.1-instruct``, ``llama-3.3-instruct``, ``minicpm5-1b``, ``tiny-llama``, ``wizardcoder-python-v1.0``, ``wizardmath-v1.0``, ``Yi``, ``Yi-1.5``, ``Yi-1.5-chat``, ``Yi-1.5-chat-16k``, ``Yi-200k``, ``Yi-chat``
 - ``codestral-v0.1``, ``mistral-instruct-v0.1``, ``mistral-instruct-v0.2``, ``mistral-instruct-v0.3``, ``mistral-large-instruct``, ``mistral-nemo-instruct``, ``mistral-v0.1``, ``openhermes-2.5``, ``seallm_v2``
 - ``Baichuan-M2``, ``codeqwen1.5``, ``codeqwen1.5-chat``, ``deepseek-r1-distill-qwen``, ``DianJin-R1``, ``fin-r1``, ``HuatuoGPT-o1-Qwen2.5``, ``KAT-V1``, ``marco-o1``, ``qwen1.5-chat``, ``qwen2-instruct``, ``qwen2.5``, ``qwen2.5-coder``, ``qwen2.5-coder-instruct``, ``qwen2.5-instruct``, ``qwen2.5-instruct-1m``, ``qwenLong-l1``, ``QwQ-32B``, ``QwQ-32B-Preview``, ``seallms-v3``, ``skywork-or1``, ``skywork-or1-preview``, ``XiYanSQL-QwenCoder-2504``
 - ``llama-3.2-vision``, ``llama-3.2-vision-instruct``
@@ -98,6 +98,7 @@ Currently, supported models include:
 - ``MiniMax-M2``, ``MiniMax-M2.5``, ``MiniMax-M2.7``
 - ``GLM-4.7-Flash``
 - ``glm-5``, ``glm-5.1``
+- ``DeepSeek-V4-Flash``, ``DeepSeek-V4-Pro``
 .. vllm_end
 
 To install Xinference and vLLM::

doc/source/models/builtin/llm/deepseek-v4-flash.rst

@@ -20,7 +20,7 @@ Model Spec 1 (pytorch, 284 Billion)
 - **Model Format:** pytorch
 - **Model Size (in billions):** 284
 - **Quantizations:** none
-- **Engines**: Transformers
+- **Engines**: vLLM, Transformers
 - **Model ID:** deepseek-ai/DeepSeek-V4-Flash
 - **Model Hubs**:  `Hugging Face <https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash>`__, `ModelScope <https://modelscope.cn/models/deepseek-ai/DeepSeek-V4-Flash>`__
 

doc/source/models/builtin/llm/deepseek-v4-pro.rst

@@ -20,7 +20,7 @@ Model Spec 1 (pytorch, 1600 Billion)
 - **Model Format:** pytorch
 - **Model Size (in billions):** 1600
 - **Quantizations:** none
-- **Engines**: Transformers
+- **Engines**: vLLM, Transformers
 - **Model ID:** deepseek-ai/DeepSeek-V4-Pro
 - **Model Hubs**:  `Hugging Face <https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro>`__, `ModelScope <https://modelscope.cn/models/deepseek-ai/DeepSeek-V4-Pro>`__
 

doc/source/models/builtin/llm/dianjin-r1.rst

@@ -7,7 +7,7 @@ DianJin-R1
 - **Context Length:** 32768
 - **Model Name:** DianJin-R1
 - **Languages:** en, zh
-- **Abilities:** chat, tools
+- **Abilities:** chat, reasoning, hybrid, tools
 - **Description:** Tongyi DianJin is a financial intelligence solution platform built by Alibaba Cloud, dedicated to providing financial business developers with a convenient artificial intelligence application development environment.
 
 Specifications

doc/source/models/builtin/llm/index.rst

@@ -187,7 +187,7 @@ The following is a list of built-in LLM in Xinference:
      - DeepSeek-VL2, an advanced series of large Mixture-of-Experts (MoE) Vision-Language Models that significantly improves upon its predecessor, DeepSeek-VL. DeepSeek-VL2 demonstrates superior capabilities across various tasks, including but not limited to visual question answering, optical character recognition, document/table/chart understanding, and visual grounding.
 
    * - :ref:`dianjin-r1 <models_llm_dianjin-r1>`
-     - chat, tools
+     - chat, reasoning, hybrid, tools
      - 32768
      - Tongyi DianJin is a financial intelligence solution platform built by Alibaba Cloud, dedicated to providing financial business developers with a convenient artificial intelligence application development environment.
 
@@ -426,6 +426,11 @@ The following is a list of built-in LLM in Xinference:
      - 32768
      - MiniCPM4 series are highly efficient large language models (LLMs) designed explicitly for end-side devices, which achieves this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems.
 
+   * - :ref:`minicpm5-1b <models_llm_minicpm5-1b>`
+     - chat, reasoning, hybrid, tools
+     - 131072
+     - MiniCPM5-1B is the first model in the MiniCPM5 series. It is a dense 1B Transformer built for on-device, local deployment, and resource-constrained scenarios, reaching 1B-class open-source SOTA. Supports hybrid thinking via enable_thinking and native XML-style tool calling (MCP-compatible).
+
    * - :ref:`minimax-m2 <models_llm_minimax-m2>`
      - chat, tools, reasoning
      - 196608
@@ -785,7 +790,7 @@ The following is a list of built-in LLM in Xinference:
 .. toctree::
    :maxdepth: 3
 
-  
+
    baichuan-2
 
    baichuan-2-chat
@@ -950,6 +955,8 @@ The following is a list of built-in LLM in Xinference:
 
    minicpm4
 
+   minicpm5-1b
+
    minimax-m2
 
    minimax-m2.5
@@ -1092,4 +1099,3 @@ The following is a list of built-in LLM in Xinference:
 
    yi-chat
 
-

doc/source/models/builtin/llm/minicpm5-1b.rst

@@ -0,0 +1,62 @@
+.. _models_llm_minicpm5-1b:
+
+========================================
+minicpm5-1b
+========================================
+
+- **Context Length:** 131072
+- **Model Name:** minicpm5-1b
+- **Languages:** en, zh
+- **Abilities:** chat, reasoning, hybrid, tools
+- **Description:** MiniCPM5-1B is the first model in the MiniCPM5 series. It is a dense 1B Transformer built for on-device, local deployment, and resource-constrained scenarios, reaching 1B-class open-source SOTA. Supports hybrid thinking via enable_thinking and native XML-style tool calling (MCP-compatible).
+
+Specifications
+^^^^^^^^^^^^^^
+
+
+Model Spec 1 (pytorch, 1 Billion)
+++++++++++++++++++++++++++++++++++++++++
+
+- **Model Format:** pytorch
+- **Model Size (in billions):** 1
+- **Quantizations:** none
+- **Engines**: vLLM, Transformers, SGLang
+- **Model ID:** openbmb/MiniCPM5-1B
+- **Model Hubs**:  `Hugging Face <https://huggingface.co/openbmb/MiniCPM5-1B>`__, `ModelScope <https://modelscope.cn/models/OpenBMB/MiniCPM5-1B>`__
+
+Execute the following command to launch the model, remember to replace ``${quantization}`` with your
+chosen quantization method from the options listed above::
+
+   xinference launch --model-engine ${engine} --model-name minicpm5-1b --size-in-billions 1 --model-format pytorch --quantization ${quantization}
+
+
+Model Spec 2 (ggufv2, 1 Billion)
+++++++++++++++++++++++++++++++++++++++++
+
+- **Model Format:** ggufv2
+- **Model Size (in billions):** 1
+- **Quantizations:** F16, Q4_K_M, Q8_0
+- **Engines**: vLLM, llama.cpp
+- **Model ID:** openbmb/MiniCPM5-1B-GGUF
+- **Model Hubs**:  `Hugging Face <https://huggingface.co/openbmb/MiniCPM5-1B-GGUF>`__, `ModelScope <https://modelscope.cn/models/OpenBMB/MiniCPM5-1B-GGUF>`__
+
+Execute the following command to launch the model, remember to replace ``${quantization}`` with your
+chosen quantization method from the options listed above::
+
+   xinference launch --model-engine ${engine} --model-name minicpm5-1b --size-in-billions 1 --model-format ggufv2 --quantization ${quantization}
+
+
+Model Spec 3 (mlx, 1 Billion)
+++++++++++++++++++++++++++++++++++++++++
+
+- **Model Format:** mlx
+- **Model Size (in billions):** 1
+- **Quantizations:** 4bit
+- **Engines**: MLX
+- **Model ID:** openbmb/MiniCPM5-1B-MLX
+- **Model Hubs**:  `Hugging Face <https://huggingface.co/openbmb/MiniCPM5-1B-MLX>`__, `ModelScope <https://modelscope.cn/models/OpenBMB/MiniCPM5-1B-MLX>`__
+
+Execute the following command to launch the model, remember to replace ``${quantization}`` with your
+chosen quantization method from the options listed above::
+
+   xinference launch --model-engine ${engine} --model-name minicpm5-1b --size-in-billions 1 --model-format mlx --quantization ${quantization}

doc/source/models/builtin/llm/qwen3.6.rst

@@ -46,7 +46,23 @@ chosen quantization method from the options listed above::
    xinference launch --model-engine ${engine} --model-name qwen3.6 --size-in-billions 27 --model-format fp8 --quantization ${quantization}
 
 
-Model Spec 3 (pytorch, 35 Billion)
+Model Spec 3 (awq, 27 Billion)
+++++++++++++++++++++++++++++++++++++++++
+
+- **Model Format:** awq
+- **Model Size (in billions):** 27
+- **Quantizations:** Int4
+- **Engines**: vLLM, Transformers, SGLang
+- **Model ID:** QuantTrio/Qwen3.6-27B-AWQ
+- **Model Hubs**:  `Hugging Face <https://huggingface.co/QuantTrio/Qwen3.6-27B-AWQ>`__, `ModelScope <https://modelscope.cn/models/tclf90/Qwen3.6-27B-AWQ>`__
+
+Execute the following command to launch the model, remember to replace ``${quantization}`` with your
+chosen quantization method from the options listed above::
+
+   xinference launch --model-engine ${engine} --model-name qwen3.6 --size-in-billions 27 --model-format awq --quantization ${quantization}
+
+
+Model Spec 4 (pytorch, 35 Billion)
 ++++++++++++++++++++++++++++++++++++++++
 
 - **Model Format:** pytorch
@@ -62,7 +78,7 @@ chosen quantization method from the options listed above::
    xinference launch --model-engine ${engine} --model-name qwen3.6 --size-in-billions 35 --model-format pytorch --quantization ${quantization}
 
 
-Model Spec 4 (fp8, 35 Billion)
+Model Spec 5 (fp8, 35 Billion)
 ++++++++++++++++++++++++++++++++++++++++
 
 - **Model Format:** fp8
@@ -78,7 +94,7 @@ chosen quantization method from the options listed above::
    xinference launch --model-engine ${engine} --model-name qwen3.6 --size-in-billions 35 --model-format fp8 --quantization ${quantization}
 
 
-Model Spec 5 (awq, 35 Billion)
+Model Spec 6 (awq, 35 Billion)
 ++++++++++++++++++++++++++++++++++++++++
 
 - **Model Format:** awq
@@ -94,7 +110,7 @@ chosen quantization method from the options listed above::
    xinference launch --model-engine ${engine} --model-name qwen3.6 --size-in-billions 35 --model-format awq --quantization ${quantization}
 
 
-Model Spec 6 (ggufv2, 35 Billion)
+Model Spec 7 (ggufv2, 35 Billion)
 ++++++++++++++++++++++++++++++++++++++++
 
 - **Model Format:** ggufv2
@@ -110,7 +126,7 @@ chosen quantization method from the options listed above::
    xinference launch --model-engine ${engine} --model-name qwen3.6 --size-in-billions 35 --model-format ggufv2 --quantization ${quantization}
 
 
-Model Spec 7 (mlx, 35 Billion)
+Model Spec 8 (mlx, 35 Billion)
 ++++++++++++++++++++++++++++++++++++++++
 
 - **Model Format:** mlx

doc/source/user_guide/backends.rst

@@ -128,7 +128,7 @@ Currently, supported model includes:
 
 .. vllm_start
 
-- ``code-llama``, ``code-llama-instruct``, ``code-llama-python``, ``deepseek``, ``deepseek-chat``, ``deepseek-coder``, ``deepseek-coder-instruct``, ``deepseek-r1-distill-llama``, ``gorilla-openfunctions-v2``, ``HuatuoGPT-o1-LLaMA-3.1``, ``llama-2``, ``llama-2-chat``, ``llama-3``, ``llama-3-instruct``, ``llama-3.1``, ``llama-3.1-instruct``, ``llama-3.3-instruct``, ``tiny-llama``, ``wizardcoder-python-v1.0``, ``wizardmath-v1.0``, ``Yi``, ``Yi-1.5``, ``Yi-1.5-chat``, ``Yi-1.5-chat-16k``, ``Yi-200k``, ``Yi-chat``
+- ``code-llama``, ``code-llama-instruct``, ``code-llama-python``, ``deepseek``, ``deepseek-chat``, ``deepseek-coder``, ``deepseek-coder-instruct``, ``deepseek-r1-distill-llama``, ``gorilla-openfunctions-v2``, ``HuatuoGPT-o1-LLaMA-3.1``, ``llama-2``, ``llama-2-chat``, ``llama-3``, ``llama-3-instruct``, ``llama-3.1``, ``llama-3.1-instruct``, ``llama-3.3-instruct``, ``minicpm5-1b``, ``tiny-llama``, ``wizardcoder-python-v1.0``, ``wizardmath-v1.0``, ``Yi``, ``Yi-1.5``, ``Yi-1.5-chat``, ``Yi-1.5-chat-16k``, ``Yi-200k``, ``Yi-chat``
 - ``codestral-v0.1``, ``mistral-instruct-v0.1``, ``mistral-instruct-v0.2``, ``mistral-instruct-v0.3``, ``mistral-large-instruct``, ``mistral-nemo-instruct``, ``mistral-v0.1``, ``openhermes-2.5``, ``seallm_v2``
 - ``Baichuan-M2``, ``codeqwen1.5``, ``codeqwen1.5-chat``, ``deepseek-r1-distill-qwen``, ``DianJin-R1``, ``fin-r1``, ``HuatuoGPT-o1-Qwen2.5``, ``KAT-V1``, ``marco-o1``, ``qwen1.5-chat``, ``qwen2-instruct``, ``qwen2.5``, ``qwen2.5-coder``, ``qwen2.5-coder-instruct``, ``qwen2.5-instruct``, ``qwen2.5-instruct-1m``, ``qwenLong-l1``, ``QwQ-32B``, ``QwQ-32B-Preview``, ``seallms-v3``, ``skywork-or1``, ``skywork-or1-preview``, ``XiYanSQL-QwenCoder-2504``
 - ``llama-3.2-vision``, ``llama-3.2-vision-instruct``
@@ -161,6 +161,7 @@ Currently, supported model includes:
 - ``MiniMax-M2``, ``MiniMax-M2.5``, ``MiniMax-M2.7``
 - ``GLM-4.7-Flash``
 - ``glm-5``, ``glm-5.1``
+- ``DeepSeek-V4-Flash``, ``DeepSeek-V4-Pro``
 .. vllm_end
 
 .. _sglang_backend:

xinference/__init__.py

@@ -37,88 +37,6 @@ def _install():
     default_router = Router.get_instance_or_empty()
     Router.set_instance(default_router)
 
-    # Compatibility shim: xoscar <= 0.9.6 (the latest PyPI release at the
-    # time of writing) does not expose ``wait_for``. xinference internals
-    # (e.g. core/worker.py, core/supervisor.py) call ``xo.wait_for`` for
-    # actor-aware timeouts that survive cancellation routed through Xoscar
-    # actor calls -- so we cannot simply alias it to ``asyncio.wait_for``
-    # (that would re-introduce the timeout-hang bug). Backport the
-    # actor-aware implementation from xoscar main instead.
-    import xoscar as _xo
-
-    if not hasattr(_xo, "wait_for"):
-        import asyncio as _asyncio
-        import logging as _logging
-        from typing import Any as _Any
-        from typing import Awaitable as _Awaitable
-
-        _wait_for_logger = _logging.getLogger("xinference.compat.xoscar")
-
-        async def _xo_wait_for(
-            fut: "_Awaitable[_Any]", timeout: "float | None" = None
-        ) -> "_Any":
-            # Mirrors xoscar.api.wait_for: asyncio.wait_for on an actor call
-            # can hang because the cancellation is delivered through a
-            # CancelMessage that may never be processed; this wrapper makes
-            # the timeout fire regardless.
-            loop = _asyncio.get_running_loop()
-            new_fut: "_asyncio.Future[_Any]" = loop.create_future()
-            task = _asyncio.ensure_future(fut)
-
-            def _on_done(f: "_asyncio.Future[_Any]") -> None:
-                if new_fut.done():
-                    return
-                if f.cancelled():
-                    new_fut.cancel()
-                elif f.exception() is not None:
-                    new_fut.set_exception(f.exception())  # type: ignore[arg-type]
-                else:
-                    new_fut.set_result(f.result())
-
-            task.add_done_callback(_on_done)
-
-            try:
-                return await _asyncio.wait_for(new_fut, timeout)
-            except _asyncio.TimeoutError:
-                if not task.done():
-                    try:
-                        task.cancel()
-                    except Exception:
-                        _wait_for_logger.warning("Failed to cancel task", exc_info=True)
-                raise
-
-        _xo.wait_for = _xo_wait_for
-
-    # Compatibility shim: older xoscar (<= 0.9.6) spawns each actor pool as
-    # a daemon process, which makes the worker process unable to spawn the
-    # model subprocess (Python raises ``AssertionError: daemonic processes
-    # are not allowed to have children``). Patch multiprocessing so the
-    # daemon flag is forced off when xoscar tries to start a child pool.
-    import multiprocessing.process as _mp_process
-
-    if not getattr(_mp_process, "_xinf_daemon_patch_applied", False):
-        _orig_start = _mp_process.BaseProcess.start
-
-        def _start_allow_daemon_children(self):
-            # Pretend the current parent process is not a daemon so that
-            # ``assert not _current_process._config.get('daemon')`` passes
-            # inside multiprocessing.Process.start(). xoscar always launches
-            # its worker pools as daemon=True; for xinference the worker
-            # must spawn model inference subprocesses, so we relax this.
-            cur = _mp_process._current_process
-            saved = cur._config.get("daemon")
-            cur._config["daemon"] = False
-            try:
-                return _orig_start(self)
-            finally:
-                if saved is None:
-                    cur._config.pop("daemon", None)
-                else:
-                    cur._config["daemon"] = saved
-
-        _mp_process.BaseProcess.start = _start_allow_daemon_children
-        _mp_process._xinf_daemon_patch_applied = True
-
 
 _install()
 del _install

xinference/core/worker.py

@@ -946,28 +946,9 @@ async def _create_subpool(
             )
             env[env_name] = ",".join([str(dev) for dev in devices])
 
-        subpool_kwargs: Dict[str, Any] = {"env": env}
-        if start_python is not None:
-            subpool_kwargs["start_python"] = start_python
-        try:
-            subpool_address = await self._main_pool.append_sub_pool(**subpool_kwargs)
-        except TypeError as e:
-            # Older xoscar (<= 0.9.6) does not accept ``start_python``.
-            # Drop it transparently so launch still works against an
-            # environment that has not been upgraded yet.
-            if "start_python" in str(e) and "start_python" in subpool_kwargs:
-                logger.warning(
-                    "Installed xoscar does not support 'start_python' "
-                    "(requires >= 0.9.7). Falling back to the default "
-                    "Python interpreter; virtual environments configured "
-                    "via 'start_python' will not take effect."
-                )
-                subpool_kwargs.pop("start_python", None)
-                subpool_address = await self._main_pool.append_sub_pool(
-                    **subpool_kwargs
-                )
-            else:
-                raise
+        subpool_address = await self._main_pool.append_sub_pool(
+            env=env, start_python=start_python
+        )
         await self._ensure_subpool_monitor()
         return subpool_address, [str(dev) for dev in devices]