xinference 1.4.0__py3-none-any.whl → 1.5.0__py3-none-any.whl

xinference/model/llm/vllm/core.py

@@ -13,12 +13,16 @@
 # limitations under the License.
 
 import asyncio
+import itertools
 import json
 import logging
 import multiprocessing
 import os
+import sys
+import threading
 import time
 import uuid
+from functools import partial
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -27,10 +31,14 @@ from typing import (
     List,
     Optional,
     Tuple,
+    Type,
     TypedDict,
     Union,
 )
 
+import xoscar as xo
+from typing_extensions import NotRequired
+
 from ....types import (
     ChatCompletion,
     ChatCompletionChunk,
@@ -73,6 +81,10 @@ class VLLMModelConfig(TypedDict, total=False):
     guided_decoding_backend: Optional[str]
     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):
@@ -161,6 +173,9 @@ if VLLM_INSTALLED and vllm.__version__ >= "0.3.0":
     VLLM_SUPPORTED_CHAT_MODELS.append("QwQ-32B")
     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")
@@ -196,6 +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")
@@ -220,6 +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__(
@@ -248,15 +267,59 @@ class VLLMModel(LLM):
         self.lora_modules = peft_model
         self.lora_requests: List[LoRARequest] = []
         self._xavier_config = None
+        # distributed inference
+        self._device_count = None
+        self._address = model_config.pop("address", None)  # type: ignore
+        self._n_worker = model_config.pop("n_worker", 1)  # type: ignore
+        self._shard = model_config.pop("shard", 0)  # type: ignore
+        self._driver_info = model_config.pop("driver_info", None)  # type: ignore
+        self._loading_thread: Optional[threading.Thread] = None
+        self._loading_error = None
+        # variables used for distributed inference and multiple GPUs
+        self._pool_addresses = None
+        self._worker_addresses: Optional[Dict[int, List[str]]] = None
+        self._all_worker_ready: Optional[threading.Event] = None
+        # used to call async
+        self._loop = None
 
     def set_xavier_config(self, value: Optional[Dict]):
         self._xavier_config = value  # type: ignore
 
+    def set_worker_addresses(self, shard: int, worker_addresses: List[str]):
+        assert self._worker_addresses is not None
+        self._worker_addresses[shard] = worker_addresses
+        if (
+            self._all_worker_ready is not None
+            and len(self._worker_addresses) == self._n_worker
+        ):
+            self._all_worker_ready.set()
+
+    @property
+    def driver_info(self) -> Optional[dict]:
+        return self._driver_info
+
+    @property
+    def need_create_pools(self):
+        return True
+
+    def set_pool_addresses(self, pool_addresses: List[str]):
+        self._pool_addresses = pool_addresses  # type: ignore
+
+    def get_pool_addresses(self) -> Optional[List[str]]:
+        return self._pool_addresses
+
+    def set_loop(self, loop: asyncio.AbstractEventLoop):
+        # loop will be passed into XinferenceDistributedExecutor,
+        # to call aynsc method with asyncio.run_coroutine_threadsafe
+        self._loop = loop  # type: ignore
+
     def load(self):
         try:
             import vllm
+            from vllm.config import VllmConfig
             from vllm.engine.arg_utils import AsyncEngineArgs
             from vllm.engine.async_llm_engine import AsyncLLMEngine
+            from vllm.executor.executor_base import ExecutorBase
             from vllm.lora.request import LoRARequest
         except ImportError:
             error_message = "Failed to import module 'vllm'"
@@ -275,6 +338,7 @@ class VLLMModel(LLM):
             # we need to set it to fork to make cupy NCCL work
             multiprocessing.set_start_method("fork", force=True)
 
+        self._device_count = self._get_cuda_count()
         self._model_config = self._sanitize_model_config(self._model_config)
         reasoning_content = self._model_config.pop("reasoning_content")
 
@@ -320,6 +384,83 @@ class VLLMModel(LLM):
             self._engine = XavierEngine.from_engine_args(
                 engine_args, xavier_config=self._xavier_config
             )
+        elif self._n_worker > 1 or (
+            self._device_count > 1 and vllm.__version__ >= "0.7.0"
+        ):
+            from .distributed_executor import XinferenceDistributedExecutor
+
+            # model across multiple workers or GPUs
+            engine_args = AsyncEngineArgs(
+                model=self.model_path,
+                enable_lora=enable_lora,
+                max_loras=max_loras,
+                **self._model_config,
+            )
+
+            assert self._loop is not None
+            self._worker_addresses = {}
+
+            def _load():
+                try:
+                    assert self._pool_addresses
+
+                    if self._shard > 0:
+                        assert self._driver_info
+                        address = self._driver_info["address"]
+
+                        coro = xo.actor_ref(address, self.raw_model_uid)
+                        model_ref = asyncio.run_coroutine_threadsafe(
+                            coro, self._loop
+                        ).result()
+                        coro = model_ref.set_worker_addresses(
+                            self._shard, self._pool_addresses
+                        )
+                        asyncio.run_coroutine_threadsafe(coro, self._loop).result()
+                    else:
+                        self.set_worker_addresses(0, self._pool_addresses)
+                        self._driver_info = {"address": self._address}
+
+                        if self._n_worker > 1:
+                            self._all_worker_ready = threading.Event()
+                            # if model across workers, wait for other workers ready
+                            self._all_worker_ready.wait()
+
+                        # gather all worker addresses
+                        worker_addresses = list(
+                            itertools.chain(
+                                *[
+                                    self._worker_addresses[shard]
+                                    for shard in range(self._n_worker)
+                                ]
+                            )
+                        )
+                        assert worker_addresses
+                        loop = self._loop
+
+                        class XinferenceAsyncLLMEngine(AsyncLLMEngine):
+                            @classmethod
+                            def _get_executor_cls(
+                                cls, engine_config: VllmConfig
+                            ) -> Type[ExecutorBase]:
+                                return partial(  # type: ignore
+                                    XinferenceDistributedExecutor,
+                                    pool_addresses=worker_addresses,
+                                    n_worker=self._n_worker,
+                                    loop=loop,
+                                )
+
+                        self._engine = XinferenceAsyncLLMEngine.from_engine_args(
+                            engine_args
+                        )
+                except:
+                    logger.exception("Creating vllm engine failed")
+                    self._loading_error = sys.exc_info()
+
+            self._loading_thread = threading.Thread(target=_load)
+            self._loading_thread.start()
+            # wait some time for init finish
+            if self._shard == 0:
+                self._loading_thread.join(1)
         else:
             engine_args = AsyncEngineArgs(
                 model=self.model_path,
@@ -332,7 +473,14 @@ class VLLMModel(LLM):
         self._check_health_task = None
         if hasattr(self._engine, "check_health"):
             # vLLM introduced `check_health` since v0.4.1
-            self._check_health_task = asyncio.create_task(self._check_healthy())
+            self._check_health_task = self._loop.create_task(self._check_healthy())
+
+    def wait_for_load(self):
+        if self._loading_thread:
+            self._loading_thread.join()
+            if self._loading_error:
+                _, err, tb = self._loading_error
+                raise err.with_traceback(tb)
 
     def stop(self):
         # though the vLLM engine will shutdown when deleted,
@@ -341,9 +489,10 @@ class VLLMModel(LLM):
         logger.info("Stopping vLLM engine")
         if self._check_health_task:
             self._check_health_task.cancel()
-        if model_executor := getattr(self._engine.engine, "model_executor", None):
-            model_executor.shutdown()
-        self._engine = None
+        if self._engine:
+            if model_executor := getattr(self._engine.engine, "model_executor", None):
+                model_executor.shutdown()
+            self._engine = None
 
     async def init_xavier(self):
         await self._engine.init_xavier()
@@ -374,22 +523,49 @@ class VLLMModel(LLM):
         if model_config is None:
             model_config = VLLMModelConfig()
 
-        cuda_count = self._get_cuda_count()
-
         model_config.setdefault("tokenizer_mode", "auto")
         model_config.setdefault("trust_remote_code", True)
-        model_config.setdefault("tensor_parallel_size", cuda_count)
+        model_config.setdefault("tensor_parallel_size", self._device_count)  # type: ignore
+        model_config.setdefault("pipeline_parallel_size", self._n_worker)  # type: ignore
         model_config.setdefault("block_size", 16)
         model_config.setdefault("swap_space", 4)
         model_config.setdefault("gpu_memory_utilization", 0.90)
         model_config.setdefault("max_num_seqs", 256)
-        model_config.setdefault("quantization", None)
+        if "model_quantization" in model_config:
+            model_config["quantization"] = model_config.pop("model_quantization")
+        else:
+            model_config.setdefault("quantization", None)
         model_config.setdefault("max_model_len", None)
         model_config.setdefault("guided_decoding_backend", "outlines")
         model_config.setdefault("reasoning_content", False)
         # 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

@@ -0,0 +1,320 @@
+# 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 asyncio
+import logging
+import os
+from functools import partial
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Union
+
+import xoscar as xo
+from vllm.executor.executor_base import DistributedExecutorBase
+from vllm.utils import _run_task_with_lock, get_distributed_init_method
+from vllm.worker.worker_base import WorkerWrapperBase
+from xoscar.utils import get_next_port
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+    from vllm.model_executor.layers.sampler import SamplerOutput
+    from vllm.sequence import ExecuteModelRequest
+
+logger = logging.getLogger(__name__)
+
+
+class WorkerActor(xo.StatelessActor):
+    def __init__(self, vllm_config: "VllmConfig", rpc_rank: int = 0, **kwargs):
+        super().__init__(**kwargs)
+        self._worker = WorkerWrapperBase(vllm_config, rpc_rank=rpc_rank)
+
+    async def __post_create__(self):
+        try:
+            # Change process title for model
+            import setproctitle
+
+            setproctitle.setproctitle(f"Xinf vLLM worker: {self._worker.rpc_rank}")
+        except ImportError:
+            pass
+
+    def __getattr__(self, item):
+        return getattr(self._worker, item)
+
+    @classmethod
+    def gen_uid(cls, rank):
+        return f"VllmWorker_{rank}"
+
+    def execute_method(self, method: Union[str, Callable], *args, **kwargs):
+        logger.debug(
+            "Calling method %s in vllm worker %s, args: %s, kwargs: %s",
+            method,
+            self.uid,
+            args,
+            kwargs,
+        )
+        if isinstance(method, str):
+            return getattr(self._worker, method)(*args, **kwargs)
+        else:
+            return method(self._worker, *args, **kwargs)
+
+
+class WorkerWrapper:
+    def __init__(
+        self,
+        loop: asyncio.AbstractEventLoop,
+        worker_actor_ref: xo.ActorRefType[WorkerActor],
+    ):
+        self._loop = loop
+        self._worker_actor_ref = worker_actor_ref
+
+    def execute_method(self, method: Union[str, Callable], *args, **kwargs):
+        coro = self._worker_actor_ref.execute_method(method, *args, **kwargs)
+        return asyncio.run_coroutine_threadsafe(coro, self._loop)
+
+    async def execute_method_async(self, method: Union[str, Callable], *args, **kwargs):
+        return await self._worker_actor_ref.execute_method(method, *args, **kwargs)
+
+    def kill(self):
+        coro = xo.destroy_actor(self._worker_actor_ref)
+        return asyncio.run_coroutine_threadsafe(coro, self._loop)
+
+
+class XinferenceDistributedExecutor(DistributedExecutorBase):
+    """Xoscar based distributed executor"""
+
+    use_ray: bool = False
+    _loop: asyncio.AbstractEventLoop
+    _pool_addresses: List[str]
+    _n_worker: int
+
+    def __init__(
+        self,
+        vllm_config: "VllmConfig",
+        pool_addresses: List[str],
+        n_worker: int,
+        loop: asyncio.AbstractEventLoop,
+        *args,
+        **kwargs,
+    ):
+        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:
+        # Create the parallel GPU workers.
+        world_size = self.parallel_config.world_size
+        tensor_parallel_size = self.parallel_config.tensor_parallel_size
+
+        self.driver_worker: Optional[WorkerActor] = None
+        # The remaining workers are Xoscar actors
+        self.workers: List[WorkerWrapper] = []
+
+        assert (
+            self._pool_addresses and len(self._pool_addresses) == world_size
+        ), f"Pool addresses(#{len(self._pool_addresses or [])} must be equal to worldsize(#{world_size})"
+
+        futures = []
+        for rank in range(world_size):
+            coro = xo.create_actor(
+                WorkerActor,
+                self.vllm_config,
+                rpc_rank=rank,
+                address=self._pool_addresses[rank],
+                uid=WorkerActor.gen_uid(rank),
+            )
+            futures.append(asyncio.run_coroutine_threadsafe(coro, self._loop))
+        refs = [fut.result() for fut in futures]
+        self.workers = [WorkerWrapper(self._loop, ref) for ref in refs[1:]]
+        self.driver_worker = WorkerActor(self.vllm_config, rpc_rank=0)
+
+        def driver_execute_method(*args, **kwargs):
+            func = partial(self.driver_worker.execute_method, *args, **kwargs)
+            return self._loop.run_in_executor(None, func)
+
+        self.driver_exec_method = driver_execute_method
+
+        # Set environment variables for the driver and workers.
+        all_args_to_update_environment_variables: List[Dict[str, str]] = [
+            dict() for _ in range(world_size)
+        ]
+
+        for args in all_args_to_update_environment_variables:
+            # some carry-over env vars from the driver
+            # TODO: refactor platform-specific env vars
+            for name in [
+                "VLLM_ATTENTION_BACKEND",
+                "TPU_CHIPS_PER_HOST_BOUNDS",
+                "TPU_HOST_BOUNDS",
+                "VLLM_USE_V1",
+                "VLLM_TRACE_FUNCTION",
+            ]:
+                if name in os.environ:
+                    args[name] = os.environ[name]
+
+        self._env_vars_for_all_workers = all_args_to_update_environment_variables
+
+        self._run_workers(
+            "update_environment_variables", self._env_vars_for_all_workers
+        )
+
+        all_kwargs = []
+        distributed_init_method = get_distributed_init_method(
+            self._pool_addresses[0].split(":", 1)[0], get_next_port()
+        )
+        for rank in range(world_size):
+            local_rank = rank % (world_size // self._n_worker)
+            kwargs = dict(
+                vllm_config=self.vllm_config,
+                local_rank=local_rank,
+                rank=rank,
+                distributed_init_method=distributed_init_method,
+                is_driver_worker=not self.parallel_config
+                or (rank % tensor_parallel_size == 0),
+            )
+            all_kwargs.append(kwargs)
+        self._run_workers("init_worker", all_kwargs)
+        self._run_workers("init_device")
+        self._run_workers(
+            "load_model",
+            max_concurrent_workers=self.parallel_config.max_parallel_loading_workers,
+        )
+
+        # This is the list of workers that are rank 0 of each TP group EXCEPT
+        # global rank 0. These are the workers that will broadcast to the
+        # rest of the workers.
+        self.tp_driver_workers: List[WorkerWrapper] = []
+        # This is the list of workers that are not drivers and not the first
+        # worker in a TP group. These are the workers that will be
+        # broadcasted to.
+        self.non_driver_workers: List[WorkerWrapper] = []
+
+        # Enforce rank order for correct rank to return final output.
+        for index, worker in enumerate(self.workers):
+            # The driver worker is rank 0 and not in self.workers.
+            rank = index + 1
+            if rank % self.parallel_config.tensor_parallel_size == 0:
+                self.tp_driver_workers.append(worker)
+            else:
+                self.non_driver_workers.append(worker)
+
+        self.pp_locks: Optional[List[asyncio.Lock]] = None
+
+    def _run_workers(
+        self,
+        method: Union[str, Callable],
+        *args,
+        async_run_tensor_parallel_workers_only: bool = False,
+        max_concurrent_workers: Optional[int] = None,
+        **kwargs,
+    ) -> Any:
+        if max_concurrent_workers:
+            raise NotImplementedError("max_concurrent_workers is not supported yet.")
+
+        workers = self.workers
+        if async_run_tensor_parallel_workers_only:
+            workers = self.non_driver_workers
+        worker_outputs = [
+            worker.execute_method(method, *args, **kwargs) for worker in workers
+        ]
+
+        if async_run_tensor_parallel_workers_only:
+            return worker_outputs
+
+        driver_worker_outputs = [
+            self.driver_worker.execute_method(method, *args, **kwargs)  # type: ignore
+        ]
+        return driver_worker_outputs + [output.result() for output in worker_outputs]
+
+    def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None:
+        """Wait for futures returned from _run_workers() with
+        async_run_remote_workers_only to complete."""
+        for result in parallel_worker_tasks:
+            result.get()
+
+    def check_health(self) -> None:
+        # Assume that the workers are healthy.
+        # TODO: check the health by checking if the workers all alive
+        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, xo.ActorNotExist):
+            # event loop closed already, ignore
+            # or actor already removed
+            pass
+
+    def __del__(self):
+        return self.shutdown()
+
+    def _driver_execute_model(
+        self, execute_model_req: Optional["ExecuteModelRequest"]
+    ) -> Optional[List["SamplerOutput"]]:
+        return self.driver_worker.execute_method("execute_model", execute_model_req)  # type: ignore
+
+    async def _driver_execute_model_async(
+        self,
+        execute_model_req: Optional["ExecuteModelRequest"] = None,
+    ) -> List["SamplerOutput"]:
+        if not self.tp_driver_workers:
+            return await self.driver_exec_method("execute_model", execute_model_req)
+
+        if self.pp_locks is None:
+            # This locks each pipeline parallel stage so multiple virtual
+            # engines can't execute on the same stage at the same time
+            # We create the locks here to avoid creating them in the constructor
+            # which uses a different asyncio loop.
+            self.pp_locks = [
+                asyncio.Lock()
+                for _ in range(self.parallel_config.pipeline_parallel_size)
+            ]
+
+        tasks = [
+            asyncio.create_task(
+                _run_task_with_lock(
+                    self.driver_exec_method,
+                    self.pp_locks[0],
+                    "execute_model",
+                    execute_model_req,
+                )
+            )
+        ]
+        for pp_rank, driver_worker in enumerate(self.tp_driver_workers, start=1):
+            tasks.append(
+                asyncio.create_task(
+                    _run_task_with_lock(
+                        driver_worker.execute_method_async,
+                        self.pp_locks[pp_rank],
+                        "execute_model",
+                        execute_model_req,
+                    )
+                )
+            )
+
+        results = await asyncio.gather(*tasks)
+
+        # Only the last PP stage has the final results.
+        return results[-1]
+
+    async def _start_worker_execution_loop(self):
+        coros = [
+            worker.execute_method_async("start_worker_execution_loop")
+            for worker in self.non_driver_workers
+        ]
+        return await asyncio.gather(*coros)

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",
@@ -106,9 +111,10 @@ def generate_rerank_description(model_spec: RerankModelSpec) -> Dict[str, List[D
     return res
 
 
-class _ModelWrapper:
+class _ModelWrapper(nn.Module):
     def __init__(self, module: nn.Module):
-        self._module = module
+        super().__init__()
+        self.model = module
         self._local_data = threading.local()
 
     @property
@@ -116,18 +122,22 @@ class _ModelWrapper:
         return getattr(self._local_data, "n_tokens", 0)
 
     @n_tokens.setter
-    def n_tokens(self, new_n_tokens):
-        self._local_data.n_tokens = new_n_tokens
-
-    def __getattr__(self, attr):
-        return getattr(self._module, attr)
+    def n_tokens(self, value):
+        self._local_data.n_tokens = value
 
-    def __call__(self, **kwargs):
-        attention_mask = kwargs["attention_mask"]
+    def forward(self, **kwargs):
+        attention_mask = kwargs.get("attention_mask")
         # when batching, the attention mask 1 means there is a token
         # thus we just sum up it to get the total number of tokens
-        self.n_tokens += attention_mask.sum().item()
-        return self._module(**kwargs)
+        if attention_mask is not None:
+            self.n_tokens += attention_mask.sum().item()
+        return self.model(**kwargs)
+
+    def __getattr__(self, attr):
+        try:
+            return super().__getattr__(attr)
+        except AttributeError:
+            return getattr(self.model, attr)
 
 
 class RerankModel: