[v1] add models & accelerator (#9579)

.gitignore

@@ -165,6 +165,9 @@ cython_debug/
 # uv
 uv.lock
 
+# macOS
+.DS_Store
+
 # custom .gitignore
 hf_cache/
 ms_cache/

data/v1_sft_demo.yaml

@@ -1,8 +1,8 @@
 identity:
-  file_name: identity.json
+  file_name: data/identity.json
   converter: alpaca
 alpaca_en_demo:
   file_name: alpaca_en_demo.json
-  dataset_dir: ~/data
+  dataset_dir: data
   converter: alpaca
   num_samples: 500

src/llamafactory/chat/vllm_engine.py

@@ -15,8 +15,8 @@
 import uuid
 from collections.abc import AsyncGenerator, AsyncIterator
 from typing import TYPE_CHECKING, Any, Optional, Union
-from packaging import version
 
+from packaging import version
 from typing_extensions import override
 
 from ..data import get_template_and_fix_tokenizer

src/llamafactory/data/mm_plugin.py

@@ -465,6 +465,7 @@ class BasePlugin(MMPluginMixin):
         self._validate_input(processor, images, videos, audios)
         return self._get_mm_inputs(images, videos, audios, processor)
 
+
 @dataclass
 class ErnieVLPlugin(BasePlugin):
     @override

src/llamafactory/v1/accelerator/distributed.py

@@ -12,16 +12,25 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ..config.model_args import ModelArguments
+from typing import Optional
-from ..extras.types import Model, Processor
+
+from torch.distributed.device_mesh import DeviceMesh
 
 
-class ModelEngine:
+class DeviceMeshManager:
-    def __init__(self, model_args: ModelArguments) -> None:
+    """Device mesh manager."""
-        self.args = model_args
 
-    def get_model(self) -> Model:
+    _instance: Optional["DeviceMeshManager"] = None
-        pass
+    _initialized: bool = False
 
-    def get_processor(self) -> Processor:
+    def __new__(cls) -> "DeviceMeshManager":
-        pass
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self) -> None:
+        if self._initialized:
+            return
+
+        self.device_mesh: Optional[DeviceMesh] = None
+        self._initialized = True

src/llamafactory/v1/accelerator/helper.py

@@ -0,0 +1,52 @@
+# Copyright 2025 the LlamaFactory team.
+#
+# 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.
+
+from functools import lru_cache
+
+import torch
+
+
+def get_current_accelerator(check_available: bool = True):
+    """Get current accelerator.
+
+    Note: this api requires torch>=2.7.0, 2.6 or lower will get an AttributeError or RuntimeError
+    """
+    if not hasattr(torch, "accelerator"):
+        raise RuntimeError("torch.accelerator is not available, please upgrade torch to 2.7.0 or higher.")
+
+    accelerator = torch.accelerator.current_accelerator(check_available=check_available)
+    if accelerator is None:
+        return torch.device("cpu")
+
+    return accelerator
+
+
+@lru_cache
+def is_torch_npu_available():
+    return get_current_accelerator().type == "npu"
+
+
+@lru_cache
+def is_torch_cuda_available():
+    return get_current_accelerator().type == "cuda"
+
+
+@lru_cache
+def is_torch_xpu_available():
+    return get_current_accelerator().type == "xpu"
+
+
+@lru_cache
+def is_torch_mps_available():
+    return get_current_accelerator().type == "mps"

src/llamafactory/v1/config/data_args.py

@@ -23,10 +23,6 @@ class DataArguments:
         default=None,
         metadata={"help": "Path to the dataset."},
     )
-    dataset_dir: str = field(
-        default="data",
-        metadata={"help": "Path to the folder containing the datasets."},
-    )
     cutoff_len: int = field(
         default=2048,
         metadata={"help": "Cutoff length for the dataset."},

src/llamafactory/v1/config/model_args.py

@@ -25,3 +25,11 @@ class ModelArguments:
         default=False,
         metadata={"help": "Trust remote code from Hugging Face."},
     )
+    use_fast_processor: bool = field(
+        default=True,
+        metadata={"help": "Use fast processor from Hugging Face."},
+    )
+    auto_model_class: str = field(
+        default="causallm",
+        metadata={"help": "Model class from Hugging Face."},
+    )

src/llamafactory/v1/config/sample_args.py

@@ -14,10 +14,20 @@
 
 
 from dataclasses import dataclass, field
+from enum import Enum
+
+
+class SampleBackend(Enum):
+    HF = "hf"
+    VLLM = "vllm"
 
 
 @dataclass
 class SampleArguments:
+    sample_backend: SampleBackend = field(
+        default=SampleBackend.HF,
+        metadata={"help": "Sampling backend, default to 'hf'."},
+    )
     max_new_tokens: int = field(
         default=128,
         metadata={"help": "Maximum number of new tokens to generate."},

src/llamafactory/v1/core/base_trainer.py

@@ -12,44 +12,51 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any
+"""The definition of trainer.
+
+Init Phase:
+
+1. Init dataloader.
+2. Init model worker.
+3. Init optimizer (deepspeed).
+4. Shard model.
+5. Init optimizer (fsdp).
+6. Init scheduler.
+
+Train Phase:
+1. Train Loop
+
+"""
 
 from ..config.training_args import TrainingArguments
-from ..extras.types import Model, Processor, Tensor, TorchDataset
+from ..extras.types import TorchDataset
-
+from .model_worker import ModelWorker
-
+from .trainer_utils.data_collator import DataCollator
-class DataCollator:
-    """Default Data collator."""
-
-    def __init__(self, processor: Processor) -> None:
-        self.processor = processor
-
-    def __call__(self, features: list[dict[str, Any]]) -> dict[str, Tensor]:
-        """Collate features into a batch."""
-        for feature in features:
-            pass
-
-        # sft: messages
-        # dpo: chosen_messages, rejected_messages
 
 
 class BaseTrainer:
     def __init__(
         self,
         args: TrainingArguments,
-        model: Model,
-        processor: Processor,
         dataset: TorchDataset,
         data_collator: DataCollator,
+        model_worker: ModelWorker,
     ) -> None:
         self.args = args
-        self.model = model
-        self.processor = processor
         self.dataset = dataset
         self.data_collator = data_collator
+        self.model_worker = model_worker
         self.optimizer = None
         self.lr_scheduler = None
 
+    def init_device_mesh(self) -> None:
+        pass
+
+    def init_model_and_optimizer(self) -> None:
+        self.model_config = self.model_worker.get_model_config()
+        # with self.dist_plugin.get_model_init_context():
+        #     self.model = self.model_worker.get_model(self.model_config)
+
     def create_dataloader(self) -> None:
         pass
 

src/llamafactory/v1/core/chat_sampler.py

@@ -12,9 +12,35 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ..config.sample_args import SampleArguments
+from abc import ABC, abstractmethod
+
+from ..config.sample_args import SampleArguments, SampleBackend
+from .model_worker import ModelWorker
+
+
+class BaseEngine(ABC):
+    @abstractmethod
+    def __init__(self, sample_args: SampleArguments, model_worker: ModelWorker) -> None: ...
+
+    @abstractmethod
+    async def generate(self):
+        pass
+
+    @abstractmethod
+    async def batch_infer(self):
+        pass
+
+
+class HuggingFaceEngine(BaseEngine):
+    def __init__(self, model_worker: ModelWorker, sample_args: SampleArguments) -> None:
+        self.model = model_worker.get_model()
+        self.processor = model_worker.get_processor()
+        self.args = sample_args
 
 
 class ChatSampler:
-    def __init__(self, sample_args: SampleArguments) -> None:
+    def __init__(self, model_worker: ModelWorker, sample_args: SampleArguments) -> None:
-        self.args = sample_args
+        if sample_args.sample_backend == SampleBackend.HF:
+            self.engine = HuggingFaceEngine(model_worker, sample_args)
+        else:
+            raise ValueError(f"Unknown sample backend: {sample_args.sample_backend}")

src/llamafactory/v1/core/data_engine.py

@@ -12,11 +12,23 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+"""The definition of data engine.
+
+Init Data engine:
+1. Parse dataset info from arguments.
+2. Load datasets according to dataset info.
+3. Build data index (and reweight samples if necessary).
+
+Get Data Sample:
+1. Get sample from data index.
+2. Convert sample to standard format.
+3. Return sample.
+"""
+
 import os
 from collections.abc import AsyncIterable, Iterable
 from typing import Any, Union
 
-from datasets import load_dataset
 from huggingface_hub import hf_hub_download
 from omegaconf import OmegaConf
 from torch.utils.data import Dataset
@@ -45,15 +57,13 @@ class DataEngine(Dataset):
 
     def get_dataset_info(self) -> None:
         """Get dataset info from data arguments."""
-        if self.args.dataset.endswith(".yaml") and os.path.isfile(
+        if self.args.dataset.endswith(".yaml") and os.path.isfile(self.args.dataset):  # local file
-            os.path.join(self.args.dataset_dir, self.args.dataset)
+            self.dataset_infos = OmegaConf.load(self.args.dataset)
-        ):  # local file
-            self.dataset_infos = OmegaConf.load(os.path.join(self.args.dataset_dir, self.args.dataset))
         elif self.args.dataset.endswith(".yaml"):  # hf hub uri, e.g. llamafactory/v1-sft-demo/dataset_info.yaml
             repo_id, filename = os.path.split(self.args.dataset)
             filepath = hf_hub_download(repo_id=repo_id, filename=filename, repo_type="dataset")
             self.dataset_infos = OmegaConf.load(filepath)
-        elif os.path.exists(os.path.join(self.args.dataset_dir, self.args.dataset)):  # local file(s)
+        elif os.path.exists(self.args.dataset):  # local file(s)
             self.dataset_infos = {"default": {"file_name": self.args.dataset}}
         else:  # hf hub dataset, e.g. llamafactory/v1-sft-demo
             self.dataset_infos = {"default": {"hf_hub_url": self.args.dataset}}
@@ -65,11 +75,13 @@ class DataEngine(Dataset):
             streaming = value.get("streaming", False)
             self.streaming |= streaming
             if "hf_hub_url" in value:
+                from datasets import load_dataset
+
                 self.datasets[key] = load_dataset(value["hf_hub_url"], split=split, streaming=streaming)
             else:  # data loader plugin
                 from ..plugins.data_plugins.loader import DataLoaderPlugin
 
-                self.datasets[key] = DataLoaderPlugin(args=self.args).auto_load_data(value)
+                self.datasets[key] = DataLoaderPlugin().auto_load_data(value)
 
     def build_data_index(self) -> None:
         """Build dataset index."""
@@ -145,11 +157,11 @@ class DataEngine(Dataset):
                 dataset_name, sample_index = selected_index
                 return self._convert_data_sample(self.datasets[dataset_name][sample_index], dataset_name)
 
-    def __iter__(self) -> Iterable:
+    def __iter__(self) -> Iterable[Sample]:
         """Get dataset iterator.
 
         Returns:
-            Iterable: Dataset iterator.
+            Iterable[Sample]: Dataset iterator.
         """
         if self.streaming:
             pass
@@ -159,11 +171,11 @@ class DataEngine(Dataset):
 
         raise NotImplementedError()
 
-    async def __aiter__(self) -> AsyncIterable:
+    async def __aiter__(self) -> AsyncIterable[Sample]:
         """Get dataset async iterator.
 
         Returns:
-            AsyncIterable: Dataset async iterator.
+            AsyncIterable[Sample]: Dataset async iterator.
         """
         if self.streaming:
             pass

src/llamafactory/v1/core/model_worker.py

@@ -0,0 +1,98 @@
+# Copyright 2025 the LlamaFactory team.
+#
+# 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.
+
+"""The definition of model worker.
+
+Init Phase:
+1. Init processor.
+2. Init model config.
+3. Init model.
+4. Init adapter.
+
+"""
+
+from typing import Optional
+
+from transformers import AutoConfig, AutoProcessor
+
+from ..config.model_args import ModelArguments
+from ..extras.types import DistModel, HFConfig, HFModel, Processor
+
+
+class ModelWorker:
+    def __init__(self, model_args: ModelArguments) -> None:
+        self.args = model_args
+        """Model arguments."""
+        self.processor: Optional[Processor] = None
+        """Tokenizer or multi-modal processor."""
+        self.model_config: Optional[HFConfig] = None
+        """Model configuration."""
+        self.unwrapped_model: Optional[HFModel] = None
+        """Unwrapped model."""
+        self.model: Optional[DistModel] = None
+        """Distributed model."""
+        self.init_processor()
+        self.init_model_config()
+        self.init_model()
+        self.init_adapter()
+
+    def init_processor(self) -> None:
+        self.processor = AutoProcessor.from_pretrained(
+            self.args.model,
+            trust_remote_code=self.args.trust_remote_code,
+            use_fast=self.args.use_fast_processor,
+        )
+
+    def init_model_config(self) -> None:
+        self.model_config = AutoConfig.from_pretrained(
+            self.args.model,
+            trust_remote_code=self.args.trust_remote_code,
+        )
+
+    def init_model(self) -> None:
+        if self.args.auto_model_class == "causallm":
+            from transformers import AutoModelForCausalLM, AutoModelForImageTextToText
+
+            if type(self.model_config) in AutoModelForImageTextToText._model_mapping.keys():
+                AutoClass = AutoModelForImageTextToText
+            else:
+                AutoClass = AutoModelForCausalLM
+        elif self.args.auto_model_class == "classification":
+            from transformers import AutoModelForTokenClassification
+
+            AutoClass = AutoModelForTokenClassification
+        else:
+            from transformers import AutoModel
+
+            AutoClass = AutoModel
+
+        self.unwrapped_model = AutoClass.from_pretrained(
+            self.args.model,
+            config=self.model_config,
+            dtype="auto",
+            device_map="cpu",
+            trust_remote_code=self.args.trust_remote_code,
+        )
+
+    def init_adapter(self) -> None:
+        pass
+
+    def get_processor(self) -> Processor:
+        return self.processor
+
+    def get_model_config(self) -> HFConfig:
+        return self.model_config
+
+    def get_model(self) -> HFModel:
+        return self.unwrapped_model

src/llamafactory/v1/core/trainer_utils/data_collator.py

@@ -0,0 +1,47 @@
+# Copyright 2025 the LlamaFactory team.
+#
+# 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.
+
+
+from typing import Any
+
+from ...extras.types import Processor, Tensor, TorchDataset
+
+
+class DataCollator:
+    """Default Data collator."""
+
+    def __init__(self, processor: Processor) -> None:
+        self.processor = processor
+
+    def __call__(self, features: list[dict[str, Any]]) -> dict[str, Tensor]:
+        """Collate features into a batch."""
+        for feature in features:
+            pass
+
+        # sft: messages
+        # dpo: chosen_messages, rejected_messages
+
+
+class DataLoader:
+    """Default DataLoader."""
+
+    def __init__(self, dataset: TorchDataset) -> None:
+        self.dataset = dataset
+        # 1. Init stateful dataloader (tokenize)
+        # 2. Add to buffer (2 * max seq len per device)
+        # 3. Yield batch indexes (micro batch * grad acc)
+        #    a ) non pack + non dynamic
+        #    b ) non pack + dynamic
+        #    c ) pack + non dynamic
+        #    d ) pack + dynamic

src/llamafactory/v1/extras/types.py

@@ -28,18 +28,24 @@ if TYPE_CHECKING:
     HFDataset = Union[datasets.Dataset, datasets.IterableDataset]
     DataCollator = transformers.DataCollator
     DataLoader = torch.utils.data.DataLoader
+    HFConfig = transformers.PretrainedConfig
     HFModel = transformers.PreTrainedModel
     DistModel = torch.nn.parallel.DistributedDataParallel
     Processor = Union[transformers.PreTrainedTokenizer, transformers.ProcessorMixin]
+    Optimizer = torch.optim.Optimizer
+    Scheduler = torch.optim.lr_scheduler.LRScheduler
 else:
     Tensor = None
     TorchDataset = None
     HFDataset = None
     DataCollator = None
     DataLoader = None
+    HFConfig = None
     HFModel = None
     DistModel = None
     Processor = None
+    Optimizer = None
+    Scheduler = None
 
 
 class DatasetInfo(TypedDict, total=False):
@@ -86,10 +92,3 @@ class DPOSample(TypedDict):
 
 
 Sample = Union[SFTSample, DPOSample]
-
-
-class Model(TypedDict):
-    hf_model: HFModel
-    """HF model."""
-    dist_model: DistModel
-    """Distributed model."""

src/llamafactory/v1/plugins/data_plugins/loader.py

@@ -19,7 +19,6 @@ from typing import Any, Literal, Optional, Union
 
 from datasets import load_dataset
 
-from ...config.data_args import DataArguments
 from ...extras.types import DatasetInfo, HFDataset
 
 
@@ -27,9 +26,6 @@ from ...extras.types import DatasetInfo, HFDataset
 class DataLoaderPlugin:
     """Plugin for loading dataset."""
 
-    args: DataArguments
-    """Data arguments."""
-
     def _get_builder_name(self, path: str) -> Literal["arrow", "csv", "json", "parquet", "text"]:
         """Get dataset builder name.
 
@@ -42,7 +38,7 @@ class DataLoaderPlugin:
         return os.path.splitext(path)[-1][1:].replace("jsonl", "json").replace("txt", "text")
 
     def auto_load_data(self, dataset_info: DatasetInfo) -> HFDataset:
-        dataset_dir = dataset_info.get("dataset_dir", self.args.dataset_dir)
+        dataset_dir = dataset_info.get("dataset_dir", ".")
         split = dataset_info.get("split", "train")
         streaming = dataset_info.get("streaming", False)
         if "file_name" in dataset_info:

src/llamafactory/v1/plugins/model_plugins/kernels/mlp/npu_fused_moe.py

@@ -18,9 +18,9 @@ import torch
 import torch.nn.functional as F
 import torch_npu
 
+from .....accelerator.helper import is_torch_npu_available
 from .....extras.packages import is_transformers_version_greater_than
 from .....extras.types import HFModel
-from ....trainer_plugins.distributed.accelerate import is_torch_npu_available
 from ..constants import DeviceType, KernelType
 from ..registry import MetaMoEKernel
 

src/llamafactory/v1/plugins/model_plugins/kernels/mlp/npu_swiglu.py

@@ -17,8 +17,8 @@ import types
 
 import torch
 
+from .....accelerator.helper import is_torch_npu_available
 from .....extras.types import HFModel
-from ....trainer_plugins.distributed.accelerate import is_torch_npu_available
 from ..constants import DeviceType, KernelType
 from ..registry import MetaSwiGluKernel
 

src/llamafactory/v1/plugins/model_plugins/kernels/registry.py

@@ -15,8 +15,8 @@
 from abc import ABC, ABCMeta, abstractmethod
 from typing import Any, Callable, Optional
 
+from ....accelerator.helper import get_current_accelerator
 from ....extras.types import HFModel
-from ...trainer_plugins.distributed.accelerate import get_available_accelerator
 from .constants import DeviceType, KernelType
 
 
@@ -206,7 +206,7 @@ def discover_kernels(model: HFModel = None) -> list[type[MetaKernel]]:
     discovered_kernels: list[type[MetaKernel]] = []
 
     # Detect current device type
-    accelerator = get_available_accelerator()
+    accelerator = get_current_accelerator()
     try:
         device_type = DeviceType(accelerator.type)
     except ValueError:
@@ -238,11 +238,11 @@ def apply_kernel(model: HFModel, kernel: type[MetaKernel], /, **kwargs) -> "HFMo
         model = AutoModelForCausalLM.from_pretrained("qwen/qwen2.5-0.5B")
         model = apply_kernel(model, NpuRMSNormKernel)
     """
-    if issubclass(kernel, MetaKernel) and kernel.device == get_available_accelerator().type:
+    if issubclass(kernel, MetaKernel) and kernel.device == get_current_accelerator().type:
         return kernel.apply(model, **kwargs)
 
     raise ValueError(
-        f"{kernel} must be a MetaKernel instance, or the kernel don't match the device type. got {kernel.device} and {get_available_accelerator().type} instead."
+        f"{kernel} must be a MetaKernel instance, or the kernel don't match the device type. got {kernel.device} and {get_current_accelerator().type} instead."
     )
 
 

src/llamafactory/v1/plugins/model_plugins/kernels/rms_norm/npu_rms_norm.py

@@ -14,8 +14,8 @@
 import re
 import types
 
+from .....accelerator.helper import is_torch_npu_available
 from .....extras.types import HFModel
-from ....trainer_plugins.distributed.accelerate import is_torch_npu_available
 from ..constants import DeviceType, KernelType
 from ..registry import MetaRMSNormKernel
 

src/llamafactory/v1/plugins/model_plugins/kernels/rope/npu_rope.py

@@ -16,8 +16,8 @@ import sys
 
 import torch
 
+from .....accelerator.helper import is_torch_npu_available
 from .....extras.types import HFModel
-from ....trainer_plugins.distributed.accelerate import is_torch_npu_available
 from ..constants import DeviceType, KernelType
 from ..registry import MetaRoPEKernel
 

src/llamafactory/v1/plugins/trainer_plugins/distributed/accelerate.py

@@ -11,37 +11,3 @@
 # 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.
-from functools import lru_cache
-
-import torch
-
-
-def get_available_accelerator():
-    """Get available accelerator in current environment.
-
-    Note: this api requires torch>=2.7.0, 2.6 or lower will get an AttributeError or RuntimeError
-    """
-    accelerator = torch.accelerator.current_accelerator()
-    if accelerator is None:
-        return torch.device("cpu")
-    return accelerator
-
-
-@lru_cache
-def is_torch_npu_available():
-    return get_available_accelerator().type == "npu"
-
-
-@lru_cache
-def is_torch_cuda_available():
-    return get_available_accelerator().type == "cuda"
-
-
-@lru_cache
-def is_torch_xpu_available():
-    return get_available_accelerator().type == "xpu"
-
-
-@lru_cache
-def is_torch_mps_available():
-    return get_available_accelerator().type == "mps"