rename package

Former-commit-id: 308edbc426

src/llamafactory/extras/callbacks.py

@@ -0,0 +1,215 @@
+import json
+import logging
+import os
+import signal
+import sys
+import time
+from concurrent.futures import ThreadPoolExecutor
+from datetime import timedelta
+from typing import TYPE_CHECKING, Any, Dict, Optional
+
+import transformers
+from transformers import TrainerCallback
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, has_length
+
+from .constants import TRAINER_LOG
+from .logging import LoggerHandler, get_logger
+from .misc import fix_valuehead_checkpoint
+
+
+if TYPE_CHECKING:
+    from transformers import TrainerControl, TrainerState, TrainingArguments
+
+
+logger = get_logger(__name__)
+
+
+class FixValueHeadModelCallback(TrainerCallback):
+    def on_save(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called after a checkpoint save.
+        """
+        if args.should_save:
+            fix_valuehead_checkpoint(
+                model=kwargs.pop("model"),
+                output_dir=os.path.join(args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, state.global_step)),
+                safe_serialization=args.save_safetensors,
+            )
+
+
+class LogCallback(TrainerCallback):
+    def __init__(self, output_dir: str) -> None:
+        r"""
+        Initializes a callback for logging training and evaluation status.
+        """
+        """ Progress """
+        self.start_time = 0
+        self.cur_steps = 0
+        self.max_steps = 0
+        self.elapsed_time = ""
+        self.remaining_time = ""
+        self.thread_pool: Optional["ThreadPoolExecutor"] = None
+        """ Status """
+        self.aborted = False
+        self.do_train = False
+        """ Web UI """
+        self.webui_mode = os.environ.get("LLAMABOARD_ENABLED", "0").lower() in ["true", "1"]
+        if self.webui_mode:
+            signal.signal(signal.SIGABRT, self._set_abort)
+            self.logger_handler = LoggerHandler(output_dir)
+            logging.root.addHandler(self.logger_handler)
+            transformers.logging.add_handler(self.logger_handler)
+
+    def _set_abort(self, signum, frame) -> None:
+        self.aborted = True
+
+    def _reset(self, max_steps: int = 0) -> None:
+        self.start_time = time.time()
+        self.cur_steps = 0
+        self.max_steps = max_steps
+        self.elapsed_time = ""
+        self.remaining_time = ""
+
+    def _timing(self, cur_steps: int) -> None:
+        cur_time = time.time()
+        elapsed_time = cur_time - self.start_time
+        avg_time_per_step = elapsed_time / cur_steps if cur_steps != 0 else 0
+        remaining_time = (self.max_steps - cur_steps) * avg_time_per_step
+        self.cur_steps = cur_steps
+        self.elapsed_time = str(timedelta(seconds=int(elapsed_time)))
+        self.remaining_time = str(timedelta(seconds=int(remaining_time)))
+
+    def _write_log(self, output_dir: str, logs: Dict[str, Any]) -> None:
+        with open(os.path.join(output_dir, TRAINER_LOG), "a", encoding="utf-8") as f:
+            f.write(json.dumps(logs) + "\n")
+
+    def _create_thread_pool(self, output_dir: str) -> None:
+        os.makedirs(output_dir, exist_ok=True)
+        self.thread_pool = ThreadPoolExecutor(max_workers=1)
+
+    def _close_thread_pool(self) -> None:
+        if self.thread_pool is not None:
+            self.thread_pool.shutdown(wait=True)
+            self.thread_pool = None
+
+    def on_init_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called at the end of the initialization of the `Trainer`.
+        """
+        if (
+            args.should_save
+            and os.path.exists(os.path.join(args.output_dir, TRAINER_LOG))
+            and args.overwrite_output_dir
+        ):
+            logger.warning("Previous trainer log in this folder will be deleted.")
+            os.remove(os.path.join(args.output_dir, TRAINER_LOG))
+
+    def on_train_begin(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called at the beginning of training.
+        """
+        if args.should_save:
+            self.do_train = True
+            self._reset(max_steps=state.max_steps)
+            self._create_thread_pool(output_dir=args.output_dir)
+
+    def on_train_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called at the end of training.
+        """
+        self._close_thread_pool()
+
+    def on_substep_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called at the end of an substep during gradient accumulation.
+        """
+        if self.aborted:
+            control.should_epoch_stop = True
+            control.should_training_stop = True
+
+    def on_step_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called at the end of a training step.
+        """
+        if self.aborted:
+            control.should_epoch_stop = True
+            control.should_training_stop = True
+
+    def on_evaluate(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called after an evaluation phase.
+        """
+        if not self.do_train:
+            self._close_thread_pool()
+
+    def on_predict(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called after a successful prediction.
+        """
+        if not self.do_train:
+            self._close_thread_pool()
+
+    def on_log(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called after logging the last logs.
+        """
+        if not args.should_save:
+            return
+
+        self._timing(cur_steps=state.global_step)
+        logs = dict(
+            current_steps=self.cur_steps,
+            total_steps=self.max_steps,
+            loss=state.log_history[-1].get("loss", None),
+            eval_loss=state.log_history[-1].get("eval_loss", None),
+            predict_loss=state.log_history[-1].get("predict_loss", None),
+            reward=state.log_history[-1].get("reward", None),
+            accuracy=state.log_history[-1].get("rewards/accuracies", None),
+            learning_rate=state.log_history[-1].get("learning_rate", None),
+            epoch=state.log_history[-1].get("epoch", None),
+            percentage=round(self.cur_steps / self.max_steps * 100, 2) if self.max_steps != 0 else 100,
+            elapsed_time=self.elapsed_time,
+            remaining_time=self.remaining_time,
+        )
+        logs = {k: v for k, v in logs.items() if v is not None}
+        if self.webui_mode and all(key in logs for key in ["loss", "learning_rate", "epoch"]):
+            logger.info(
+                "{{'loss': {:.4f}, 'learning_rate': {:2.4e}, 'epoch': {:.2f}}}".format(
+                    logs["loss"], logs["learning_rate"], logs["epoch"]
+                )
+            )
+
+        if self.thread_pool is not None:
+            self.thread_pool.submit(self._write_log, args.output_dir, logs)
+
+    def on_prediction_step(
+        self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs
+    ):
+        r"""
+        Event called after a prediction step.
+        """
+        if self.do_train:
+            return
+
+        if self.aborted:
+            sys.exit(0)
+
+        if not args.should_save:
+            return
+
+        eval_dataloader = kwargs.pop("eval_dataloader", None)
+        if has_length(eval_dataloader):
+            if self.max_steps == 0:
+                self._reset(max_steps=len(eval_dataloader))
+                self._create_thread_pool(output_dir=args.output_dir)
+
+            self._timing(cur_steps=self.cur_steps + 1)
+            if self.cur_steps % 5 == 0 and self.thread_pool is not None:
+                logs = dict(
+                    current_steps=self.cur_steps,
+                    total_steps=self.max_steps,
+                    percentage=round(self.cur_steps / self.max_steps * 100, 2) if self.max_steps != 0 else 100,
+                    elapsed_time=self.elapsed_time,
+                    remaining_time=self.remaining_time,
+                )
+                self.thread_pool.submit(self._write_log, args.output_dir, logs)

src/llamafactory/extras/logging.py

@@ -0,0 +1,68 @@
+import logging
+import os
+import sys
+from concurrent.futures import ThreadPoolExecutor
+
+from .constants import RUNNING_LOG
+
+
+class LoggerHandler(logging.Handler):
+    r"""
+    Logger handler used in Web UI.
+    """
+
+    def __init__(self, output_dir: str) -> None:
+        super().__init__()
+        formatter = logging.Formatter(
+            fmt="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S"
+        )
+        self.setLevel(logging.INFO)
+        self.setFormatter(formatter)
+
+        os.makedirs(output_dir, exist_ok=True)
+        self.running_log = os.path.join(output_dir, RUNNING_LOG)
+        if os.path.exists(self.running_log):
+            os.remove(self.running_log)
+
+        self.thread_pool = ThreadPoolExecutor(max_workers=1)
+
+    def _write_log(self, log_entry: str) -> None:
+        with open(self.running_log, "a", encoding="utf-8") as f:
+            f.write(log_entry + "\n\n")
+
+    def emit(self, record) -> None:
+        if record.name == "httpx":
+            return
+
+        log_entry = self.format(record)
+        self.thread_pool.submit(self._write_log, log_entry)
+
+    def close(self) -> None:
+        self.thread_pool.shutdown(wait=True)
+        return super().close()
+
+
+def get_logger(name: str) -> logging.Logger:
+    r"""
+    Gets a standard logger with a stream hander to stdout.
+    """
+    formatter = logging.Formatter(
+        fmt="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S"
+    )
+    handler = logging.StreamHandler(sys.stdout)
+    handler.setFormatter(formatter)
+
+    logger = logging.getLogger(name)
+    logger.setLevel(logging.INFO)
+    logger.addHandler(handler)
+
+    return logger
+
+
+def reset_logging() -> None:
+    r"""
+    Removes basic config of root logger. (unused in script)
+    """
+    root = logging.getLogger()
+    list(map(root.removeHandler, root.handlers))
+    list(map(root.removeFilter, root.filters))

src/llamafactory/extras/misc.py

@@ -0,0 +1,228 @@
+import gc
+import os
+from typing import TYPE_CHECKING, Dict, Tuple
+
+import torch
+from peft import PeftModel
+from transformers import InfNanRemoveLogitsProcessor, LogitsProcessorList, PreTrainedModel
+from transformers.utils import (
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_NAME,
+    is_torch_bf16_gpu_available,
+    is_torch_cuda_available,
+    is_torch_mps_available,
+    is_torch_npu_available,
+    is_torch_xpu_available,
+)
+from transformers.utils.versions import require_version
+
+from .constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
+from .logging import get_logger
+
+
+_is_fp16_available = is_torch_npu_available() or is_torch_cuda_available()
+try:
+    _is_bf16_available = is_torch_bf16_gpu_available()
+except Exception:
+    _is_bf16_available = False
+
+
+if TYPE_CHECKING:
+    from trl import AutoModelForCausalLMWithValueHead
+
+    from ..hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+class AverageMeter:
+    r"""
+    Computes and stores the average and current value.
+    """
+
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+
+def check_dependencies() -> None:
+    if os.environ.get("DISABLE_VERSION_CHECK", "0").lower() in ["true", "1"]:
+        logger.warning("Version checking has been disabled, may lead to unexpected behaviors.")
+    else:
+        require_version("transformers>=4.37.2", "To fix: pip install transformers>=4.37.2")
+        require_version("datasets>=2.14.3", "To fix: pip install datasets>=2.14.3")
+        require_version("accelerate>=0.27.2", "To fix: pip install accelerate>=0.27.2")
+        require_version("peft>=0.10.0", "To fix: pip install peft>=0.10.0")
+        require_version("trl>=0.8.1", "To fix: pip install trl>=0.8.1")
+
+
+def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
+    r"""
+    Returns the number of trainable parameters and number of all parameters in the model.
+    """
+    trainable_params, all_param = 0, 0
+    for param in model.parameters():
+        num_params = param.numel()
+        # if using DS Zero 3 and the weights are initialized empty
+        if num_params == 0 and hasattr(param, "ds_numel"):
+            num_params = param.ds_numel
+
+        # Due to the design of 4bit linear layers from bitsandbytes, multiply the number of parameters by 2
+        if param.__class__.__name__ == "Params4bit":
+            if hasattr(param, "quant_storage") and hasattr(param.quant_storage, "itemsize"):
+                num_bytes = param.quant_storage.itemsize
+            elif hasattr(param, "element_size"):  # for older pytorch version
+                num_bytes = param.element_size()
+            else:
+                num_bytes = 1
+
+            num_params = num_params * 2 * num_bytes
+
+        all_param += num_params
+        if param.requires_grad:
+            trainable_params += num_params
+
+    return trainable_params, all_param
+
+
+def fix_valuehead_checkpoint(
+    model: "AutoModelForCausalLMWithValueHead", output_dir: str, safe_serialization: bool
+) -> None:
+    r"""
+    The model is already unwrapped.
+
+    There are three cases:
+    1. full tuning without ds_zero3: state_dict = {"model.layers.*": ..., "v_head.summary.*": ...}
+    2. lora tuning without ds_zero3: state_dict = {"v_head.summary.*": ...}
+    3. under deepspeed zero3: state_dict = {"pretrained_model.model.layers.*": ..., "v_head.summary.*": ...}
+
+    We assume `stage3_gather_16bit_weights_on_model_save=true`.
+    """
+    if not isinstance(model.pretrained_model, (PreTrainedModel, PeftModel)):
+        return
+
+    if safe_serialization:
+        from safetensors import safe_open
+        from safetensors.torch import save_file
+
+        path_to_checkpoint = os.path.join(output_dir, SAFE_WEIGHTS_NAME)
+        with safe_open(path_to_checkpoint, framework="pt", device="cpu") as f:
+            state_dict: Dict[str, torch.Tensor] = {key: f.get_tensor(key) for key in f.keys()}
+    else:
+        path_to_checkpoint = os.path.join(output_dir, WEIGHTS_NAME)
+        state_dict: Dict[str, torch.Tensor] = torch.load(path_to_checkpoint, map_location="cpu")
+
+    decoder_state_dict = {}
+    v_head_state_dict = {}
+    for name, param in state_dict.items():
+        if name.startswith("v_head."):
+            v_head_state_dict[name] = param
+        else:
+            decoder_state_dict[name.replace("pretrained_model.", "")] = param
+
+    os.remove(path_to_checkpoint)
+    model.pretrained_model.save_pretrained(
+        output_dir, state_dict=decoder_state_dict or None, safe_serialization=safe_serialization
+    )
+
+    if safe_serialization:
+        save_file(v_head_state_dict, os.path.join(output_dir, V_HEAD_SAFE_WEIGHTS_NAME), metadata={"format": "pt"})
+    else:
+        torch.save(v_head_state_dict, os.path.join(output_dir, V_HEAD_WEIGHTS_NAME))
+
+    logger.info("Value head model saved at: {}".format(output_dir))
+
+
+def get_current_device() -> torch.device:
+    r"""
+    Gets the current available device.
+    """
+    if is_torch_xpu_available():
+        device = "xpu:{}".format(os.environ.get("LOCAL_RANK", "0"))
+    elif is_torch_npu_available():
+        device = "npu:{}".format(os.environ.get("LOCAL_RANK", "0"))
+    elif is_torch_mps_available():
+        device = "mps:{}".format(os.environ.get("LOCAL_RANK", "0"))
+    elif is_torch_cuda_available():
+        device = "cuda:{}".format(os.environ.get("LOCAL_RANK", "0"))
+    else:
+        device = "cpu"
+
+    return torch.device(device)
+
+
+def get_device_count() -> int:
+    r"""
+    Gets the number of available GPU devices.
+    """
+    if not torch.cuda.is_available():
+        return 0
+
+    return torch.cuda.device_count()
+
+
+def get_logits_processor() -> "LogitsProcessorList":
+    r"""
+    Gets logits processor that removes NaN and Inf logits.
+    """
+    logits_processor = LogitsProcessorList()
+    logits_processor.append(InfNanRemoveLogitsProcessor())
+    return logits_processor
+
+
+def infer_optim_dtype(model_dtype: torch.dtype) -> torch.dtype:
+    r"""
+    Infers the optimal dtype according to the model_dtype and device compatibility.
+    """
+    if _is_bf16_available and model_dtype == torch.bfloat16:
+        return torch.bfloat16
+    elif _is_fp16_available:
+        return torch.float16
+    else:
+        return torch.float32
+
+
+def has_tokenized_data(path: os.PathLike) -> bool:
+    r"""
+    Checks if the path has a tokenized dataset.
+    """
+    return os.path.isdir(path) and len(os.listdir(path)) > 0
+
+
+def torch_gc() -> None:
+    r"""
+    Collects GPU memory.
+    """
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+
+
+def try_download_model_from_ms(model_args: "ModelArguments") -> str:
+    if not use_modelscope() or os.path.exists(model_args.model_name_or_path):
+        return model_args.model_name_or_path
+
+    try:
+        from modelscope import snapshot_download
+
+        revision = "master" if model_args.model_revision == "main" else model_args.model_revision
+        return snapshot_download(model_args.model_name_or_path, revision=revision, cache_dir=model_args.cache_dir)
+    except ImportError:
+        raise ImportError("Please install modelscope via `pip install modelscope -U`")
+
+
+def use_modelscope() -> bool:
+    return bool(int(os.environ.get("USE_MODELSCOPE_HUB", "0")))

src/llamafactory/extras/packages.py

@@ -0,0 +1,76 @@
+import importlib.metadata
+import importlib.util
+from typing import TYPE_CHECKING
+
+from packaging import version
+
+
+if TYPE_CHECKING:
+    from packaging.version import Version
+
+
+def _is_package_available(name: str) -> bool:
+    return importlib.util.find_spec(name) is not None
+
+
+def _get_package_version(name: str) -> "Version":
+    try:
+        return version.parse(importlib.metadata.version(name))
+    except Exception:
+        return version.parse("0.0.0")
+
+
+def is_fastapi_available():
+    return _is_package_available("fastapi")
+
+
+def is_flash_attn2_available():
+    return _is_package_available("flash_attn") and _get_package_version("flash_attn") > version.parse("2.0.0")
+
+
+def is_galore_available():
+    return _is_package_available("galore_torch")
+
+
+def is_gradio_available():
+    return _is_package_available("gradio")
+
+
+def is_jieba_available():
+    return _is_package_available("jieba")
+
+
+def is_matplotlib_available():
+    return _is_package_available("matplotlib")
+
+
+def is_nltk_available():
+    return _is_package_available("nltk")
+
+
+def is_pillow_available():
+    return _is_package_available("PIL")
+
+
+def is_requests_available():
+    return _is_package_available("requests")
+
+
+def is_rouge_available():
+    return _is_package_available("rouge_chinese")
+
+
+def is_sdpa_available():
+    return _get_package_version("torch") > version.parse("2.1.1")
+
+
+def is_starlette_available():
+    return _is_package_available("sse_starlette")
+
+
+def is_uvicorn_available():
+    return _is_package_available("uvicorn")
+
+
+def is_vllm_available():
+    return _is_package_available("vllm")

src/llamafactory/extras/ploting.py

@@ -0,0 +1,87 @@
+import json
+import math
+import os
+from typing import Any, Dict, List
+
+from transformers.trainer import TRAINER_STATE_NAME
+
+from .logging import get_logger
+from .packages import is_matplotlib_available
+
+
+if is_matplotlib_available():
+    import matplotlib.figure
+    import matplotlib.pyplot as plt
+
+
+logger = get_logger(__name__)
+
+
+def smooth(scalars: List[float]) -> List[float]:
+    r"""
+    EMA implementation according to TensorBoard.
+    """
+    if len(scalars) == 0:
+        return []
+
+    last = scalars[0]
+    smoothed = []
+    weight = 1.8 * (1 / (1 + math.exp(-0.05 * len(scalars))) - 0.5)  # a sigmoid function
+    for next_val in scalars:
+        smoothed_val = last * weight + (1 - weight) * next_val
+        smoothed.append(smoothed_val)
+        last = smoothed_val
+    return smoothed
+
+
+def gen_loss_plot(trainer_log: List[Dict[str, Any]]) -> "matplotlib.figure.Figure":
+    r"""
+    Plots loss curves in LlamaBoard.
+    """
+    plt.close("all")
+    plt.switch_backend("agg")
+    fig = plt.figure()
+    ax = fig.add_subplot(111)
+    steps, losses = [], []
+    for log in trainer_log:
+        if log.get("loss", None):
+            steps.append(log["current_steps"])
+            losses.append(log["loss"])
+
+    ax.plot(steps, losses, color="#1f77b4", alpha=0.4, label="original")
+    ax.plot(steps, smooth(losses), color="#1f77b4", label="smoothed")
+    ax.legend()
+    ax.set_xlabel("step")
+    ax.set_ylabel("loss")
+    return fig
+
+
+def plot_loss(save_dictionary: os.PathLike, keys: List[str] = ["loss"]) -> None:
+    r"""
+    Plots loss curves and saves the image.
+    """
+    plt.switch_backend("agg")
+    with open(os.path.join(save_dictionary, TRAINER_STATE_NAME), "r", encoding="utf-8") as f:
+        data = json.load(f)
+
+    for key in keys:
+        steps, metrics = [], []
+        for i in range(len(data["log_history"])):
+            if key in data["log_history"][i]:
+                steps.append(data["log_history"][i]["step"])
+                metrics.append(data["log_history"][i][key])
+
+        if len(metrics) == 0:
+            logger.warning(f"No metric {key} to plot.")
+            continue
+
+        plt.figure()
+        plt.plot(steps, metrics, color="#1f77b4", alpha=0.4, label="original")
+        plt.plot(steps, smooth(metrics), color="#1f77b4", label="smoothed")
+        plt.title("training {} of {}".format(key, save_dictionary))
+        plt.xlabel("step")
+        plt.ylabel(key)
+        plt.legend()
+        figure_path = os.path.join(save_dictionary, "training_{}.png".format(key.replace("/", "_")))
+        plt.savefig(figure_path, format="png", dpi=100)
+        print("Figure saved at:", figure_path)