lightseekorg--tokenspeed/python/tokenspeed/runtime/configs/load_config.py

# Copyright (c) 2026 LightSeek Foundation
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

"""Model weight loading configuration."""

import enum
import json
from dataclasses import dataclass, field

from tokenspeed.runtime.utils import get_colorful_logger

logger = get_colorful_logger(__name__)


class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    SHARDED_STATE = "sharded_state"
    MISTRAL = "mistral"
    EXTENSIBLE = "extensible"


@dataclass
class LoadConfig:
    """
    download_dir: Directory to download and load the weights, default to the
        default cache directory of huggingface.
    load_format: The format of the model weights to load:
        "auto" will try to load the weights in the safetensors format and
            fall back to the pytorch bin format if safetensors format is
            not available.
        "pt" will load the weights in the pytorch bin format.
        "safetensors" will load the weights in the safetensors format.
        "npcache" will load the weights in pytorch format and store
            a numpy cache to speed up the loading.
        "dummy" will initialize the weights with random values, which is
            mainly for profiling.
    ignore_patterns: The list of patterns to ignore when loading the model.
        Default to "original/**/*" to avoid repeated loading of llama's
        checkpoints.
    decryption_key_file: If set, decrypts the output files with a password read
        from this file (after PBKDF2).
    """

    load_format: str | LoadFormat = LoadFormat.AUTO
    download_dir: str | None = None
    model_loader_extra_config: str | dict | None = field(default_factory=dict)
    ignore_patterns: list[str] | str | None = None
    decryption_key_file: str | None = None
    weight_loader_prefetch_checkpoints: bool = False
    weight_loader_prefetch_num_threads: int = 4

    ext_yaml: str | None = None

    def __post_init__(self) -> None:
        model_loader_extra_config = self.model_loader_extra_config or {}
        if isinstance(model_loader_extra_config, str):
            self.model_loader_extra_config = json.loads(model_loader_extra_config)
        self._verify_load_format()

        if self.ignore_patterns is not None and len(self.ignore_patterns) > 0:
            logger.info(
                "Ignoring the following patterns when downloading weights: %s",
                self.ignore_patterns,
            )
        else:
            self.ignore_patterns = ["original/**/*"]

    def _verify_load_format(self) -> None:
        if not isinstance(self.load_format, str):
            return

        load_format = self.load_format.lower()
        self.load_format = LoadFormat(load_format)