ipex-llm/python/llm/src/ipex_llm/optimize.py

#
# Copyright 2016 The BigDL Authors.
#
# 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 torch
import os
import json
from .transformers import ggml_convert_low_bit
from torch.nn.modules import Module
from torch.nn.modules.module import _IncompatibleKeys
from accelerate import init_empty_weights
from accelerate.utils import set_module_tensor_to_device
from ipex_llm.ggml.quantize import ggml_tensor_qtype
from ipex_llm.utils.common import invalidInputError
from ipex_llm.transformers.utils import extract_local_archive_file, get_local_shard_files
import transformers
import warnings
from transformers import PreTrainedModel
from .utils.common import MuteHFLogger
from .utils.lazy_load_torch import LazyLoadTensors
from contextlib import ExitStack, contextmanager


# Simulate the Hugging Face format
PYTORCH_MODEL_NAME = "pytorch_model.bin"
CONFIG_NAME = "bigdl_config.json"


def _save_low_bit(self, save_dir, *args, **kwargs):
    invalidInputError(self._bigdl_config.get("bigdl_transformers_low_bit", False),
                      f"Detected this model is not a low-bit model, please use from_pretrained's"
                      f" load_in_4bit or load_in_low_bit parameter to load a 4-bit model first.")
    os.makedirs(save_dir, exist_ok=True)
    model_path = os.path.join(save_dir, PYTORCH_MODEL_NAME)
    if isinstance(self, PreTrainedModel):
        # We borrowed this method to adapt to Transformer model cases
        # as much as possible, and later we may merge these two situations
        kwargs['safe_serialization'] = False
        self.save_pretrained(save_dir, *args, **kwargs)
    else:
        # TODO: For the lowbit model still larger than 8GB,
        #       save it into shards.
        torch.save(self.state_dict(), model_path, *args, **kwargs)
    with open(os.path.join(save_dir, CONFIG_NAME), "w") as json_file:
        json.dump(self._bigdl_config, json_file)


# Under `init_empty_weights()`, we need to disable all actions
# that may lead to any parameter allocation", otherwise may need to error:
# NotImplementedError: Cannot copy out of meta tensor; no data!
class DisableTorchAllocTensor():
    def __init__(self) -> None:
        self._old_torch_load_state_dict = Module.load_state_dict
        self._old_torch_to_device = Module.to
        self._old_torch_load_from_state_dict = Module._load_from_state_dict
        # Chatglm2 init weights manually,
        # and `skip_init` init on `cpu` by default
        self._old_skip_init = torch.nn.utils.skip_init

    def __enter__(self):
        Module.load_state_dict = lambda *args, **kwargs: _IncompatibleKeys([], [])
        Module._load_from_state_dict = lambda *args, **kwargs: None
        Module.to = lambda self, *args, **kwargs: self

        def skip_init_on_meta(module_cls, *args, **kwargs):
            kwargs['device'] = 'meta'
            return self._old_skip_init(module_cls, *args, **kwargs)
        torch.nn.utils.skip_init = skip_init_on_meta

    def __exit__(self, exc_type, exc_value, traceback):
        Module.load_state_dict = self._old_torch_load_state_dict
        Module._load_from_state_dict = self._old_torch_load_from_state_dict
        Module.to = self._old_torch_to_device
        torch.nn.utils.skip_init = self._old_skip_init


class ContextManagers:
    """
    Wrapper for `contextlib.ExitStack` which enters a collection of context managers.
    Adaptation of `ContextManagers` in the `fastcore` library.
    """

    def __init__(self, context_managers):
        self.context_managers = context_managers
        self.stack = ExitStack()

    def __enter__(self):
        for context_manager in self.context_managers:
            self.stack.enter_context(context_manager)

    def __exit__(self, *args, **kwargs):
        self.stack.__exit__(*args, **kwargs)


def low_bit_sanity_check(model_path):
    invalidInputError(os.path.isdir(model_path),
                      "model_path should be a valid directory path.")
    invalidInputError(os.path.isfile(os.path.join(model_path, CONFIG_NAME)),
                      "bigdl_config.json should be under your model directory,"
                      "please check your input path.")
    with open(os.path.join(model_path, CONFIG_NAME), 'r') as f:
        _config = json.load(f)

    low_bit = _config.get("bigdl_transformers_low_bit", None)
    invalidInputError(low_bit,
                      "Detect this model is not a low-bit model, Please use `optimize_model`"
                      " with low_bit to get a low-bit model , and "
                      " serialize the model using save_low_bit first.")
    return low_bit


@contextmanager
def low_memory_init():
    init_contexts = []
    init_contexts.extend([init_empty_weights(), DisableTorchAllocTensor()])
    # Load everything except Tensors' parameters
    init_contexts.append(LazyLoadTensors())
    # As we have muted the `torch.load`, this will trigger a key missing warning in hf
    # but this matters not for we will load again later.
    init_contexts.append(MuteHFLogger(logger=transformers.modeling_utils.logger))
    with ContextManagers(init_contexts):
        yield


def load_low_bit(model, model_path):
    """
    Load the optimized pytorch model.

    :param model: The PyTorch model instance
    :param model_path: The path of saved optimized model

    :return: The optimized model.

    >>> # Example 1:
    >>> # Take ChatGLM2-6B model as an example
    >>> # Make sure you have saved the optimized model by calling 'save_low_bit'
    >>> from ipex_llm.optimize import low_memory_init, load_low_bit
    >>> with low_memory_init(): # Fast and low cost by loading model on meta device
    >>>     model = AutoModel.from_pretrained(saved_dir,
    >>>                                       torch_dtype="auto",
    >>>                                       trust_remote_code=True)
    >>> model = load_low_bit(model, saved_dir) # Load the optimized model

    >>> # Example 2:
    >>> # If the model doesn't fit 'low_memory_init' method,
    >>> # alternatively, you can obtain the model instance through traditional loading method.
    >>> # Take OpenAI Whisper model as an example
    >>> # Make sure you have saved the optimized model by calling 'save_low_bit'
    >>> from ipex_llm.optimize import load_low_bit
    >>> model = whisper.load_model('tiny') # A model instance through traditional loading method
    >>> model = load_low_bit(model, saved_dir) # Load the optimized model
    """
    low_bit = low_bit_sanity_check(model_path)
    invalidInputError(isinstance(model, torch.nn.Module),
                      "model should be a instance of "
                      f"`torch.nn.Module`, but got {type(model)} at last.")
    if low_bit:
        invalidInputError(isinstance(model, torch.nn.Module),
                          "model should be an instance of `torch.nn.Module`, "
                          f"but got {type(model)} at last.")
        invalidInputError(model.device.type in ('cpu', 'meta'),
                          "Expect model on device `cpu` or `meta`, "
                          f"but got device type {model.device.type}")
        qtype = ggml_tensor_qtype[low_bit]
        model = ggml_convert_low_bit(model, qtype=qtype, convert_shape_only=True)

    resolved_archive_file, is_sharded = extract_local_archive_file(model_path, subfolder="")
    if is_sharded:
        # For now only shards transformers models
        # can run in this branch.
        resolved_archive_file, _ = \
            get_local_shard_files(model_path,
                                  resolved_archive_file,
                                  subfolder="")
    else:
        resolved_archive_file = [os.path.join(model_path, PYTORCH_MODEL_NAME)]

    for model_file in resolved_archive_file:
        state_dict = torch.load(model_file)
        for param_name, param in state_dict.items():
            set_module_tensor_to_device(model, param_name, "cpu", param)
    return model


def optimize_model(model, low_bit='sym_int4', optimize_llm=True, modules_to_not_convert=None,
                   cpu_embedding=False, lightweight_bmm=False, **kwargs):
    """
    A method to optimize any pytorch model.

    :param model: The original PyTorch model (nn.module)
    :param low_bit: str value, options are ``'sym_int4'``, ``'asym_int4'``, ``'sym_int5'``,
                    ``'asym_int5'``, ``'sym_int8'``, ``'nf3'``, ``'nf4'``, ``'fp4'``,
                    ``'fp8'``, ``'fp8_e4m3'``, ``'fp8_e5m2'``, ``'fp16'`` or ``'bf16'``,
                    ``'sym_int4'`` means symmetric int 4, ``'asym_int4'`` means
                    asymmetric int 4, ``'nf4'`` means 4-bit NormalFloat, etc.
                    Relevant low bit optimizations will be applied to the model.
    :param optimize_llm: Whether to further optimize llm model. Default to be ``True``.
    :param modules_to_not_convert: list of str value, modules (nn.Module) that are skipped
        when conducting model optimizations. Default to be ``None``.
    :param cpu_embedding: Whether to replace the Embedding layer, may need to set it
        to ``True`` when running BigDL-LLM on GPU on Windows. Default to be ``False``.
    :param lightweight_bmm: Whether to replace the torch.bmm ops, may need to set it
        to ``True`` when running BigDL-LLM on GPU on Windows. Default to be ``False``.

    :return: The optimized model.

    >>> # Take OpenAI Whisper model as an example
    >>> from ipex_llm import optimize_model
    >>> model = whisper.load_model('tiny') # Load whisper model under pytorch framework
    >>> model = optimize_model(model) # With only one line code change
    >>> # Use the optimized model without other API change
    >>> result = model.transcribe(audio, verbose=True, language="English")
    >>> # (Optional) you can also save the optimized model by calling 'save_low_bit'
    >>> model.save_low_bit(saved_dir)
    """
    invalidInputError(low_bit in ggml_tensor_qtype,
                      f"Unknown load_in_low_bit value: {low_bit}, expected:"
                      f" sym_int4, asym_int4, sym_int5, asym_int5 or sym_int8.")
    invalidInputError(isinstance(model, torch.nn.Module),
                      "model should be an instance of "
                      f"`torch.nn.Module`, but got {type(model)} at last.")
    # To adapt vLLM models
    if hasattr(model, 'device'):
        invalidInputError(model.device.type in ('cpu', 'meta'),
                          "Expect model on device `cpu` or `meta`, "
                          f"but got device type {model.device.type}")
    if kwargs.pop("replace_embedding", False):
        warnings.warn("replace_embedding is deprecated and will be removed in a future version,"
                      " please use cpu_embedding instead.", FutureWarning)
        cpu_embedding = True
    if low_bit == "fp16":
        torch_dtype = kwargs.get("torch_dtype", None)
        if torch_dtype is not None and torch_dtype != torch.float16:
            invalidInputError(False,
                              "Please use torch_dtype=torch.float16 when setting low_bit='fp16'.")
        else:
            torch_dtype = torch.float16
    else:
        torch_dtype = kwargs.get("torch_dtype", "auto")
    qtype = ggml_tensor_qtype[low_bit]
    model = ggml_convert_low_bit(model,
                                 qtype=qtype,
                                 torch_dtype=torch_dtype,
                                 optimize_model=optimize_llm,
                                 modules_to_not_convert=modules_to_not_convert,
                                 cpu_embedding=cpu_embedding,
                                 lightweight_bmm=lightweight_bmm,
                                 enable_xetla=kwargs.pop("enable_xetla", False))
    # add save_low_bit to pretrained model dynamically
    import types
    model._bigdl_config = dict()
    model._bigdl_config["bigdl_transformers_low_bit"] = low_bit
    model.save_low_bit = types.MethodType(_save_low_bit, model)
    return model