diff --git a/docker/transformers-all-latest-gpu/Dockerfile b/docker/transformers-all-latest-gpu/Dockerfile index 0bdf94ef93e9d7..c1502651f64e38 100644 --- a/docker/transformers-all-latest-gpu/Dockerfile +++ b/docker/transformers-all-latest-gpu/Dockerfile @@ -45,6 +45,9 @@ RUN python3 -m pip install -U "itsdangerous<2.1.0" RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate +# Add bitsandbytes for mixed int8 testing +RUN python3 -m pip install -i https://test.pypi.org/simple/ bitsandbytes==0.31.5 + RUN python3 -m pip install --no-cache-dir decord # When installing in editable mode, `transformers` is not recognized as a package. diff --git a/docs/source/en/main_classes/model.mdx b/docs/source/en/main_classes/model.mdx index fd19b3db52b734..d384f78dc8e177 100644 --- a/docs/source/en/main_classes/model.mdx +++ b/docs/source/en/main_classes/model.mdx @@ -133,6 +133,45 @@ model = AutoModel.from_config(config) Due to Pytorch design, this functionality is only available for floating dtypes. +### `bitsandbytes` integration for Int8 mixed-precision matrix decomposition + +From the paper `GPT3.int8() : 8-bit Matrix Multiplication for Transformers at Scale`, we suport HuggingFace 🤗 integration for all models in the Hub with few lines of code. +For models trained in half-precision (aka, either `float16` or `bfloat16`) or full precision. This method aims to reduce `nn.Linear` size by 2 (if trained in half precision) or by 4 if trained in full precision, without affecting too much quality by operating on the outliers in half-precision. +This technique is useful and works well for billion scale models (>1B parameters) therefore we advice you to use it only for models of that scale. This method has been tested for 2-billion to 176-billion scale models and supports only PyTorch models. + +![HFxbitsandbytes.png](https://s3.amazonaws.com/moonup/production/uploads/1659861207959-62441d1d9fdefb55a0b7d12c.png) + +Int8 mixed-precision matrix decomposition works by separating a matrix multiplication into two streams: (1) and systematic feature outlier stream matrix multiplied in fp16 (0.01%), (2) a regular stream of int8 matrix multiplication (99.9%). With this method, int8 inference with no predictive degradation is possible for very large models (>=176B parameters). +Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], but there are some exceptional systematic outliers that are very differently distributed for large models. These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6, but a lower threshold might be needed for more unstable models (small models, fine-tuning). + +Note also that you would require a GPU to run mixed-8bit models as the kernels has been compiled for GPUs only. Make sure that you have enough GPU RAM to store the quarter (or half if your model is natively in half precision) of the model before using this feature. + +Below are some notes to help you use this module, or follow this demo on Google colab: [![Open In Google Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1qOjXfQIAULfKvZqwCen8-MoWKGdSatZ4?usp=sharing) + +#### Requirements + +- Make sure you run that on a NVIDIA GPU that supports 8-bit tensor cores (Turing or Ampere GPUs - e.g. T4, RTX20s RTX30s, A40-A100). Note that previous generations of NVIDIA GPUs do not support 8-bit tensor cores. +- Install the correct version of `bitsandbytes` by running: +`pip install -i https://test.pypi.org/simple/ bitsandbytes` +- Install `accelerate`: +`pip install accelerate` + +#### Running mixed-int8 models + +After carefully installing the required libraries, the way to load your mixed 8-bit model is as follows: +```py +model_name = "bigscience/bloom-2b5" +model_8bit = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True) +``` +The implementation supports multi-GPU setup thanks to `accelerate` as backend. If you want to control the GPU memory you want to allocate for each GPU, you can use the `max_memory` argument as follows: +(If allocating `1GB` into GPU-0 and `2GB` into GPU-1, you can use `max_memory={0:"1GB", 1:"2GB"}`) +```py +max_memory_mapping = {0: "1GB", 1: "2GB"} +model_name = "bigscience/bloom-3b" +model_8bit = AutoModelForCausalLM.from_pretrained( + model_name, device_map="auto", load_in_8bit=True, max_memory=max_memory_mapping +) +``` ## ModuleUtilsMixin diff --git a/src/transformers/modeling_utils.py b/src/transformers/modeling_utils.py index baade15a9875af..5587e58608a5f5 100644 --- a/src/transformers/modeling_utils.py +++ b/src/transformers/modeling_utils.py @@ -85,10 +85,8 @@ else: get_balanced_memory = None -if is_safetensors_available(): - from safetensors import safe_open - from safetensors.torch import load_file as safe_load_file - from safetensors.torch import save_file as safe_save_file +if is_bitsandbytes_available(): + from .utils.bitsandbytes import get_key_to_not_convert, replace_8bit_linear, set_module_8bit_tensor_to_device logger = logging.get_logger(__name__) @@ -1789,7 +1787,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P https://test.pypi.org/simple/ bitsandbytes-cudaXXX` where XXX is your CUDA version (e.g. 11.6 = 116). Make also sure that you have enough GPU RAM to store half of the model size since the 8bit modules are not compiled and adapted for CPUs. - load_in_8bit_threshold (`float`, *optional*, defaults to 6): + int8_threshold (`float`, *optional*, defaults to 6): Works together with `load_in_8bit`. This corresponds to the outlier threshold for outlier detection as described in `GPT3.int8() : 8-bit Matrix Multiplication for Transformers at Scale` paper. Any hidden states value that is above this threshold will be considered an outlier and the operation on those @@ -1799,9 +1797,6 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P quantization works well for values of magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6, but a lower threshold might be needed for more unstable models (small models, fine-tuning). - load_in_8bit_skip_modules (`List[str]`, *optional*): - An explicit list of the modules that we do not want to convert in 8-bit. This is useful for models such - as Jukebox that has several heads in different places and not necessarily at the last position. subfolder (`str`, *optional*, defaults to `""`): In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can specify the folder name here. @@ -1893,8 +1888,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P offload_folder = kwargs.pop("offload_folder", None) offload_state_dict = kwargs.pop("offload_state_dict", False) load_in_8bit = kwargs.pop("load_in_8bit", False) - load_in_8bit_threshold = kwargs.pop("load_in_8bit_threshold", 6.0) - load_in_8bit_skip_modules = kwargs.pop("load_in_8bit_skip_modules", None) + int8_threshold = kwargs.pop("int8_threshold", 6.0) subfolder = kwargs.pop("subfolder", "") commit_hash = kwargs.pop("_commit_hash", None) @@ -2207,18 +2201,11 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P model = cls(config, *model_args, **model_kwargs) if load_in_8bit: - from .utils.bitsandbytes import get_keys_to_not_convert, replace_8bit_linear - logger.info("Detected 8-bit loading: activating 8-bit loading for this model") - # We keep some modules such as the lm_head in their original dtype for numerical stability reasons - if load_in_8bit_skip_modules is None: - modules_to_not_convert = get_keys_to_not_convert(model) - else: - modules_to_not_convert = load_in_8bit_skip_modules - model = replace_8bit_linear( - model, threshold=load_in_8bit_threshold, modules_to_not_convert=modules_to_not_convert - ) + # We never convert lm_head or any last modules for numerical stability reasons + modules_to_not_convert = get_key_to_not_convert(model) + model = replace_8bit_linear(model, threshold=int8_threshold, modules_to_not_convert=modules_to_not_convert) if isinstance(device_map, str): if model._no_split_modules is None: @@ -2249,18 +2236,12 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P ) if load_in_8bit: - # The LM head / tied weights or any last module can stay on disk / CPU + # The LM head can stay on disk / CPU device_map_without_lm_head = { - key: device_map[key] for key in device_map.keys() if key not in modules_to_not_convert + key: device_map[key] for key in device_map.keys() if key != modules_to_not_convert } if "cpu" in device_map_without_lm_head.values() or "disk" in device_map_without_lm_head.values(): - raise ValueError( - """ - Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit - the quantized model. If you have set a value for `max_memory` you should increase that. To have - an idea of the modules that are set on the CPU or RAM you can print model.hf_device_map. - """ - ) + raise ValueError("8-bit operations on `bitsandbytes` are not supported under CPU!") del device_map_without_lm_head if from_tf: diff --git a/src/transformers/utils/bitsandbytes.py b/src/transformers/utils/bitsandbytes.py index b2339efd626953..ee4e52d421fd09 100644 --- a/src/transformers/utils/bitsandbytes.py +++ b/src/transformers/utils/bitsandbytes.py @@ -1,5 +1,3 @@ -from copy import deepcopy - from transformers.utils import is_accelerate_available, is_bitsandbytes_available @@ -11,7 +9,6 @@ if is_accelerate_available(): from accelerate import init_empty_weights - from accelerate.utils import find_tied_parameters def set_module_8bit_tensor_to_device(module, tensor_name, device, value=None): @@ -114,7 +111,7 @@ def replace_8bit_linear(model, threshold=6.0, modules_to_not_convert="lm_head"): if len(list(module.children())) > 0: replace_8bit_linear(module, threshold, modules_to_not_convert) - if isinstance(module, nn.Linear) and name not in modules_to_not_convert: + if isinstance(module, nn.Linear) and name != modules_to_not_convert: with init_empty_weights(): model._modules[name] = bnb.nn.Linear8bitLt( module.in_features, @@ -126,38 +123,20 @@ def replace_8bit_linear(model, threshold=6.0, modules_to_not_convert="lm_head"): return model -def get_keys_to_not_convert(model): +def get_key_to_not_convert(model): r""" An utility function to get the key of the module to keep in full precision if any For example for CausalLM modules - we may want to keep the lm_head in full precision for numerical stability reasons. For other architectures, we want - to keep the tied weights of the model. The function will return a list of the keys of the modules to not convert in - int8. + we may want to keep the lm_head in full precision for numerical stability reasons. Parameters: model (`torch.nn.Module`): Input model """ - # Create a copy of the model and tie the weights, then - # check if it contains tied weights - tied_model = deepcopy(model) # this has 0 cost since it is done inside `init_empty_weights` context manager` - tied_model.tie_weights() - - tied_keys = list(find_tied_parameters(tied_model).values()) - has_tied_params = len(tied_keys) > 0 - - # Check if it is a base model - is_base_model = not hasattr(model, model.base_model_prefix) - # Ignore this for base models (BertModel, GPT2Model, etc.) - if (not has_tied_params) and is_base_model: + if not hasattr(model, model.base_model_prefix): return "" # otherwise they have an attached head list_modules = list(model.named_parameters()) - list_last_module = [list_modules[-1][0]] - - # add last module together with tied weights - intersection = set(list_last_module) - set(tied_keys) - list_untouched = tied_keys + list(intersection) - - return [module_name.split(".")[0] for module_name in list_untouched] + last_name = list_modules[-1][0] + return last_name.split(".")[0] diff --git a/tests/mixed_int8/README.md b/tests/mixed_int8/README.md index 7a0f86dbb25639..3a188b62658ae7 100644 --- a/tests/mixed_int8/README.md +++ b/tests/mixed_int8/README.md @@ -1,120 +1,37 @@ # Testing mixed int8 quantization -![HFxbitsandbytes.png](https://s3.amazonaws.com/moonup/production/uploads/1660567705337-62441d1d9fdefb55a0b7d12c.png) - -The following is the recipe on how to effectively debug `bitsandbytes` integration on Hugging Face `transformers`. - -## Library requirements - -+ `transformers>=4.22.0` -+ `accelerate>=0.12.0` -+ `bitsandbytes>=0.31.5`. ## Hardware requirements -The following instructions are tested with 2 NVIDIA-Tesla T4 GPUs. To run successfully `bitsandbytes` you would need a 8-bit core tensor supported GPU. Note that Turing, Ampere or newer architectures - e.g. T4, RTX20s RTX30s, A40-A100, A6000 should be supported. +I am using a setup of 2 GPUs that are NVIDIA-Tesla T4 15GB ## Virutal envs -```bash -conda create --name int8-testing python==3.8 -pip install bitsandbytes>=0.31.5 -pip install accelerate>=0.12.0 -pip install transformers>=4.23.0 -``` -if `transformers>=4.23.0` is not released yet, then use: -``` -pip install git+https://github.com/huggingface/transformers.git -``` - -## Troubleshooting - -A list of common errors: +```conda create --name int8-testing python==3.8``` +```git clone https://github.com/younesbelkada/transformers.git && git checkout integration-8bit``` +```pip install -e ".[dev]"``` +```pip install -i https://test.pypi.org/simple/ bitsandbytes``` +```pip install git+https://github.com/huggingface/accelerate.git@e0212893ea6098cc0a7a3c7a6eb286a9104214c1``` -### Torch does not correctly do the operations on GPU -First check that: +## Trobleshooting -```py -import torch - -vec = torch.randn(1, 2, 3).to(0) -``` +```conda create --name int8-testing python==3.8``` +```pip install -i https://test.pypi.org/simple/ bitsandbytes``` +```conda install pytorch torchvision torchaudio -c pytorch``` +```git clone https://github.com/younesbelkada/transformers.git && git checkout integration-8bit``` +```pip install -e ".[dev]"``` +```pip install git+https://github.com/huggingface/accelerate.git@b52b793ea8bac108ba61192eead3cf11ca02433d``` -Works without any error. If not, install torch using `conda` like: +### Check driver settings: -```bash -conda create --name int8-testing python==3.8 -conda install pytorch torchvision torchaudio cudatoolkit=11.6 -c pytorch -c conda-forge -pip install bitsandbytes>=0.31.5 -pip install accelerate>=0.12.0 -pip install transformers>=4.23.0 -``` -For the latest pytorch instructions please see [this](https://pytorch.org/get-started/locally/) - -and the snippet above should work. - -### ` bitsandbytes operations are not supported under CPU!` - -This happens when some Linear weights are set to the CPU when using `accelerate`. Please check carefully `model.hf_device_map` and make sure that there is no `Linear` module that is assigned to CPU. It is fine to have the last module (usually the Lm_head) set on CPU. - -### `To use the type as a Parameter, please correct the detach() semantics defined by __torch_dispatch__() implementation.` - -Use the latest version of `accelerate` with a command such as: `pip install -U accelerate` and the problem should be solved. - -### `Parameter has no attribue .CB` - -Same solution as above. - -### `RuntimeError: CUDA error: an illegal memory access was encountered ... consider passing CUDA_LAUNCH_BLOCKING=1` - -Run your script by pre-pending `CUDA_LAUNCH_BLOCKING=1` and you should observe an error as described in the next section. - -### `CUDA illegal memory error: an illegal memory access at line...`: - -Check the CUDA verisons with: ``` nvcc --version ``` -and confirm it is the same version as the one detected by `bitsandbytes`. If not, run: -``` -ls -l $CONDA_PREFIX/lib/libcudart.so -``` -or -``` -ls -l $LD_LIBRARY_PATH -``` -Check if `libcudart.so` has a correct symlink that is set. Sometimes `nvcc` detects the correct CUDA version but `bitsandbytes` doesn't. You have to make sure that the symlink that is set for the file `libcudart.so` is redirected to the correct CUDA file. - -Here is an example of a badly configured CUDA installation: - -`nvcc --version` gives: - -![Screenshot 2022-08-15 at 15.12.23.png](https://s3.amazonaws.com/moonup/production/uploads/1660569220888-62441d1d9fdefb55a0b7d12c.png) -which means that the detected CUDA version is 11.3 but `bitsandbytes` outputs: - -![image.png](https://s3.amazonaws.com/moonup/production/uploads/1660569284243-62441d1d9fdefb55a0b7d12c.png) - -First check: - -```bash -echo $LD_LIBRARY_PATH ``` - -If this contains multiple paths separated by `:`. Then you have to make sure that the correct CUDA version is set. By doing: - -```bash -ls -l $path/libcudart.so -``` - -On each path (`$path`) separated by `:`. -If not, simply run -```bash -ls -l $LD_LIBRARY_PATH/libcudart.so +ls -l $CONDA_PREFIX/lib/libcudart.so ``` -and you can see - -![Screenshot 2022-08-15 at 15.12.33.png](https://s3.amazonaws.com/moonup/production/uploads/1660569176504-62441d1d9fdefb55a0b7d12c.png) +### Recurrent bugs -If you see that the file is linked to the wrong CUDA version (here 10.2), find the correct location for `libcudart.so` (`find --name libcudart.so`) and replace the environment variable `LD_LIBRARY_PATH` with the one containing the correct `libcudart.so` file. \ No newline at end of file +Sometimes you have to run a "dummy" inference pass when dealing with a multi-GPU setup. Checkout the ```test_multi_gpu_loading``` and the ```test_pipeline``` functions. diff --git a/tests/mixed_int8/test_mixed_int8.py b/tests/mixed_int8/test_mixed_int8.py index 2911d67748809a..0cd7ca16411c19 100644 --- a/tests/mixed_int8/test_mixed_int8.py +++ b/tests/mixed_int8/test_mixed_int8.py @@ -15,14 +15,7 @@ import gc import unittest -from transformers import ( - AutoModel, - AutoModelForCausalLM, - AutoModelForSeq2SeqLM, - AutoModelForSequenceClassification, - AutoTokenizer, - pipeline, -) +from transformers import AutoModel, AutoModelForCausalLM, AutoModelForSequenceClassification, AutoTokenizer, pipeline from transformers.testing_utils import ( is_torch_available, require_accelerate, @@ -113,21 +106,12 @@ def setUp(self): super().setUp() # model_name self.model_name = "bigscience/bloom-560m" - self.seq_to_seq_name = "t5-small" - - # Different types of model - + # Models and tokenizer self.base_model = AutoModel.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto") - # Sequence classification model self.sequence_model = AutoModelForSequenceClassification.from_pretrained( self.model_name, load_in_8bit=True, device_map="auto" ) - # CausalLM model self.model_8bit = AutoModelForCausalLM.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto") - # Seq2seq model - self.seq_to_seq_model = AutoModelForSeq2SeqLM.from_pretrained( - self.seq_to_seq_name, load_in_8bit=True, device_map="auto" - ) def tearDown(self): r""" @@ -137,7 +121,6 @@ def tearDown(self): del self.base_model del self.sequence_model del self.model_8bit - del self.seq_to_seq_model gc.collect() torch.cuda.empty_cache() @@ -155,7 +138,6 @@ def test_correct_head_class(self): # Other heads should be nn.Parameter self.assertTrue(self.model_8bit.lm_head.weight.__class__ == torch.nn.Parameter) self.assertTrue(self.sequence_model.score.weight.__class__ == torch.nn.Parameter) - self.assertTrue(self.seq_to_seq_model.lm_head.weight.__class__ == torch.nn.Parameter) class MixedInt8TestPipeline(BaseMixedInt8Test):