bitsandbytes - Linear8bitLt integration into transformers models

docs/source/en/main_classes/model.mdx

@@ -105,7 +105,7 @@ You can also write your own device map following the same format (a dictionary l
 device_map = {"shared": 0, "encoder": 0, "decoder": 1, "lm_head": 1}
 ```
 
-Another way to minimize the memory impact of your model is to instantiate it at a lower precision dtype (like `torch.float16`).
+Another way to minimize the memory impact of your model is to instantiate it at a lower precision dtype (like `torch.float16`) or use direct quantization techniques as described below.
 
 ### Model Instantiation dtype
 
@@ -133,6 +133,31 @@ model = AutoModel.from_config(config)
 
 Due to Pytorch design, this functionality is only available for floating dtypes.
 
+### Mixed int8 quantization 🪶
+
+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`). This method aims to reduce `nn.Linear` size by 2, without affecting too much quality by operating on the outliers in half-precision. Therefore the method is termed as mixed 8bit quantization.
+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. 
+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 T4 or A100 GPU that supports 8-bit tensor cores. 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-cudaXXX` with `XXX` corresponding to your `CUDA` version. E.g, if you have `CUDA` 11.3 installed on your machine, run `pip install -i https://test.pypi.org/simple/ bitsandbytes-cuda113`
+- 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:
+```
+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. 
 
 
 ## ModuleUtilsMixin

src/transformers/__init__.py

@@ -454,6 +454,7 @@
         "is_vision_available",
         "logging",
     ],
+    "utils.bitsandbytes": [],
 }
 
 # sentencepiece-backed objects

src/transformers/modeling_utils.py

@@ -65,6 +65,7 @@
     has_file,
     hf_bucket_url,
     is_accelerate_available,
+    is_bitsandbytes_available,
     is_offline_mode,
     is_remote_url,
     logging,
@@ -82,6 +83,9 @@
         set_module_tensor_to_device,
     )
 
+if is_bitsandbytes_available():
+    from .utils.bitsandbytes import replace_8bit_linear, set_module_8bit_tensor_to_device
+
 logger = logging.get_logger(__name__)
 
 
@@ -490,6 +494,7 @@ def _load_state_dict_into_meta_model(
     state_dict_folder=None,
     state_dict_index=None,
     dtype=None,
+    load_in_8bit=False,
 ):
     """
     This is somewhat similar to `_load_state_dict_into_model`, but deals with a model that has some or all of its
@@ -548,13 +553,14 @@ def _load_state_dict_into_meta_model(
                 # TODO: group all errors and raise at the end.
                 raise ValueError(f"{param_name} doesn't have any device set.")
             param_device = device_map[module_name]
-
         if param_device == "disk":
             offload_index = offload_weight(param, param_name, offload_folder, offload_index)
         elif param_device == "cpu" and state_dict_index is not None:
             state_dict_index = offload_weight(param, param_name, state_dict_folder, state_dict_index)
-        else:
+        elif not load_in_8bit:
             set_module_tensor_to_device(model, param_name, param_device, value=param)
+        else:
+            set_module_8bit_tensor_to_device(model, param_name, param_device, value=param)
 
     return error_msgs, offload_index, state_dict_index
 
@@ -1564,6 +1570,24 @@ def save_pretrained(
             url = self._push_to_hub(repo, commit_message=commit_message)
             logger.info(f"Model pushed to the hub in this commit: {url}")
 
+    def get_memory_footprint(self, return_buffers=True):
+        r"""
+        Get the memory footprint of a model. This will return the memory footprint of the current model in bytes.
+        Useful to benchmark the memory footprint of the current model and design some tests. Solution inspired from the
+        PyTorch discussions: https://discuss.pytorch.org/t/gpu-memory-that-model-uses/56822/2
+
+        Arguments:
+            return_buffers (`bool`, *optional*):
+                Whether to return the size of the buffer tensors in the computation of the memory footprint. Buffers
+                are tensors that do not require gradients and not registered as parameters. E.g. mean and std in batch
+                norm layers. Please see: https://discuss.pytorch.org/t/what-pytorch-means-by-buffers/120266/2
+        """
+        mem = sum([param.nelement() * param.element_size() for param in self.parameters()])
+        if return_buffers:
+            mem_bufs = sum([buf.nelement() * buf.element_size() for buf in self.buffers()])
+            mem = mem + mem_bufs
+        return mem
+
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
         r"""
@@ -1691,6 +1715,18 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 If `True`, will temporarily offload the CPU state dict to the hard drive to avoid getting out of CPU
                 RAM if the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to
                 `True` when there is some disk offload.
+            load_in_8bit (`bool`, *optional*):
+                If `True`, will convert the loaded model into mixed-8bit quantized model. To use this feature please
+                install `bitsandbytes` compiled with your CUDA version by running `pip install -i
+                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.
+            int8_threshold (`int`, *optional*):
+                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 (#TODO provide
+                link to paper). Any hidden states value that is above this threshold will be considered an outlier and
+                the operation on those values will be done in fp16. A lower value means more outliers will be detected,
+                therefore more operations in fp16. Set to `6.0` by default as described in the paper.
             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.
@@ -1779,9 +1815,10 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
         device_map = kwargs.pop("device_map", None)
         max_memory = kwargs.pop("max_memory", None)
         offload_folder = kwargs.pop("offload_folder", None)
-        offload_state_dict = kwargs.pop("offload_state_dict", None)
+        offload_state_dict = kwargs.pop("offload_state_dict", False)
+        load_in_8bit = kwargs.pop("load_in_8bit", False)
+        int8_threshold = kwargs.pop("int8_threshold", 6.0)
         subfolder = kwargs.pop("subfolder", "")
-
         if device_map is not None:
             if low_cpu_mem_usage is None:
                 low_cpu_mem_usage = True
@@ -1801,6 +1838,30 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                     "Using `low_cpu_mem_usage=True` or a `device_map` requires Accelerate: `pip install accelerate`"
                 )
 
+        if load_in_8bit:
+            if not (is_accelerate_available() and is_bitsandbytes_available()):
+                raise ImportError(
+                    "Using `load_in_8bit=True` requires Accelerate: `pip install accelerate` and the latest version of"
+                    " bitsandbytes `pip install -i https://test.pypi.org/simple/ bitsandbytes-cudaXXX` with `XXX`"
+                    " correspondingto your CUDA version (e.g. `pip install -i https://test.pypi.org/simple/"
+                    " bitsandbytes-cuda113` for CUDA 11.3)"
+                )
+            if torch_dtype == "auto" or torch_dtype != torch.float16:
+                torch_dtype = (
+                    torch.float16
+                )  # We force the `dtype` to be float16, this is a requirement from `bitsandbytes`
+                logger.info("Loading the model in mixed int8 - forcing the weights to be casted in float16")
+            if device_map is None:
+                raise ValueError(
+                    "A device map needs to be passed to run convert models into mixed-int8 format. Please run"
+                    "`.from_pretrained` with `device_map='auto'`"
+                )
+            if from_tf or from_flax:
+                raise ValueError(
+                    "Converting into mixed 8-bit weights from tf/flax weights is currently not supported, please make"
+                    " sure the weights are in PyTorch format."
+                )
+
         from_pt = not (from_tf | from_flax)
 
         user_agent = {"file_type": "model", "framework": "pytorch", "from_auto_class": from_auto_class}
@@ -2083,22 +2144,35 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
 
             logger.info("Detected DeepSpeed ZeRO-3: activating zero.init() for this model")
             init_contexts = [deepspeed.zero.Init(config_dict_or_path=deepspeed_config())] + init_contexts
+        elif load_in_8bit:
+            init_contexts = [init_empty_weights()]  # Force enable init empty weights
+            logger.info("Detected 8-bit loading: activating 8-bit loading for this model")
         elif low_cpu_mem_usage:
             init_contexts.append(init_empty_weights())
 
         with ContextManagers(init_contexts):
             model = cls(config, *model_args, **model_kwargs)
 
+        if load_in_8bit:
+            model = replace_8bit_linear(model, threshold=int8_threshold)
+
         if device_map == "auto":
             if model._no_split_modules is None:
                 raise ValueError(f"{model.__class__.__name__} does not support `device_map='auto'` yet.")
             no_split_modules = model._no_split_modules
             # Make sure tied weights are tied before creating the device map.
             model.tie_weights()
             device_map = infer_auto_device_map(
-                model, no_split_module_classes=no_split_modules, dtype=torch_dtype, max_memory=max_memory
+                model,
+                no_split_module_classes=no_split_modules,
+                dtype=torch_dtype if not load_in_8bit else torch.int8,
+                max_memory=max_memory,
             )
 
+            if load_in_8bit:
+                if "cpu" in device_map.values() or "disk" in device_map.values():
+                    raise ValueError("8-bit operations on `bitsandbytes` are not supported under CPU!")
+
         if from_tf:
             if resolved_archive_file.endswith(".index"):
                 # Load from a TensorFlow 1.X checkpoint - provided by original authors
@@ -2150,6 +2224,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 offload_folder=offload_folder,
                 offload_state_dict=offload_state_dict,
                 dtype=torch_dtype,
+                load_in_8bit=load_in_8bit,
             )
 
         # make sure token embedding weights are still tied if needed
@@ -2190,6 +2265,7 @@ def _load_pretrained_model(
         offload_folder=None,
         offload_state_dict=None,
         dtype=None,
+        load_in_8bit=False,
     ):
         if device_map is not None and "disk" in device_map.values():
             if offload_folder is None:
@@ -2255,7 +2331,10 @@ def _fix_key(key):
                     key = ".".join(key.split(".")[1:])
                 param = model_state_dict[key]
                 if param.device == torch.device("meta"):
-                    set_module_tensor_to_device(model, key, "cpu", torch.empty(*param.size()))
+                    if not load_in_8bit:
+                        set_module_tensor_to_device(model, key, "cpu", torch.empty(*param.size()))
+                    else:
+                        set_module_8bit_tensor_to_device(model, key, "cpu", torch.empty(*param.size()))
 
         # retrieve unintialized modules and initialize before maybe overriding that with the pretrained weights.
         if _fast_init:
@@ -2364,6 +2443,7 @@ def _find_mismatched_keys(
                         state_dict_folder=state_dict_folder,
                         state_dict_index=state_dict_index,
                         dtype=dtype,
+                        load_in_8bit=load_in_8bit,
                     )
                     error_msgs += new_error_msgs
                 else:

src/transformers/utils/bitsandbytes.py

@@ -0,0 +1,117 @@
+from transformers.utils import is_accelerate_available, is_bitsandbytes_available
+
+
+if is_bitsandbytes_available():
+    import torch
+    import torch.nn as nn
+
+    import bitsandbytes as bnb
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+
+
+def set_module_8bit_tensor_to_device(module, tensor_name, device, value=None):
+    """
+    A helper function to set a given tensor (parameter of buffer) of a module on a specific device (note that doing
+    `param.to(device)` creates a new tensor not linked to the parameter, which is why we need this function). The
+    function is adapted from `set_module_tensor_to_device` function from accelerate that is adapted to support the
+    class `Int8Params` from `bitsandbytes`.
+
+    Args:
+        module (`torch.nn.Module`):
+            The module in which the tensor we want to move lives.
+        tensor_name (`str`):
+            The full name of the parameter/buffer.
+        device (`int`, `str` or `torch.device`):
+            The device on which to set the tensor.
+        value (`torch.Tensor`, *optional*):
+            The value of the tensor (useful when going from the meta device to any other device).
+    """
+    # Recurse if needed
+    if "." in tensor_name:
+        splits = tensor_name.split(".")
+        for split in splits[:-1]:
+            new_module = getattr(module, split)
+            if new_module is None:
+                raise ValueError(f"{module} has no attribute {split}.")
+            module = new_module
+        tensor_name = splits[-1]
+
+    if tensor_name not in module._parameters and tensor_name not in module._buffers:
+        raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.")
+    is_buffer = tensor_name in module._buffers
+    old_value = getattr(module, tensor_name)
+
+    if old_value.device == torch.device("meta") and device not in ["meta", torch.device("meta")] and value is None:
+        raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {device}.")
+
+    if is_buffer:
+        has_fp16_weights = None
+    else:
+        has_fp16_weights = getattr(module._parameters[tensor_name], "has_fp16_weights", None)
+
+    if has_fp16_weights is not None:
+        param = module._parameters[tensor_name]
+        if param.device.type != "cuda":
+            if value is None:
+                new_value = old_value.to(device)
+            elif isinstance(value, torch.Tensor):
+                new_value = value.to("cpu")
+                if value.dtype == torch.int8:
+                    raise ValueError(
+                        "You cannot load weights that are saved in int8 using `load_in_8bit=True`, make sure you are",
+                        " using `load_in_8bit=True` on float32/float16/bfloat16 weights.",
+                    )
+            else:
+                new_value = torch.tensor(value, device="cpu")
+            new_value = bnb.nn.Int8Params(new_value, requires_grad=False, has_fp16_weights=has_fp16_weights).to(device)
+            module._parameters[tensor_name] = new_value
+    else:
+        if value is None:
+            new_value = old_value.to(device)
+        elif isinstance(value, torch.Tensor):
+            new_value = value.to(device)
+        else:
+            new_value = torch.tensor(value, device=device)
+
+        if is_buffer:
+            module._buffers[tensor_name] = new_value
+        else:
+            new_value = nn.Parameter(new_value, requires_grad=old_value.requires_grad)
+            module._parameters[tensor_name] = new_value
+
+
+def replace_8bit_linear(model, threshold=6.0):
+    """
+    A helper function to replace all `torch.nn.Linear` modules by `bnb.nn.Linear8bit` modules from the `bitsandbytes`
+    library. This will enable running your models using mixed int8 precision as described by the paper `GPT3.int8():
+    8-bit Matrix Multiplication for Transformers at Scale`. Make sure `bitsandbytes` compiled with the correct CUDA
+    version of your hardware is installed before running this function. `pip install -i https://test.pypi.org/simple/
+    bitsandbytes-cudaXXX` with `XXX` is your CUDA version (e.g., 11.6 = 116)
+
+    The function will be run recursively and replace all `torch.nn.Linear` modules except for the `lm_head` that should
+    be kept as a `torch.nn.Linear` module. The replacement is done under `init_empty_weights` context manager so no
+    CPU/GPU memory is required to run this function.
+
+    Parameters:
+        model (`torch.nn.Module`):
+            Input model or `torch.nn.Module` as the function is run recursively.
+        threshold (`float`, *optional*):
+            `int8_threshold` for outlier detection as described in the formentioned paper. This parameters is set to
+            `6.0` as described by the paper.
+    """
+    for n, module in model.named_children():
+        if len(list(module.children())) > 0:
+            replace_8bit_linear(module, threshold)
+
+        if isinstance(module, nn.Linear) and n != "lm_head":
+            with init_empty_weights():
+                model._modules[n] = bnb.nn.Linear8bitLt(
+                    module.in_features,
+                    module.out_features,
+                    module.bias is not None,
+                    has_fp16_weights=False,
+                    threshold=threshold,
+                )
+    return model

tests/mixed_int8/README.md

@@ -0,0 +1,13 @@
+# Testing mixed int8 quantization
+
+## Hardware requirements
+
+I am using a setup of 2 GPUs that are NVIDIA-Tesla T4 15GB
+
+## Virutal envs
+
+```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-cuda114```
+```pip install git+https://github.com/huggingface/accelerate.git@e0212893ea6098cc0a7a3c7a6eb286a9104214c1```