bitsandbytes - Linear8bitLt integration into transformers models (

huggingface#17901) * first commit * correct replace function * add final changes - works like charm! - cannot implement tests yet - tested * clean up a bit * add bitsandbytes dependencies * working version - added import function - added bitsandbytes utils file * small fix * small fix - fix import issue * fix import issues * Apply suggestions from code review Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * refactor a bit - move bitsandbytes utils to utils - change comments on functions * reformat docstring - reformat docstring on init_empty_weights_8bit * Update src/transformers/__init__.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * revert bad formatting * change to bitsandbytes * refactor a bit - remove init8bit since it is useless * more refactoring - fixed init empty weights issue - added threshold param * small hack to make it work * Update src/transformers/modeling_utils.py * Update src/transformers/modeling_utils.py * revmoe the small hack * modify utils file * make style + refactor a bit * create correctly device map * add correct dtype for device map creation * Apply suggestions from code review Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * apply suggestions - remove with torch.grad - do not rely on Python bool magic! * add docstring - add docstring for new kwargs * add docstring - comment `replace_8bit_linear` function - fix weird formatting * - added more documentation - added new utility function for memory footprint tracking - colab demo to add * few modifs - typo doc - force cast into float16 when load_in_8bit is enabled * added colab link * add test architecture + docstring a bit * refactor a bit testing class * make style + refactor a bit * enhance checks - add more checks - start writing saving test * clean up a bit * male style * add more details on doc * add more tests - still needs to fix 2 tests * replace by "or" - could not fix it from GitHub GUI Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * refactor a bit testing code + add readme * make style * fix import issue * Update src/transformers/modeling_utils.py Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com> * add few comments * add more doctring + make style * more docstring * raise error when loaded in 8bit * make style * add warning if loaded on CPU * add small sanity check * fix small comment * add bitsandbytes on dockerfile * Improve documentation - improve documentation from comments * add few comments * slow tests pass on the VM but not on the CI VM * Fix merge conflict * make style * another test should pass on a multi gpu setup * fix bad import in testing file * Fix slow tests - remove dummy batches - no more CUDA illegal memory errors * odify dockerfile * Update docs/source/en/main_classes/model.mdx * Update Dockerfile * Update model.mdx * Update Dockerfile * Apply suggestions from code review * few modifications - lm head can stay on disk/cpu - change model name so that test pass * change test value - change test value to the correct output - torch bmm changed to baddmm in bloom modeling when merging * modify installation guidelines * Apply suggestions from code review Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * replace `n`by `name` * merge `load_in_8bit` and `low_cpu_mem_usage` * first try - keep the lm head in full precision * better check - check the attribute `base_model_prefix` instead of computing the number of parameters * added more tests * Update src/transformers/utils/bitsandbytes.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Merge branch 'integration-8bit' of https://github.com/younesbelkada/transformers into integration-8bit * improve documentation - fix typos for installation - change title in the documentation Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com>
oneraghavan · Sep 26, 2022 · 7444876 · 7444876
1 parent 489bfb1
commit 7444876
Show file tree

Hide file tree

Showing 9 changed files with 534 additions and 8 deletions.
diff --git 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
@@ -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 +133,45 @@ 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/__init__.py b/src/transformers/__init__.py
@@ -462,6 +462,7 @@
         "is_vision_available",
         "logging",
     ],
+    "utils.bitsandbytes": [],
 }
 
 # sentencepiece-backed objects

diff --git a/src/transformers/modeling_utils.py b/src/transformers/modeling_utils.py
@@ -61,6 +61,7 @@
     copy_func,
     has_file,
     is_accelerate_available,
+    is_bitsandbytes_available,
     is_offline_mode,
     logging,
     replace_return_docstrings,
@@ -83,6 +84,9 @@
     else:
         get_balanced_memory = None
 
+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__)
 
 
@@ -501,6 +505,7 @@ 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
@@ -561,13 +566,14 @@ 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
 
@@ -1578,6 +1584,24 @@ def save_pretrained(
                 save_directory, repo_id, files_timestamps, commit_message=commit_message, token=token
             )
 
+    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*, defaults to `True`):
+                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"""
@@ -1707,6 +1731,22 @@ 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*, defaults to `False`):
+                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 (`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
+                values will be done in fp16. 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).
             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.
@@ -1796,15 +1836,16 @@ 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 trust_remote_code is True:
             logger.warning(
                 "The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is"
                 " ignored."
             )
-
         if device_map is not None:
             if low_cpu_mem_usage is None:
                 low_cpu_mem_usage = True
@@ -1824,6 +1865,28 @@ 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` or"
+                    " pip install bitsandbytes` "
+                )
+            if torch_dtype == "auto" or torch_dtype != torch.float16:
+                # We force the `dtype` to be float16, this is a requirement from `bitsandbytes`
+                torch_dtype = torch.float16
+                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}
@@ -2063,12 +2126,19 @@ 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 low_cpu_mem_usage:
+        elif load_in_8bit or 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:
+            logger.info("Detected 8-bit loading: activating 8-bit loading for this model")
+
+            # 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:
                 raise ValueError(f"{model.__class__.__name__} does not support `device_map='{device_map}'` yet.")
@@ -2091,9 +2161,21 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
             # 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:
+                # 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 != modules_to_not_convert
+                }
+                if "cpu" in device_map_without_lm_head.values() or "disk" in device_map_without_lm_head.values():
+                    raise ValueError("8-bit operations on `bitsandbytes` are not supported under CPU!")
+                del device_map_without_lm_head
+
         if from_tf:
             if resolved_archive_file.endswith(".index"):
                 # Load from a TensorFlow 1.X checkpoint - provided by original authors
@@ -2145,6 +2227,7 @@ 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
@@ -2185,6 +2268,7 @@ 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:
@@ -2250,7 +2334,10 @@ 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:
@@ -2359,6 +2446,7 @@ 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: