Add new quant method (huggingface#32047)

* Add new quant method

* update

* fix multi-device

* add test

* add offload

* style

* style

* add simple example

* initial doc

* docstring

* style again

* works ?

* better docs

* switch to non persistant

* remove print

* fix init

* code review

docs/source/en/_toctree.yml

@@ -157,6 +157,8 @@
     title: EETQ
   - local: quantization/hqq
     title: HQQ
+  - local: quantization/fbgemm_fp8
+    title: FBGEMM_FP8
   - local: quantization/optimum
     title: Optimum
   - local: quantization/contribute

docs/source/en/main_classes/quantization.md

@@ -56,3 +56,8 @@ Learn how to quantize models in the [Quantization](../quantization) guide.
 ## HqqConfig
 
 [[autodoc]] HqqConfig
+
+## FbgemmFp8Config
+
+[[autodoc]] FbgemmFp8Config
+

docs/source/en/quantization/fbgemm_fp8.md

@@ -0,0 +1,58 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# FBGEMM FP8
+
+With FBGEMM FP8 quantization method, you can quantize your model in FP8 (W8A8):
+- the weights will be quantized in 8bit (FP8) per channel
+- the activation will be quantized in 8bit (FP8) per token
+
+It relies on the [FBGEMM](https://github.com/pytorch/FBGEMM) library which provides efficient low-precision general matrix multiplication for small batch sizes and support for accuracy-loss minimizing techniques such as row-wise quantization and outlier-aware quantization. 
+
+> [!TIP]
+> You need a GPU with compute capability>=9 (e.g. H100) 
+
+Before you begin, make sure the following libraries are installed with their latest version:
+
+```bash
+pip install --upgrade accelerate fbgemm-gpu torch
+```
+
+If you are having issues with fbgemm-gpu and torch library, you might need to install the nighlty release. You can follow the instruction [here](https://pytorch.org/FBGEMM/fbgemm_gpu-development/InstallationInstructions.html#fbgemm-gpu-install-libraries:~:text=found%20here.-,Install%20the%20FBGEMM_GPU%20Package,-Install%20through%20PyTorch)
+
+
+```py
+from transformers import FbgemmFp8Config, AutoModelForCausalLM, AutoTokenizer
+
+model_name = "meta-llama/Meta-Llama-3-8B"
+quantization_config = FbgemmFp8Config()
+quantized_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", quantization_config=quantization_config)
+
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+output = quantized_model.generate(**input_ids, max_new_tokens=10)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+A quantized model can be saved via "saved_pretrained" and be reused again via the "from_pretrained".
+
+```py
+quant_path = "/path/to/save/quantized/model"
+model.save_pretrained(quant_path)
+model = AutoModelForCausalLM.from_pretrained(quant_path, device_map="auto")
+```

docs/source/en/quantization/overview.md

@@ -55,4 +55,5 @@ Use the table below to help you decide which quantization method to use.
 | [GPTQ](./gptq)                                | 🔴                       | 🔴   | 🟢        | 🟢              | 🔴                     | 🔴                       | 2 - 3 - 4 - 8          | 🟢                                   | 🟢            | 🟢                      | https://github.com/AutoGPTQ/AutoGPTQ        |
 | [HQQ](./hqq)                                 | 🟢                       | 🟢    | 🟢        | 🔴              | 🔴                     | 🟢                       | 1 - 8          | 🟢                                   | 🔴            | 🟢                      | https://github.com/mobiusml/hqq/            |
 | [Quanto](./quanto)                              | 🟢                       | 🟢   | 🟢        | 🔴              | 🟢                     | 🟢                       | 2 / 4 / 8      | 🔴                                   | 🔴            | 🟢                      | https://github.com/huggingface/quanto       |
+| [FBGEMM_FP8](./fbgemm_fp8.md)                              | 🟢                       | 🔴    | 🟢        | 🔴              | 🔴                      | 🔴                        | 8      | 🔴                                   | 🟢            | 🟢                      | https://github.com/pytorch/FBGEMM       |
 

src/transformers/__init__.py

@@ -934,6 +934,7 @@
         "AwqConfig",
         "BitsAndBytesConfig",
         "EetqConfig",
+        "FbgemmFp8Config",
         "GPTQConfig",
         "HqqConfig",
         "QuantoConfig",
@@ -5671,6 +5672,7 @@
         AwqConfig,
         BitsAndBytesConfig,
         EetqConfig,
+        FbgemmFp8Config,
         GPTQConfig,
         HqqConfig,
         QuantoConfig,

src/transformers/integrations/__init__.py

@@ -45,6 +45,7 @@
         "unset_hf_deepspeed_config",
     ],
     "eetq": ["replace_with_eetq_linear"],
+    "fbgemm_fp8": ["FbgemmFp8Linear", "replace_with_fbgemm_fp8_linear"],
     "ggml": [
         "GGUF_CONFIG_MAPPING",
         "GGUF_TENSOR_MAPPING",
@@ -126,6 +127,7 @@
         unset_hf_deepspeed_config,
     )
     from .eetq import replace_with_eetq_linear
+    from .fbgemm_fp8 import FbgemmFp8Linear, replace_with_fbgemm_fp8_linear
     from .ggml import (
         GGUF_CONFIG_MAPPING,
         GGUF_TENSOR_MAPPING,

src/transformers/integrations/fbgemm_fp8.py

@@ -0,0 +1,161 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# 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.
+
+from ..utils import is_accelerate_available, is_fbgemm_gpu_available, is_torch_available, logging
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+
+if is_fbgemm_gpu_available():
+    import fbgemm_gpu.experimental.gen_ai  # noqa: F401
+
+logger = logging.get_logger(__name__)
+
+
+class FbgemmFp8Linear(torch.nn.Module):
+    def __init__(self, in_features, out_features, bias, weight_dtype=torch.float32):
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+
+        self.register_buffer("weight", torch.zeros((out_features, in_features), dtype=torch.float8_e4m3fn))
+        self.register_buffer("weight_scale", torch.zeros((out_features, 1), dtype=weight_dtype))
+        self.register_buffer("input_scale_ub", torch.zeros([1], dtype=torch.float), persistent=False)
+
+        if bias:
+            self.register_buffer("bias", torch.zeros((self.out_features), dtype=weight_dtype))
+        else:
+            self.bias = None
+
+    def forward(self, x):
+        num_tokens = None
+        # x_quantized and x_scale are not necessarily on the same device as x, this is an issue.
+        # https://github.com/pytorch/FBGEMM/blob/e08af8539c391437f447173863df0f3f6f6f1855/fbgemm_gpu/experimental/gen_ai/src/quantize/quantize.cu#L1237C3-L1237C45
+        x_quantized, x_scale = torch.ops.fbgemm.quantize_fp8_per_row(
+            x.view(-1, x.shape[-1]), num_tokens, self.input_scale_ub
+        )
+        # moving x_quantized, x_scale here creates glibberish output ... However, if we move the output, it works
+        # x_quantized, x_scale = x_quantized.to(x.device), x_scale.to(x.device)
+
+        # The computation still happens on the device where self.weight is even if x_quantized is not on the same device as self.weight
+        output = torch.ops.fbgemm.f8f8bf16_rowwise(
+            x_quantized, self.weight, x_scale, self.weight_scale, use_fast_accum=True
+        )
+        output = output + self.bias if self.bias is not None else output
+        # Hacky for now, we have the output to the device of x
+        output = output.to(x.device)
+        del x_quantized, x_scale
+        return output
+
+
+def _replace_with_fbgemm_fp8_linear(
+    model,
+    modules_to_not_convert=None,
+    current_key_name=None,
+    quantization_config=None,
+    has_been_replaced=False,
+    pre_quantized=False,
+):
+    """
+    Private method that wraps the recursion for module replacement.
+
+    Returns the converted model and a boolean that indicates if the conversion has been successfull or not.
+    """
+    if current_key_name is None:
+        current_key_name = []
+
+    for name, module in model.named_children():
+        current_key_name.append(name)
+
+        if (isinstance(module, nn.Linear)) and name not in modules_to_not_convert:
+            # Check if the current key is not in the `modules_to_not_convert`
+            current_key_name_str = ".".join(current_key_name)
+            if not any(
+                (key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert
+            ):
+                with init_empty_weights(include_buffers=True):
+                    in_features = module.in_features
+                    out_features = module.out_features
+                    model._modules[name] = FbgemmFp8Linear(
+                        in_features,
+                        out_features,
+                        module.bias is not None,
+                    )
+                    has_been_replaced = True
+
+                    # Force requires grad to False to avoid unexpected errors
+                    model._modules[name].requires_grad_(False)
+                # set non persistant buffer outside of init_empty_weights
+                model._modules[name].input_scale_ub = torch.tensor(
+                    [quantization_config.activation_scale_ub], dtype=torch.float
+                )
+        if len(list(module.children())) > 0:
+            _, has_been_replaced = _replace_with_fbgemm_fp8_linear(
+                module,
+                modules_to_not_convert,
+                current_key_name,
+                quantization_config,
+                has_been_replaced=has_been_replaced,
+                pre_quantized=pre_quantized,
+            )
+        # Remove the last key for recursion
+        current_key_name.pop(-1)
+    return model, has_been_replaced
+
+
+def replace_with_fbgemm_fp8_linear(
+    model, modules_to_not_convert=None, current_key_name=None, quantization_config=None, pre_quantized=False
+):
+    """
+    A helper function to replace all `torch.nn.Linear` modules by `FbgemmFp8Linear` modules.
+    This will enable running your models using high performance fp8 kernel from FBGEMM library.
+
+    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. Each weight will be quantized along the channel.
+
+    Parameters:
+        model (`torch.nn.Module`):
+            Input model or `torch.nn.Module` as the function is run recursively.
+        modules_to_not_convert (`List[`str`]`, *optional*, defaults to `["lm_head"]`):
+            Names of the modules to not convert in `FP8Linear`. In practice we keep the `lm_head` in full precision
+            for numerical stability reasons.
+        current_key_name (`List[`str`]`, *optional*):
+            An array to track the current key of the recursion. This is used to check whether the current key (part of
+            it) is not in the list of modules to not convert (for instances modules that are offloaded to `cpu` or
+            `disk`).
+    """
+
+    modules_to_not_convert = ["lm_head"] if modules_to_not_convert is None else modules_to_not_convert
+
+    if quantization_config.modules_to_not_convert is not None:
+        modules_to_not_convert.extend(quantization_config.modules_to_not_convert)
+    modules_to_not_convert = list(set(modules_to_not_convert))
+    model, has_been_replaced = _replace_with_fbgemm_fp8_linear(
+        model, modules_to_not_convert, current_key_name, quantization_config, pre_quantized=pre_quantized
+    )
+
+    if not has_been_replaced:
+        logger.warning(
+            "You are loading your model using FP8 quantization but no linear modules were found in your model."
+            " Please double check your model architecture, or submit an issue on github if you think this is"
+            " a bug."
+        )
+
+    return model

src/transformers/modeling_utils.py

@@ -868,7 +868,7 @@ def _load_state_dict_into_meta_model(
 
         # We convert floating dtypes to the `dtype` passed. We want to keep the buffers/params
         # in int/uint/bool and not cast them.
-        if dtype is not None and torch.is_floating_point(param):
+        if dtype is not None and torch.is_floating_point(param) and param.dtype != torch.float8_e4m3fn:
             if (
                 keep_in_fp32_modules is not None
                 and any(
@@ -894,7 +894,6 @@ def _load_state_dict_into_meta_model(
             old_param = getattr(old_param, split)
             if old_param is None:
                 break
-
         if old_param is not None:
             if dtype is None:
                 param = param.to(old_param.dtype)
@@ -3955,6 +3954,14 @@ def from_pretrained(
                 and hf_quantizer.quantization_config.quant_method == QuantizationMethod.HQQ
             ):
                 device_map_kwargs["force_hooks"] = True
+            if (
+                hf_quantizer is not None
+                and hf_quantizer.quantization_config.quant_method == QuantizationMethod.FBGEMM_FP8
+                and isinstance(device_map, dict)
+                and ("cpu" in device_map.values() or "disk" in device_map.values())
+            ):
+                device_map_kwargs["offload_buffers"] = True
+
             if not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
                 dispatch_model(model, **device_map_kwargs)
 
@@ -4105,7 +4112,6 @@ def _fix_key(key):
         if cls._keys_to_ignore_on_load_unexpected is not None:
             for pat in cls._keys_to_ignore_on_load_unexpected:
                 unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
-
         if hf_quantizer is not None:
             missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix)
 

src/transformers/quantizers/auto.py

@@ -20,6 +20,7 @@
     AwqConfig,
     BitsAndBytesConfig,
     EetqConfig,
+    FbgemmFp8Config,
     GPTQConfig,
     HqqConfig,
     QuantizationConfigMixin,
@@ -31,6 +32,7 @@
 from .quantizer_bnb_4bit import Bnb4BitHfQuantizer
 from .quantizer_bnb_8bit import Bnb8BitHfQuantizer
 from .quantizer_eetq import EetqHfQuantizer
+from .quantizer_fbgemm_fp8 import FbgemmFp8HfQuantizer
 from .quantizer_gptq import GptqHfQuantizer
 from .quantizer_hqq import HqqHfQuantizer
 from .quantizer_quanto import QuantoHfQuantizer
@@ -45,6 +47,7 @@
     "quanto": QuantoHfQuantizer,
     "eetq": EetqHfQuantizer,
     "hqq": HqqHfQuantizer,
+    "fbgemm_fp8": FbgemmFp8HfQuantizer,
 }
 
 AUTO_QUANTIZATION_CONFIG_MAPPING = {
@@ -56,6 +59,7 @@
     "aqlm": AqlmConfig,
     "quanto": QuantoConfig,
     "hqq": HqqConfig,
+    "fbgemm_fp8": FbgemmFp8Config,
 }
 
 
@@ -156,8 +160,11 @@ def merge_quantization_configs(
         if isinstance(quantization_config, dict):
             quantization_config = AutoQuantizationConfig.from_dict(quantization_config)
 
-        if isinstance(quantization_config, (GPTQConfig, AwqConfig)) and quantization_config_from_args is not None:
-            # special case for GPTQ / AWQ config collision
+        if (
+            isinstance(quantization_config, (GPTQConfig, AwqConfig, FbgemmFp8Config))
+            and quantization_config_from_args is not None
+        ):
+            # special case for GPTQ / AWQ / FbgemmFp8 config collision
             loading_attr_dict = quantization_config_from_args.get_loading_attributes()
             for attr, val in loading_attr_dict.items():
                 setattr(quantization_config, attr, val)