add_exllamav2 (huggingface#1419)

* add_exllamav2 * style * fix doc * fix doc * raise error * Update docs/source/llm_quantization/usage_guides/quantization.mdx Co-authored-by: fxmarty <[email protected]> * update doc * update min version of autogptq --------- Co-authored-by: fxmarty <[email protected]>
SunMarc · Oct 24, 2023 · aba7f46 · aba7f46
1 parent f6adbef
commit aba7f46
Show file tree

Hide file tree

Showing 4 changed files with 107 additions and 20 deletions.
diff --git a/docs/source/llm_quantization/usage_guides/quantization.mdx b/docs/source/llm_quantization/usage_guides/quantization.mdx
@@ -76,7 +76,7 @@ quantized_model = load_quantized_model(empty_model, save_folder=save_folder, dev
 
 ### Exllama kernels for faster inference
 
-For 4-bit model, you can use the exllama kernels in order to a faster inference speed. It is activated by default. If you want to change its value, you just need to pass `disable_exllama` in [`~optimum.gptq.load_quantized_model`]. In order to use these kernels, you need to have the entire model on gpus.
+With the release of the exllamav2 kernel, you can get faster inference speed compared to the exllama kernels for 4-bit model. It is activated by default: `disable_exllamav2=False` in [`~optimum.gptq.load_quantized_model`]. In order to use these kernels, you need to have the entire model on gpus.
 
 ```py
 from optimum.gptq import GPTQQuantizer, load_quantized_model
@@ -86,10 +86,25 @@ from accelerate import init_empty_weights
 with init_empty_weights():
     empty_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
 empty_model.tie_weights()
-quantized_model = load_quantized_model(empty_model, save_folder=save_folder, device_map="auto", disable_exllama=False)
+quantized_model = load_quantized_model(empty_model, save_folder=save_folder, device_map="auto")
 ```
 
-Note that only 4-bit models are supported with exllama kernels for now. Furthermore, it is recommended to disable the exllama kernel when you are finetuning your model with peft.
+If you wish to use exllama kernels, you will have to disable the exllamav2 kernel and activate the exllama kernel:
+
+```py
+from optimum.gptq import GPTQQuantizer, load_quantized_model
+import torch
+
+from accelerate import init_empty_weights
+with init_empty_weights():
+    empty_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
+empty_model.tie_weights()
+quantized_model = load_quantized_model(empty_model, save_folder=save_folder, device_map="auto", disable_exllama=False, disable_exllamav2=True)
+```
+
+Note that only 4-bit models are supported with exllama/exllamav2 kernels for now. Furthermore, it is recommended to disable the exllama/exllamav2 kernel when you are finetuning your model with peft.
+
+You can find the benchmark of these kernels [here](https://github.com/huggingface/optimum/tree/main/tests/benchmark#gptq-benchmark)
 
 #### Fine-tune a quantized model
 

diff --git a/optimum/gptq/quantizer.py b/optimum/gptq/quantizer.py
@@ -69,7 +69,8 @@ def __init__(
         module_name_preceding_first_block: Optional[List[str]] = None,
         batch_size: int = 1,
         pad_token_id: Optional[int] = None,
-        disable_exllama: bool = False,
+        disable_exllama: bool = True,
+        disable_exllamav2: bool = False,
         max_input_length: Optional[int] = None,
         *args,
         **kwargs,
@@ -107,8 +108,10 @@ def __init__(
                 The batch size of the dataset
             pad_token_id (`Optional[int]`, defaults to `None`):
                 The pad token id. Needed to prepare the dataset when `batch_size` > 1.
-            disable_exllama (`bool`, defaults to `False`):
+            disable_exllama (`bool`, defaults to `True`):
                 Whether to use exllama backend. Only works with `bits` = 4.
+            disable_exllamav2 (`bool`, defaults to `False`):
+                Whether to use exllamav2 backend. Only works with `bits` = 4.
             max_input_length (`Optional[int]`, defaults to `None`):
                 The maximum input length. This is needed to initialize a buffer that depends on the maximum expected input length.
                 It is specific to the exllama backend with act-order.
@@ -128,6 +131,7 @@ def __init__(
         self.batch_size = batch_size
         self.pad_token_id = pad_token_id
         self.disable_exllama = disable_exllama
+        self.disable_exllamav2 = disable_exllamav2
         self.max_input_length = max_input_length
         self.quant_method = QuantizationMethod.GPTQ
 
@@ -137,6 +141,10 @@ def __init__(
             raise ValueError("group_size must be greater than 0 or equal to -1")
         if not (0 < self.damp_percent < 1):
             raise ValueError("damp_percent must between 0 and 1.")
+        if not self.disable_exllamav2 and not self.disable_exllama:
+            raise ValueError(
+                "disable_exllamav2 and disable_exllama are both set to `False`. Please disable one of the kernels."
+            )
 
     def to_dict(self):
         """
@@ -205,6 +213,7 @@ def _replace_by_quant_layers(self, module: nn.Module, names: List[str], name: st
             group_size=self.group_size,
             bits=self.bits,
             disable_exllama=self.disable_exllama,
+            disable_exllamav2=self.disable_exllamav2,
         )
         if isinstance(module, QuantLinear):
             return
@@ -440,13 +449,21 @@ def tmp(_, input, output):
             layer_inputs, layer_outputs = layer_outputs, []
             torch.cuda.empty_cache()
 
-        if self.bits == 4 and not self.disable_exllama:
+        if self.bits == 4:
+            # device not on gpu
             if device == torch.device("cpu") or (has_device_map and any(d in devices for d in ["cpu", "disk"])):
-                logger.warning(
-                    "Found modules on cpu/disk. Using Exllama backend requires all the modules to be on GPU. Setting `disable_exllama=True`"
-                )
-                self.disable_exllama = True
-            elif self.desc_act:
+                if not self.disable_exllama:
+                    logger.warning(
+                        "Found modules on cpu/disk. Using Exllama backend requires all the modules to be on GPU. Setting `disable_exllama=True`"
+                    )
+                    self.disable_exllama = True
+                if not self.disable_exllamav2:
+                    logger.warning(
+                        "Found modules on cpu/disk. Using Exllamav2 backend requires all the modules to be on GPU. Setting `disable_exllamav2=True`"
+                    )
+                    self.disable_exllamav2 = True
+            # act order and exllama
+            elif self.desc_act and not self.disable_exllama:
                 logger.warning(
                     "Using Exllama backend with act_order will reorder the weights offline, thus you will not be able to save the model with the right weights."
                     "Setting `disable_exllama=True`. You should only use Exllama backend with act_order for inference. "
@@ -475,13 +492,13 @@ def post_init_model(self, model):
             model (`nn.Module`):
                 The input model
         """
-        if self.bits == 4 and not self.disable_exllama:
+        if self.bits == 4 and (not self.disable_exllama or not self.disable_exllamav2):
             if get_device(model) == torch.device("cpu") or (
                 hasattr(model, "hf_device_map") and any(d in model.hf_device_map for d in ["cpu", "disk"])
             ):
                 raise ValueError(
-                    "Found modules on cpu/disk. Using Exllama backend requires all the modules to be on GPU."
-                    "You can deactivate exllama backend by setting `disable_exllama=True` in the quantization config object"
+                    "Found modules on cpu/disk. Using Exllama or Exllamav2 backend requires all the modules to be on GPU."
+                    "You can deactivate exllama backend by setting `disable_exllama=True` or `disable_exllamav2=True` in the quantization config object"
                 )
 
         class StoreAttr(object):
@@ -514,6 +531,7 @@ def pack_model(
             group_size=self.group_size,
             bits=self.bits,
             disable_exllama=self.disable_exllama,
+            disable_exllamav2=self.disable_exllamav2,
         )
         logger.info("Packing model...")
         layers = get_layers(model)
@@ -579,7 +597,8 @@ def load_quantized_model(
     offload_folder: Optional[str] = None,
     offload_buffers: Optional[str] = None,
     offload_state_dict: bool = False,
-    disable_exllama: bool = False,
+    disable_exllama: bool = True,
+    disable_exllamav2: bool = False,
     max_input_length: Optional[int] = None,
 ):
     """
@@ -615,6 +634,8 @@ def load_quantized_model(
             picked contains `"disk"` values.
         disable_exllama (`bool`, defaults to `False`):
             Whether to use exllama backend. Only works with `bits` = 4.
+        disable_exllama (`bool`, defaults to `False`):
+            Whether to use exllamav2 backend. Only works with `bits` = 4.
         max_input_length (`Optional[int]`, defaults to `None`):
             The maximum input length. This is needed to initialize a buffer that depends on the maximum expected input length.
             It is specific to the exllama backend with act-order.
@@ -648,6 +669,7 @@ def load_quantized_model(
         ) from err
     quantizer = GPTQQuantizer.from_dict(quantize_config_dict)
     quantizer.disable_exllama = disable_exllama
+    quantizer.disable_exllamav2 = disable_exllamav2
     quantizer.max_input_length = max_input_length
 
     model = quantizer.convert_model(model)

diff --git a/optimum/utils/import_utils.py b/optimum/utils/import_utils.py
@@ -110,11 +110,11 @@ def is_timm_available():
 def is_auto_gptq_available():
     if _auto_gptq_available:
         version_autogptq = packaging.version.parse(importlib_metadata.version("auto_gptq"))
-        if AUTOGPTQ_MINIMUM_VERSION <= version_autogptq:
+        if AUTOGPTQ_MINIMUM_VERSION < version_autogptq:
             return True
         else:
             raise ImportError(
-                f"Found an incompatible version of auto-gptq. Found version {version_autogptq}, but only {AUTOGPTQ_MINIMUM_VERSION} and above are supported"
+                f"Found an incompatible version of auto-gptq. Found version {version_autogptq}, but only version above {AUTOGPTQ_MINIMUM_VERSION} are supported"
             )
 
 

diff --git a/tests/gptq/test_quantization.py b/tests/gptq/test_quantization.py
@@ -46,6 +46,7 @@ class GPTQTest(unittest.TestCase):
     group_size = 128
     desc_act = False
     disable_exllama = True
+    disable_exllamav2 = True
 
     dataset = [
         "auto-gptq is an easy-to-use model quantization library with user-friendly apis, based on GPTQ algorithm."
@@ -69,6 +70,7 @@ def setUpClass(cls):
             group_size=cls.group_size,
             desc_act=cls.desc_act,
             disable_exllama=cls.disable_exllama,
+            disable_exllamav2=cls.disable_exllamav2,
         )
 
         cls.quantized_model = cls.quantizer.quantize_model(cls.model_fp16, cls.tokenizer)
@@ -96,6 +98,7 @@ def test_quantized_layers_class(self):
             group_size=self.group_size,
             bits=self.bits,
             disable_exllama=self.disable_exllama,
+            disable_exllamav2=self.disable_exllamav2,
         )
         self.assertTrue(self.quantized_model.transformer.h[0].mlp.dense_4h_to_h.__class__ == QuantLinear)
 
@@ -133,13 +136,18 @@ def test_serialization(self):
                 )
             empty_model.tie_weights()
             quantized_model_from_saved = load_quantized_model(
-                empty_model, save_folder=tmpdirname, device_map={"": 0}, disable_exllama=self.disable_exllama
+                empty_model,
+                save_folder=tmpdirname,
+                device_map={"": 0},
+                disable_exllama=self.disable_exllama,
+                disable_exllamav2=self.disable_exllamav2,
             )
             self.check_inference_correctness(quantized_model_from_saved)
 
 
 class GPTQTestExllama(GPTQTest):
     disable_exllama = False
+    disable_exllamav2 = True
     EXPECTED_OUTPUTS = set()
     EXPECTED_OUTPUTS.add("Hello my name is John, I am a professional photographer and I")
     EXPECTED_OUTPUTS.add("Hello my name is jay and i am a student at university.")
@@ -153,6 +161,7 @@ class GPTQTestActOrder(GPTQTest):
     EXPECTED_OUTPUTS.add("Hello my name is nathalie, I am a young girl from")
 
     disable_exllama = True
+    disable_exllamav2 = True
     desc_act = True
 
     def test_generate_quality(self):
@@ -178,7 +187,7 @@ def test_exllama_serialization(self):
                 )
             empty_model.tie_weights()
             quantized_model_from_saved = load_quantized_model(
-                empty_model, save_folder=tmpdirname, device_map={"": 0}, disable_exllama=False
+                empty_model, save_folder=tmpdirname, device_map={"": 0}, disable_exllama=False, disable_exllamav2=True
             )
             self.check_inference_correctness(quantized_model_from_saved)
 
@@ -197,7 +206,12 @@ def test_exllama_max_input_length(self):
                 )
             empty_model.tie_weights()
             quantized_model_from_saved = load_quantized_model(
-                empty_model, save_folder=tmpdirname, device_map={"": 0}, disable_exllama=False, max_input_length=4028
+                empty_model,
+                save_folder=tmpdirname,
+                device_map={"": 0},
+                disable_exllama=False,
+                max_input_length=4028,
+                disable_exllamav2=True,
             )
 
             prompt = "I am in Paris and" * 1000
@@ -213,6 +227,42 @@ def test_exllama_max_input_length(self):
             quantized_model_from_saved.generate(**inp, num_beams=1, min_new_tokens=3, max_new_tokens=3)
 
 
+class GPTQTestExllamav2(GPTQTest):
+    desc_act = False
+    disable_exllama = True
+    disable_exllamav2 = True
+
+    def test_generate_quality(self):
+        # don't need to test
+        pass
+
+    def test_serialization(self):
+        # don't need to test
+        pass
+
+    def test_exllama_serialization(self):
+        """
+        Test the serialization of the model and the loading of the quantized weights with exllamav2 kernel
+        """
+        from accelerate import init_empty_weights
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            self.quantizer.save(self.quantized_model, tmpdirname)
+            self.quantized_model.config.save_pretrained(tmpdirname)
+            with init_empty_weights():
+                empty_model = AutoModelForCausalLM.from_config(
+                    AutoConfig.from_pretrained(self.model_name), torch_dtype=torch.float16
+                )
+            empty_model.tie_weights()
+            quantized_model_from_saved = load_quantized_model(
+                empty_model,
+                save_folder=tmpdirname,
+                device_map={"": 0},
+                disable_exllamav2=False,
+            )
+            self.check_inference_correctness(quantized_model_from_saved)
+
+
 class GPTQUtilsTest(unittest.TestCase):
     """
     Test utilities