improve code format

auto_gptq/modeling/_base.py

@@ -17,11 +17,11 @@ from transformers import AutoConfig, AutoModelForCausalLM, PreTrainedModel
 from transformers.utils.hub import PushToHubMixin, cached_file, create_repo, create_commit, CommitOperationAdd
 from transformers.utils.generic import ContextManagers
 from transformers.modeling_utils import no_init_weights
+from xformers.ops.fmha import AttentionOp
 
 from ._const import *
 from ._utils import *
 from ..nn_modules.qlinear import GeneralQuantLinear
-from ..nn_modules._fused_base import FusedBaseAttentionModule, FusedBaseMLPModule
 from ..quantization import GPTQ
 from ..utils.data_utils import collate_data
 from ..utils.import_utils import (
@@ -73,7 +73,7 @@ class BaseQuantizeConfig(PushToHubMixin):
         quantize_config_filename = "quantize_config.json"
         if os.path.isdir(save_dir):  # Local
             resolved_config_file = join(save_dir, quantize_config_filename)
-        else: # Remote
+        else:  # Remote
             resolved_config_file = cached_file(
                 save_dir,
                 quantize_config_filename,
@@ -89,10 +89,9 @@ class BaseQuantizeConfig(PushToHubMixin):
                 _raise_exceptions_for_connection_errors=False,
                 _commit_hash=commit_hash,
             )
-        
         with open(resolved_config_file, "r", encoding="utf-8") as f:
             return cls(**json.load(f))
-                
+
     def to_dict(self):
         return {
             "bits": self.bits,
@@ -114,9 +113,6 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
     inside_layer_modules: List[List[str]] = None
     lm_head_name: str = "lm_head"
 
-    fused_attn_module_type: Optional[FusedBaseAttentionModule] = None
-    fused_mlp_module_type: Optional[FusedBaseMLPModule] = None
-
     def __init__(
         self,
         model: PreTrainedModel,
@@ -210,14 +206,14 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         if self.quantized:
             raise EnvironmentError("can't execute quantize because the model is quantized.")
         if use_triton and not TRITON_AVAILABLE:
-            logger.warning("triton is not installed, reset use_triton to False")
+            logger.warning("Triton is not installed, reset use_triton to False")
             use_triton = False
 
         device_map = self.hf_device_map
         if device_map:
             for name, device in device_map.items():
                 if device == "cpu":
-                    logger.info(f"truly offloading {name} to cpu with hook.")
+                    logger.info(f"Truly offloading {name} to cpu with hook.")
                     module = get_module_by_name_suffix(self.model, name)
                     remove_hook_from_module(module, recurse=True)
                     accelerate.cpu_offload_with_hook(module, CUDA_0)
@@ -437,10 +433,10 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
     def forward(self, *args, **kwargs):
         return self.model(*args, **kwargs)
 
-    def generate(self, **kwargs):
+    def generate(self, *args, **kwargs):
         """shortcut for model.generate"""
-        with torch.inference_mode(), torch.amp.autocast(device_type=self.device.type):
+        with torch.no_grad(), torch.amp.autocast(device_type=self.device.type):
-            return self.model.generate(**kwargs)
+            return self.model.generate(*args, **kwargs)
 
     def prepare_inputs_for_generation(self, *args, **kwargs):
         """shortcut for model.prepare_inputs_for_generation"""
@@ -489,9 +485,14 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
             create_pr (`bool`, *optional*, defaults to `False`):
                 Whether or not to create a PR with the uploaded files or directly commit.
         """
-        if (self.quantize_config.model_name_or_path is None or not isdir(self.quantize_config.model_name_or_path)) and save_dir is None:
+        if (
-            raise ValueError("Quantized model should be saved first, or you can provide save_dir to make sure model is saved to local disk before uploading.")
+            self.quantize_config.model_name_or_path is None or not isdir(self.quantize_config.model_name_or_path)
-        
+        ) and save_dir is None:
+            raise ValueError(
+                "Quantized model should be saved first, or you can provide save_dir to "
+                "make sure model is saved to local disk before uploading."
+            )
+
         if save_dir is not None:
             logger.info(f"Saving model to {save_dir}")
             self.save_quantized(save_dir, use_safetensors, safetensors_metadata)
@@ -517,16 +518,30 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                 repo_type="model",
             )
 
-    def save_quantized(self, save_dir: str, use_safetensors: bool = False, safetensors_metadata: Optional[Dict[str, str]] = None):
+    def save_quantized(
+        self,
+        save_dir: str,
+        use_safetensors: bool = False,
+        safetensors_metadata: Optional[Dict[str, str]] = None
+    ):
         """save quantized model and configs to local disk"""
         os.makedirs(save_dir, exist_ok=True)
 
         if not self.quantized:
-            raise EnvironmentError("can only save quantized model, please execute .quantize first.")
+            raise TypeError("Can only save quantized model, please execute .quantize() method first.")
+        if self.injected_fused_attention or self.injected_fused_mlp:
+            raise TypeError(
+                "At least one of attention modules and mlp modules are injected with fused ops, "
+                "please disable 'inject_fused_attention' and 'inject_fused_mlp' at model loading stage, "
+                "and don't call ._fuse_attention() and ._fuse_mlp() methods before calling this method."
+            )
 
         self.model.to(CPU)
 
-        model_base_name = self.quantize_config.model_file_base_name or f"gptq_model-{self.quantize_config.bits}bit-{self.quantize_config.group_size}g"
+        model_base_name = (
+            self.quantize_config.model_file_base_name or
+            f"gptq_model-{self.quantize_config.bits}bit-{self.quantize_config.group_size}g"
+        )
         if use_safetensors:
             model_save_name = model_base_name + ".safetensors"
             state_dict = self.model.state_dict()
@@ -546,13 +561,23 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                             new_key = str(key)
                             new_value = str(value)
                         except Exception as e:
-                            raise TypeError(f"safetensors_metadata: both keys and values must be strings and an error occured when trying to convert them: {e}")
+                            raise TypeError(
+                                f"safetensors_metadata: both keys and values must be strings and "
+                                f"an error occured when trying to convert them: {e}"
+                            )
                         if new_key in new_safetensors_metadata:
-                            logger.warning(f"After converting safetensors_metadata keys to strings, the key '{new_key}' is duplicated. Ensure that all your metadata keys are strings to avoid overwriting.")
+                            logger.warning(
+                                f"After converting safetensors_metadata keys to strings, the key "
+                                f"'{new_key}' is duplicated. Ensure that all your metadata keys are "
+                                f"strings to avoid overwriting."
+                            )
                         new_safetensors_metadata[new_key] = new_value
                 safetensors_metadata = new_safetensors_metadata
                 if converted_keys:
-                    logger.debug(f"One or more safetensors_metadata keys or values had to be converted to str(). Final safetensors_metadata: {safetensors_metadata}")
+                    logger.debug(
+                        f"One or more safetensors_metadata keys or values had to be converted to str(). "
+                        f"Final safetensors_metadata: {safetensors_metadata}"
+                    )
 
             # Format is required to enable Accelerate to load the metadata
             # otherwise it raises an OSError
@@ -576,9 +601,15 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         self.quantize_config.model_name_or_path = save_dir
         self.quantize_config.model_file_base_name = model_base_name
 
-    def save_pretrained(self, save_dir: str, use_safetensors: bool = False, safetensors_metadata: Optional[Dict[str, str]] = None, **kwargs):
+    def save_pretrained(
+        self,
+        save_dir: str,
+        use_safetensors: bool = False,
+        safetensors_metadata: Optional[Dict[str, str]] = None,
+        **kwargs
+    ):
         """alias of save_quantized"""
-        logger.warning("you are using save_pretrained, which will re-direct to save_quantized.")
+        logger.warning("You are using save_pretrained, which will re-direct to save_quantized.")
         self.save_quantized(save_dir, use_safetensors, safetensors_metadata)
 
     @classmethod
@@ -672,9 +703,8 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                     model.seqlen = model_config[key]
                     break
         else:
-            logger.warning("can't get model's sequence length from model config, will set to 4096.")
+            logger.warning("Can't get model's sequence length from model config, will set to 4096.")
             model.seqlen = 4096
-        model.eval()
 
         return cls(model, False, quantize_config)
 
@@ -688,8 +718,8 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         low_cpu_mem_usage: bool = False,
         use_triton: bool = False,
         torch_dtype: torch.dtype = torch.float16,
-        inject_fused_attention: bool = True,
+        inject_fused_attention: bool = False,
-        inject_fused_mlp: bool = True,
+        inject_fused_mlp: bool = False,
         use_cuda_fp16: bool = True,
         quantize_config: Optional[BaseQuantizeConfig] = None,
         model_basename: Optional[str] = None,
@@ -697,11 +727,12 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         trust_remote_code: bool = False,
         warmup_triton: bool = False,
         trainable: bool = False,
+        attn_op: Optional[AttentionOp] = None,
         disable_exllama: bool = False,
         **kwargs
     ):
         """load quantized model from local disk"""
-       
+
         # Parameters related to loading from Hugging Face Hub
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
@@ -725,9 +756,9 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
             "_raise_exceptions_for_missing_entries": False,
             "_commit_hash": commit_hash,
         }
-            
+
         if use_triton and not TRITON_AVAILABLE:
-            logger.warning("Triton is not installed, reset use_triton to False.")
+            logger.warning("Triton is not installed, reset use_triton to False")
             use_triton = False
         if not disable_exllama and not EXLLAMA_KERNELS_AVAILABLE:
             logger.warning(
@@ -746,22 +777,33 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                 "2. You are using pytorch without CUDA support.\n"
                 "3. CUDA and nvcc are not installed in your device."
             )
+        if any([inject_fused_attention, inject_fused_mlp]) and trainable:
+            logger.warning(
+                "Neither fused attention nor fused mlp is tested under trainable mode, "
+                "this may cause unexpected behavior or lead to error if you are training "
+                "a quantized model with fused ops, please consider disabling 'inject_fused_attention' "
+                "and 'inject_fused_mlp'."
+            )
 
         # == step1: prepare configs and file names == #
-        config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=trust_remote_code, **cached_file_kwargs)
+        config = AutoConfig.from_pretrained(
+            model_name_or_path,
+            trust_remote_code=trust_remote_code,
+            **cached_file_kwargs
+        )
 
         if config.model_type not in SUPPORTED_MODELS:
             raise TypeError(f"{config.model_type} isn't supported yet.")
 
         if quantize_config is None:
             quantize_config = BaseQuantizeConfig.from_pretrained(model_name_or_path, **cached_file_kwargs, **kwargs)
-        
+
         if model_basename is None:
             if quantize_config.model_file_base_name:
                 model_basename = quantize_config.model_file_base_name
             else:
                 model_basename = f"gptq_model-{quantize_config.bits}bit-{quantize_config.group_size}g"
-        
+
         quantize_config.model_name_or_path = model_name_or_path
         quantize_config.model_file_base_name = model_basename
 
@@ -786,18 +828,20 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                 resolved_archive_file = cached_file(model_name_or_path, model_basename + ext, **cached_file_kwargs)
                 if resolved_archive_file is not None:
                     break
-        
+
-        if resolved_archive_file is None: # Could not find a model file to use
+        if resolved_archive_file is None:  # Could not find a model file to use
             raise FileNotFoundError(f"Could not find model in {model_name_or_path}")
-                
+
         model_save_name = resolved_archive_file
 
         if not disable_exllama and trainable:
             logger.warning("QuantLinear with exllama backend not support trainable mode yet, Switch to the pytorch backend.")
             disable_exllama = True
-            
         elif not use_triton and trainable:
-            logger.warning("QuantLinear with cuda backend not support trainable mode yet, Switch to the pytorch backend.")
+            logger.warning(
+                "QuantLinear with cuda backend not support trainable mode yet, will switch to pytorch backend, "
+                "this may cause very slow inference speed, disable trainable if you are not training model."
+            )
 
         # == step2: convert model to gptq-model (replace Linear with QuantLinear) == #
         def skip(*args, **kwargs):
@@ -881,7 +925,11 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         )
         model = simple_dispatch_model(model, device_map)
 
-        # == step4: set seqlen == #
+        # == step4: post init model == #
+        # Any post-initialization that require device information, for example buffers initialization on device.
+        model = autogptq_post_init(model, use_act_order=quantize_config.desc_act)
+
+        # == step5: set seqlen == #
         model_config = model.config.to_dict()
         seq_len_keys = ["max_position_embeddings", "seq_length", "n_positions"]
         if any([k in model_config for k in seq_len_keys]):
@@ -890,50 +938,62 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                     model.seqlen = model_config[key]
                     break
         else:
-            logger.warning("can't get model's sequence length from model config, will set to 4096.")
+            logger.warning("Can't get model's sequence length from model config, will set to 4096.")
             model.seqlen = 4096
 
-        # == step5: (optional) inject optimized module == #
+        # == step6: (optional) inject optimized module == #
         if inject_fused_attention:
-            if cls.fused_attn_module_type is None:
+            try:
+                cls._fuse_attention(model, attn_op, trainable)
+            except NotImplementedError:
                 inject_fused_attention = False
-                logger.warning(f"{cls.__name__} hasn't fused attention module yet, will skip inject fused attention.")
+                logger.warning(
-            else:
+                    f"{cls.__name__} doesn't support fusing attention yet, will skip inject fused attention."
-                cls.fused_attn_module_type.inject_to_model(
-                    model,
-                    use_triton=use_triton,
-                    group_size=quantize_config.group_size,
-                    use_cuda_fp16=use_cuda_fp16,
-                    desc_act=quantize_config.desc_act,
-                    trainable=trainable,
-                    bits=quantize_config.bits,
-                    disable_exllama=disable_exllama,
                 )
+            except:
+                logger.error(
+                    f"Inject fused attention failed, you can set 'inject_fused_attention' to False to "
+                    f"bypass the error for now and report it on github."
+                )
+                raise
         if inject_fused_mlp:
-            if cls.fused_mlp_module_type is None:
+            try:
+                cls._fuse_mlp(model, trainable)
+            except NotImplementedError:
                 inject_fused_mlp = False
-                logger.warning(f"{cls.__name__} hasn't fused mlp module yet, will skip inject fused mlp.")
+                logger.warning(
-            else:
+                    f"{cls.__name__} doesn't support fusing mlp yet, will skip inject fused mlp."
-                cls.fused_mlp_module_type.inject_to_model(
-                    model,
-                    use_triton=use_triton
                 )
+            except:
+                logger.error(
+                    f"Inject fused mlp failed, you can set 'inject_fused_mlp' to False to "
+                    f"bypass the error for now and report it on github."
+                )
+                raise
+        if inject_fused_attention or inject_fused_mlp:
+            logger.warning(
+                "You are using at least one of 'inject_fused_attention' and 'inject_fused_mlp' "
+                "modes, which are now marked as experimental features, feel free to open an issue "
+                "or ask any question about those two features on github if you encounter unexpected "
+                "behaviors and errors."
+            )
 
-        # Any post-initialization that require device information, for example buffers initialization on device.
+        # == step7: (optional) warmup triton == #
-        model = autogptq_post_init(model, use_act_order=quantize_config.desc_act)
-
-        model.eval()
-        # == step6: (optional) warmup triton == #
         if use_triton and warmup_triton:
-            from ..nn_modules.qlinear.qlinear_triton import QuantLinear
+            cls.warmup_triton(model)
-            QuantLinear.warmup(model, seqlen=model.seqlen)
 
-            if inject_fused_mlp and cls.fused_mlp_module_type is not None:
+        # == step8: convert all QuantLinear to sub-class of torch.nn.Linear
-                cls.fused_mlp_module_type.warmup(model, seqlen=model.seqlen)
+        # note if _fuse_attention() and _fuse_mlp() is implemented,
-
+        # all QuantLinear will be converted to sub-class of torch.nn.Linear at injection stage
-        # == step7: make model compatible with peft
+        GeneralQuantLinear.convert_to_torch_linear(
-        cls.make_sure_compatible_with_peft(
+            model,
-            model, use_triton, quantize_config.desc_act, quantize_config.group_size, bits=quantize_config.bits
+            dynamically_import_QuantLinear(
+                use_triton,
+                quantize_config.desc_act,
+                quantize_config.group_size,
+                quantize_config.bits,
+                disable_exllama
+            )
         )
 
         return cls(
@@ -942,39 +1002,35 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
             quantize_config,
             is_triton_backend=use_triton,
             injected_fused_attention=inject_fused_attention,
-            injected_fused_mlp=inject_fused_mlp and use_triton,
+            injected_fused_mlp=inject_fused_mlp,
             trainable=trainable
         )
 
-    def warmup_triton(self, enabled: bool = True):
+    @staticmethod
+    def _fuse_attention(
+        model: PreTrainedModel,
+        attn_op: Optional[AttentionOp] = None,
+        trainable: bool = False
+    ) -> None:
+        raise NotImplementedError()
+
+    @staticmethod
+    def _fuse_mlp(
+        model: PreTrainedModel,
+        trainable: bool = False
+    ) -> None:
+        raise NotImplementedError()
+
+    @staticmethod
+    def warmup_triton(model: nn.Module, enabled: bool = True) -> None:
         if not enabled:
             return
         if not TRITON_AVAILABLE:
-            logger.warning(f"triton is not available, skip warmup stage directly.")
+            logger.warning(f"Triton is not available, skip warmup stage directly.")
             return
 
         from ..nn_modules.qlinear.qlinear_triton import QuantLinear
-        QuantLinear.warmup(self.model, seqlen=self.model.seqlen)
+        QuantLinear.warmup(model, seqlen=model.seqlen)
-
-        if self.fused_mlp_module_type is not None:
-            self.fused_mlp_module_type.warmup(self.model, seqlen=self.model.seqlen)
-
-    def enable_trainable_mode(self, enabled: bool = True):
-        if not self.is_triton_backend and enabled:
-            raise NotImplementedError("For now, trainable mode only supports triton backend.")
-        for n, m in self.model.named_modules():
-            if hasattr(m, "trainable"):
-                setattr(m, "trainable", enabled)
-
-    def disable_trainable_mode(self):
-        self.enable_trainable_mode(enabled=False)
-
-    @staticmethod
-    def make_sure_compatible_with_peft(model: PreTrainedModel, use_triton: bool, desc_act: bool, group_size: int, bits: int):
-        GeneralQuantLinear.inject_to_model(
-            model,
-            dynamically_import_QuantLinear(use_triton, desc_act, group_size, bits=bits)
-        )
 
     def __getattr__(self, item):
         try:

auto_gptq/modeling/_const.py

@@ -1,9 +1,5 @@
-from packaging.version import parse as parse_version
-
 from torch import device
 
-from ..utils.import_utils import compare_transformers_version
-
 CPU = device("cpu")
 CUDA_0 = device("cuda:0")
 
@@ -20,9 +16,8 @@ SUPPORTED_MODELS = [
     "RefinedWeb",
     "baichuan",
     "internlm",
+    "llama",
     "qwen",
 ]
-if compare_transformers_version("v4.28.0", op="ge"):
-    SUPPORTED_MODELS.append("llama")
 
 __all__ = ["CPU", "CUDA_0", "SUPPORTED_MODELS"]

auto_gptq/modeling/auto.py

@@ -1,6 +1,8 @@
 from inspect import signature
 from typing import Dict, Optional, Union
 
+from xformers.ops.fmha import AttentionOp
+
 from ._base import BaseQuantizeConfig, BaseGPTQForCausalLM
 from ._utils import check_and_get_model_type
 from .bloom import BloomGPTQForCausalLM
@@ -81,6 +83,7 @@ class AutoGPTQForCausalLM:
         trust_remote_code: bool = False,
         warmup_triton: bool = False,
         trainable: bool = False,
+        attn_op: Optional[AttentionOp] = None,
         disable_exllama: bool = False,
         **kwargs
     ) -> BaseGPTQForCausalLM:
@@ -121,6 +124,7 @@ class AutoGPTQForCausalLM:
             trust_remote_code=trust_remote_code,
             warmup_triton=warmup_triton,
             trainable=trainable,
+            attn_op=attn_op,
             disable_exllama=disable_exllama,
             **keywords
         )

auto_gptq/modeling/baichuan.py

@@ -1,4 +1,19 @@
+from copy import deepcopy
+from typing import Optional
+
+import xformers.ops as xop
+from torch.cuda import empty_cache
+from transformers import PreTrainedModel
+from xformers.ops.fmha import AttentionOp
+
 from ._base import *
+from ..nn_modules.fused_modules.attention import build_rope_cache, FusedAttentionWithRoPE
+from ..nn_modules.fused_modules.linear import FusedGeneralQuantLinear
+from ..nn_modules.fused_modules.mlp import FusedGatedMLP
+
+
+class BaiChuanFusedAttentionWithRope(FusedAttentionWithRoPE):
+    pass
 
 
 class BaiChuanGPTQForCausalLM(BaseGPTQForCausalLM):
@@ -12,5 +27,49 @@ class BaiChuanGPTQForCausalLM(BaseGPTQForCausalLM):
         ["mlp.down_proj"]
     ]
 
+    @staticmethod
+    def _fuse_attention(
+        model: PreTrainedModel,
+        attn_op: Optional[AttentionOp] = None,
+        trainable: bool = False
+    ) -> None:
+        model_config = model.config
+        num_heads = model_config.num_attention_heads
+        scale = (model_config.hidden_size // num_heads) ** -0.5
+        layers = model.model.layers
+
+        rope_cache = build_rope_cache(
+            rotary_dim=model_config.hidden_size // num_heads,
+            max_position=model_config.max_position_embeddings,
+            device=model.device,
+            dtype=model.dtype
+        )
+
+        for layer in layers:
+            old_attn = layer.self_attn
+            attn_device = old_attn.W_pack.qweight.data.device
+            new_qkv_proj = FusedGeneralQuantLinear(old_attn.W_pack)
+            new_out_proj = FusedGeneralQuantLinear(old_attn.o_proj)
+            new_attn = BaiChuanFusedAttentionWithRope(
+                qkv_proj=new_qkv_proj,
+                out_proj=new_out_proj,
+                cos_sin_cache=rope_cache if attn_device == model.device else deepcopy(rope_cache).to(attn_device),
+                num_query_heads=num_heads,
+                num_key_heads=num_heads,
+                num_value_heads=num_heads,
+                attn_dropout=0.0,
+                resid_dropout=0.0,
+                scale=scale,
+                attention_ops=attn_op,
+                outputs_handler=(lambda x, y, z: (x, z, y)),
+                training=trainable
+            )
+
+            layer.self_attn = new_attn
+
+            del old_attn
+
+        empty_cache()
+
 
 __all__ = ["BaiChuanGPTQForCausalLM"]

auto_gptq/modeling/gpt_bigcode.py

@@ -1,4 +1,4 @@
-from auto_gptq.modeling import BaseGPTQForCausalLM
+from ._base import BaseGPTQForCausalLM
 
 
 class GPTBigCodeGPTQForCausalLM(BaseGPTQForCausalLM):
@@ -14,4 +14,5 @@ class GPTBigCodeGPTQForCausalLM(BaseGPTQForCausalLM):
         ["mlp.c_proj"]
     ]
 
-__all__ = ["GPTBigCodeGPTQForCausalLM"]
+
+__all__ = ["GPTBigCodeGPTQForCausalLM"]

auto_gptq/modeling/gptj.py

@@ -1,5 +1,136 @@
+from copy import deepcopy
+from typing import Callable, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import xformers.ops as xop
+from torch.cuda import empty_cache
+from transformers import PreTrainedModel
+from transformers.activations import ACT2FN
+from transformers.models.gptj.modeling_gptj import apply_rotary_pos_emb
+from xformers.ops.fmha import AttentionOp
+
 from ._base import *
-from ..nn_modules.fused_gptj_attn import FusedGPTJAttentionForQuantizedModel
+from ..nn_modules.fused_modules.linear import FusedGeneralQuantLinear
+from ..nn_modules.fused_modules.attention import FusedAttention
+from ..nn_modules.fused_modules.mlp import FusedMLP
+
+
+class GPTJFusedAttention(FusedAttention):
+    def __init__(
+        self,
+        qkv_proj: nn.Linear,
+        out_proj: nn.Linear,
+        embed_positions: torch.Tensor,
+        rotary_dim: Optional[int],
+        num_query_heads: int,
+        num_key_heads: int,
+        num_value_heads: int,
+        attn_dropout: float = 0.0,
+        resid_dropout: float = 0.0,
+        scale: Optional[float] = None,
+        attention_ops: Optional[xop.AttentionOp] = None,
+        outputs_handler: Optional[Callable] = None,
+        training: bool = False,
+    ):
+        super(GPTJFusedAttention, self).__init__(
+            qkv_proj,
+            out_proj,
+            num_query_heads,
+            num_key_heads,
+            num_value_heads,
+            attn_dropout,
+            resid_dropout,
+            scale,
+            attention_ops,
+            outputs_handler,
+            training
+        )
+        self.embed_positions = embed_positions
+        self.rotary_dim = rotary_dim
+
+    def _get_embed_positions(self, position_ids: torch.Tensor):
+        return self.embed_positions.repeat(position_ids.shape[0], 1, 1)
+
+    def _apply_rotary(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        position_ids: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        bsz, seq_len = key.shape[:2]
+
+        dtype = query.dtype
+        query = query.view(bsz, seq_len, self.num_query_heads, -1).to(dtype=torch.float)
+        key = key.view(bsz, seq_len, self.num_key_heads, -1).to(dtype=torch.float)
+
+        embed_positions = self._get_embed_positions(position_ids)
+
+        repeated_position_ids = position_ids.unsqueeze(-1).repeat(1, 1, embed_positions.shape[-1])
+        sincos = torch.gather(embed_positions, 1, repeated_position_ids)
+        sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1)
+
+        if self.rotary_dim is not None:
+            k_rot = key[:, :, :, : self.rotary_dim]
+            k_pass = key[:, :, :, self.rotary_dim:]
+
+            q_rot = query[:, :, :, : self.rotary_dim]
+            q_pass = query[:, :, :, self.rotary_dim:]
+
+            k_rot = apply_rotary_pos_emb(k_rot, sin, cos)
+            q_rot = apply_rotary_pos_emb(q_rot, sin, cos)
+
+            key = torch.cat([k_rot, k_pass], dim=-1)
+            query = torch.cat([q_rot, q_pass], dim=-1)
+        else:
+            key = apply_rotary_pos_emb(key, sin, cos)
+            query = apply_rotary_pos_emb(query, sin, cos)
+
+        return query.view(bsz, seq_len, -1).to(dtype=dtype), key.view(bsz, seq_len, -1).to(dtype=dtype)
+
+    def _build_attn_bias(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None
+    ) -> Optional[xop.AttentionBias]:
+        return xop.LowerTriangularMask()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = False,
+        **kwargs
+    ):
+        bsz, seq_len = hidden_states.shape[:2]
+
+        q, k, v = self.qkv_proj(hidden_states).chunk(chunks=3, dim=-1)
+
+        if position_ids is not None:
+            q, k = self._apply_rotary(q, k, position_ids)
+
+        attn_bias = self._build_attn_bias(hidden_states, attention_mask) if layer_past is None else None
+        attn_out, present = self._attn(
+            bsz,
+            seq_len,
+            q,
+            k,
+            v,
+            attn_bias,
+            use_cache,
+            layer_past
+        )
+
+        out = self.out_proj(attn_out)
+        out = self.resid_dropout(out)
+
+        outputs = (out, present, None)
+        if self.outputs_handler:
+            outputs = self.outputs_handler(*outputs)
+
+        return outputs
 
 
 class GPTJGPTQForCausalLM(BaseGPTQForCausalLM):
@@ -13,7 +144,75 @@ class GPTJGPTQForCausalLM(BaseGPTQForCausalLM):
         ["mlp.fc_out"]
     ]
 
-    fused_attn_module_type = FusedGPTJAttentionForQuantizedModel
+    @staticmethod
+    def _fuse_attention(
+        model: PreTrainedModel,
+        attn_op: Optional[AttentionOp] = None,
+        trainable: bool = False
+    ) -> None:
+        model_config = model.config
+        num_heads = model_config.n_head
+        scale = (model_config.hidden_size // num_heads) ** -0.5
+        layers = model.transformer.h
+
+        for layer in layers:
+            old_attn = layer.attn
+            device = old_attn.q_proj.qweight.data.device
+            new_qkv_proj = FusedGeneralQuantLinear.fuse(
+                old_attn.q_proj,
+                old_attn.k_proj,
+                old_attn.v_proj
+            )
+            new_out_proj = FusedGeneralQuantLinear(old_attn.out_proj)
+            new_attn = GPTJFusedAttention(
+                qkv_proj=new_qkv_proj,
+                out_proj=new_out_proj,
+                embed_positions=old_attn.embed_positions.to(device),
+                rotary_dim=old_attn.rotary_dim,
+                num_query_heads=num_heads,
+                num_key_heads=num_heads,
+                num_value_heads=num_heads,
+                attn_dropout=model_config.attn_pdrop,
+                resid_dropout=model_config.resid_pdrop,
+                scale=scale,
+                attention_ops=attn_op,
+                outputs_handler=None,
+                training=trainable
+            )
+
+            layer.attn = new_attn
+
+            del old_attn
+
+        empty_cache()
+
+    @staticmethod
+    def _fuse_mlp(
+        model: PreTrainedModel,
+        trainable: bool = False
+    ) -> None:
+        model_config = model.config
+        act = ACT2FN[model_config.activation_function]
+        out_dropout = model_config.resid_pdrop
+        layers = model.transformer.h
+
+        for layer in layers:
+            old_mlp = layer.mlp
+            new_mlp = FusedMLP(
+                input_proj=FusedGeneralQuantLinear(old_mlp.fc_in),
+                out_proj=FusedGeneralQuantLinear(old_mlp.fc_out),
+                activation=act,
+                in_dropout=0.0,
+                out_dropout=out_dropout,
+                training=trainable,
+                residual=False
+            )
+
+            layer.mlp = new_mlp
+
+            del old_mlp
+
+        empty_cache()
 
 
 __all__ = ["GPTJGPTQForCausalLM"]

auto_gptq/modeling/llama.py

@@ -1,16 +1,18 @@
-from logging import getLogger
+from copy import deepcopy
+from typing import Optional
+
+from torch.cuda import empty_cache
+from transformers import PreTrainedModel
+from xformers.ops.fmha import AttentionOp
 
 from ._base import *
-from ..utils.import_utils import compare_transformers_version
+from ..nn_modules.fused_modules.attention import build_rope_cache, FusedAttentionWithRoPE
+from ..nn_modules.fused_modules.linear import FusedGeneralQuantLinear
+from ..nn_modules.fused_modules.mlp import FusedGatedMLP
 
-if compare_transformers_version("v4.28.0", op="ge"):
-    from ..nn_modules.fused_llama_attn import FusedLlamaAttentionForQuantizedModel
-    from ..nn_modules.fused_llama_mlp import FusedLlamaMLPForQuantizedModel
-else:
-    FusedLlamaAttentionForQuantizedModel = None
-    FusedLlamaMLPForQuantizedModel = None
 
-logger = getLogger(__name__)
+class LlamaFusedAttentionWithRoPE(FusedAttentionWithRoPE):
+    pass
 
 
 class LlamaGPTQForCausalLM(BaseGPTQForCausalLM):
@@ -24,8 +26,61 @@ class LlamaGPTQForCausalLM(BaseGPTQForCausalLM):
         ["mlp.down_proj"]
     ]
 
-    fused_attn_module_type = FusedLlamaAttentionForQuantizedModel
+    @staticmethod
-    fused_mlp_module_type = FusedLlamaMLPForQuantizedModel
+    def _fuse_attention(
+        model: PreTrainedModel,
+        attn_op: Optional[AttentionOp] = None,
+        trainable: bool = False
+    ) -> None:
+        model_config = model.config
+        num_heads = model_config.num_attention_heads
+        scale = (model_config.hidden_size // num_heads) ** -0.5
+        layers = model.model.layers
+
+        rope_cache = build_rope_cache(
+            rotary_dim=model_config.hidden_size // num_heads,
+            max_position=model_config.max_position_embeddings,
+            base=10000,
+            device=model.device,
+            dtype=model.dtype
+        )
+
+        for layer in layers:
+            old_attn = layer.self_attn
+            attn_device = old_attn.q_proj.qweight.data.device
+            new_qkv_proj = FusedGeneralQuantLinear.fuse(
+                old_attn.q_proj,
+                old_attn.k_proj,
+                old_attn.v_proj
+            )
+            new_out_proj = FusedGeneralQuantLinear(old_attn.o_proj)
+            new_attn = LlamaFusedAttentionWithRoPE(
+                qkv_proj=new_qkv_proj,
+                out_proj=new_out_proj,
+                cos_sin_cache=rope_cache if attn_device == model.device else deepcopy(rope_cache).to(attn_device),
+                num_query_heads=num_heads,
+                num_key_heads=num_heads,
+                num_value_heads=num_heads,
+                attn_dropout=0.0,
+                resid_dropout=0.0,
+                scale=scale,
+                attention_ops=attn_op,
+                outputs_handler=(lambda x, y, z: (x, z, y)),
+                training=trainable
+            )
+
+            layer.self_attn = new_attn
+
+            del old_attn
+
+        empty_cache()
+
+    # @staticmethod
+    # def _fuse_mlp(
+    #     model: PreTrainedModel,
+    #     trainable: bool = False
+    # ) -> None:
+    #     pass
 
 
 __all__ = ["LlamaGPTQForCausalLM"]

auto_gptq/nn_modules/fused_modules/attention.py

@@ -0,0 +1,394 @@
+import math
+from typing import Callable, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import xformers.ops as xop
+from vllm import pos_encoding_ops as vllm_pos_encoding_ops
+from xformers.ops.fmha.attn_bias import LowerTriangularMask, LowerTriangularMaskWithTensorBias
+
+
+POTENTIAL_KV_CACHE_NAMES = (
+    "past_key_value",
+    "layer_past",
+    "kv_cache"
+)
+
+
+def _try_to_get_kv_cache(**kwargs) -> Optional[Tuple[torch.Tensor, torch.Tensor]]:
+    kv_cache = None
+    for name in POTENTIAL_KV_CACHE_NAMES:
+        if name in kwargs:
+            return kwargs[name]
+    return kv_cache
+
+
+def build_rope_cache(
+    rotary_dim: int,
+    max_position: int = 2048,
+    base: int = 10000,
+    device: torch.device = torch.device("cuda:0"),
+    dtype: torch.dtype = torch.float16
+):  # TODO: support multiple scaling strategies
+    inv_freq = (1.0 / (base ** (torch.arange(0, rotary_dim, 2, device=device, dtype=dtype) / rotary_dim)))
+    t = torch.arange(max_position, device=device, dtype=dtype)
+    freqs = torch.einsum("i,j -> ij", t, inv_freq)
+    cos = freqs.cos()
+    sin = freqs.sin()
+    cache = torch.cat((cos, sin), dim=-1)
+
+    return cache
+
+
+def build_alibi_slopes(
+    num_heads: int,
+    device: torch.device = torch.device("cuda:0"),
+    dtype: torch.dtype = torch.float16
+):
+    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
+    base = torch.tensor(
+        2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=device, dtype=torch.float32
+    )
+    powers = torch.arange(1, 1 + closest_power_of_2, device=device, dtype=torch.int32)
+    slopes = torch.pow(base, powers)
+
+    if closest_power_of_2 != num_heads:
+        extra_base = torch.tensor(
+            2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=device, dtype=torch.float32
+        )
+        num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
+        extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=device, dtype=torch.int32)
+        slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
+
+    slopes = slopes.to(dtype)
+
+    return slopes
+
+
+def attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_ops: Optional[xop.AttentionOp] = (xop.fmha.flash.FwOp(), None),
+    attention_bias: Optional[xop.AttentionBias] = None,
+    p: float = 0.0,
+    scale: Optional[float] = None
+):
+    if value.shape[2] != query.shape[2]:
+        # MQA expand
+        if value.shape[2] == 1:
+            pass  # TODO
+        # GQA reshape
+        else:
+            original_shape = value.shape
+            pass  # TODO
+
+    return xop.memory_efficient_attention(
+        query=query,
+        key=key,
+        value=value,
+        attn_bias=attention_bias,
+        p=p,
+        scale=scale,
+        op=attention_ops
+    )
+
+
+class FusedAttention(nn.Module):
+    def __init__(
+        self,
+        qkv_proj: nn.Linear,
+        out_proj: nn.Linear,
+        num_query_heads: int,
+        num_key_heads: int,
+        num_value_heads: int,
+        attn_dropout: float = 0.0,
+        resid_dropout: float = 0.0,
+        scale: Optional[float] = None,
+        attention_ops: Optional[xop.AttentionOp] = None,
+        outputs_handler: Optional[Callable] = None,
+        training: bool = False,
+    ):
+        super(FusedAttention, self).__init__()
+
+        self.qkv_proj = qkv_proj
+        self.out_proj = out_proj
+
+        self.num_query_heads = num_query_heads
+        self.num_key_heads = num_key_heads
+        self.num_value_heads = num_value_heads
+
+        self.attn_dropout = attn_dropout if training else 0.0
+        self.scale = scale
+
+        self.attention_ops = attention_ops
+
+        self.outputs_handler = outputs_handler
+
+        self.resid_dropout = nn.Dropout(resid_dropout if training else 0.0)
+
+    def _build_attn_bias(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None
+    ) -> Optional[xop.AttentionBias]:
+        return None
+
+    def _attn(
+        self,
+        batch_size: int,
+        seq_len: int,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        attention_bias: Optional[xop.AttentionBias] = None,
+        use_cache: bool = False,
+        kv_cache: Optional[Tuple[torch.Tensor, torch.Tensor]] = None
+    ):
+        q = q.view(batch_size, seq_len, self.num_query_heads, -1).transpose(1, 2)
+        k = k.view(batch_size, seq_len, self.num_key_heads, -1).transpose(1, 2)
+        v = v.view(batch_size, seq_len, self.num_value_heads, -1).transpose(1, 2)
+
+        if kv_cache is not None:
+            k_cache, v_cache = kv_cache
+            k = torch.cat((k_cache, k), dim=2)
+            v = torch.cat((v_cache, v), dim=2)
+
+        present = None
+        if use_cache:
+            present = (k, v)
+
+        attn_out = attention(
+            query=q.transpose(1, 2),
+            key=k.transpose(1, 2),
+            value=v.transpose(1, 2),
+            attention_ops=self.attention_ops,
+            attention_bias=attention_bias,
+            p=self.attn_dropout,
+            scale=self.scale
+        ).view(batch_size, seq_len, -1)
+
+        return attn_out, present
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs
+    ):
+        bsz, seq_len = hidden_states.shape[:2]
+        use_cache = kwargs.get("use_cache", False)
+        kv_cache = _try_to_get_kv_cache(**kwargs)
+
+        q, k, v = self.qkv_proj(hidden_states).chunk(chunks=3, dim=-1)
+
+        attn_bias = self._build_attn_bias(hidden_states, attention_mask)
+        attn_out, present = self._attn(
+            bsz,
+            seq_len,
+            q,
+            k,
+            v,
+            attn_bias,
+            use_cache,
+            kv_cache
+        )
+
+        out = self.out_proj(attn_out)
+        out = self.resid_dropout(out)
+
+        outputs = (out, present, None)
+        if self.outputs_handler:
+            outputs = self.outputs_handler(*outputs)
+
+        return outputs
+
+
+class FusedAttentionWithRoPE(FusedAttention):
+    def __init__(
+        self,
+        qkv_proj: nn.Linear,
+        out_proj: nn.Linear,
+        cos_sin_cache: torch.Tensor,
+        num_query_heads: int,
+        num_key_heads: int,
+        num_value_heads: int,
+        attn_dropout: float = 0.0,
+        resid_dropout: float = 0.0,
+        scale: Optional[float] = None,
+        attention_ops: Optional[xop.AttentionOp] = None,
+        outputs_handler: Optional[Callable] = None,
+        training: bool = False,
+    ):
+        super(FusedAttentionWithRoPE, self).__init__(
+            qkv_proj=qkv_proj,
+            out_proj=out_proj,
+            num_query_heads=num_query_heads,
+            num_key_heads=num_key_heads,
+            num_value_heads=num_value_heads,
+            attn_dropout=attn_dropout,
+            resid_dropout=resid_dropout,
+            scale=scale,
+            attention_ops=attention_ops,
+            outputs_handler=outputs_handler,
+            training=training
+        )
+
+        self.register_buffer("cos_sin_cache", cos_sin_cache, persistent=False)
+
+    def _build_attn_bias(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None
+    ) -> Optional[xop.AttentionBias]:
+        return LowerTriangularMask()
+
+    def _apply_rotary_embedding(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        position_ids: Optional[torch.Tensor] = None
+    ):
+        bsz, seq_len, hidden_size = query.shape
+
+        if position_ids is not None:
+            query = query.view(bsz * seq_len, -1)
+            key = key.view(bsz * seq_len, -1)
+            vllm_pos_encoding_ops.rotary_embedding_neox(
+                position_ids.view(-1).to(query.device),
+                query,
+                key,
+                hidden_size // self.num_query_heads,
+                self.cos_sin_cache,
+            )
+            query = query.view(bsz, seq_len, -1)
+            key = key.view(bsz, seq_len, -1)
+
+        return query, key
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs
+    ):
+        bsz, seq_len = hidden_states.shape[:2]
+        position_ids = kwargs.get("position_ids", None)
+        use_cache = kwargs.get("use_cache", False)
+        kv_cache = _try_to_get_kv_cache(**kwargs)
+
+        q, k, v = self.qkv_proj(hidden_states).chunk(chunks=3, dim=-1)
+
+        q, k = self._apply_rotary_embedding(q, k, position_ids)
+
+        attn_bias = self._build_attn_bias(hidden_states, attention_mask) if kv_cache is None else None
+        attn_out, present = self._attn(
+            bsz,
+            seq_len,
+            q,
+            k,
+            v,
+            attn_bias,
+            use_cache,
+            kv_cache
+        )
+
+        out = self.out_proj(attn_out)
+        out = self.resid_dropout(out)
+
+        outputs = (out, present, None)
+        if self.outputs_handler:
+            outputs = self.outputs_handler(*outputs)
+
+        return outputs
+
+
+class FusedAttentionWithALiBi(FusedAttention):
+    def __init__(
+        self,
+        qkv_proj: nn.Linear,
+        out_proj: nn.Linear,
+        alibi_slopes: torch.Tensor,
+        num_query_heads: int,
+        num_key_heads: int,
+        num_value_heads: int,
+        attn_dropout: float = 0.0,
+        resid_dropout: float = 0.0,
+        scale: Optional[float] = None,
+        attention_ops: Optional[xop.AttentionOp] = None,
+        outputs_handler: Optional[Callable] = None,
+        training: bool = False,
+    ):
+        super(FusedAttentionWithALiBi, self).__init__(
+            qkv_proj=qkv_proj,
+            out_proj=out_proj,
+            num_query_heads=num_query_heads,
+            num_key_heads=num_key_heads,
+            num_value_heads=num_value_heads,
+            attn_dropout=attn_dropout,
+            resid_dropout=resid_dropout,
+            scale=scale,
+            attention_ops=attention_ops,
+            outputs_handler=outputs_handler,
+            training=training
+        )
+
+        self.register_buffer("alibi_slopes", alibi_slopes, persistent=False)
+
+    def _build_attn_bias(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None
+    ) -> Optional[xop.AttentionBias]:  # adopt from vllm
+        bsz, seq_len = hidden_states.shape[:2]
+
+        bias = torch.arange(seq_len)
+        bias = bias[None, :] - bias[:, None]
+        bias = bias.to(hidden_states.device)
+
+        # When using custom attention bias, xformers requires the bias to
+        # be sliced from a tensor whose length is a multiple of 8.
+        padded_len = (seq_len + 7) // 8 * 8
+        bias = torch.empty(
+            self.num_query_heads,
+            padded_len,
+            padded_len,
+            device=self.alibi_slopes.device,
+        )[:, :seq_len, :seq_len].copy_(bias)
+        bias.mul_(self.alibi_slopes[:, None, None])
+        bias = LowerTriangularMaskWithTensorBias(bias.unsqueeze(0).repeat(bsz, 1, 1, 1))
+
+        return bias
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs
+    ):
+        bsz, seq_len = hidden_states.shape[:2]
+        use_cache = kwargs.get("use_cache", False)
+        kv_cache = _try_to_get_kv_cache(**kwargs)
+
+        q, k, v = self.qkv_proj(hidden_states).chunk(chunks=3, dim=-1)
+
+        attn_bias = self._build_attn_bias(hidden_states, attention_mask) if kv_cache is None else None
+        attn_out, present = self._attn(
+            bsz,
+            seq_len,
+            q,
+            k,
+            v,
+            attn_bias,
+            use_cache,
+            kv_cache
+        )
+
+        out = self.out_proj(attn_out)
+        out = self.resid_dropout(out)
+
+        outputs = (out, present, None)
+        if self.outputs_handler:
+            outputs = self.outputs_handler(*outputs)
+
+        return outputs

auto_gptq/nn_modules/fused_modules/linear.py

@@ -0,0 +1,97 @@
+import torch
+
+from ..qlinear import GeneralQuantLinear
+from ..qlinear.qlinear_cuda import QuantLinear as CudaQuantLinear
+from ..qlinear.qlinear_cuda_old import QuantLinear as OldCudaQuantLinear
+try:
+    from ..qlinear.qlinear_triton import QuantLinear as TritonQuantLinear
+except:
+    TritonQuantLinear = None
+try:
+    from ..qlinear.qlinear_exllama import QuantLinear as ExllamaQuantLinear
+except:
+    ExllamaQuantLinear = None
+
+
+class FusedGeneralQuantLinear(GeneralQuantLinear):
+    def __init__(self, quant_linear_module):
+        super(FusedGeneralQuantLinear, self).__init__(quant_linear_module)
+
+    @classmethod
+    def fuse(
+        cls,
+        q_proj,
+        k_proj=None,
+        v_proj=None,
+    ):
+        qweights, qzeros, scales, g_idx, bias = [], [], [], [], []
+        outfeatures = 0
+        for module in [q_proj, k_proj, v_proj]:
+            if module is not None:
+                qweights.append(module.qweight)
+                qzeros.append(module.qzeros)
+                scales.append(module.scales)
+                g_idx.append(module.g_idx)
+                bias.append(module.bias)
+                outfeatures += module.outfeatures
+
+        if bias[0] is None:
+            bias = None
+
+        if len(qweights) > 1:
+            qweights = torch.cat(qweights, dim=1)
+            qzeros = torch.cat(qzeros, dim=1)
+            scales = torch.cat(scales, dim=1)
+            g_idx = torch.cat(g_idx, dim=0)
+            if bias is not None:
+                bias = torch.cat(bias, dim=0)
+
+        qlinear_args = (
+            q_proj.bits,
+            q_proj.group_size,
+            q_proj.infeatures,
+            outfeatures,
+            bias is not None
+        )
+        qlinear_kwargs = {"trainable": q_proj.trainable}
+        if isinstance(q_proj, (OldCudaQuantLinear, CudaQuantLinear)):
+            qlinear_kwargs["kernel_switch_threshold"] = q_proj.kernel_switch_threshold
+            if isinstance(q_proj, OldCudaQuantLinear):
+                qlinear_kwargs["use_cuda_fp16"] = q_proj.use_cuda_fp16
+                QuantLinear = OldCudaQuantLinear
+            else:
+                QuantLinear = CudaQuantLinear
+        elif isinstance(q_proj, TritonQuantLinear):
+            QuantLinear = TritonQuantLinear
+        else:
+            QuantLinear = ExllamaQuantLinear
+        fused_proj = QuantLinear(*qlinear_args, **qlinear_kwargs)
+
+        fused_proj.qweight = qweights
+        fused_proj.qzeros = qzeros
+        fused_proj.scales = scales
+        fused_proj.g_idx = g_idx
+        fused_proj.bias = bias
+
+        if isinstance(q_proj, ExllamaQuantLinear):
+            if not hasattr(q_proj, "_use_act_order"):
+                raise AttributeError(
+                    "q_proj doesn't have attribute _use_act_order, please execute "
+                    "auto_gptq.modeling._utils.autogptq_post_init function before "
+                    "fuse quant linears."
+                )
+            if q_proj._use_act_order:
+                # TODO: support it. The issue lies maybe in the line:
+                # int groups = qzeros.size(0);
+                # in exllama_ext.cpp
+                raise ValueError(
+                    "Exllama kernel does not support layer fusion with act-order. "
+                    "Please either use inject_fused_attention=False or disable_exllama=True."
+                )
+            fused_proj._use_act_order = q_proj._use_act_order
+            fused_proj.g_idx = None
+            fused_proj.post_init()
+
+        del q_proj, k_proj, v_proj
+
+        return cls(fused_proj)

auto_gptq/nn_modules/fused_modules/mlp.py

@@ -0,0 +1,168 @@
+from functools import partial
+from typing import Union, Callable, Optional
+
+import torch
+import torch.nn as nn
+from functorch.compile import memory_efficient_fusion
+from torch.nn import functional as F
+
+
+def act_dropout(
+    hidden_states: torch.Tensor,
+    activation: Union[Callable, nn.Module],
+    dropout: float = 0.0
+):
+    hidden_states = activation(hidden_states)
+    return hidden_states if dropout == 0.0 else F.dropout(hidden_states, dropout)
+
+
+def dropout_res(
+    hidden_states: torch.Tensor,
+    residual: torch.Tensor,
+    dropout: float = 0.0
+):
+    hidden_states = hidden_states if dropout == 0.0 else F.dropout(hidden_states, dropout)
+    return torch.add(hidden_states, residual)
+
+
+def act_dropout_res(
+    hidden_states: torch.Tensor,
+    residual: torch.Tensor,
+    activation: Union[Callable, nn.Module],
+    dropout: float = 0.0
+):
+    hidden_states = activation(hidden_states)
+    hidden_states = hidden_states if dropout == 0.0 else F.dropout(hidden_states, dropout)
+    return torch.add(hidden_states, residual)
+
+
+class NVFusedActDropoutRes(nn.Module):
+    def __init__(
+        self,
+        activation: Optional[Union[Callable, nn.Module]] = None,
+        dropout: float = 0.0,
+        residual: bool = False,
+        is_cuda: bool = False
+    ):
+        super(NVFusedActDropoutRes, self).__init__()
+
+        fn = partial(F.dropout, p=dropout)
+        if activation is not None and residual:
+            fn = partial(act_dropout_res, activation=activation, dropout=dropout)
+        elif activation is not None:
+            fn = partial(act_dropout, activation=activation, dropout=dropout)
+        elif residual:
+            fn = partial(dropout_res, dropout=dropout)
+
+        self.fn = fn
+        if is_cuda:
+            self.fn = memory_efficient_fusion(self.fn)
+
+        self.residual = residual
+
+    def forward(self, hidden_states: torch.Tensor, residual: Optional[torch.Tensor] = None):
+        if self.residual:
+            return self.fn(hidden_states, residual)
+        else:
+            return self.fn(hidden_states)
+
+
+class FusedMLP(nn.Module):
+    def __init__(
+        self,
+        input_proj: nn.Linear,
+        out_proj: nn.Linear,
+        activation: Optional[Union[Callable, nn.Module]] = None,
+        in_dropout: float = 0.0,
+        out_dropout: float = 0.0,
+        training: bool = False,
+        residual: bool = False
+    ):
+        super(FusedMLP, self).__init__()
+
+        if activation is None:
+            activation = nn.Identity()
+
+        is_cuda = input_proj.weight.data.device.type == "cuda"
+
+        self.input_proj = input_proj
+        self.fused_op1 = NVFusedActDropoutRes(
+            activation=activation,
+            dropout=in_dropout if training else 0.0,
+            residual=False,
+            is_cuda=is_cuda
+        )
+        self.out_proj = out_proj
+        self.fused_op2 = NVFusedActDropoutRes(
+            activation=None,
+            dropout=out_dropout if training else 0.0,
+            residual=residual,
+            is_cuda=is_cuda
+        )
+
+    def forward(self, hidden_states: torch.Tensor, residual: Optional[torch.Tensor] = None):
+        return self.fused_op2(self.out_proj(self.fused_op1(self.input_proj(hidden_states))), residual)
+
+
+def gated_act_dropout(
+    gate_states: torch.Tensor,
+    up_states: torch.Tensor,
+    activation: Union[Callable, nn.Module],
+    dropout: float = 0.0
+):
+    hidden_states = activation(gate_states) * up_states
+    return hidden_states if dropout == 0.0 else F.dropout(hidden_states, dropout)
+
+
+class NVFusedGatedActDropout(nn.Module):
+    def __init__(
+        self,
+        activation: Optional[Union[Callable, nn.Module]] = None,
+        dropout: float = 0.0,
+        is_cuda: bool = False
+    ):
+        super(NVFusedGatedActDropout, self).__init__()
+
+        fn = partial(F.dropout, p=dropout)
+        if activation is not None:
+            fn = partial(gated_act_dropout, activation=activation, dropout=dropout)
+
+        self.fn = fn
+        if is_cuda:
+            self.fn = memory_efficient_fusion(self.fn)
+
+    def forward(self, gate_states: torch.Tensor, up_states):
+        if isinstance(self.fn, nn.Dropout):
+            return self.fn(gate_states * up_states)
+        return self.fn(gate_states, up_states)
+
+
+class FusedGatedMLP(nn.Module):
+    def __init__(
+        self,
+        input_proj: nn.Linear,
+        out_proj: nn.Linear,
+        activation: Optional[Union[Callable, nn.Module]] = None,
+        in_dropout: float = 0.0,
+        out_dropout: float = 0.0,
+        training: bool = False
+    ):
+        super(FusedGatedMLP, self).__init__()
+
+        if activation is None:
+            activation = nn.Identity()
+
+        self.input_proj = input_proj
+        self.fused_op = NVFusedGatedActDropout(
+            activation=activation,
+            dropout=in_dropout if training else 0.0,
+            is_cuda=input_proj.weight.data.device.type == "cuda"
+        )
+        self.out_proj = out_proj
+        self.out_dropout = nn.Dropout(out_dropout)
+
+        self.intermediate_size = self.input_proj.out_features // 2
+
+    def forward(self, hidden_states: torch.Tensor):
+        hidden_states = self.input_proj(hidden_states)
+        return self.out_dropout(self.out_proj(self.fused_op(*hidden_states.chunk(chunks=2, dim=-1))))

auto_gptq/nn_modules/qlinear/__init__.py

@@ -6,7 +6,7 @@ class GeneralQuantLinear(nn.Linear):
         super().__init__(
             in_features=quant_linear_module.infeatures,
             out_features=quant_linear_module.outfeatures,
-            bias=True
+            bias=quant_linear_module.bias is not None
         )
         self.infeatures = quant_linear_module.infeatures
         self.outfeatures = quant_linear_module.outfeatures
@@ -18,28 +18,47 @@ class GeneralQuantLinear(nn.Linear):
 
         self.weight.data = quant_linear_module.qweight
         self.register_buffer('qweight', quant_linear_module.qweight)
-        self.bias.data = quant_linear_module.bias
+        if quant_linear_module.bias is not None:
-
+            self.bias.data = quant_linear_module.bias
-        self.qweight.requires_grad = False
-        self.bias.requires_grad = False
 
         self.register_buffer('qzeros', quant_linear_module.qzeros)
         self.register_buffer('scales', quant_linear_module.scales)
         self.register_buffer('g_idx', quant_linear_module.g_idx)
 
+        # arg of qlinear_cuda and qlinear_cuda_old
         if hasattr(quant_linear_module, "wf"):
             self.wf = quant_linear_module.wf
+        # arg of qlinaer_cuda and qlinear_cuda_old
         if hasattr(quant_linear_module, "kernel_switch_threshold"):
             self.kernel_switch_threshold = quant_linear_module.kernel_switch_threshold
+        # arg of qlinaer_cuda and qlinear_cuda_old
         if hasattr(quant_linear_module, "autogptq_cuda_available"):
             self.autogptq_cuda_available = quant_linear_module.autogptq_cuda_available
+        # arg of qlinaer_cuda and qlinear_cuda_old
+        if hasattr(quant_linear_module, "autogptq_cuda"):
+            self.autogptq_cuda = quant_linear_module.autogptq_cuda
+        # arg of qlinear_cuda_old
+        if hasattr(quant_linear_module, "half_indim"):
+            self.half_indim = quant_linear_module.half_indim
+        # arg of qlinear_cuda_old
+        if hasattr(quant_linear_module, "use_cuda_fp16"):
+            self.use_cuda_fp16 = quant_linear_module.use_cuda_fp16
+        # args of qlinear_exllama
+        if hasattr(quant_linear_module, "_use_act_order"):
+            self._use_act_order = quant_linear_module._use_act_order
+        # arg of qlinaer_exllama
+        if hasattr(quant_linear_module, "width"):
+            self.width = quant_linear_module.width
+        # arg of qlinear_exllama
+        if hasattr(quant_linear_module, "q4"):
+            self.q4 = quant_linear_module.q4
 
         self.trainable = quant_linear_module.trainable
 
         self.forward = quant_linear_module.forward
 
     @classmethod
-    def inject_to_model(cls, model, target_module_type):
+    def convert_to_torch_linear(cls, model: nn.Module, target_module_type: "QuantLinear"):
         for name, m in model.named_modules():
             if not isinstance(m, target_module_type):
                 continue

auto_gptq/nn_modules/qlinear/qlinear_cuda.py

@@ -36,8 +36,6 @@ class QuantLinear(nn.Module):
         global _autogptq_cuda_available
         if bits not in [2, 3, 4, 8]:
             raise NotImplementedError("Only 2,3,4,8 bits are supported.")
-        if trainable:
-            _autogptq_cuda_available = False
 
         self.infeatures = infeatures
         self.outfeatures = outfeatures
@@ -198,7 +196,7 @@ class QuantLinear(nn.Module):
         x = x.reshape(-1, x.shape[-1])
         if self.autogptq_cuda_available and (
             self.kernel_switch_threshold == 0 or x.shape[0] < self.kernel_switch_threshold
-        ):
+        ) and not self.trainable:
             out = torch.zeros((x.shape[0], self.outfeatures), device=x.device, dtype=torch.float32)
             if self.bits == 2:
                 self.autogptq_cuda.vecquant2matmul(x.float(), self.qweight, out, self.scales.float(), self.qzeros, self.g_idx)

auto_gptq/nn_modules/qlinear/qlinear_cuda_old.py

@@ -36,8 +36,7 @@ class QuantLinear(nn.Module):
         global _autogptq_cuda_available
         if bits not in [2, 3, 4, 8]:
             raise NotImplementedError("Only 2,3,4,8 bits are supported.")
-        if trainable:
+
-            _autogptq_cuda_available = False
         self.infeatures = infeatures
         self.outfeatures = outfeatures
         self.bits = bits
@@ -198,7 +197,7 @@ class QuantLinear(nn.Module):
         x = x.reshape(-1, x.shape[-1])
         if self.autogptq_cuda_available is True and (
             self.kernel_switch_threshold is False or x.shape[0] < self.kernel_switch_threshold
-        ):
+        ) and not self.trainable:
             out = torch.zeros(x.shape[0], out_shape[-1], dtype=torch.float, device=x.device)
             if self.use_cuda_fp16:
                 x = x.half()

auto_gptq/utils/peft_utils.py

@@ -354,7 +354,7 @@ def get_gptq_peft_model(
     train_mode: bool = False
 ):
     if train_mode and not model.trainable:
-        model.enable_trainable_mode()
+        raise TypeError("model is not trainable, please load model with 'trainable=True'")
     if train_mode and not peft_config:
         raise ValueError("peft_config not specified when in train mode.")
     if not train_mode and not model_id:

examples/benchmark/generation_speed.py

@@ -87,30 +87,29 @@ def load_data(data_path, tokenizer, n_samples, max_new_tokens):
         outputs = examples["output"]
 
         prompts = []
-        texts = []
+        outs = []
         input_ids = []
         attention_mask = []
         for istr, inp, opt in zip(instructions, inputs, outputs):
             if inp:
                 prompt = f"Instruction:\n{istr}\nInput:\n{inp}\nOutput:\n"
-                text = prompt + opt
             else:
                 prompt = f"Instruction:\n{istr}\nOutput:\n"
-                text = prompt + opt
             if len(tokenizer(prompt)["input_ids"]) >= tokenizer.model_max_length - max_new_tokens:
                 continue
 
-            tokenized_data = tokenizer(text)
+            tokenized_data = tokenizer(prompt)
 
             input_ids.append(tokenized_data["input_ids"][: tokenizer.model_max_length])
             attention_mask.append(tokenized_data["attention_mask"][: tokenizer.model_max_length])
             prompts.append(prompt)
-            texts.append(text)
+            outs.append(opt)
 
         return {
             "input_ids": input_ids,
             "attention_mask": attention_mask,
-            "prompt": prompts
+            "prompt": prompts,
+            "output": outs
         }
 
     dataset = Dataset.from_generator(dummy_gen)
@@ -236,9 +235,9 @@ def main():
     parser.add_argument("--use_triton", action="store_true")
     parser.add_argument("--use_safetensors", action="store_true")
     parser.add_argument("--use_fast_tokenizer", action="store_true")
+    parser.add_argument("--inject_fused_attention", action="store_true")
+    parser.add_argument("--inject_fused_mlp", action="store_true")
     parser.add_argument("--disable_exllama", action="store_true")
-    parser.add_argument("--no_inject_fused_attention", action="store_true")
-    parser.add_argument("--no_inject_fused_mlp", action="store_true")
     parser.add_argument("--num_samples", type=int, default=10)
     parser.add_argument("--per_gpu_max_memory", type=int, default=None)
     parser.add_argument("--cpu_max_memory", type=int, default=None)
@@ -277,8 +276,8 @@ def main():
         use_triton=args.use_triton,
         use_safetensors=args.use_safetensors,
         use_fast_tokenizer=args.use_fast_tokenizer,
-        inject_fused_attention=not args.no_inject_fused_attention,
+        inject_fused_attention=args.inject_fused_attention,
-        inject_fused_mlp=not args.no_inject_fused_mlp,
+        inject_fused_mlp=args.inject_fused_mlp,
         disable_exllama=args.disable_exllama
     )
     end = time.time()
@@ -289,7 +288,7 @@ def main():
 
     if args.use_triton:
         logger.info("warmup triton, this may take a while.")
-        model.warmup_triton()
+        model.warmup_triton(model)
 
     logger.info("loading data")
     examples = load_data(

examples/benchmark/perplexity.py

@@ -37,12 +37,18 @@ if __name__ == "__main__":
     parser.add_argument("--use_safetensors", action="store_true", help="Whether to use safetensors model file")
     parser.add_argument("--use_fast_tokenizer", action="store_true", help="Wheter to use fast tokenizer")
     parser.add_argument("--trust_remote_code", action="store_true", help="Whether to use remote code")
+    parser.add_argument("--inject_fused_attention", action="store_true", help="Whether to inject fused attention")
+    parser.add_argument("--inject_fused_mlp", action="store_true", help="Whether to inject fused mlp")
     parser.add_argument("--disable_exllama", action="store_true", help="Whether to use disable exllama kernel")
     args = parser.parse_args()
 
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
-    tokenizer = AutoTokenizer.from_pretrained(args.model_name, use_fast=args.use_fast_tokenizer)
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.model_name,
+        use_fast=args.use_fast_tokenizer,
+        trust_remote_code=args.trust_remote_code
+    )
     if not tokenizer.pad_token_id:
         tokenizer.pad_token_id = tokenizer.eos_token_id
 
@@ -68,8 +74,8 @@ if __name__ == "__main__":
             model_basename=args.model_basename,
             use_safetensors=args.use_safetensors,
             trust_remote_code=args.trust_remote_code,
-            inject_fused_mlp=False,
+            inject_fused_mlp=args.inject_fused_mlp,
-            inject_fused_attention=False,
+            inject_fused_attention=args.inject_fused_attention,
             disable_exllama=args.disable_exllama
         )
     else:

setup.py

@@ -68,7 +68,10 @@ requirements = [
     "datasets",
     "numpy",
     "rouge",
-    "torch>=1.13.0",
+    "torch>=2.0.1",
+    "functorch",
+    "xformers>=0.0.20",
+    "vllm",
     "safetensors",
     "transformers>=4.31.0",
     "peft"