wip for new datasets abstractions

* auto gptq support

* more tweaks and add yml

* remove old gptq docker

* don't need explicit peft install for tests

* fix setup.py to use extra index url

install torch for tests
fix cuda version for autogptq index
set torch in requirements so that it installs properly
move gptq install around to work with github cicd

* gptq doesn't play well with sample packing

* address pr feedback

* remove torch install for now

* set quantization_config from model config

* Fix the implementation for getting quant config from model config

.github/workflows/main.yml

@@ -23,11 +23,6 @@ jobs:
             python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.9"
-            pytorch: 2.0.1
-            axolotl_extras: gptq
     runs-on: self-hosted
     steps:
       - name: Checkout
@@ -73,11 +68,6 @@ jobs:
             pytorch: 2.0.1
             axolotl_extras:
             is_latest: true
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.9"
-            pytorch: 2.0.1
-            axolotl_extras: gptq
     runs-on: self-hosted
     steps:
       - name: Checkout

.github/workflows/tests.yml

@@ -24,7 +24,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip install -e .[peft]
+          pip install -e .
           pip install -r requirements-tests.txt
 
       - name: Run tests

docker/Dockerfile

@@ -11,14 +11,13 @@ RUN apt-get update && \
 
 WORKDIR /workspace
 
-RUN pip3 install "peft @ git+https://github.com/huggingface/peft.git@main"
 RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN cd axolotl && \
     if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[flash-attn,$AXOLOTL_EXTRAS]; \
+        pip install -e .[flash-attn,gptq,$AXOLOTL_EXTRAS]; \
     else \
-        pip install -e .[flash-attn]; \
+        pip install -e .[flash-attn,gptq]; \
     fi
 
 # fix so that git fetch/pull from remote works

examples/gptq-lora-7b/README.md

@@ -1,8 +0,0 @@
-# LLaMa 7B using LoRA
-
-This is a good place to start for beginners. This will run on an NVIDIA RTX4090 with no other changes needed.
-
-```shell
-accelerate launch scripts/finetune.py examples/gptq-lora-7b/config.yml
-
-```

examples/gptq-lora-7b/config.yml

@@ -1,63 +0,0 @@
-base_model: Neko-Institute-of-Science/LLaMA-7B-4bit-128g
-base_model_config: Neko-Institute-of-Science/LLaMA-7B-4bit-128g
-model_type: LlamaForCausalLM
-tokenizer_type: LlamaTokenizer
-trust_remote_code:
-load_in_8bit: true
-gptq: true
-datasets:
-  - path: vicgalle/alpaca-gpt4
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.02
-adapter:
-lora_model_dir:
-sequence_len: 2048
-max_packed_sequence_len:
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - v_proj
-lora_fan_in_fan_out: false
-wandb_project: llama-7b-lora-int4
-wandb_entity:
-wandb_watch:
-wandb_run_id:
-wandb_log_model:
-output_dir: ./llama-7b-lora-int4
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 3
-optimizer: adamw_bnb_8bit
-torchdistx_path:
-lr_scheduler: cosine
-learning_rate: 0.0000002
-train_on_inputs: false
-group_by_length: false
-fp16: true
-bf16: false
-tf32: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 5
-xformers_attention:
-flash_attention:
-gradient_checkpointing: true
-gptq_groupsize: 128
-gptq_model_v1: false
-warmup_steps: 20
-eval_steps: 110
-save_steps: 660
-debug:
-deepspeed:
-weight_decay: 0.0001
-fsdp:
-fsdp_config:
-tokens:
-  pad_token: "<pad>"
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/llama-2/gptq-lora.yml

@@ -0,0 +1,76 @@
+base_model: TheBloke/Llama-2-7B-GPTQ
+base_model_config: TheBloke/Llama-2-7B-GPTQ
+is_llama_derived_model: false
+gptq: true
+gptq_bits: 4
+model_type: AutoModelForCausalLM
+tokenizer_type: LlamaTokenizer
+tokenizer_use_fast: true
+tokenizer_legacy: true
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+push_dataset_to_hub:
+hf_use_auth_token: true
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+adapter: lora
+lora_model_dir:
+sequence_len: 4096
+sample_packing:
+lora_r: 8
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - k_proj
+  - o_proj
+  - q_proj
+  - v_proj
+lora_target_linear:
+lora_fan_in_fan_out:
+wandb_project:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+output_dir: ./model-out
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 3
+optimizer: adamw_torch
+adam_beta2: 0.95
+adam_eps: 0.00001
+max_grad_norm: 1.0
+torchdistx_path:
+lr_scheduler: cosine
+lr_quadratic_warmup: true
+learning_rate: 0.000017
+train_on_inputs: false
+group_by_length: false
+bf16: false
+fp16: false
+float16: true
+tf32: true
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention:
+sdp_attention:
+flash_optimum:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 100
+eval_steps:
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.1
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

requirements.txt

@@ -1,3 +1,7 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+--extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+torch==2.0.1
+auto-gptq
 packaging
 peft @ git+https://github.com/huggingface/peft.git
 transformers @ git+https://github.com/huggingface/transformers.git

scripts/finetune.py

@@ -24,7 +24,7 @@ from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process
-from axolotl.utils.models import load_model_config, load_tokenizer
+from axolotl.utils.models import load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
 from axolotl.utils.wandb import setup_wandb_env_vars
 
@@ -216,15 +216,6 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
             else:
                 cfg[k] = kwargs[k]
 
-    model_config = load_model_config(cfg)
-
-    # figure out if the model is llama
-    cfg.is_llama_derived_model = (
-        (hasattr(model_config, "model_type") and model_config.model_type == "llama")
-        or cfg.is_llama_derived_model
-        or "llama" in cfg.base_model
-        or (cfg.model_type and "llama" in cfg.model_type.lower())
-    )
     validate_config(cfg)
 
     normalize_config(cfg)

setup.py

@@ -2,15 +2,27 @@
 
 from setuptools import find_packages, setup
 
-install_requires = []
-with open("./requirements.txt", encoding="utf-8") as requirements_file:
-    # don't include peft yet until we check the int4
-    # need to manually install peft for now...
-    reqs = [r.strip() for r in requirements_file.readlines() if "peft" not in r]
-    reqs = [r for r in reqs if "flash-attn" not in r]
-    reqs = [r for r in reqs if r and r[0] != "#"]
-    for r in reqs:
-        install_requires.append(r)
+
+def parse_requirements():
+    _install_requires = []
+    _dependency_links = []
+    with open("./requirements.txt", encoding="utf-8") as requirements_file:
+        lines = [
+            r.strip() for r in requirements_file.readlines() if "auto-gptq" not in r
+        ]
+        for line in lines:
+            if line.startswith("--extra-index-url"):
+                # Handle custom index URLs
+                _, url = line.split()
+                _dependency_links.append(url)
+            elif "flash-attn" not in line and line and line[0] != "#":
+                # Handle standard packages
+                _install_requires.append(line)
+    return _install_requires, _dependency_links
+
+
+install_requires, dependency_links = parse_requirements()
+
 
 setup(
     name="axolotl",
@@ -19,12 +31,10 @@ setup(
     package_dir={"": "src"},
     packages=find_packages(),
     install_requires=install_requires,
+    dependency_links=dependency_links,
     extras_require={
         "gptq": [
-            "alpaca_lora_4bit @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
-        ],
-        "gptq_triton": [
-            "alpaca_lora_4bit[triton] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
+            "auto-gptq",
         ],
         "flash-attn": [
             "flash-attn==2.0.8",
@@ -32,8 +42,5 @@ setup(
         "extras": [
             "deepspeed",
         ],
-        "peft": [
-            "peft @ git+https://github.com/huggingface/peft.git",
-        ],
     },
 )

src/axolotl/core/datasets.py

@@ -0,0 +1,144 @@
+import logging
+from dataclasses import dataclass, field
+from enum import Enum
+from pathlib import Path
+from typing import Any, Dict, Generator, List, Optional, Union
+
+from datasets import Dataset as Dataset_ds
+from datasets import DatasetDict, IterableDataset, load_dataset, load_from_disk
+from huggingface_hub import hf_hub_download
+
+logger = logging.getLogger("axolotl")
+
+
+class DsType(Enum):
+    JSON = "json"
+    ARROW = "arrow"
+    PARQUET = "parquet"
+
+
+@dataclass
+class DatasetConfiguration:
+    path: str
+    type: str
+    name: Optional[str] = field(
+        default=None,
+        metadata={"help": "the name of the dataset configuration to load."},
+    )
+    ds_type: Optional[DsType] = None
+    data_files: Optional[Union[str, List[str]]] = None
+    shards: Optional[int] = None
+    test_size: Optional[float] = None
+
+    @staticmethod
+    def from_dict(d: Dict[str, Any]) -> Generator["DatasetConfiguration", None, None]:
+        if "name" in d and isinstance(d["name"], list):
+            name = d.pop("name")
+            for n in name:
+                yield DatasetConfiguration(
+                    **d,
+                    name=n,
+                )
+
+
+def load_dataset_from_local(config: DatasetConfiguration) -> Optional[Dataset_ds]:
+    local_path = Path(config.path)
+    if not local_path.exists():
+        return None
+    ds = None
+    if local_path.is_dir():
+        if config.ds_type:
+            # TODO dirs with arrow or parquet files could be loaded with `load_from_disk`
+            ds = load_from_disk(config.path)
+        else:
+            ds = load_dataset(
+                config.path,
+                name=config.name,
+                data_files=config.data_files,
+                streaming=False,
+                split=None,
+            )
+    elif local_path.is_file():
+        ds_type = "json"
+        if config.ds_type:
+            ds_type = config.ds_type.value
+        elif "parquet" in config.path:
+            ds_type = "parquet"
+        elif "arrow" in config.path:
+            ds_type = "arrow"
+        ds = load_dataset(
+            ds_type,
+            name=config.name,
+            data_files=config.path,
+            streaming=False,
+            split=None,  # is this correct?
+        )
+    if not ds:
+        raise ValueError(
+            "unhandled dataset load: local path exists, but is neither a directory or a file"
+        )
+    return ds
+
+
+# TODO should this be a DatasetDict?
+class Dataset(Dataset_ds):
+    _config: DatasetConfiguration
+
+    def __init__(self, *args, config: DatasetConfiguration = None, **kwargs):
+        self._config = config
+        super().__init__(*args, **kwargs)
+
+    @staticmethod
+    def from_config(
+        config: DatasetConfiguration,
+        token: bool = False,
+        default_test_size: float = 0.1,
+    ):
+        ds = load_dataset_from_local(config)
+        if not ds:
+            try:
+                ds = load_dataset(
+                    config.path,
+                    name=config.name,
+                    data_files=config.data_files,
+                    token=token,
+                )
+            except FileNotFoundError:
+                pass
+        if not ds:
+            fp = hf_hub_download(
+                repo_id=config.path,
+                repo_type="dataset",
+                filename=config.data_files,
+                token=token,
+            )
+            ds = load_dataset(
+                "json", name=config.name, data_files=fp, streaming=False, split=None
+            )
+        if not ds:
+            raise ValueError("unhandled dataset load")
+        test_size = config.test_size if config.test_size else default_test_size
+        # determine if the dataset is pre-tokenized
+        check_ds = ds["train"] if isinstance(ds, DatasetDict) and "train" in ds else ds
+        is_ds_tokenized = False
+        if "input_ids" in check_ds.features:
+            is_ds_tokenized = True
+            if "attention_mask" not in check_ds.features:
+                logger.warning("`attention_mask` missing from pre-tokenized dataset")
+            if "labels" not in check_ds.features:
+                logger.warning("`labels` missing from pre-tokenized dataset")
+        if test_size and (not isinstance(ds, DatasetDict) or "test" not in ds):
+            ds.train_test_split(test_size=test_size, shuffle=False)
+            pass
+
+
+class DatasetCollection:
+    datasets: List[Dataset] = []
+
+    def __init__(self, datasets: Union[Dataset, List[Dataset]]):
+        self.datasets = datasets if isinstance(datasets, list) else [datasets]
+
+    def __iter__(self):
+        for ds in self.datasets:
+            for d in ds:
+                yield d

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -2,7 +2,9 @@
 
 # copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py
 
+import logging
 import warnings
+from functools import partial
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -33,6 +35,9 @@ except ImportError:
     )
 
 
+LOG = logging.getLogger("axolotl")
+
+
 def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
     transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
         _prepare_decoder_attention_mask
@@ -44,6 +49,34 @@ def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
             llama_model_forward
         )
 
+    try:
+        from flash_attn.losses.cross_entropy import CrossEntropyLoss
+
+        LOG.info("patching with flash_attn.losses.cross_entropy")
+        transformers.models.llama.modeling_llama.CrossEntropyLoss = partial(
+            CrossEntropyLoss, inplace_backward=True
+        )
+    except ImportError:
+        LOG.info(
+            "optimized flash-attention CrossEntropyLoss not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=xentropy_cuda_lib&subdirectory=csrc/xentropy'`)"
+        )
+
+    try:
+        from flash_attn.ops.rms_norm import RMSNorm
+
+        class LlamaRMSNorm(RMSNorm):
+            """Patched LLamaRMSNorm"""
+
+            def __init__(self, hidden_size, eps=1e-6):
+                super().__init__(hidden_size, eps=eps)
+
+        LOG.info("patching with flash_attn.ops.rms_norm")
+        transformers.models.llama.modeling_llama.LlamaRMSNorm = LlamaRMSNorm
+    except ImportError:
+        LOG.info(
+            "optimized flash-attention RMSNorm not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=dropout_layer_norm&subdirectory=csrc/layer_norm'`)"
+        )
+
 
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
 # requires the attention mask to be the same as the key_padding_mask

src/axolotl/prompters.py

@@ -309,10 +309,6 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
         )
 
     def build_prompt(self, source) -> Generator[str, None, None]:
-        # ignore the system prompt if provided
-        if source[0]["from"] == "system":
-            source.pop(0)
-
         if len(source) < 2:
             # If there isn't a back and forth conversation, ignore it
             # also happens on the data splitting leaving empty conversations
@@ -321,6 +317,12 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
             )
 
         conv = self._conversation.copy()
+
+        # Add the conversation system prompt if provided, otherwise use the default one
+        if source[0]["from"] == "system":
+            conv.system = source[0]["value"]
+            source.pop(0)
+
         roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
 
         try:

src/axolotl/utils/callbacks.py

@@ -27,6 +27,7 @@ from axolotl.utils.distributed import (
     barrier,
     gather_scalar_from_all_ranks,
     get_world_size,
+    is_distributed,
     is_main_process,
     zero_first,
 )
@@ -270,10 +271,13 @@ def bench_eval_callback_factory(trainer, tokenizer):
                 lambda: len(data_loader), get_world_size()
             )
 
-            if not is_main_process():
+            if is_distributed() and not is_main_process():
                 dist.gather_object(local_bench_names, dst=0)
             else:
-                dist.gather_object(local_bench_names, gathered_bench_names, dst=0)
+                if is_distributed():
+                    dist.gather_object(local_bench_names, gathered_bench_names, dst=0)
+                else:
+                    gathered_bench_names = [local_bench_names]
                 bench_loss = sum(loss_bench_ranks) / sum(len_data_loader_ranks)
                 results = {f"{bench_split}_bench_loss": bench_loss}
 

src/axolotl/utils/config.py

@@ -6,6 +6,7 @@ import os
 import torch
 
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.models import load_model_config
 
 LOG = logging.getLogger("axolotl")
 
@@ -69,6 +70,16 @@ def normalize_config(cfg):
     else:
         cfg.torch_dtype = torch.float32
 
+    model_config = load_model_config(cfg)
+
+    # figure out if the model is llama
+    cfg.is_llama_derived_model = (
+        (hasattr(model_config, "model_type") and model_config.model_type == "llama")
+        or cfg.is_llama_derived_model
+        or "llama" in cfg.base_model
+        or (cfg.model_type and "llama" in cfg.model_type.lower())
+    )
+
     log_gpu_memory_usage(LOG, "baseline", cfg.device)
 
 
@@ -97,9 +108,7 @@ def validate_config(cfg):
             "To calculate the equivalent gradient_accumulation_steps, divide batch_size / micro_batch_size / number of gpus.",
         )
     if cfg.load_4bit:
-        raise ValueError(
-            "cfg.load_4bit parameter has been deprecated and replaced by cfg.gptq"
-        )
+        raise ValueError("cfg.load_4bit parameter has been deprecated")
 
     if cfg.adapter == "qlora":
         if cfg.merge_lora:

src/axolotl/utils/distributed.py

@@ -74,6 +74,8 @@ def gather_scalar_from_all_ranks(fn, world_size=1):  # pylint: disable=invalid-n
     - A list of computed values from all ranks if on the gathering rank, otherwise None.
     """
     value_scalar = fn()
+    if not is_distributed():
+        return [value_scalar]
     value_tensor = torch.tensor(value_scalar, device=dist.get_rank()).float()
 
     if not is_main_process():

src/axolotl/utils/models.py

@@ -4,19 +4,19 @@
 import logging
 import math
 import os
-from pathlib import Path
 from typing import Optional, Tuple  # noqa: F401
 
 import bitsandbytes as bnb
 import torch
 import transformers
 from optimum.bettertransformer import BetterTransformer
-from peft import PeftConfig
+from peft import PeftConfig, prepare_model_for_kbit_training
 from transformers import (  # noqa: F401
     AutoConfig,
     AutoModelForCausalLM,
     AutoTokenizer,
     BitsAndBytesConfig,
+    GPTQConfig,
     LlamaConfig,
     PreTrainedModel,
     PreTrainedTokenizerBase,
@@ -155,32 +155,17 @@ def load_model(
         LOG.info("patching _expand_mask")
         hijack_expand_mask()
 
-    try:
-        if cfg.gptq:
-            from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
-                replace_peft_model_with_int4_lora_model,
-            )
-
-            replace_peft_model_with_int4_lora_model()
-    except Exception as err:
-        LOG.exception(err)
-        raise err
-
-    if not cfg.gptq and (
-        (cfg.adapter == "lora" and load_in_8bit)
-        or (cfg.adapter == "qlora" and cfg.load_in_4bit)
-    ):
-        try:
-            from peft import prepare_model_for_kbit_training
-        except ImportError:
-            # For backward compatibility
-            from peft import (
-                prepare_model_for_int8_training as prepare_model_for_kbit_training,
-            )
-
     model_kwargs = {}
     if cfg.model_revision:
         model_kwargs["revision"] = cfg.model_revision
+    if cfg.gptq:
+        model_config = load_model_config(cfg)
+        if hasattr(model_config, "quantization_config"):
+            LOG.warning("model config does not contain quantization_config information")
+        else:
+            model_kwargs["quantization_config"] = GPTQConfig(
+                **model_config.quantization_config
+            )
     if cfg.adapter == "qlora" and cfg.load_in_4bit:
         model_kwargs["quantization_config"] = BitsAndBytesConfig(
             load_in_4bit=True,
@@ -191,45 +176,7 @@ def load_model(
             bnb_4bit_quant_type="nf4",
         )
     try:
-        if cfg.gptq and cfg.is_llama_derived_model:
-            from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
-            from huggingface_hub import snapshot_download
-
-            try:
-                snapshot_download_kwargs = {}
-                if cfg.base_model_ignore_patterns:
-                    snapshot_download_kwargs[
-                        "ignore_patterns"
-                    ] = cfg.base_model_ignore_patterns
-                cache_model_path = Path(
-                    snapshot_download(base_model, **snapshot_download_kwargs)
-                )
-                files = (
-                    list(cache_model_path.glob("*.pt"))
-                    + list(cache_model_path.glob("*.safetensors"))
-                    + list(cache_model_path.glob("*.bin"))
-                )
-                if len(files) > 0:
-                    model_path = str(files[0])
-                else:
-                    LOG.warning(
-                        "unable to find a cached model file, this will likely fail..."
-                    )
-                    model_path = str(cache_model_path)
-            except Exception:  # pylint: disable=broad-exception-caught
-                model_path = cfg.base_model
-            model, _ = load_llama_model_4bit_low_ram(
-                base_model_config if base_model_config else base_model,
-                model_path,
-                device_map=cfg.device_map,
-                half=cfg.fp16,
-                groupsize=cfg.gptq_groupsize if cfg.gptq_groupsize else -1,
-                is_v1_model=cfg.gptq_model_v1
-                if cfg.gptq_model_v1 is not None
-                else True,
-            )
-            load_in_8bit = False
-        elif cfg.is_llama_derived_model and not cfg.trust_remote_code:
+        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
             from transformers import LlamaForCausalLM
 
             config_kwargs = {}
@@ -275,15 +222,24 @@ def load_model(
         #     )
         #     model.train() # sets to train instead of eval mode
         elif model_type and not cfg.trust_remote_code:
-            model = getattr(transformers, model_type).from_pretrained(
-                base_model,
-                device_map=cfg.device_map,
-                load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
-                load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=cfg.torch_dtype,
-                trust_remote_code=cfg.trust_remote_code or False,
-                **model_kwargs,
-            )
+            if cfg.gptq:
+                model = AutoModelForCausalLM.from_pretrained(
+                    base_model,
+                    device_map=cfg.device_map,
+                    torch_dtype=cfg.torch_dtype,
+                    trust_remote_code=cfg.trust_remote_code or False,
+                    **model_kwargs,
+                )
+            else:
+                model = getattr(transformers, model_type).from_pretrained(
+                    base_model,
+                    device_map=cfg.device_map,
+                    load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
+                    load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
+                    torch_dtype=cfg.torch_dtype,
+                    trust_remote_code=cfg.trust_remote_code or False,
+                    **model_kwargs,
+                )
         else:
             config = AutoConfig.from_pretrained(
                 base_model,
@@ -359,11 +315,12 @@ def load_model(
                 module.to(torch.float32)
 
     needs_fa2_dtype = cfg.adapter or cfg.fsdp
-    if not cfg.gptq and (
-        (cfg.adapter == "lora" and load_in_8bit)
-        or (cfg.adapter == "qlora" and cfg.load_in_4bit)
+    if (cfg.adapter == "lora" and load_in_8bit) or (
+        cfg.adapter == "qlora" and cfg.load_in_4bit
     ):
         LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
+        if cfg.gradient_checkpointing:
+            model.gradient_checkpointing_enable()
         model = prepare_model_for_kbit_training(
             model, use_gradient_checkpointing=cfg.gradient_checkpointing
         )
@@ -385,22 +342,10 @@ def load_model(
     if cfg.ddp and not load_in_8bit:
         model.to(f"cuda:{cfg.local_rank}")
 
-    if cfg.gptq:
-        # Scales to half
-        LOG.info("Fitting 4bit scales and zeros to half")
-        for _, module in model.named_modules():
-            if "Autograd4bitQuantLinear" in str(type(module)) or "Linear4bitLt" in str(
-                type(module)
-            ):
-                if hasattr(module, "is_v1_model") and module.is_v1_model:
-                    module.zeros = module.zeros.half()
-                module.scales = module.scales.half()
-                module.bias = module.bias.half()
-
     if (
         torch.cuda.device_count() > 1
         and int(os.getenv("WORLD_SIZE", "1")) > 1
-        and (cfg.gptq or cfg.load_in_4bit)
+        and (cfg.load_in_4bit)
     ):
         # llama is PROBABLY model parallelizable, but the default isn't that it is
         # so let's only set it for the 4bit, see

src/axolotl/utils/trainer.py

@@ -514,23 +514,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         training_arguments_kwargs["seed"] = cfg.seed
 
     if cfg.gradient_checkpointing:
-        if cfg.gptq:
-            from alpaca_lora_4bit.gradient_checkpointing import (
-                apply_gradient_checkpointing,
-            )
-
-            gradient_checkpointing_ratio = (
-                cfg.gradient_checkpointing_ratio
-                if cfg.gradient_checkpointing_ratio
-                else 1.0
-            )
-            apply_gradient_checkpointing(
-                model, checkpoint_ratio=gradient_checkpointing_ratio
-            )
-        else:
-            training_arguments_kwargs[
-                "gradient_checkpointing"
-            ] = cfg.gradient_checkpointing
+        training_arguments_kwargs["gradient_checkpointing"] = cfg.gradient_checkpointing
     if cfg.fsdp:
         training_arguments_kwargs["fsdp"] = cfg.fsdp
         if cfg.fsdp_config: