remove torch install for now

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

.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

README.md

@@ -163,8 +163,6 @@ accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
   ```
   </details>
 
-- Windows: Please use WSL or Docker!
-
 ### Dataset
 
 Axolotl supports a variety of dataset formats. Below are some of the formats you can use.
@@ -625,11 +623,6 @@ fsdp_config:
 # Deepspeed config path
 deepspeed:
 
-# Advanced DDP Arguments
-ddp_timeout:
-ddp_bucket_cap_mb:
-ddp_broadcast_buffers:
-
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:
 

deepspeed/zero3.json

@@ -35,7 +35,10 @@
     "type": "AdamW",
     "params": {
       "lr": "auto",
-      "betas": "auto",
+      "betas": [
+        0.9,
+        0.95
+      ],
       "eps": 1e-8,
       "weight_decay": "auto"
     }

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

@@ -1,10 +1,13 @@
+--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
 bitsandbytes>=0.41.1
 accelerate @ git+https://github.com/huggingface/accelerate@2a289f6108e77a77a4efffb3f6316bc98538413b
 addict
-evaluate
 fire
 PyYAML>=6.0
 datasets

scripts/finetune.py

@@ -4,7 +4,9 @@ import importlib
 import logging
 import os
 import random
+import signal
 import sys
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Union
 
@@ -15,17 +17,17 @@ import yaml
 
 # add src to the pythonpath so we don't need to pip install this
 from art import text2art
+from optimum.bettertransformer import BetterTransformer
 from transformers import GenerationConfig, TextStreamer
 
-from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.logging_config import configure_logging
-from axolotl.train import TrainDatasetMeta, train
 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_model, load_model_config, load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
+from axolotl.utils.trainer import setup_trainer
 from axolotl.utils.wandb import setup_wandb_env_vars
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
@@ -38,13 +40,26 @@ LOG = logging.getLogger("axolotl.scripts")
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 
 
+@dataclass
+class TrainerCliArgs:
+    """
+    dataclass representing the various non-training arguments
+    """
+
+    debug: bool = field(default=False)
+    inference: bool = field(default=False)
+    merge_lora: bool = field(default=False)
+    prepare_ds_only: bool = field(default=False)
+    prompter: Optional[str] = field(default=None)
+    shard: bool = field(default=False)
+
+
 def print_axolotl_text_art(suffix=None):
     font = "nancyj"
     ascii_text = "  axolotl"
     if suffix:
         ascii_text += f"  x  {suffix}"
     ascii_art = text2art(" axolotl", font=font)
-
     if is_main_process():
         print(ascii_art)
 
@@ -58,45 +73,9 @@ def get_multi_line_input() -> Optional[str]:
     return instruction
 
 
-def do_merge_lora(
+def do_inference(cfg, model, tokenizer, prompter: Optional[str]):
-    *,
+    if prompter == "None":
-    cfg: DictDefault,
+        prompter = None
-    cli_args: TrainerCliArgs,
-):
-    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
-    safe_serialization = cfg.save_safetensors is True
-
-    LOG.info("running merge of LoRA with base model")
-    model = model.merge_and_unload()
-    model.to(dtype=torch.float16)
-
-    if cfg.local_rank == 0:
-        LOG.info("saving merged model")
-        model.save_pretrained(
-            str(Path(cfg.output_dir) / "merged"),
-            safe_serialization=safe_serialization,
-        )
-        tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
-
-
-def shard(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-):
-    model, _ = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
-    safe_serialization = cfg.save_safetensors is True
-    LOG.debug("Re-saving model w/ sharding")
-    model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
-
-
-def do_inference(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-):
-    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
-    prompter = cli_args.prompter
     default_tokens = {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
 
     for token, symbol in default_tokens.items():
@@ -197,6 +176,141 @@ def check_not_in(list1: List[str], list2: Union[Dict[str, Any], List[str]]) -> b
     return not any(el in list2 for el in list1)
 
 
+def train(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+):
+    # load the tokenizer first
+    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
+    tokenizer = load_tokenizer(cfg)
+
+    if not (
+        cli_args.shard or cli_args.merge_lora or cli_args.inference
+    ):  # don't need to load dataset for these
+        train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
+
+    if cli_args.debug or cfg.debug:
+        LOG.info("check_dataset_labels...")
+        check_dataset_labels(
+            train_dataset.select(
+                [random.randrange(0, len(train_dataset) - 1) for _ in range(5)]  # nosec
+            ),
+            tokenizer,
+        )
+
+    if cli_args.prepare_ds_only:
+        LOG.info("Finished preparing dataset. Exiting...")
+        return
+
+    # Load the model and tokenizer
+    LOG.info("loading model and (optionally) peft_config...")
+    model, peft_config = load_model(cfg, tokenizer, inference=cli_args.inference)
+
+    safe_serialization = cfg.save_safetensors is True
+
+    if cli_args.merge_lora and cfg.adapter is not None:
+        LOG.info("running merge of LoRA with base model")
+        model = model.merge_and_unload()
+        model.to(dtype=torch.float16)
+
+        if cfg.local_rank == 0:
+            LOG.info("saving merged model")
+            model.save_pretrained(
+                str(Path(cfg.output_dir) / "merged"),
+                safe_serialization=safe_serialization,
+            )
+            tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
+        return
+
+    if cli_args.inference:
+        LOG.debug("Running inference on model")
+        do_inference(cfg, model, tokenizer, prompter=cli_args.prompter)
+        return
+
+    if cli_args.shard:
+        LOG.debug("Re-saving model w/ sharding")
+        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+        return
+
+    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
+        possible_checkpoints = [
+            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
+        ]
+        if len(possible_checkpoints) > 0:
+            sorted_paths = sorted(
+                possible_checkpoints,
+                key=lambda path: int(path.split("-")[-1]),
+            )
+            cfg.resume_from_checkpoint = sorted_paths[-1]
+            LOG.info(
+                f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
+            )
+    resume_from_checkpoint = cfg.resume_from_checkpoint
+
+    trainer = setup_trainer(
+        cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
+    )
+
+    model.config.use_cache = False
+
+    if torch.__version__ >= "2" and sys.platform != "win32":
+        LOG.info("Compiling torch model")
+        model = torch.compile(model)
+
+    # go ahead and presave, so we have the adapter config available to inspect
+    if peft_config:
+        LOG.info(f"Pre-saving adapter config to {cfg.output_dir}")
+        peft_config.save_pretrained(cfg.output_dir)
+
+    # In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
+    if cfg.local_rank == 0:
+
+        def terminate_handler(_, __, model):
+            if cfg.flash_optimum:
+                model = BetterTransformer.reverse(model)
+            model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+            sys.exit(0)
+
+        signal.signal(
+            signal.SIGINT, lambda signum, frame: terminate_handler(signum, frame, model)
+        )
+
+    LOG.info("Starting trainer...")
+    if cfg.group_by_length:
+        LOG.info("hang tight... sorting dataset for group_by_length")
+
+    if not Path(cfg.output_dir).is_dir():
+        os.makedirs(cfg.output_dir, exist_ok=True)
+    tokenizer.save_pretrained(cfg.output_dir)
+    if cfg.flash_optimum:
+        with torch.backends.cuda.sdp_kernel(
+            enable_flash=True, enable_math=True, enable_mem_efficient=True
+        ):
+            trainer.train(resume_from_checkpoint=resume_from_checkpoint)
+    else:
+        trainer.train(resume_from_checkpoint=resume_from_checkpoint)
+
+    LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
+
+    if cfg.relora_steps:
+        if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
+            model = model.merge_and_unload()
+        else:
+            # final model weights have already been saved by `ReLoRACallback.on_train_end`
+            return
+
+    # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
+    # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
+    if cfg.fsdp:
+        trainer.save_model(cfg.output_dir)
+    elif cfg.local_rank == 0:
+        if cfg.flash_optimum:
+            model = BetterTransformer.reverse(model)
+
+        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+
+
 def load_cfg(config: Path = Path("examples/"), **kwargs):
     if Path(config).is_dir():
         config = choose_config(config)
@@ -233,55 +347,15 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
     return cfg
 
 
-def load_datasets(
+def do_train(config: Path = Path("examples/"), **kwargs):
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-) -> TrainDatasetMeta:
-    tokenizer = load_tokenizer(cfg)
-
-    train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
-
-    if cli_args.debug or cfg.debug:
-        LOG.info("check_dataset_labels...")
-        check_dataset_labels(
-            train_dataset.select(
-                [
-                    random.randrange(0, len(train_dataset) - 1)  # nosec
-                    for _ in range(cli_args.debug_num_examples)
-                ]
-            ),
-            tokenizer,
-            num_examples=cli_args.debug_num_examples,
-            text_only=cli_args.debug_text_only,
-        )
-
-    return TrainDatasetMeta(
-        train_dataset=train_dataset,
-        eval_dataset=eval_dataset,
-        total_num_steps=total_num_steps,
-    )
-
-
-def do_cli(config: Path = Path("examples/"), **kwargs):
     print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
     parser = transformers.HfArgumentParser((TrainerCliArgs))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
-    if parsed_cli_args.inference:
+    train(cfg=parsed_cfg, cli_args=parsed_cli_args)
-        do_inference(cfg=parsed_cfg, cli_args=parsed_cli_args)
-    elif parsed_cli_args.merge_lora:
-        do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
-    elif parsed_cli_args.shard:
-        shard(cfg=parsed_cfg, cli_args=parsed_cli_args)
-    else:
-        dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
-        if parsed_cli_args.prepare_ds_only:
-            return
-        train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 
 if __name__ == "__main__":
-    fire.Fire(do_cli)
+    fire.Fire(do_train)

setup.py

@@ -2,15 +2,27 @@
 
 from setuptools import find_packages, setup
 
-install_requires = []
+
-with open("./requirements.txt", encoding="utf-8") as requirements_file:
+def parse_requirements():
-    # don't include peft yet until we check the int4
+    _install_requires = []
-    # need to manually install peft for now...
+    _dependency_links = []
-    reqs = [r.strip() for r in requirements_file.readlines() if "peft" not in r]
+    with open("./requirements.txt", encoding="utf-8") as requirements_file:
-    reqs = [r for r in reqs if "flash-attn" not in r]
+        lines = [
-    reqs = [r for r in reqs if r and r[0] != "#"]
+            r.strip() for r in requirements_file.readlines() if "auto-gptq" not in r
-    for r in reqs:
+        ]
-        install_requires.append(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",
+            "auto-gptq",
-        ],
-        "gptq_triton": [
-            "alpaca_lora_4bit[triton] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
         ],
         "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/common/cli.py

@@ -1,43 +0,0 @@
-"""
-shared module for cli specific things
-"""
-
-import logging
-from dataclasses import dataclass, field
-from typing import Optional
-
-from axolotl.logging_config import configure_logging
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model, load_tokenizer
-
-configure_logging()
-LOG = logging.getLogger("axolotl.common.cli")
-
-
-@dataclass
-class TrainerCliArgs:
-    """
-    dataclass representing the various non-training arguments
-    """
-
-    debug: bool = field(default=False)
-    debug_text_only: bool = field(default=False)
-    debug_num_examples: int = field(default=5)
-    inference: bool = field(default=False)
-    merge_lora: bool = field(default=False)
-    prepare_ds_only: bool = field(default=False)
-    prompter: Optional[str] = field(default=None)
-    shard: bool = field(default=False)
-
-
-def load_model_and_tokenizer(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-):
-    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
-    tokenizer = load_tokenizer(cfg)
-    LOG.info("loading model and (optionally) peft_config...")
-    model, _ = load_model(cfg, tokenizer, inference=cli_args.inference)
-
-    return model, tokenizer

src/axolotl/train.py

@@ -1,139 +0,0 @@
-"""Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
-
-import logging
-import os
-import signal
-import sys
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Optional
-
-import torch
-
-# add src to the pythonpath so we don't need to pip install this
-from datasets import Dataset
-from optimum.bettertransformer import BetterTransformer
-
-from axolotl.common.cli import TrainerCliArgs
-from axolotl.logging_config import configure_logging
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model, load_tokenizer
-from axolotl.utils.trainer import setup_trainer
-
-project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
-src_dir = os.path.join(project_root, "src")
-sys.path.insert(0, src_dir)
-
-configure_logging()
-LOG = logging.getLogger("axolotl.train")
-
-
-@dataclass
-class TrainDatasetMeta:
-    """
-    dataclass to capture the dataset specific options for training
-    """
-
-    train_dataset: Dataset
-    eval_dataset: Optional[Dataset] = None
-    total_num_steps: Optional[int] = None
-
-
-def train(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-    dataset_meta: TrainDatasetMeta,
-):
-    # load the tokenizer first
-    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
-    tokenizer = load_tokenizer(cfg)
-
-    train_dataset = dataset_meta.train_dataset
-    eval_dataset = dataset_meta.eval_dataset
-    total_num_steps = dataset_meta.total_num_steps
-
-    # Load the model and tokenizer
-    LOG.info("loading model and (optionally) peft_config...")
-    model, peft_config = load_model(cfg, tokenizer, inference=cli_args.inference)
-
-    safe_serialization = cfg.save_safetensors is True
-
-    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
-        possible_checkpoints = [
-            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
-        ]
-        if len(possible_checkpoints) > 0:
-            sorted_paths = sorted(
-                possible_checkpoints,
-                key=lambda path: int(path.split("-")[-1]),
-            )
-            cfg.resume_from_checkpoint = sorted_paths[-1]
-            LOG.info(
-                f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
-            )
-    resume_from_checkpoint = cfg.resume_from_checkpoint
-
-    trainer = setup_trainer(
-        cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
-    )
-
-    model.config.use_cache = False
-
-    if torch.__version__ >= "2" and sys.platform != "win32":
-        LOG.info("Compiling torch model")
-        model = torch.compile(model)
-
-    # go ahead and presave, so we have the adapter config available to inspect
-    if peft_config:
-        LOG.info(f"Pre-saving adapter config to {cfg.output_dir}")
-        peft_config.save_pretrained(cfg.output_dir)
-
-    # In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
-    if cfg.local_rank == 0:
-
-        def terminate_handler(_, __, model):
-            if cfg.flash_optimum:
-                model = BetterTransformer.reverse(model)
-            model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
-            sys.exit(0)
-
-        signal.signal(
-            signal.SIGINT, lambda signum, frame: terminate_handler(signum, frame, model)
-        )
-
-    LOG.info("Starting trainer...")
-    if cfg.group_by_length:
-        LOG.info("hang tight... sorting dataset for group_by_length")
-
-    if not Path(cfg.output_dir).is_dir():
-        os.makedirs(cfg.output_dir, exist_ok=True)
-    tokenizer.save_pretrained(cfg.output_dir)
-    if cfg.flash_optimum:
-        with torch.backends.cuda.sdp_kernel(
-            enable_flash=True, enable_math=True, enable_mem_efficient=True
-        ):
-            trainer.train(resume_from_checkpoint=resume_from_checkpoint)
-    else:
-        trainer.train(resume_from_checkpoint=resume_from_checkpoint)
-
-    LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
-
-    if cfg.relora_steps:
-        if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
-            model = model.merge_and_unload()
-        else:
-            # final model weights have already been saved by `ReLoRACallback.on_train_end`
-            return model, tokenizer
-
-    # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
-    # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
-    if cfg.fsdp:
-        trainer.save_model(cfg.output_dir)
-    elif cfg.local_rank == 0:
-        if cfg.flash_optimum:
-            model = BetterTransformer.reverse(model)
-
-        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
-
-    return model, tokenizer

src/axolotl/utils/callbacks.py

@@ -1,20 +1,9 @@
 """Callbacks for Trainer class"""
 
-from __future__ import annotations
-
 import logging
 import os
-from typing import TYPE_CHECKING, Dict, List
 
-import evaluate
-import numpy as np
-import pandas as pd
-import torch
-import torch.distributed as dist
-from accelerate.state import PartialState
-from datasets import load_dataset
 from optimum.bettertransformer import BetterTransformer
-from tqdm import tqdm
 from transformers import (
     TrainerCallback,
     TrainerControl,
@@ -24,18 +13,8 @@ from transformers import (
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, IntervalStrategy
 
 from axolotl.utils.bench import log_gpu_memory_usage
-from axolotl.utils.distributed import (
-    gather_scalar_from_all_ranks,
-    get_world_size,
-    is_main_process,
-)
-
-if TYPE_CHECKING:
-    from axolotl.utils.trainer import AxolotlTrainingArguments
 
 LOG = logging.getLogger("axolotl.callbacks")
-IGNORE_INDEX = -100
-dist_state = PartialState()
 
 
 class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-methods
@@ -117,199 +96,3 @@ class GPUStatsCallback(
             log_gpu_memory_usage(LOG, "while training", self.cfg.device)
             self.logged = True
         return control
-
-
-def bench_eval_callback_factory(trainer, tokenizer):
-    accuracy = evaluate.load("accuracy")
-    abcd_idx = [
-        tokenizer("A", add_special_tokens=False).input_ids[0],
-        tokenizer("B", add_special_tokens=False).input_ids[0],
-        tokenizer("C", add_special_tokens=False).input_ids[0],
-        tokenizer("D", add_special_tokens=False).input_ids[0],
-        tokenizer("E", add_special_tokens=False).input_ids[0],
-        tokenizer("F", add_special_tokens=False).input_ids[0],
-        tokenizer("G", add_special_tokens=False).input_ids[0],
-    ]
-    bench_split = "eval"
-
-    def transform_bench_subject(example):
-        # Split on ':' and trim whitespace
-        parts = example["subject"].split(":")
-        first_part = (
-            parts[0].strip().lower().replace("-", "_")
-        )  # Lowercase the first part
-        second_part = (
-            parts[1].strip().replace("-", "_") if len(parts) > 1 else "all"
-        )  # Replace hyphens with underscores
-
-        # Return the transformed values
-        return {"name": first_part, "subject": second_part}
-
-    if trainer.args.bench_dataset == "mmlu-zs":
-        bench_dataset = load_dataset(
-            "openaccess-ai-collective/mmlu-evals",
-            data_files={
-                "eval": "zero_shot_mmlu_val.json",
-                "test": "zero_shot_mmlu_test.json",
-            },
-        )
-        # bench_dataset = bench_dataset.remove_columns("subject")
-    # MMLU Five-shot (Eval/Test only)
-    elif trainer.args.bench_dataset in ["mmlu", "mmlu-fs"]:
-        bench_dataset = load_dataset(
-            "openaccess-ai-collective/mmlu-evals",
-            data_files={
-                "eval": "five_shot_mmlu_val.json",
-                "test": "five_shot_mmlu_test.json",
-            },
-        )
-        # bench_dataset = bench_dataset.remove_columns('subject')
-    elif "/" in trainer.args.bench_dataset:
-        bench_ds = trainer.args.bench_dataset
-        bench_ds_name = "/".join(bench_ds.split("/", 2)[:2])
-        bench_ds_data_file = "/".join(bench_ds.split("/", 2)[2:])
-        bench_dataset = load_dataset(
-            bench_ds_name,
-            data_files={
-                "eval": bench_ds_data_file,
-            },
-        )
-        bench_dataset["eval"] = bench_dataset["eval"].map(transform_bench_subject)
-    else:
-        raise ValueError(
-            f"unhandled value `{trainer.args.bench_dataset}` for bench_dataset training args"
-        )
-    bench_dataset = bench_dataset[trainer.args.bench_split]
-    if trainer.args.max_bench_samples is not None:
-        bench_dataset = bench_dataset.select(range(trainer.args.max_bench_samples))
-
-    def tokenize_evals(example):
-        source = f"{tokenizer.bos_token}{example['input']}"
-        target = f"{example['output']}{tokenizer.eos_token}"
-
-        tokenized_source = tokenizer(
-            source,
-            max_length=2048,
-            truncation=True,
-            add_special_tokens=False,
-        )
-        tokenized_target = tokenizer(
-            target,
-            max_length=2048,
-            truncation=True,
-            add_special_tokens=False,
-        )
-        input_ids = tokenized_source["input_ids"] + tokenized_target["input_ids"]
-        labels = [IGNORE_INDEX] * len(tokenized_source["input_ids"]) + tokenized_target[
-            "input_ids"
-        ]
-
-        return {
-            "input_ids": input_ids,
-            "labels": labels,
-            "subject": example["subject"],
-        }
-
-    with dist_state.main_process_first():
-        bench_dataset = bench_dataset.map(tokenize_evals)
-        bench_dataset = bench_dataset.filter(lambda x: x["labels"][-2] in abcd_idx)
-
-    class BenchEvalCallback(TrainerCallback):
-        """
-        TrainerCallback that runs the MMLU evals
-        """
-
-        def on_evaluate(
-            self,
-            args: AxolotlTrainingArguments,
-            state: TrainerState,  # pylint: disable=unused-argument
-            control: TrainerControl,  # pylint: disable=unused-argument
-            metrics: Dict[str, float],  # pylint: disable=unused-argument
-            **kwargs,  # pylint: disable=unused-argument
-        ):
-            data_loader = trainer.get_bench_dataloader(
-                bench_dataset.remove_columns(["input", "subject", "output", "name"])
-            )
-            trainer.model.eval()
-            preds, refs = [], []
-            loss_bench = 0
-            for batch in tqdm(data_loader, total=len(data_loader)):
-                (loss, logits, labels) = trainer.prediction_step(
-                    trainer.model,
-                    batch,
-                    prediction_loss_only=False,
-                )
-                # There are two tokens, the output, and eos token.
-                for i, logit in enumerate(logits):
-                    label_non_zero_id = (batch["labels"][i] != IGNORE_INDEX).nonzero()[
-                        0
-                    ][0]
-                    logit_abcd = logit[label_non_zero_id - 1][abcd_idx]
-                    preds.append(torch.argmax(logit_abcd).item())
-                labels = labels[labels != IGNORE_INDEX].view(-1, 2)[:, 0]
-                refs += [
-                    abcd_idx.index(label) if label in abcd_idx else -1
-                    for label in labels.tolist()
-                ]
-                loss_bench += loss.item()
-            # Extract results by subject.
-            bench_name = bench_dataset["name"]
-            bench_names: dict = {s: {"refs": [], "preds": []} for s in set(bench_name)}
-            for s, p, r in zip(bench_name, preds, refs):  # pylint: disable=invalid-name
-                bench_names[s]["preds"].append(p)
-                bench_names[s]["refs"].append(r)
-            dist_state.wait_for_everyone()
-            local_bench_names = bench_names
-            gathered_bench_names: List[Dict] = [{} for _ in range(get_world_size())]
-            # Gather results from all GPUs to GPU 0
-
-            loss_bench_ranks = gather_scalar_from_all_ranks(
-                lambda: loss_bench, get_world_size()
-            )
-            len_data_loader_ranks = gather_scalar_from_all_ranks(
-                lambda: len(data_loader), get_world_size()
-            )
-
-            if not is_main_process():
-                dist.gather_object(local_bench_names, dst=0)
-            else:
-                dist.gather_object(local_bench_names, gathered_bench_names, dst=0)
-                bench_loss = sum(loss_bench_ranks) / sum(len_data_loader_ranks)
-                results = {f"{bench_split}_bench_loss": bench_loss}
-
-                # Combine results from all GPUs
-                combined_bench_names: Dict[str, Dict[str, List]] = {}
-                for bench_name in gathered_bench_names:
-                    for name, data in bench_name.items():
-                        if name not in combined_bench_names:
-                            combined_bench_names[name] = {"refs": [], "preds": []}
-                        combined_bench_names[name]["refs"].extend(data["refs"])
-                        combined_bench_names[name]["preds"].extend(data["preds"])
-
-                bench_scores = []
-                bench_refs = []
-                bench_preds = []
-                for (
-                    bench_name
-                ) in combined_bench_names:  # pylint: disable=consider-using-dict-items
-                    bench_score = accuracy.compute(
-                        references=combined_bench_names[bench_name]["refs"],
-                        predictions=combined_bench_names[bench_name]["preds"],
-                    )["accuracy"]
-                    bench_refs.extend(combined_bench_names[bench_name]["refs"])
-                    bench_preds.extend(combined_bench_names[bench_name]["preds"])
-                    if not pd.isna(bench_score):
-                        results[
-                            f"{bench_split}_bench_accuracy_{bench_name}"
-                        ] = bench_score
-                        bench_scores.append(bench_score)
-                    else:
-                        results[f"{bench_split}_bench_accuracy_{bench_name}"] = 0.0
-                        bench_scores.append(0.0)
-                results[f"{bench_split}_bench_average_accuracy"] = np.mean(bench_scores)
-                results[f"{bench_split}_bench_total_accuracy"] = accuracy.compute(
-                    references=bench_refs, predictions=bench_preds
-                )["accuracy"]
-                trainer.log(results)
-
-    return BenchEvalCallback

src/axolotl/utils/config.py

@@ -97,9 +97,7 @@ 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(
+        raise ValueError("cfg.load_4bit parameter has been deprecated")
-            "cfg.load_4bit parameter has been deprecated and replaced by cfg.gptq"
-        )
 
     if cfg.adapter == "qlora":
         if cfg.merge_lora:

src/axolotl/utils/data.py

@@ -7,7 +7,6 @@ from pathlib import Path
 from typing import Tuple, Union
 
 import torch
-from accelerate.state import PartialState
 from datasets import (
     Dataset,
     DatasetDict,
@@ -43,6 +42,7 @@ from axolotl.prompters import (
     SummarizeTLDRPrompter,
 )
 from axolotl.utils.dict import DictDefault
+from axolotl.utils.distributed import is_main_process, zero_first
 from axolotl.utils.trainer import (
     calculate_total_num_steps,
     process_datasets_for_packing,
@@ -50,12 +50,11 @@ from axolotl.utils.trainer import (
 
 LOG = logging.getLogger("axolotl")
 DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"
-state = PartialState()
 
 
 def prepare_dataset(cfg, tokenizer):
     if not cfg.pretraining_dataset:
-        with state.main_process_first():
+        with zero_first(is_main_process()):
             train_dataset, eval_dataset = load_prepare_datasets(
                 tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
             )
@@ -70,7 +69,7 @@ def prepare_dataset(cfg, tokenizer):
         train_dataset = train_dataset.with_format("torch")
         eval_dataset = None
 
-    with state.main_process_first():
+    with zero_first(is_main_process()):
         train_dataset, eval_dataset = process_datasets_for_packing(
             cfg, train_dataset, eval_dataset
         )
@@ -508,7 +507,7 @@ def load_prepare_datasets(
             to_hash_test.encode(), usedforsecurity=False
         ).hexdigest()
 
-        with state.main_process_first():
+        with zero_first(is_main_process()):
             dataset = dataset.train_test_split(
                 test_size=cfg.val_set_size,
                 shuffle=False,

src/axolotl/utils/distributed.py

@@ -1,27 +1,27 @@
 """
 utility helpers for distributed checks
 """
-import torch
+from contextlib import contextmanager
+
 import torch.distributed as dist
-from accelerate import DistributedType
+from accelerate import Accelerator
-from accelerate.state import PartialState
-from accelerate.utils import wait_for_everyone
 
 accelerate = None  # pylint: disable=invalid-name
 
-state = PartialState()
+
+def load_accelerate():
+    global accelerate  # pylint: disable=global-statement
+    accelerate = Accelerator()
 
 
 def is_distributed():
     """
     Check if distributed training is initialized.
     """
-    return state.distributed_type in (
+    global accelerate  # pylint: disable=global-statement
-        DistributedType.MULTI_GPU,
+    if not accelerate:
-        DistributedType.MULTI_CPU,
+        accelerate = Accelerator()
-        DistributedType.DEEPSPEED,
+    return dist.is_available() and dist.is_initialized()
-        DistributedType.FSDP,
-    )
 
 
 def barrier():
@@ -29,48 +29,27 @@ def barrier():
     Acts as a barrier to wait for all processes. This ensures that all processes
     reach the barrier before proceeding further.
     """
-    wait_for_everyone()
+    if is_distributed():
+        dist.barrier()
 
 
-def is_main_process() -> bool:
+def is_main_process():
     """
     Check if the current process is the main process.
     If not in distributed mode, always return True.
     """
-    return state.is_main_process
+    if not is_distributed():
+        return True
+    return dist.get_rank() == 0
 
 
-def get_world_size() -> int:
+@contextmanager
-    return state.num_processes
+def zero_first(is_main):
-
-
-def gather_scalar_from_all_ranks(fn, world_size=1):  # pylint: disable=invalid-name
     """
-    Run a callable 'fn' on all ranks and gather the results on the specified rank.
+    runs the wrapped context so that rank 0 runs first before other ranks
-
-    Args:
-    - fn (callable): A function that computes the value. This should not have any side effects.
-    - rank (int, optional): The rank that gathers the values. Default is 0.
-    - world_size (int, optional): Total number of processes in the current distributed setup.
-
-    Returns:
-    - A list of computed values from all ranks if on the gathering rank, otherwise None.
     """
-    value_scalar = fn()
+    if not is_main:  # other ranks wait first
-    value_tensor = torch.tensor(value_scalar, device=dist.get_rank()).float()
+        barrier()
-
+    yield
-    if not state.is_main_process:
+    if is_main:  # then rank 0 waits after it has run the context
-        dist.gather(value_tensor, dst=0)
+        barrier()
-    else:
-        gathered_tensors = [torch.zeros_like(value_tensor) for _ in range(world_size)]
-        dist.gather(value_tensor, gather_list=gathered_tensors, dst=0)
-
-        # Convert tensors back to their original type (int or float)
-        gathered_values = []
-        for tensor in gathered_tensors:
-            if tensor == tensor.int():
-                gathered_values.append(int(tensor.item()))
-            else:
-                gathered_values.append(float(tensor.item()))
-        return gathered_values
-    return None

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 +155,15 @@ 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:
+        # TODO we should figure out how read the models config.json first
+        model_kwargs["quantization_config"] = GPTQConfig(
+            bits=cfg.gptq_bits,
+            disable_exllama=True,
+        )
     if cfg.adapter == "qlora" and cfg.load_in_4bit:
         model_kwargs["quantization_config"] = BitsAndBytesConfig(
             load_in_4bit=True,
@@ -191,45 +174,7 @@ def load_model(
             bnb_4bit_quant_type="nf4",
         )
     try:
-        if cfg.gptq and cfg.is_llama_derived_model:
+        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
-            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:
             from transformers import LlamaForCausalLM
 
             config_kwargs = {}
@@ -275,15 +220,24 @@ 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(
+            if cfg.gptq:
-                base_model,
+                model = AutoModelForCausalLM.from_pretrained(
-                device_map=cfg.device_map,
+                    base_model,
-                load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
+                    device_map=cfg.device_map,
-                load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
+                    torch_dtype=cfg.torch_dtype,
-                torch_dtype=cfg.torch_dtype,
+                    trust_remote_code=cfg.trust_remote_code or False,
-                trust_remote_code=cfg.trust_remote_code or False,
+                    **model_kwargs,
-                **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 +313,12 @@ def load_model(
                 module.to(torch.float32)
 
     needs_fa2_dtype = cfg.adapter or cfg.fsdp
-    if not cfg.gptq and (
+    if (cfg.adapter == "lora" and load_in_8bit) or (
-        (cfg.adapter == "lora" and load_in_8bit)
+        cfg.adapter == "qlora" and cfg.load_in_4bit
-        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
         )
@@ -371,7 +326,7 @@ def load_model(
 
     # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
     # convert them back to fp16/bf16 for flash-attn compatibility.
-    if needs_fa2_dtype or (cfg.flash_attention and cfg.is_llama_derived_model):
+    if needs_fa2_dtype and (cfg.flash_attention and cfg.is_llama_derived_model):
         LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
         for name, module in model.named_modules():
             if "norm" in name:
@@ -385,22 +340,10 @@ 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/tokenization.py

@@ -8,13 +8,13 @@ from termcolor import colored
 LOG = logging.getLogger("axolotl")
 
 
-def check_dataset_labels(dataset, tokenizer, num_examples=5, text_only=False):
+def check_dataset_labels(dataset, tokenizer):
     # the dataset is already shuffled, so let's just check the first 5 elements
-    for idx in range(num_examples):
+    for idx in range(5):
-        check_example_labels(dataset[idx], tokenizer, text_only=text_only)
+        check_example_labels(dataset[idx], tokenizer)
 
 
-def check_example_labels(example, tokenizer, text_only=False):
+def check_example_labels(example, tokenizer):
     # Get the input_ids, labels, and attention_mask from the dataset
     input_ids = example["input_ids"]
     labels = example["labels"]
@@ -29,10 +29,8 @@ def check_example_labels(example, tokenizer, text_only=False):
         decoded_input_token = tokenizer.decode(input_id)
         # Choose the color based on whether the label has the ignore value or not
         color = "red" if label_id == -100 else ("yellow" if label_id == 0 else "green")
-        colored_token = colored(decoded_input_token, color) + (
+        colored_token = colored(decoded_input_token, color) + colored(
-            not text_only
+            f"({label_id}, {mask}, {input_id})", "white"
-            and colored(f"({label_id}, {mask}, {input_id})", "white")
-            or ""
         )
         colored_tokens.append(colored_token)
 

src/axolotl/utils/trainer.py

@@ -12,15 +12,9 @@ from typing import Optional, Union
 
 import numpy as np
 import torch.cuda
-import transformers
 from datasets import Dataset, set_caching_enabled
 from torch.optim.lr_scheduler import OneCycleLR
-from torch.utils.data import (
+from torch.utils.data import DataLoader, DistributedSampler, RandomSampler
-    DataLoader,
-    DistributedSampler,
-    RandomSampler,
-    SequentialSampler,
-)
 from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
 from transformers.trainer_pt_utils import SequentialDistributedSampler
 
@@ -29,7 +23,6 @@ from axolotl.utils.callbacks import (
     GPUStatsCallback,
     SaveBetterTransformerModelCallback,
     SavePeftModelCallback,
-    bench_eval_callback_factory,
 )
 from axolotl.utils.collators import DataCollatorForSeq2Seq
 from axolotl.utils.dataloader import MultipackDistributedDataloader
@@ -134,27 +127,6 @@ class AxolotlTrainingArguments(TrainingArguments):
         default=None,
         metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
     )
-    bench_split: Optional[str] = field(
-        default="eval", metadata={"help": "The benchmark split to run on"}
-    )
-    bench_dataset: Optional[str] = field(
-        default="pharaouk/dharma-1/dharma_1_mini.json",
-        metadata={
-            "help": "Benchmark dataset to use: options are `mmlu-zs`, `mmlu-fs`, or the full path to the dataset file"
-        },
-    )
-    do_bench_eval: Optional[bool] = field(
-        default=False, metadata={"help": "Whether to run the Benchmark evaluation."}
-    )
-    max_bench_samples: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "If set, only evaluates on `max_bench_samples` of the benchmark dataset."
-        },
-    )
-    bench_source_max_len: int = field(
-        default=2048, metadata={"help": "Maximum source sequence length for bench."}
-    )
 
 
 class AxolotlTrainer(Trainer):
@@ -164,10 +136,6 @@ class AxolotlTrainer(Trainer):
 
     args = None  # type: AxolotlTrainingArguments
 
-    def __init__(self, *args, bench_data_collator=None, **kwargs):
-        self.bench_data_collator = bench_data_collator
-        super().__init__(*args, **kwargs)
-
     def create_scheduler(
         self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
     ):
@@ -258,31 +226,6 @@ class AxolotlTrainer(Trainer):
             )
         return super().get_eval_dataloader(eval_dataset)
 
-    def _get_bench_sampler(
-        self, bench_dataset: Dataset
-    ) -> Optional[torch.utils.data.Sampler]:
-        if self.args.world_size <= 1:
-            return SequentialSampler(bench_dataset)
-        return None
-
-    def get_bench_dataloader(
-        self,
-        bench_dataset: Dataset,
-    ) -> Union[DataLoader, MultipackDistributedDataloader]:
-        dataloader_params = {
-            "batch_size": self.args.eval_batch_size,
-            "collate_fn": self.bench_data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-        }
-
-        if not isinstance(bench_dataset, torch.utils.data.IterableDataset):
-            dataloader_params["sampler"] = self._get_bench_sampler(bench_dataset)
-            dataloader_params["drop_last"] = self.args.dataloader_drop_last
-
-        return DataLoader(bench_dataset, **dataloader_params)
-        # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
-
     def compute_loss(self, model, inputs, return_outputs=False):
         # use one's weighted cross entropy loss calc
         # if self.args.sample_packing:
@@ -361,7 +304,7 @@ def add_position_ids(sample):
 
 
 def drop_long_seq(sample, sequence_len=2048):
-    return len(sample["input_ids"]) <= sequence_len and len(sample["input_ids"]) > 0
+    return len(sample["input_ids"]) <= sequence_len
 
 
 @contextmanager
@@ -401,16 +344,6 @@ def calculate_total_num_steps(cfg, train_dataset, tokenizer):
             LOG.info(f"📝 UPDATE CONFIG WITH: `total_num_tokens: {total_num_tokens}`")
             cfg.total_num_tokens = total_num_tokens
 
-        if not cfg.total_supervised_tokens:
-            total_supervised_tokens = (
-                train_dataset.data.column("labels")
-                .to_pandas()
-                .apply(lambda x: np.sum(np.array(x) != -100))
-                .sum()
-            )
-            LOG.info(f"`total_supervised_tokens: {total_supervised_tokens}`")
-            cfg.total_supervised_tokens = total_supervised_tokens
-
         if cfg.sample_packing_eff_est:
             total_num_steps = (
                 # match count to len est in dataloader
@@ -514,23 +447,7 @@ 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:
+        training_arguments_kwargs["gradient_checkpointing"] = cfg.gradient_checkpointing
-            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
     if cfg.fsdp:
         training_arguments_kwargs["fsdp"] = cfg.fsdp
         if cfg.fsdp_config:
@@ -584,20 +501,6 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
             "steps" if cfg.save_steps else "epoch"
         )
 
-    if cfg.do_bench_eval:
-        training_arguments_kwargs["do_bench_eval"] = cfg.do_bench_eval
-        if cfg.bench_dataset:
-            training_arguments_kwargs["bench_dataset"] = cfg.bench_dataset
-
-    # DDP Config
-    if cfg.ddp_timeout:
-        training_arguments_kwargs["ddp_timeout"] = cfg.ddp_timeout
-    # see https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html
-    if cfg.ddp_bucket_cap_mb:
-        training_arguments_kwargs["ddp_bucket_cap_mb"] = cfg.ddp_bucket_cap_mb
-    if cfg.ddp_broadcast_buffers is not None:
-        training_arguments_kwargs["ddp_broadcast_buffers"] = cfg.ddp_broadcast_buffers
-
     training_args = AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
         max_steps=total_num_steps if cfg.max_steps else -1,
         max_seq_length=cfg.sequence_len,
@@ -710,16 +613,8 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
             return_tensors="pt",
             **data_collator_kwargs,
         ),
-        bench_data_collator=transformers.DataCollatorForSeq2Seq(
-            tokenizer,
-            return_tensors="pt",
-            **data_collator_kwargs,
-        ),
         callbacks=callbacks,
         **trainer_kwargs,
     )
 
-    if cfg.do_bench_eval:
-        trainer.add_callback(bench_eval_callback_factory(trainer, tokenizer))
-
     return trainer