add sync_model_states parameter to fix resume from checkpoint with fsdp

fix formatting for linter
fixes FSDP resume from checkpoint (unpacked only)
chore: fix linter
chore: lint

.github/workflows/main.yml

@@ -23,6 +23,11 @@ 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
@@ -68,6 +73,11 @@ 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 .
+          pip install -e .[peft]
           pip install -r requirements-tests.txt
 
       - name: Run tests

README.md

@@ -665,6 +665,7 @@ fsdp:
 fsdp_config:
   fsdp_offload_params: true
   fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sync_module_states: true
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 

docker/Dockerfile

@@ -11,13 +11,14 @@ 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 [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[flash-attn,gptq,$AXOLOTL_EXTRAS]; \
+        pip install -e .[flash-attn,$AXOLOTL_EXTRAS]; \
     else \
-        pip install -e .[flash-attn,gptq]; \
+        pip install -e .[flash-attn]; \
     fi
 
 # fix so that git fetch/pull from remote works

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

@@ -0,0 +1,8 @@
+# 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

@@ -0,0 +1,63 @@
+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

@@ -1,76 +0,0 @@
-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

@@ -1,13 +1,10 @@
---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

setup.py

@@ -2,27 +2,15 @@
 
 from setuptools import find_packages, setup
 
-
-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()
-
+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)
 
 setup(
     name="axolotl",
@@ -31,10 +19,12 @@ setup(
     package_dir={"": "src"},
     packages=find_packages(),
     install_requires=install_requires,
-    dependency_links=dependency_links,
     extras_require={
         "gptq": [
-            "auto-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",
         ],
         "flash-attn": [
             "flash-attn==2.0.8",
@@ -42,5 +32,8 @@ setup(
         "extras": [
             "deepspeed",
         ],
+        "peft": [
+            "peft @ git+https://github.com/huggingface/peft.git",
+        ],
     },
 )

src/axolotl/monkeypatch/fsdp.py

@@ -0,0 +1,45 @@
+"""
+Monkeypatch to fix fsdp set state when no previous state was set
+"""
+
+import contextlib
+from typing import Generator, Optional
+
+import torch
+from torch import nn
+from torch.distributed.fsdp.api import (
+    OptimStateDictConfig,
+    StateDictConfig,
+    StateDictType,
+)
+from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel
+
+
+@staticmethod
+@contextlib.contextmanager
+def state_dict_type_patch(
+    module: nn.Module,
+    state_dict_type: StateDictType,
+    state_dict_config: Optional[StateDictConfig] = None,
+    optim_state_dict_config: Optional[OptimStateDictConfig] = None,
+) -> Generator:
+    prev_state_dict_settings = FullyShardedDataParallel.set_state_dict_type(
+        module,
+        state_dict_type,
+        state_dict_config,
+        optim_state_dict_config,
+    )
+    yield
+    if prev_state_dict_settings.state_dict_type:
+        FullyShardedDataParallel.set_state_dict_type(
+            module,
+            prev_state_dict_settings.state_dict_type,
+            prev_state_dict_settings.state_dict_config,
+            prev_state_dict_settings.optim_state_dict_config,
+        )
+
+
+def replace_fsdp_state_dict_type():
+    torch.distributed.fsdp.fully_sharded_data_parallel.FullyShardedDataParallel.state_dict_type = (
+        state_dict_type_patch
+    )

src/axolotl/utils/callbacks.py

@@ -1,9 +1,19 @@
 """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 datasets import load_dataset
 from optimum.bettertransformer import BetterTransformer
+from tqdm import tqdm
 from transformers import (
     TrainerCallback,
     TrainerControl,
@@ -13,8 +23,19 @@ 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 (
+    barrier,
+    gather_scalar_from_all_ranks,
+    get_world_size,
+    is_main_process,
+    zero_first,
+)
+
+if TYPE_CHECKING:
+    from axolotl.utils.trainer import AxolotlTrainingArguments
 
 LOG = logging.getLogger("axolotl.callbacks")
+IGNORE_INDEX = -100
 
 
 class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-methods
@@ -96,3 +117,192 @@ 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 zero_first(is_main_process()):
+        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)
+            barrier()
+            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 = {"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 = []
+                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"]
+                    if not pd.isna(bench_score):
+                        results[
+                            f"bench_{bench_split}_accuracy_{bench_name}"
+                        ] = bench_score
+                        bench_scores.append(bench_score)
+                    else:
+                        results[f"bench_{bench_split}_accuracy_{bench_name}"] = 0.0
+                        bench_scores.append(0.0)
+                results[f"bench_{bench_split}_accuracy"] = np.mean(bench_scores)
+                trainer.log(results)
+
+    return BenchEvalCallback

src/axolotl/utils/config.py

@@ -97,7 +97,9 @@ 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")
+        raise ValueError(
+            "cfg.load_4bit parameter has been deprecated and replaced by cfg.gptq"
+        )
 
     if cfg.adapter == "qlora":
         if cfg.merge_lora:
@@ -150,6 +152,16 @@ def validate_config(cfg):
     if (cfg.base_model and "falcon" in cfg.base_model.lower()) and cfg.fsdp:
         raise ValueError("FSDP is not supported for falcon models")
 
+    if (
+        cfg.fsdp
+        and cfg.fsdp_config
+        and cfg.fsdp_config.fsdp_state_dict_type
+        and not cfg.fsdp_config.fsdp_sync_module_states
+    ):
+        LOG.warning(
+            "We recommend setting fsdp_config.fsdp_sync_module_states to `true`"
+        )
+
     if (
         cfg.base_model and "mpt" in cfg.base_model.lower()
     ) and cfg.gradient_checkpointing:

src/axolotl/utils/distributed.py

@@ -1,8 +1,10 @@
 """
 utility helpers for distributed checks
 """
+import os
 from contextlib import contextmanager
 
+import torch
 import torch.distributed as dist
 from accelerate import Accelerator
 
@@ -43,6 +45,10 @@ def is_main_process():
     return dist.get_rank() == 0
 
 
+def get_world_size():
+    return int(os.getenv("WORLD_SIZE", "1"))
+
+
 @contextmanager
 def zero_first(is_main):
     """
@@ -53,3 +59,35 @@ def zero_first(is_main):
     yield
     if is_main:  # then rank 0 waits after it has run the context
         barrier()
+
+
+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.
+
+    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()
+    value_tensor = torch.tensor(value_scalar, device=dist.get_rank()).float()
+
+    if not is_main_process():
+        dist.gather(value_tensor, dst=0)
+    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, prepare_model_for_kbit_training
+from peft import PeftConfig
 from transformers import (  # noqa: F401
     AutoConfig,
     AutoModelForCausalLM,
     AutoTokenizer,
     BitsAndBytesConfig,
-    GPTQConfig,
     LlamaConfig,
     PreTrainedModel,
     PreTrainedTokenizerBase,
@@ -155,15 +155,32 @@ 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,
@@ -174,7 +191,45 @@ def load_model(
             bnb_4bit_quant_type="nf4",
         )
     try:
-        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
+        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:
             from transformers import LlamaForCausalLM
 
             config_kwargs = {}
@@ -220,24 +275,15 @@ def load_model(
         #     )
         #     model.train() # sets to train instead of eval mode
         elif model_type and not cfg.trust_remote_code:
-            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,
-                )
+            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,
@@ -313,12 +359,11 @@ def load_model(
                 module.to(torch.float32)
 
     needs_fa2_dtype = cfg.adapter or cfg.fsdp
-    if (cfg.adapter == "lora" and load_in_8bit) or (
-        cfg.adapter == "qlora" and cfg.load_in_4bit
+    if not cfg.gptq and (
+        (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
         )
@@ -340,10 +385,22 @@ 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.load_in_4bit)
+        and (cfg.gptq or 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

@@ -12,9 +12,15 @@ 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 DataLoader, DistributedSampler, RandomSampler
+from torch.utils.data import (
+    DataLoader,
+    DistributedSampler,
+    RandomSampler,
+    SequentialSampler,
+)
 from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
 from transformers.trainer_pt_utils import SequentialDistributedSampler
 
@@ -23,6 +29,7 @@ from axolotl.utils.callbacks import (
     GPUStatsCallback,
     SaveBetterTransformerModelCallback,
     SavePeftModelCallback,
+    bench_eval_callback_factory,
 )
 from axolotl.utils.collators import DataCollatorForSeq2Seq
 from axolotl.utils.dataloader import MultipackDistributedDataloader
@@ -127,6 +134,27 @@ 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):
@@ -136,6 +164,10 @@ 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
     ):
@@ -226,6 +258,31 @@ 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:
@@ -414,6 +471,9 @@ def setup_fsdp_envs(cfg):
         os.environ[
             "FSDP_TRANSFORMER_CLS_TO_WRAP"
         ] = cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap
+    from axolotl.monkeypatch.fsdp import replace_fsdp_state_dict_type
+
+    replace_fsdp_state_dict_type()
 
 
 def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
@@ -447,7 +507,23 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         training_arguments_kwargs["seed"] = cfg.seed
 
     if cfg.gradient_checkpointing:
-        training_arguments_kwargs["gradient_checkpointing"] = 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
     if cfg.fsdp:
         training_arguments_kwargs["fsdp"] = cfg.fsdp
         if cfg.fsdp_config:
@@ -501,6 +577,11 @@ 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
+
     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,
@@ -613,8 +694,16 @@ 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