chore: lint

* support for custom lr groups for non-embedding modules

invert name check for group modules
include lr_groups in training args
additional conditional for creating optimizer
fix regular params as w weight decay
fix lookup and add docs

* address pr feedback

.github/workflows/lint.yml

@@ -1,6 +1,7 @@
 name: lint
 on:
   # check on PRs, and manual triggers
+  merge_group:
   pull_request:
       paths:
        - '**.py'

.github/workflows/main.yml

@@ -25,7 +25,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.3.1
             axolotl_extras: mamba-ssm
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -36,6 +35,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -92,7 +92,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.3.1
             axolotl_extras:
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -103,6 +102,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/multi-gpu-e2e.yml

@@ -52,7 +52,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==0.63.64 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV

.github/workflows/tests-nightly.yml

@@ -129,7 +129,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==0.63.64 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV

.github/workflows/tests.yml

@@ -1,6 +1,7 @@
 name: Tests
 on:
   # check on push/merge to main, PRs, and manual triggers
+  merge_group:
   push:
     branches:
       - "main"
@@ -60,6 +61,15 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-${{ hashFiles('**/conftest.py') }}
+
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -100,6 +110,15 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
@@ -115,6 +134,15 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-${{ hashFiles('**/conftest.py') }}
+
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -156,6 +184,15 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   docker-e2e-tests-1st:
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
@@ -183,7 +220,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==0.63.64 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -229,7 +266,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==0.63.64 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV

.pre-commit-config.yaml

@@ -23,7 +23,7 @@ repos:
     hooks:
     - id: flake8
 -   repo: https://github.com/PyCQA/pylint
-    rev: v2.17.4
+    rev: v3.3.0
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy

.pylintrc

@@ -1,5 +1,5 @@
 [MASTER]
-init-hook="from pylint.config import find_pylintrc; import os, sys; sys.path.append(os.path.dirname(find_pylintrc()))"
+init-hook="from pylint.config import find_default_config_files; import sys; sys.path.append(next(find_default_config_files()).parent.as_posix())"
 
 [TYPECHECK]
 
@@ -12,3 +12,4 @@ generated-members=numpy.*, torch.*
 disable=missing-function-docstring, line-too-long, import-error,
     too-many-arguments, too-many-locals, too-many-statements, too-many-branches, too-few-public-methods,
     too-many-instance-attributes, fixme, import-outside-toplevel, logging-fstring-interpolation,
+    too-many-positional-arguments, possibly-used-before-assignment

README.md

@@ -217,7 +217,7 @@ If you love axolotl, consider sponsoring the project by reaching out directly to
 
 ---
 
-- [Modal](https://www.modal.com?utm_source=github&utm_medium=github&utm_campaign=axolotl) Modal lets you run data/AI jobs in the cloud, by just writing a few lines of Python. Customers use Modal to deploy Gen AI models at large scale, fine-tune LLM models, run protein folding simulations, and much more.
+- [Modal](https://modal.com/) Modal lets you run data/AI jobs in the cloud, by just writing a few lines of Python. Customers use Modal to deploy Gen AI models at large scale, fine-tune LLM models, run protein folding simulations, and much more.
 
 ---
 
@@ -519,8 +519,8 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
       train_on_split: validation
 
       # loading from s3 or gcs
-      # s3 creds will be loaded from the system default and gcs only supports public access
-    - path: s3://path_to_ds # Accepts folder with arrow/parquet or file path like above. Supports s3, gcs.
+      # s3 creds will be loaded from the system default / gcs will attempt to load from gcloud creds, google metadata service, or anon
+    - path: s3://path_to_ds # Accepts folder with arrow/parquet or file path like above
       ...
 
       # Loading Data From a Public URL

cicd/Dockerfile.jinja

@@ -8,6 +8,7 @@ ENV PYTORCH_VERSION="{{ PYTORCH_VERSION }}"
 ENV GITHUB_REF="{{ GITHUB_REF }}"
 ENV GITHUB_SHA="{{ GITHUB_SHA }}"
 ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
+ENV HF_HOME="{{ HF_HOME }}"
 
 RUN apt-get update && \
     apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev

cicd/cicd.sh

@@ -6,5 +6,6 @@ python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
 # pytest -v --durations=10 -n8 --dist loadfile /workspace/axolotl/tests/patched/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/
+pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
-pytest -v --durations=10 --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/
+pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/

cicd/multigpu.py

@@ -28,6 +28,7 @@ df_args = {
     "CUDA": os.environ.get("CUDA", "121"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "HF_HOME": "/workspace/data/huggingface-cache/hub",
 }
 
 dockerfile_contents = df_template.render(**df_args)
@@ -48,6 +49,12 @@ cicd_image = (
 
 app = App("Axolotl CI/CD", secrets=[])
 
+hf_cache_volume = modal.Volume.from_name(
+    "axolotl-ci-hf-hub-cache", create_if_missing=True
+)
+VOLUME_CONFIG = {
+    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
+}
 
 N_GPUS = int(os.environ.get("N_GPUS", 2))
 GPU_CONFIG = modal.gpu.H100(count=N_GPUS)
@@ -67,6 +74,7 @@ def run_cmd(cmd: str, run_folder: str):
     timeout=60 * 60,
     cpu=8.0,
     memory=131072 * N_GPUS,
+    volumes=VOLUME_CONFIG,
 )
 def cicd_pytest():
     run_cmd("./cicd/multigpu.sh", "/workspace/axolotl")

cicd/tests.py

@@ -29,6 +29,7 @@ df_args = {
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
+    "HF_HOME": "/workspace/data/huggingface-cache/hub",
 }
 
 dockerfile_contents = df_template.render(**df_args)
@@ -50,6 +51,12 @@ cicd_image = (
 
 app = App("Axolotl CI/CD", secrets=[])
 
+hf_cache_volume = modal.Volume.from_name(
+    "axolotl-ci-hf-hub-cache", create_if_missing=True
+)
+VOLUME_CONFIG = {
+    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
+}
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
 GPU_CONFIG = modal.gpu.A10G(count=N_GPUS)
@@ -69,6 +76,7 @@ def run_cmd(cmd: str, run_folder: str):
     timeout=60 * 60,
     cpu=8.0,
     memory=131072,
+    volumes=VOLUME_CONFIG,
 )
 def cicd_pytest():
     run_cmd("./cicd/cicd.sh", "/workspace/axolotl")

docker/Dockerfile-cloud

@@ -20,7 +20,8 @@ RUN apt install --yes --no-install-recommends openssh-server tmux && \
     printf "\n[[ -z \"\$TMUX\"  ]] && { tmux attach-session -t ssh_tmux || tmux new-session -s ssh_tmux; exit; }\n" >> ~/.bashrc && \
     printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \
     chmod +x /workspace/axolotl/scripts/cloud-entrypoint.sh && \
-    chmod +x /root/cloud-entrypoint.sh
+    chmod +x /root/cloud-entrypoint.sh && \
+    echo 'set-option -g history-limit 5000' >> ~/.tmux.conf
 
 ENTRYPOINT ["/root/cloud-entrypoint.sh"]
 CMD ["sleep", "infinity"]

docs/config.qmd

@@ -244,6 +244,8 @@ total_num_tokens:
 sample_packing_group_size: 100000
 # The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
 sample_packing_bin_size: 200
+# whether to concatenate samples during pretraining
+pretraining_sample_concatenation:
 
 # Use batch flattening for speedups when not using sample_packing
 batch_flattening:
@@ -358,10 +360,11 @@ warmup_ratio: 0.05  # cannot use with warmup_steps
 learning_rate: 0.00003
 lr_quadratic_warmup:
 logging_steps:
-eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
+eval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps
 evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps
-save_strategy: # Set to `"no"` to skip checkpoint saves
-save_steps: # Leave empty to save at each epoch
+eval_strategy: # Set to `"no"` to skip evaluation, `"epoch"` at end of each epoch, leave empty to infer from `eval_steps`.
+save_strategy: # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of each epoch, `"best"` when better result is achieved, leave empty to infer from `save_steps`.
+save_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps
 saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
 # Maximum number of iterations to train for. It precedes num_epochs which means that

docs/dataset-formats/pretraining.qmd

@@ -19,7 +19,14 @@ For pretraining, there is no prompt template or roles.  The only required field
 Axolotl usually loads the entire dataset into memory. This will be challenging for large datasets. Use the following config to enable streaming:
 
 ```{.yaml filename="config.yaml"}
-pretraining_dataset: # hf path only
+pretraining_dataset:
+  - name:
+    path:
+    split:
+    text_column: # column in dataset with the data, usually `text`
+    type: pretrain
+    trust_remote_code:
+    skip: # number of rows of data to skip over from the beginning
 ...
 ```
 

docs/lr_groups.qmd

@@ -0,0 +1,29 @@
+---
+title: Learning Rate Groups
+description: "Setting different learning rates by module name"
+---
+
+## Background
+
+Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of
+modules in a model.
+
+## Example
+
+```yaml
+lr_groups:
+  - name: o_proj
+    modules:
+      - self_attn.o_proj.weight
+    lr: 1e-6
+  - name: q_proj
+    modules:
+      - model.layers.2.self_attn.q_proj.weight
+    lr: 1e-5
+
+learning_rate: 2e-5
+```
+
+In this example, we have a default learning rate of 2e-5 across the entire model, but we have a separate learning rate
+of 1e-6 for all the self attention `o_proj` modules across all layers, and a learning are of 1e-5 to the 3rd layer's
+self attention `q_proj` module.

examples/cloud/modal.yaml

@@ -1,15 +0,0 @@
-volumes:
-  - name: axolotl-data
-    mount: /workspace/data
-  - name: axolotl-artifacts
-    mount: /workspace/artifacts
-secrets:
-  - HF_TOKEN
-  - WANDB_API_KEY
-branch: cli-cloud-modal
-gpu: h100
-gpu_count: 1
-memory: 128
-timeout: 86400
-timeout_preprocess: 14400
-memory_preprocess: 32

lm_eval-kd.yaml

@@ -1,11 +0,0 @@
-lm_eval_model: axolotl-ai-co/numina-8b-ep1-exp1
-lm_eval_tasks:
-  - leaderboard_math_hard
-lm_eval_batch_size: 64
-
-apply_chat_template: false
-wandb_project: numina-kd-experiment
-wandb_entity: axolotl-ai
-bf16: true
-flash_attention: true
-output_dir: ./outputs/model-evals-out

requirements.txt

@@ -2,7 +2,7 @@
 
 # START section of dependencies that don't install on Darwin/MacOS
 bitsandbytes==0.45.0
-triton>=2.3.0
+triton>=3.0.0
 mamba-ssm==1.2.0.post1
 flash-attn==2.7.0.post2
 xformers>=0.0.23.post1
@@ -13,19 +13,18 @@ liger-kernel==0.5.2
 packaging==23.2
 
 peft==0.14.0
-transformers==4.47.1
-tokenizers>=0.20.1
-accelerate==1.2.1
-datasets==3.1.0
+transformers==4.48.1
+tokenizers>=0.21.0
+accelerate==1.3.0
+datasets==3.2.0
 deepspeed==0.16.1
-trl==0.12.1
+trl==0.13.0
 
 optimum==1.16.2
 hf_transfer
 sentencepiece
 gradio==3.50.2
 
-modal==0.70.5
 pydantic==2.6.3
 addict
 fire
@@ -62,4 +61,4 @@ antlr4-python3-runtime==4.13.2
 torchao==0.7.0
 schedulefree==1.3.0
 
-axolotl-contribs-lgpl==0.0.2
+axolotl-contribs-lgpl==0.0.3

scripts/chat_datasets.py

@@ -30,7 +30,7 @@ def parse_dataset(dataset=None, split="train"):
         )
     ds_cfg["field_messages"] = field_messages
 
-    message_fields = features["conversations"][0].keys()
+    message_fields = features[field_messages][0].keys()
     message_field_role = None
     for key in ["from", "role"]:
         if key in message_fields:

scripts/finetune.py

@@ -1,52 +0,0 @@
-"""Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
-import logging
-from pathlib import Path
-
-import fire
-import transformers
-
-from axolotl.cli import (
-    check_accelerate_default_config,
-    check_user_token,
-    do_inference,
-    do_merge_lora,
-    load_cfg,
-    load_datasets,
-    print_axolotl_text_art,
-)
-from axolotl.cli.shard import shard
-from axolotl.common.cli import TrainerCliArgs
-from axolotl.train import train
-
-LOG = logging.getLogger("axolotl.scripts.finetune")
-
-
-def do_cli(config: Path = Path("examples/"), **kwargs):
-    print_axolotl_text_art()
-    LOG.warning(
-        str(
-            PendingDeprecationWarning(
-                "scripts/finetune.py will be replaced with calling axolotl.cli.train"
-            )
-        )
-    )
-    parsed_cfg = load_cfg(config, **kwargs)
-    check_accelerate_default_config()
-    check_user_token()
-    parser = transformers.HfArgumentParser((TrainerCliArgs))
-    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
-        return_remaining_strings=True
-    )
-    if parsed_cli_args.inference:
-        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)
-        train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
-
-
-if __name__ == "__main__":
-    fire.Fire(do_cli)

scripts/motd

@@ -1,15 +1,10 @@
 
-     #@@ #@@      @@# @@#
-    @@  @@          @@  @@           =@@#                               @@                 #@    =@@#.
-    @@    #@@@@@@@@@    @@           #@#@=                              @@                 #@     .=@@
-      #@@@@@@@@@@@@@@@@@            =@# @#     ##=     ##    =####=+    @@      =#####+  =#@@###.   @@
-    @@@@@@@@@@/  +@@/  +@@          #@  =@=     #@=   @@   =@#+  +#@#   @@    =@#+  +#@#   #@.      @@
-    @@@@@@@@@@  ##@@  ##@@         =@#   @#      =@# @#    @@      @@   @@    @@      #@   #@       @@
-     @@@@@@@@@@@@@@@@@@@@          #@=+++#@=      =@@#     @@      @@   @@    @@      #@   #@       @@
-                                  =@#=====@@     =@# @#    @@      @@   @@    @@      #@   #@       @@
-    @@@@@@@@@@@@@@@@  @@@@        #@      #@=   #@=  +@@   #@#    =@#   @@.   =@#    =@#   #@.      @@
-                                 =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
-    @@@@  @@@@@@@@@@@@@@@@
+                                 dP            dP   dP
+                                 88            88   88
+      .d8888b. dP.  .dP .d8888b. 88 .d8888b. d8888P 88
+      88'  `88  `8bd8'  88'  `88 88 88'  `88   88   88
+      88.  .88  .d88b.  88.  .88 88 88.  .88   88   88
+      `88888P8 dP'  `dP `88888P' dP `88888P'   dP   dP
 
 Welcome to the axolotl cloud image! If the you've mounted a disk to /workspace and the axolotl directory ie empty, run the following commands:
 

setup.py

@@ -1,4 +1,5 @@
 """setup.py for axolotl"""
+
 import ast
 import os
 import platform
@@ -29,15 +30,30 @@ def parse_requirements():
             elif not is_extras and line and line[0] != "#":
                 # Handle standard packages
                 _install_requires.append(line)
-
     try:
         xformers_version = [req for req in _install_requires if "xformers" in req][0]
+        triton_version = [req for req in _install_requires if "triton" in req][0]
         torchao_version = [req for req in _install_requires if "torchao" in req][0]
         autoawq_version = [req for req in _install_requires if "autoawq" in req][0]
-
         if "Darwin" in platform.system():
-            # don't install xformers on MacOS
-            _install_requires.pop(_install_requires.index(xformers_version))
+            # skip packages not compatible with OSX
+            skip_packages = [
+                "bitsandbytes",
+                "triton",
+                "mamba-ssm",
+                "flash-attn",
+                "xformers",
+                "autoawq",
+                "liger-kernel",
+            ]
+            _install_requires = [
+                req
+                for req in _install_requires
+                if re.split(r"[>=<]", req)[0].strip() not in skip_packages
+            ]
+            print(
+                _install_requires, [req in skip_packages for req in _install_requires]
+            )
         else:
             # detect the version of torch already installed
             # and set it so dependencies don't clobber the torch version
@@ -73,6 +89,8 @@ def parse_requirements():
                     _install_requires.append("xformers==0.0.28.post1")
             elif (major, minor) >= (2, 3):
                 _install_requires.pop(_install_requires.index(torchao_version))
+                _install_requires.pop(_install_requires.index(triton_version))
+                _install_requires.append("triton>=2.3.1")
                 if patch == 0:
                     _install_requires.pop(_install_requires.index(xformers_version))
                     _install_requires.append("xformers>=0.0.26.post1")

src/axolotl/cli/__init__.py

@@ -1,568 +1,5 @@
-"""Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
+"""Axolotl CLI module initialization."""
 
-import importlib
-import json
-import logging
-import math
 import os
-import random
-import sys
-import tempfile
-from pathlib import Path
-from threading import Thread
-from typing import Any, Dict, List, Optional, Union
-from urllib.parse import urlparse
-
-import requests
-import torch
-import yaml
-
-# add src to the pythonpath so we don't need to pip install this
-from accelerate.commands.config import config_args
-from art import text2art
-from huggingface_hub import HfApi
-from huggingface_hub.utils import LocalTokenNotFoundError
-from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
-from transformers.utils import is_torch_bf16_gpu_available
-from transformers.utils.import_utils import _is_package_available
-
-from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
-from axolotl.logging_config import configure_logging
-from axolotl.train import TrainDatasetMeta
-from axolotl.utils.chat_templates import (
-    get_chat_template,
-    get_chat_template_from_config,
-)
-from axolotl.utils.comet_ import setup_comet_env_vars
-from axolotl.utils.config import (
-    normalize_cfg_datasets,
-    normalize_config,
-    prepare_plugins,
-    validate_config,
-)
-from axolotl.utils.data import load_prepare_dpo_datasets, prepare_dataset
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.distributed import is_main_process
-from axolotl.utils.mlflow_ import setup_mlflow_env_vars
-from axolotl.utils.models import load_processor, load_tokenizer
-from axolotl.utils.tokenization import check_dataset_labels
-from axolotl.utils.trainer import prepare_opinionated_env, prepare_optim_env
-from axolotl.utils.wandb_ import setup_wandb_env_vars
-
-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.scripts")
 
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
-
-AXOLOTL_LOGO = """
-     #@@ #@@      @@# @@#
-    @@  @@          @@  @@           =@@#                               @@                 #@    =@@#.
-    @@    #@@@@@@@@@    @@           #@#@=                              @@                 #@     .=@@
-      #@@@@@@@@@@@@@@@@@            =@# @#     ##=     ##    =####=+    @@      =#####+  =#@@###.   @@
-    @@@@@@@@@@/  +@@/  +@@          #@  =@=     #@=   @@   =@#+  +#@#   @@    =@#+  +#@#   #@.      @@
-    @@@@@@@@@@  ##@@  ##@@         =@#   @#      =@# @#    @@      @@   @@    @@      #@   #@       @@
-     @@@@@@@@@@@@@@@@@@@@          #@=+++#@=      =@@#     @@      @@   @@    @@      #@   #@       @@
-                                  =@#=====@@     =@# @#    @@      @@   @@    @@      #@   #@       @@
-    @@@@@@@@@@@@@@@@  @@@@        #@      #@=   #@=  +@@   #@#    =@#   @@.   =@#    =@#   #@.      @@
-                                 =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
-    @@@@  @@@@@@@@@@@@@@@@
-"""
-
-
-def print_legacy_axolotl_text_art(suffix=None):
-    font = "nancyj"
-    ascii_text = "  axolotl"
-    if suffix:
-        ascii_text += f"  x  {suffix}"
-    ascii_art = text2art(ascii_text, font=font)
-
-    if is_main_process():
-        print(ascii_art)
-
-    print_dep_versions()
-
-
-def print_axolotl_text_art(
-    **kwargs,  # pylint: disable=unused-argument
-):
-    if is_main_process():
-        print(AXOLOTL_LOGO)
-
-
-def print_dep_versions():
-    packages = ["accelerate", "peft", "transformers", "trl", "torch", "bitsandbytes"]
-    max_len = max(len(pkg) for pkg in packages)
-    if is_main_process():
-        print("*" * 40)
-        print("**** Axolotl Dependency Versions *****")
-        for pkg in packages:
-            pkg_version = _is_package_available(pkg, return_version=True)
-            print(f"{pkg: >{max_len}}: {pkg_version[1]: <15}")
-        print("*" * 40)
-
-
-def check_remote_config(config: Union[str, Path]):
-    # Check if the config is a valid HTTPS URL to a .yml or .yaml file
-    if not (isinstance(config, str) and config.startswith("https://")):
-        return config  # Return the original value if it's not a valid URL
-
-    filename = os.path.basename(urlparse(config).path)
-    temp_dir = tempfile.mkdtemp()
-
-    try:
-        response = requests.get(config, timeout=30)
-        response.raise_for_status()  # Check for HTTP errors
-
-        content = response.content
-        try:
-            # Try parsing as JSON first to catch cases where JSON content is mistakenly considered YAML
-            json.loads(content)
-            # Log a warning but do not raise an error; JSON is technically valid YAML - this can happen when you forget to point to a raw github link
-            LOG.warning(
-                f"Warning: The content of the file at {config} is JSON, which is technically valid YAML but might not be intended."
-            )
-        except json.JSONDecodeError:
-            # If it's not valid JSON, verify it's valid YAML
-            try:
-                yaml.safe_load(content)
-            except yaml.YAMLError as err:
-                raise ValueError(
-                    f"Failed to parse the content at {config} as YAML: {err}"
-                ) from err
-
-        # Write the content to a file if it's valid YAML (or JSON treated as YAML)
-        output_path = Path(temp_dir) / filename
-        with open(output_path, "wb") as file:
-            file.write(content)
-        LOG.info(
-            f"Using the following config obtained from {config}: \n\n{content.decode('utf-8')}\n"
-        )
-        return output_path
-
-    except requests.RequestException as err:
-        # This catches all requests-related exceptions including HTTPError
-        raise RuntimeError(f"Failed to download {config}: {err}") from err
-    except Exception as err:
-        # Catch-all for any other exceptions
-        raise err
-
-
-def get_multi_line_input() -> Optional[str]:
-    print("Give me an instruction (Ctrl + D to submit): ")
-    instruction = ""
-    for line in sys.stdin:
-        instruction += line  # pylint: disable=consider-using-join
-    # instruction = pathlib.Path("/proc/self/fd/0").read_text()
-    return instruction
-
-
-def do_merge_lora(
-    *,
-    cfg: DictDefault,
-    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(progressbar=True)
-    try:
-        model.to(dtype=cfg.torch_dtype)
-    except RuntimeError:
-        pass
-    model.generation_config.do_sample = True
-
-    if cfg.local_rank == 0:
-        LOG.info(f"saving merged model to: {str(Path(cfg.output_dir) / 'merged')}")
-        model.save_pretrained(
-            str(Path(cfg.output_dir) / "merged"),
-            safe_serialization=safe_serialization,
-            progressbar=True,
-        )
-        tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
-
-
-def do_inference(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-):
-    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
-    prompter = cli_args.prompter
-
-    prompter_module = None
-    chat_template_str = None
-    if prompter:
-        prompter_module = getattr(
-            importlib.import_module("axolotl.prompters"), prompter
-        )
-    elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template)
-    elif cfg.datasets[0].type == "chat_template":
-        chat_template_str = get_chat_template_from_config(
-            cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer
-        )
-
-    model = model.to(cfg.device, dtype=cfg.torch_dtype)
-
-    while True:
-        print("=" * 80)
-        # support for multiline inputs
-        instruction = get_multi_line_input()
-        if not instruction:
-            return
-
-        if prompter_module:
-            prompt: str = next(
-                prompter_module().build_prompt(instruction=instruction.strip("\n"))
-            )
-        else:
-            prompt = instruction.strip()
-
-        if chat_template_str:
-            batch = tokenizer.apply_chat_template(
-                [
-                    {
-                        "role": "user",
-                        "content": prompt,
-                    }
-                ],
-                return_tensors="pt",
-                add_special_tokens=True,
-                add_generation_prompt=True,
-                chat_template=chat_template_str,
-                tokenize=True,
-                return_dict=True,
-            )
-        else:
-            batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
-
-        print("=" * 40)
-        model.eval()
-        with torch.no_grad():
-            generation_config = GenerationConfig(
-                repetition_penalty=1.1,
-                max_new_tokens=1024,
-                temperature=0.9,
-                top_p=0.95,
-                top_k=40,
-                bos_token_id=tokenizer.bos_token_id,
-                eos_token_id=tokenizer.eos_token_id,
-                pad_token_id=tokenizer.pad_token_id,
-                do_sample=True,
-                use_cache=True,
-                return_dict_in_generate=True,
-                output_attentions=False,
-                output_hidden_states=False,
-                output_scores=False,
-            )
-            streamer = TextStreamer(tokenizer)
-            generated = model.generate(
-                inputs=batch["input_ids"].to(cfg.device),
-                generation_config=generation_config,
-                streamer=streamer,
-            )
-        print("=" * 40)
-        print(tokenizer.decode(generated["sequences"].cpu().tolist()[0]))
-
-
-def do_inference_gradio(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-):
-    import gradio as gr
-
-    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
-    prompter = cli_args.prompter
-
-    prompter_module = None
-    chat_template_str = None
-    if prompter:
-        prompter_module = getattr(
-            importlib.import_module("axolotl.prompters"), prompter
-        )
-    elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
-
-    model = model.to(cfg.device, dtype=cfg.torch_dtype)
-
-    def generate(instruction):
-        if not instruction:
-            return
-        if prompter_module:
-            # pylint: disable=stop-iteration-return
-            prompt: str = next(
-                prompter_module().build_prompt(instruction=instruction.strip("\n"))
-            )
-        else:
-            prompt = instruction.strip()
-
-        if chat_template_str:
-            batch = tokenizer.apply_chat_template(
-                [
-                    {
-                        "role": "user",
-                        "content": prompt,
-                    }
-                ],
-                return_tensors="pt",
-                add_special_tokens=True,
-                add_generation_prompt=True,
-                chat_template=chat_template_str,
-                tokenize=True,
-                return_dict=True,
-            )
-        else:
-            batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
-
-        model.eval()
-        with torch.no_grad():
-            generation_config = GenerationConfig(
-                repetition_penalty=1.1,
-                max_new_tokens=cfg.get("gradio_max_new_tokens", 1024),
-                temperature=cfg.get("gradio_temperature", 0.9),
-                top_p=0.95,
-                top_k=40,
-                bos_token_id=tokenizer.bos_token_id,
-                eos_token_id=tokenizer.eos_token_id,
-                pad_token_id=tokenizer.pad_token_id,
-                do_sample=True,
-                use_cache=True,
-                return_dict_in_generate=True,
-                output_attentions=False,
-                output_hidden_states=False,
-                output_scores=False,
-            )
-            streamer = TextIteratorStreamer(tokenizer)
-            generation_kwargs = {
-                "inputs": batch["input_ids"].to(cfg.device),
-                "attention_mask": batch["attention_mask"].to(cfg.device),
-                "generation_config": generation_config,
-                "streamer": streamer,
-            }
-
-            thread = Thread(target=model.generate, kwargs=generation_kwargs)
-            thread.start()
-
-            all_text = ""
-
-            for new_text in streamer:
-                all_text += new_text
-                yield all_text
-
-    demo = gr.Interface(
-        fn=generate,
-        inputs="textbox",
-        outputs="text",
-        title=cfg.get("gradio_title", "Axolotl Gradio Interface"),
-    )
-
-    demo.queue().launch(
-        show_api=False,
-        share=cfg.get("gradio_share", True),
-        server_name=cfg.get("gradio_server_name", "127.0.0.1"),
-        server_port=cfg.get("gradio_server_port", None),
-    )
-
-
-def choose_config(path: Path):
-    yaml_files = list(path.glob("*.yml"))
-
-    if not yaml_files:
-        raise ValueError(
-            "No YAML config files found in the specified directory. Are you using a .yml extension?"
-        )
-
-    if len(yaml_files) == 1:
-        print(f"Using default YAML file '{yaml_files[0]}'")
-        return str(yaml_files[0])
-
-    print("Choose a YAML file:")
-    for idx, file in enumerate(yaml_files):
-        print(f"{idx + 1}. {file}")
-
-    chosen_file = None
-    while chosen_file is None:
-        try:
-            choice = int(input("Enter the number of your choice: "))
-            if 1 <= choice <= len(yaml_files):
-                chosen_file = str(yaml_files[choice - 1])
-            else:
-                print("Invalid choice. Please choose a number from the list.")
-        except ValueError:
-            print("Invalid input. Please enter a number.")
-
-    return chosen_file
-
-
-def check_not_in(list1: List[str], list2: Union[Dict[str, Any], List[str]]) -> bool:
-    return not any(el in list2 for el in list1)
-
-
-def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs):
-    config = check_remote_config(config)
-    if Path(config).is_dir():
-        config = choose_config(Path(config))
-
-    # load the config from the yaml file
-    with open(config, encoding="utf-8") as file:
-        cfg: DictDefault = DictDefault(yaml.safe_load(file))
-    # if there are any options passed in the cli, if it is something that seems valid from the yaml,
-    # then overwrite the value
-    cfg_keys = cfg.keys()
-    for k, _ in kwargs.items():
-        # if not strict, allow writing to cfg even if it's not in the yml already
-        if k in cfg_keys or not cfg.strict:
-            # handle booleans
-            if isinstance(cfg[k], bool):
-                cfg[k] = bool(kwargs[k])
-            else:
-                cfg[k] = kwargs[k]
-
-    cfg.axolotl_config_path = config
-
-    try:
-        device_props = torch.cuda.get_device_properties("cuda")
-        gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)
-    except:  # pylint: disable=bare-except # noqa: E722
-        gpu_version = None
-
-    prepare_plugins(cfg)
-
-    cfg = validate_config(
-        cfg,
-        capabilities={
-            "bf16": is_torch_bf16_gpu_available(),
-            "n_gpu": int(os.environ.get("WORLD_SIZE", 1)),
-            "compute_capability": gpu_version,
-        },
-        env_capabilities={
-            "torch_version": str(torch.__version__).split("+", maxsplit=1)[0],
-        },
-    )
-
-    prepare_optim_env(cfg)
-
-    prepare_opinionated_env(cfg)
-
-    normalize_config(cfg)
-
-    normalize_cfg_datasets(cfg)
-
-    setup_wandb_env_vars(cfg)
-
-    setup_mlflow_env_vars(cfg)
-
-    setup_comet_env_vars(cfg)
-
-    return cfg
-
-
-def load_datasets(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,
-) -> TrainDatasetMeta:
-    tokenizer = load_tokenizer(cfg)
-    processor = load_processor(cfg, tokenizer=tokenizer) if cfg.processor_type else None
-
-    train_dataset, eval_dataset, total_num_steps, prompters = prepare_dataset(
-        cfg,
-        tokenizer,
-        processor=processor,
-    )
-
-    if (
-        cli_args.debug
-        or cfg.debug
-        or cli_args.debug_text_only
-        or int(cli_args.debug_num_examples) > 0
-    ):
-        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,
-        )
-
-        LOG.info("printing prompters...")
-        for prompter in prompters:
-            LOG.info(prompter)
-
-    return TrainDatasetMeta(
-        train_dataset=train_dataset,
-        eval_dataset=eval_dataset,
-        total_num_steps=total_num_steps,
-    )
-
-
-def load_rl_datasets(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,  # pylint: disable=unused-argument
-) -> TrainDatasetMeta:
-    train_dataset, eval_dataset = load_prepare_dpo_datasets(cfg)
-    total_num_steps = int(
-        math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
-    )
-
-    if cli_args.debug or cfg.debug:
-        LOG.info("check_dataset_labels...")
-
-        tokenizer = load_tokenizer(cfg)
-        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,
-            rl_mode=True,
-        )
-
-    return TrainDatasetMeta(
-        train_dataset=train_dataset,
-        eval_dataset=eval_dataset,
-        total_num_steps=total_num_steps,
-    )
-
-
-def check_accelerate_default_config():
-    if Path(config_args.default_yaml_config_file).exists():
-        LOG.warning(
-            f"accelerate config file found at {config_args.default_yaml_config_file}. This can lead to unexpected errors"
-        )
-
-
-def check_user_token():
-    # Skip check if HF_HUB_OFFLINE is set to True
-    if os.getenv("HF_HUB_OFFLINE") == "1":
-        LOG.info(
-            "Skipping HuggingFace token verification because HF_HUB_OFFLINE is set to True. Only local files will be used."
-        )
-        return True
-
-    # Verify if token is valid
-    api = HfApi()
-    try:
-        user_info = api.whoami()
-        return bool(user_info)
-    except LocalTokenNotFoundError:
-        LOG.warning(
-            "Error verifying HuggingFace token. Remember to log in using `huggingface-cli login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
-        )
-        return False

src/axolotl/cli/args.py

@@ -0,0 +1,43 @@
+"""Module for axolotl CLI command arguments."""
+
+from dataclasses import dataclass, field
+from typing import Optional
+
+
+@dataclass
+class PreprocessCliArgs:
+    """Dataclass with CLI arguments for `axolotl preprocess` command."""
+
+    debug: bool = field(default=False)
+    debug_text_only: bool = field(default=False)
+    debug_num_examples: int = field(default=1)
+    prompter: Optional[str] = field(default=None)
+    download: Optional[bool] = field(default=True)
+
+
+@dataclass
+class TrainerCliArgs:
+    """Dataclass with CLI arguments for `axolotl train` command."""
+
+    debug: bool = field(default=False)
+    debug_text_only: bool = field(default=False)
+    debug_num_examples: int = field(default=0)
+    merge_lora: bool = field(default=False)
+    prompter: Optional[str] = field(default=None)
+    shard: bool = field(default=False)
+
+
+@dataclass
+class EvaluateCliArgs:
+    """Dataclass with CLI arguments for `axolotl evaluate` command."""
+
+    debug: bool = field(default=False)
+    debug_text_only: bool = field(default=False)
+    debug_num_examples: int = field(default=0)
+
+
+@dataclass
+class InferenceCliArgs:
+    """Dataclass with CLI arguments for `axolotl inference` command."""
+
+    prompter: Optional[str] = field(default=None)

src/axolotl/cli/art.py

@@ -0,0 +1,23 @@
+"""Axolotl ASCII logo utils."""
+
+from axolotl.utils.distributed import is_main_process
+
+AXOLOTL_LOGO = """
+     #@@ #@@      @@# @@#
+    @@  @@          @@  @@           =@@#                               @@                 #@    =@@#.
+    @@    #@@@@@@@@@    @@           #@#@=                              @@                 #@     .=@@
+      #@@@@@@@@@@@@@@@@@            =@# @#     ##=     ##    =####=+    @@      =#####+  =#@@###.   @@
+    @@@@@@@@@@/  +@@/  +@@          #@  =@=     #@=   @@   =@#+  +#@#   @@    =@#+  +#@#   #@.      @@
+    @@@@@@@@@@  ##@@  ##@@         =@#   @#      =@# @#    @@      @@   @@    @@      #@   #@       @@
+     @@@@@@@@@@@@@@@@@@@@          #@=+++#@=      =@@#     @@      @@   @@    @@      #@   #@       @@
+                                  =@#=====@@     =@# @#    @@      @@   @@    @@      #@   #@       @@
+    @@@@@@@@@@@@@@@@  @@@@        #@      #@=   #@=  +@@   #@#    =@#   @@.   =@#    =@#   #@.      @@
+                                 =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
+    @@@@  @@@@@@@@@@@@@@@@
+"""
+
+
+def print_axolotl_text_art():
+    """Prints axolotl ASCII art."""
+    if is_main_process():
+        print(AXOLOTL_LOGO)

src/axolotl/cli/checks.py

@@ -0,0 +1,50 @@
+"""Various checks for Axolotl CLI."""
+
+import logging
+import os
+from pathlib import Path
+
+from accelerate.commands.config import config_args
+from huggingface_hub import HfApi
+from huggingface_hub.utils import LocalTokenNotFoundError
+
+from axolotl.logging_config import configure_logging
+
+configure_logging()
+LOG = logging.getLogger(__name__)
+
+
+def check_accelerate_default_config() -> None:
+    """Logs at warning level if no accelerate config file is found."""
+    if Path(config_args.default_yaml_config_file).exists():
+        LOG.warning(
+            f"accelerate config file found at {config_args.default_yaml_config_file}. This can lead to unexpected errors"
+        )
+
+
+def check_user_token() -> bool:
+    """Checks for HF user info. Check is skipped if HF_HUB_OFFLINE=1.
+
+    Returns:
+        Boolean indicating successful check (i.e., HF_HUB_OFFLINE=1 or HF user info is retrieved).
+
+    Raises:
+        LocalTokenNotFoundError: If HF user info can't be retrieved.
+    """
+    # Skip check if HF_HUB_OFFLINE is set to True
+    if os.getenv("HF_HUB_OFFLINE") == "1":
+        LOG.info(
+            "Skipping HuggingFace token verification because HF_HUB_OFFLINE is set to True. Only local files will be used."
+        )
+        return True
+
+    # Verify if token is valid
+    api = HfApi()
+    try:
+        user_info = api.whoami()
+        return bool(user_info)
+    except LocalTokenNotFoundError:
+        LOG.warning(
+            "Error verifying HuggingFace token. Remember to log in using `huggingface-cli login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
+        )
+        return False

src/axolotl/cli/cloud/__init__.py

@@ -1,56 +0,0 @@
-"""
-launch axolotl in supported cloud platforms
-"""
-from pathlib import Path
-from typing import Union
-
-import yaml
-
-from axolotl.cli import print_axolotl_text_art
-from axolotl.cli.cloud.modal_ import ModalCloud
-from axolotl.utils.dict import DictDefault
-
-
-def load_cloud_cfg(cloud_config: Union[Path, str]) -> DictDefault:
-    """Load and validate cloud configuration."""
-    # Load cloud configuration.
-    with open(cloud_config, encoding="utf-8") as file:
-        cloud_cfg: DictDefault = DictDefault(yaml.safe_load(file))
-    return cloud_cfg
-
-
-def do_cli_preprocess(
-    cloud_config: Union[Path, str],
-    config: Union[Path, str] = Path("examples/"),
-) -> None:
-    print_axolotl_text_art()
-    cloud_cfg = load_cloud_cfg(cloud_config)
-    cloud = ModalCloud(cloud_cfg)
-    with open(config, "r", encoding="utf-8") as file:
-        config_yaml = file.read()
-    cloud.preprocess(config_yaml)
-
-
-def do_cli_train(
-    cloud_config: Union[Path, str],
-    config: Union[Path, str] = Path("examples/"),
-    accelerate: bool = True,
-) -> None:
-    print_axolotl_text_art()
-    cloud_cfg = load_cloud_cfg(cloud_config)
-    cloud = ModalCloud(cloud_cfg)
-    with open(config, "r", encoding="utf-8") as file:
-        config_yaml = file.read()
-    cloud.train(config_yaml, accelerate=accelerate)
-
-
-def do_cli_lm_eval(
-    cloud_config: Union[Path, str],
-    config: Union[Path, str] = Path("examples/"),
-) -> None:
-    print_axolotl_text_art()
-    cloud_cfg = load_cloud_cfg(cloud_config)
-    cloud = ModalCloud(cloud_cfg)
-    with open(config, "r", encoding="utf-8") as file:
-        config_yaml = file.read()
-    cloud.lm_eval(config_yaml)

src/axolotl/cli/cloud/base.py

@@ -1,18 +0,0 @@
-"""
-base class for cloud platforms from cli
-"""
-from abc import ABC, abstractmethod
-
-
-class Cloud(ABC):
-    """
-    Abstract base class for cloud platforms.
-    """
-
-    @abstractmethod
-    def preprocess(self, config_yaml: str, *args, **kwargs) -> None:
-        pass
-
-    @abstractmethod
-    def train(self, config_yaml: str, accelerate: bool = True) -> str:
-        pass

src/axolotl/cli/cloud/modal_.py

@@ -1,272 +0,0 @@
-"""
-Modal Cloud support from CLI
-"""
-import copy
-import json
-import os
-import subprocess  # nosec B404
-from pathlib import Path
-from random import randint
-
-import modal
-
-from axolotl.cli.cloud.base import Cloud
-
-
-def run_cmd(cmd: str, run_folder: str, volumes=None):
-    """Run a command inside a folder, with Modal Volume reloading before and commit on success."""
-    # Ensure volumes contain latest files.
-    if volumes:
-        for _, vol in volumes.items():
-            vol.reload()
-
-    # modal workaround so it doesn't use the automounted axolotl
-    new_env = copy.deepcopy(os.environ)
-    if "PYTHONPATH" in new_env:
-        del new_env["PYTHONPATH"]
-
-    # Propagate errors from subprocess.
-    if exit_code := subprocess.call(  # nosec B603
-        cmd.split(), cwd=run_folder, env=new_env
-    ):
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
-
-    # Commit writes to volume.
-    if volumes:
-        for _, vol in volumes.items():
-            vol.commit()
-
-
-class ModalCloud(Cloud):
-    """
-    Modal Cloud implementation.
-    """
-
-    def __init__(self, config, app=None):
-        self.config = config
-        if not app:
-            app = modal.App()
-        self.app = app
-
-        self.volumes = {}
-        if config.volumes:
-            for volume_config in config.volumes:
-                _, mount, vol = self.create_volume(volume_config)
-                self.volumes[mount] = (vol, volume_config)
-
-    def get_env(self):
-        res = {
-            "HF_DATASETS_CACHE": "/workspace/data/huggingface-cache/datasets",
-            "HF_HUB_CACHE": "/workspace/data/huggingface-cache/hub",
-        }
-
-        for key in self.config.get("env", []):
-            if isinstance(key, str):
-                if val := os.environ.get(key, ""):
-                    res[key] = val
-            elif isinstance(key, dict):
-                (key_, val) = list(key.items())[0]
-                res[key_] = val
-        return res
-
-    def get_image(self):
-        docker_tag = "main-py3.11-cu124-2.5.1"
-        if self.config.docker_tag:
-            docker_tag = self.config.docker_tag
-        docker_image = f"axolotlai/axolotl:{docker_tag}"
-
-        # grab the sha256 hash from docker hub for this image+tag
-        # this ensures that we always get the latest image for this tag, even if it's already cached
-        try:
-            manifest = subprocess.check_output(  # nosec B602
-                f"docker manifest inspect {docker_image}",
-                shell=True,
-            ).decode("utf-8")
-            sha256_hash = json.loads(manifest)["manifests"][0]["digest"]
-        except subprocess.CalledProcessError:
-            sha256_hash = None
-
-        # create the image
-        if sha256_hash:
-            image = modal.Image.from_registry(f"axolotlai/axolotl@{sha256_hash}")
-        else:
-            image = modal.Image.from_registry(docker_image)
-
-        # branch
-        if self.config.branch:
-            image = image.dockerfile_commands(
-                [
-                    # Random id for cache busting of branch commits
-                    f"RUN echo '{str(randint(0, 1000000))}'",  # nosec B311
-                    f"RUN cd /workspace/axolotl && git fetch && git checkout {self.config.branch}",
-                    "RUN cd /workspace/ && git clone https://github.com/winglian/lm-evaluation-harness.git && cd lm-evaluation-harness && pip install -e .[math]",
-                ]
-            )
-
-        if env := self.get_env():
-            image = image.env(env)
-
-        image = image.pip_install("fastapi==0.110.0", "pydantic==2.6.3")
-
-        return image
-
-    def get_secrets(self):
-        res = []
-        if self.config.secrets:
-            for key in self.config.get("secrets", []):
-                # pylint: disable=duplicate-code
-                if isinstance(key, str):
-                    if val := os.environ.get(key, ""):
-                        res.append(modal.Secret.from_dict({key: val}))
-                elif isinstance(key, dict):
-                    (key_, val) = list(key.items())[0]
-                    res.append(modal.Secret.from_dict({key_: val}))
-        return res
-
-    def create_volume(self, volume_config):
-        name = volume_config.name
-        mount = volume_config.mount
-        return name, mount, modal.Volume.from_name(name, create_if_missing=True)
-
-    def get_ephemeral_disk_size(self):
-        return 1000 * 525  # 1 TiB
-
-    def get_preprocess_timeout(self):
-        if self.config.timeout_preprocess:
-            return int(self.config.timeout_preprocess)
-        return 60 * 60 * 3  # 3 hours
-
-    def get_preprocess_memory(self):
-        memory = 128  # default to 128GiB
-        if self.config.memory:
-            memory = int(self.config.memory)
-        if self.config.memory_preprocess:
-            memory = int(self.config.memory_preprocess)
-        return 1024 * memory
-
-    def get_preprocess_env(self):
-        return self.app.function(
-            image=self.get_image(),
-            volumes={k: v[0] for k, v in self.volumes.items()},
-            cpu=8.0,
-            ephemeral_disk=self.get_ephemeral_disk_size(),
-            memory=self.get_preprocess_memory(),
-            timeout=self.get_preprocess_timeout(),
-            secrets=self.get_secrets(),
-        )
-
-    def preprocess(self, config_yaml: str, *args, **kwargs):
-        modal_fn = self.get_preprocess_env()(_preprocess)
-        with modal.enable_output():
-            with self.app.run(detach=True):
-                modal_fn.remote(
-                    config_yaml,
-                    volumes={k: v[0] for k, v in self.volumes.items()},
-                    *args,
-                    **kwargs,
-                )
-
-    def get_train_timeout(self):
-        if self.config.timeout:
-            return int(self.config.timeout)
-        return 60 * 60 * 24  # 24 hours
-
-    def get_train_gpu(self):  # pylint: disable=too-many-return-statements
-        count = self.config.gpu_count or 1
-        family = self.config.gpu.lower() or "l40s"
-
-        if family == "l40s":
-            return modal.gpu.L40S(count=count)
-        if family == "a100":
-            return modal.gpu.A100(count=count, size="40GB")
-        if family == "a100-80gb":
-            return modal.gpu.A100(count=count, size="80GB")
-        if family in ["a10", "a10g"]:
-            return modal.gpu.A10G(count=count)
-        if family == "h100":
-            return modal.gpu.H100(count=count)
-        if family == "t4":
-            return modal.gpu.T4(count=count)
-        if family == "l4":
-            return modal.gpu.L4(count=count)
-        raise ValueError(f"Unsupported GPU family: {family}")
-
-    def get_train_memory(self):
-        memory = 128  # default to 128GiB
-        if self.config.memory:
-            memory = int(self.config.memory)
-        return 1024 * memory
-
-    def get_train_env(self):
-        return self.app.function(
-            image=self.get_image(),
-            volumes={k: v[0] for k, v in self.volumes.items()},
-            cpu=16.0,
-            gpu=self.get_train_gpu(),
-            memory=self.get_train_memory(),
-            timeout=self.get_train_timeout(),
-            secrets=self.get_secrets(),
-        )
-
-    def train(self, config_yaml: str, accelerate: bool = True):
-        modal_fn = self.get_train_env()(_train)
-        with modal.enable_output():
-            with self.app.run(detach=True):
-                modal_fn.remote(
-                    config_yaml,
-                    accelerate=accelerate,
-                    volumes={k: v[0] for k, v in self.volumes.items()},
-                )
-
-    def lm_eval(self, config_yaml: str):
-        modal_fn = self.get_train_env()(_lm_eval)
-        with modal.enable_output():
-            with self.app.run(detach=True):
-                modal_fn.remote(
-                    config_yaml,
-                    volumes={k: v[0] for k, v in self.volumes.items()},
-                )
-
-
-def _preprocess(config_yaml: str, volumes=None):
-    Path("/workspace/artifacts/axolotl").mkdir(parents=True, exist_ok=True)
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
-        f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
-    run_cmd(
-        "axolotl preprocess /workspace/artifacts/axolotl/config.yaml --dataset-processes=8",
-        run_folder,
-        volumes,
-    )
-
-
-def _train(config_yaml: str, accelerate: bool = True, volumes=None):
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
-        f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
-    if accelerate:
-        accelerate_args = "--accelerate"
-    else:
-        accelerate_args = "--no-accelerate"
-    run_cmd(
-        f"axolotl train {accelerate_args} /workspace/artifacts/axolotl/config.yaml",
-        run_folder,
-        volumes,
-    )
-
-
-def _lm_eval(config_yaml: str, volumes=None):
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
-        f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
-    run_cmd(
-        "axolotl lm-eval /workspace/artifacts/axolotl/config.yaml",
-        run_folder,
-        volumes,
-    )

src/axolotl/cli/config.py

@@ -0,0 +1,217 @@
+"""Configuration loading and processing."""
+
+import json
+import logging
+import os
+import tempfile
+from pathlib import Path
+from typing import Union
+from urllib.parse import urlparse
+
+import requests
+import torch
+import yaml
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.integrations.base import PluginManager
+from axolotl.utils.comet_ import setup_comet_env_vars
+from axolotl.utils.config import (
+    normalize_cfg_datasets,
+    normalize_config,
+    validate_config,
+)
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.mlflow_ import setup_mlflow_env_vars
+from axolotl.utils.trainer import prepare_opinionated_env, prepare_optim_env
+from axolotl.utils.wandb_ import setup_wandb_env_vars
+
+LOG = logging.getLogger(__name__)
+
+
+def check_remote_config(config: Union[str, Path]) -> Union[str, Path]:
+    """
+    First, determines if the passed config is a valid HTTPS URL. Then, attempts to query
+    for it and parse its content, first as JSON, then as YAML (YAML is preferred).
+    Finally, the parsed content is written to a local file and its path is returned.
+
+    Args:
+        config: HTTPS URL to a YAML or JSON file.
+
+    Returns:
+        Either the original `config` if it's not a valid HTTPS URL, or the path to the
+        downloaded remote config.
+
+    Raises:
+        ValueError: If the remote configuration is neither valid JSON or YAML.
+        RuntimeError: If some request-related exception occurs from the file download.
+        Exception: Catch-all for any other exception.
+    """
+    # Check if the config is a valid HTTPS URL to a .yml or .yaml file
+    if not (isinstance(config, str) and config.startswith("https://")):
+        return config  # Return the original value if it's not a valid URL
+
+    filename = os.path.basename(urlparse(config).path)
+    temp_dir = tempfile.mkdtemp()
+
+    try:
+        response = requests.get(config, timeout=30)
+        response.raise_for_status()  # Check for HTTP errors
+
+        content = response.content
+        try:
+            # Try parsing as JSON first to catch cases where JSON content is mistakenly
+            # considered YAML.
+            json.loads(content)
+
+            # Log a warning but do not raise an error; JSON is technically valid YAML.
+            # This can happen when you forget to point to a raw GitHub link.
+            LOG.warning(
+                f"Warning: The content of the file at {config} is JSON, which is technically valid YAML but might not be intended."
+            )
+        except json.JSONDecodeError:
+            # If it's not valid JSON, verify it's valid YAML
+            try:
+                yaml.safe_load(content)
+            except yaml.YAMLError as err:
+                raise ValueError(
+                    f"Failed to parse the content at {config} as YAML: {err}"
+                ) from err
+
+        # Write the content to a file if it's valid YAML (or JSON treated as YAML)
+        output_path = Path(temp_dir) / filename
+        with open(output_path, "wb") as file:
+            file.write(content)
+        LOG.info(
+            f"Using the following config obtained from {config}: \n\n{content.decode('utf-8')}\n"
+        )
+        return output_path
+
+    except requests.RequestException as err:
+        # This catches all requests-related exceptions including HTTPError
+        raise RuntimeError(f"Failed to download {config}: {err}") from err
+    except Exception as err:
+        # Catch-all for any other exceptions
+        raise err
+
+
+def choose_config(path: Path) -> str:
+    """
+    Helper method for choosing a `axolotl` config YAML file (considering only files
+    ending with `.yml` or `.yaml`). If more than one config file exists in the passed
+    `path`, the user is prompted to choose one.
+
+    Args:
+        path: Directory in which config file(s) are stored.
+
+    Returns:
+        Path to either (1) the sole YAML file, or (2) if more than one YAML files exist,
+        the user-selected YAML file.
+
+    Raises:
+        ValueError: If no YAML files are found in the given `path`.
+    """
+    yaml_files = list(path.glob("*.yml")) + list(path.glob("*.yaml"))
+
+    if not yaml_files:
+        raise ValueError(
+            "No YAML config files found in the specified directory. Are you using a .yml extension?"
+        )
+
+    if len(yaml_files) == 1:
+        print(f"Using default YAML file '{yaml_files[0]}'")
+        return str(yaml_files[0])
+
+    print("Choose a YAML file:")
+    for idx, file in enumerate(yaml_files):
+        print(f"{idx + 1}. {file}")
+
+    chosen_file = None
+    while chosen_file is None:
+        try:
+            choice = int(input("Enter the number of your choice: "))
+            if 1 <= choice <= len(yaml_files):
+                chosen_file = str(yaml_files[choice - 1])
+            else:
+                print("Invalid choice. Please choose a number from the list.")
+        except ValueError:
+            print("Invalid input. Please enter a number.")
+
+    return chosen_file
+
+
+def prepare_plugins(cfg: DictDefault):
+    """
+    Registers the plugins for the given configuration.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+    """
+    if cfg.get("plugins"):
+        plugin_manager = PluginManager.get_instance()
+        for plugin_name in cfg["plugins"]:
+            plugin_manager.register(plugin_name)
+
+
+def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefault:
+    """
+    Loads the `axolotl` configuration stored at `config`, validates it, and performs
+    various setup.
+
+    Args:
+        config: Path (local or remote) to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments to override config file values.
+
+    Returns:
+        `DictDefault` mapping configuration keys to values.
+    """
+    config = check_remote_config(config)
+    if Path(config).is_dir():
+        config = choose_config(Path(config))
+
+    # Load the config from the yaml file
+    with open(config, encoding="utf-8") as file:
+        cfg: DictDefault = DictDefault(yaml.safe_load(file))
+
+    # If there are any options passed in the cli, if it is something that seems valid
+    # from the yaml, then overwrite the value
+    cfg_keys = cfg.keys()
+    for k, _ in kwargs.items():
+        # if not strict, allow writing to cfg even if it's not in the yml already
+        if k in cfg_keys or not cfg.strict:
+            # handle booleans
+            if isinstance(cfg[k], bool):
+                cfg[k] = bool(kwargs[k])
+            else:
+                cfg[k] = kwargs[k]
+
+    cfg.axolotl_config_path = config
+
+    try:
+        device_props = torch.cuda.get_device_properties("cuda")
+        gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)
+    except:  # pylint: disable=bare-except # noqa: E722
+        gpu_version = None
+
+    prepare_plugins(cfg)
+
+    cfg = validate_config(
+        cfg,
+        capabilities={
+            "bf16": is_torch_bf16_gpu_available(),
+            "n_gpu": int(os.environ.get("WORLD_SIZE", 1)),
+            "compute_capability": gpu_version,
+        },
+        env_capabilities={
+            "torch_version": str(torch.__version__).split("+", maxsplit=1)[0]
+        },
+    )
+
+    prepare_optim_env(cfg)
+    prepare_opinionated_env(cfg)
+    normalize_config(cfg)
+    normalize_cfg_datasets(cfg)
+    setup_wandb_env_vars(cfg)
+    setup_mlflow_env_vars(cfg)
+    setup_comet_env_vars(cfg)
+
+    return cfg

src/axolotl/cli/evaluate.py

@@ -1,6 +1,5 @@
-"""
-CLI to run training on a model
-"""
+"""CLI to run evaluation on a model."""
+
 import logging
 from pathlib import Path
 from typing import Union
@@ -9,35 +8,48 @@ import fire
 from dotenv import load_dotenv
 from transformers.hf_argparser import HfArgumentParser
 
-from axolotl.cli import (
-    check_accelerate_default_config,
-    check_user_token,
-    load_cfg,
-    load_datasets,
-    load_rl_datasets,
-    print_axolotl_text_art,
-)
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.cli.art import print_axolotl_text_art
+from axolotl.cli.checks import check_accelerate_default_config, check_user_token
+from axolotl.cli.config import load_cfg
+from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.evaluate import evaluate
+from axolotl.utils.dict import DictDefault
 
-LOG = logging.getLogger("axolotl.cli.evaluate")
+LOG = logging.getLogger(__name__)
 
 
-def do_evaluate(cfg, cli_args) -> None:
+def do_evaluate(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
+    """
+    Evaluates a `transformers` model by first loading the dataset(s) specified in the
+    `axolotl` config, and then calling `axolotl.evaluate.evaluate`, which computes
+    evaluation metrics on the given dataset(s) and writes them to disk.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        cli_args: CLI arguments.
+    """
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
 
-    if cfg.rl:  # and cfg.rl != "orpo":
-        dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
+    if cfg.rl:
+        dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-    evaluate(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+    evaluate(cfg=cfg, dataset_meta=dataset_meta)
 
 
 def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
+    """
+    Parses `axolotl` config, CLI args, and calls `do_evaluate`.
+
+    Args:
+        config: Path to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments to override config file values.
+    """
     # pylint: disable=duplicate-code
     parsed_cfg = load_cfg(config, **kwargs)
     parser = HfArgumentParser(TrainerCliArgs)

src/axolotl/cli/inference.py

@@ -1,32 +1,267 @@
-"""
-CLI to run inference on a trained model
-"""
+"""CLI to run inference on a trained model."""
+
+import importlib
+import logging
+import sys
 from pathlib import Path
+from threading import Thread
 from typing import Union
 
 import fire
+import torch
 import transformers
 from dotenv import load_dotenv
+from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
 
-from axolotl.cli import (
-    do_inference,
-    do_inference_gradio,
-    load_cfg,
-    print_axolotl_text_art,
+from axolotl.cli.args import InferenceCliArgs
+from axolotl.cli.art import print_axolotl_text_art
+from axolotl.cli.config import load_cfg
+from axolotl.cli.utils import load_model_and_tokenizer
+from axolotl.utils.chat_templates import (
+    get_chat_template,
+    get_chat_template_from_config,
 )
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.utils.dict import DictDefault
+
+LOG = logging.getLogger(__name__)
 
 
-def do_cli(config: Union[Path, str] = Path("examples/"), gradio=False, **kwargs):
+def get_multi_line_input() -> str:
+    """
+    Gets multi-line input from terminal.
+
+    Returns:
+        Possibly multi-line, possibly empty stdin input as a string.
+    """
+    print("Give me an instruction (Ctrl + D to submit): ")
+
+    instruction = ""
+    for line in sys.stdin:
+        instruction += line  # pylint: disable=consider-using-join
+
+    return instruction
+
+
+def do_inference(
+    *,
+    cfg: DictDefault,
+    cli_args: InferenceCliArgs,
+):
+    """
+    Runs inference on the command line in a loop. User input is accepted, a chat template
+    is (optionally) applied, and the model specified in the `axolotl` config is used to
+    generate completions according to a default generation config.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        cli_args: Inference-specific CLI arguments.
+    """
+    model, tokenizer = load_model_and_tokenizer(cfg=cfg, inference=True)
+    prompter = cli_args.prompter
+
+    prompter_module = None
+    chat_template_str = None
+    if prompter:
+        prompter_module = getattr(
+            importlib.import_module("axolotl.prompters"), prompter
+        )
+    elif cfg.chat_template:
+        chat_template_str = get_chat_template(cfg.chat_template)
+    elif cfg.datasets[0].type == "chat_template":
+        chat_template_str = get_chat_template_from_config(
+            cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer
+        )
+
+    model = model.to(cfg.device, dtype=cfg.torch_dtype)
+
+    while True:
+        print("=" * 80)
+        # support for multiline inputs
+        instruction = get_multi_line_input()
+        if not instruction:
+            return
+
+        if prompter_module:
+            prompt: str = next(
+                prompter_module().build_prompt(instruction=instruction.strip("\n"))
+            )
+        else:
+            prompt = instruction.strip()
+
+        if chat_template_str:
+            batch = tokenizer.apply_chat_template(
+                [
+                    {
+                        "role": "user",
+                        "content": prompt,
+                    }
+                ],
+                return_tensors="pt",
+                add_special_tokens=True,
+                add_generation_prompt=True,
+                chat_template=chat_template_str,
+                tokenize=True,
+                return_dict=True,
+            )
+        else:
+            batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
+
+        print("=" * 40)
+        model.eval()
+        with torch.no_grad():
+            generation_config = GenerationConfig(
+                repetition_penalty=1.1,
+                max_new_tokens=1024,
+                temperature=0.9,
+                top_p=0.95,
+                top_k=40,
+                bos_token_id=tokenizer.bos_token_id,
+                eos_token_id=tokenizer.eos_token_id,
+                pad_token_id=tokenizer.pad_token_id,
+                do_sample=True,
+                use_cache=True,
+                return_dict_in_generate=True,
+                output_attentions=False,
+                output_hidden_states=False,
+                output_scores=False,
+            )
+            streamer = TextStreamer(tokenizer)
+            generated = model.generate(
+                inputs=batch["input_ids"].to(cfg.device),
+                generation_config=generation_config,
+                streamer=streamer,
+            )
+        print("=" * 40)
+        print(tokenizer.decode(generated["sequences"].cpu().tolist()[0]))
+
+
+def do_inference_gradio(
+    *,
+    cfg: DictDefault,
+    cli_args: InferenceCliArgs,
+):
+    """
+    Runs inference in a Gradio interface. User input is accepted, a chat template is
+    (optionally) applied, and the model specified in the `axolotl` config is used to
+    generate completions according to a default generation config.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        cli_args: Inference-specific CLI arguments.
+    """
+    import gradio as gr
+
+    model, tokenizer = load_model_and_tokenizer(cfg=cfg, inference=True)
+    prompter = cli_args.prompter
+
+    prompter_module = None
+    chat_template_str = None
+    if prompter:
+        prompter_module = getattr(
+            importlib.import_module("axolotl.prompters"), prompter
+        )
+    elif cfg.chat_template:
+        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
+
+    model = model.to(cfg.device, dtype=cfg.torch_dtype)
+
+    def generate(instruction):
+        if not instruction:
+            return
+        if prompter_module:
+            # pylint: disable=stop-iteration-return
+            prompt: str = next(
+                prompter_module().build_prompt(instruction=instruction.strip("\n"))
+            )
+        else:
+            prompt = instruction.strip()
+
+        if chat_template_str:
+            batch = tokenizer.apply_chat_template(
+                [
+                    {
+                        "role": "user",
+                        "content": prompt,
+                    }
+                ],
+                return_tensors="pt",
+                add_special_tokens=True,
+                add_generation_prompt=True,
+                chat_template=chat_template_str,
+                tokenize=True,
+                return_dict=True,
+            )
+        else:
+            batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
+
+        model.eval()
+        with torch.no_grad():
+            generation_config = GenerationConfig(
+                repetition_penalty=1.1,
+                max_new_tokens=cfg.get("gradio_max_new_tokens", 1024),
+                temperature=cfg.get("gradio_temperature", 0.9),
+                top_p=0.95,
+                top_k=40,
+                bos_token_id=tokenizer.bos_token_id,
+                eos_token_id=tokenizer.eos_token_id,
+                pad_token_id=tokenizer.pad_token_id,
+                do_sample=True,
+                use_cache=True,
+                return_dict_in_generate=True,
+                output_attentions=False,
+                output_hidden_states=False,
+                output_scores=False,
+            )
+            streamer = TextIteratorStreamer(tokenizer)
+            generation_kwargs = {
+                "inputs": batch["input_ids"].to(cfg.device),
+                "attention_mask": batch["attention_mask"].to(cfg.device),
+                "generation_config": generation_config,
+                "streamer": streamer,
+            }
+
+            thread = Thread(target=model.generate, kwargs=generation_kwargs)
+            thread.start()
+
+            all_text = ""
+
+            for new_text in streamer:
+                all_text += new_text
+                yield all_text
+
+    demo = gr.Interface(
+        fn=generate,
+        inputs="textbox",
+        outputs="text",
+        title=cfg.get("gradio_title", "Axolotl Gradio Interface"),
+    )
+
+    demo.queue().launch(
+        show_api=False,
+        share=cfg.get("gradio_share", True),
+        server_name=cfg.get("gradio_server_name", "127.0.0.1"),
+        server_port=cfg.get("gradio_server_port", None),
+    )
+
+
+def do_cli(
+    config: Union[Path, str] = Path("examples/"), gradio: bool = False, **kwargs
+) -> None:
+    """
+    Parses axolotl config, CLI args, and calls `do_inference` or `do_inference_gradio`.
+
+    Args:
+        config: Path to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments to override config file values.
+    """
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     parsed_cfg = load_cfg(config, inference=True, **kwargs)
     parsed_cfg.sample_packing = False
-    parser = transformers.HfArgumentParser((TrainerCliArgs))
+    parser = transformers.HfArgumentParser(InferenceCliArgs)
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
-    parsed_cli_args.inference = True
 
     if gradio:
         do_inference_gradio(cfg=parsed_cfg, cli_args=parsed_cli_args)

src/axolotl/cli/main.py

@@ -1,19 +1,20 @@
-"""CLI definition for various axolotl commands."""
+"""Click CLI definitions for various axolotl commands."""
 # pylint: disable=redefined-outer-name
+
 import subprocess  # nosec B404
 from typing import Optional
 
 import click
 
 import axolotl
+from axolotl.cli.args import EvaluateCliArgs, PreprocessCliArgs, TrainerCliArgs
 from axolotl.cli.utils import (
     add_options_from_config,
     add_options_from_dataclass,
     build_command,
     fetch_from_github,
+    filter_none_kwargs,
 )
-from axolotl.common.cli import EvaluateCliArgs, PreprocessCliArgs, TrainerCliArgs
-from axolotl.integrations.lm_eval.cli import lm_eval
 from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.config.models.input.v0_4_1 import AxolotlInputConfig
 
@@ -26,59 +27,58 @@ def cli():
 
 @cli.command()
 @click.argument("config", type=click.Path(exists=True, path_type=str))
-@click.option("--cloud", default=None, type=click.Path(exists=True, path_type=str))
 @add_options_from_dataclass(PreprocessCliArgs)
 @add_options_from_config(AxolotlInputConfig)
-def preprocess(config: str, cloud: Optional[str] = None, **kwargs):
-    """Preprocess datasets before training."""
-    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+@filter_none_kwargs
+def preprocess(config: str, **kwargs) -> None:
+    """
+    Preprocess datasets before training.
 
-    if cloud:
-        from axolotl.cli.cloud import do_cli_preprocess
+    Args:
+        config: Path to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
+            config options.
+    """
+    from axolotl.cli.preprocess import do_cli
 
-        do_cli_preprocess(cloud_config=cloud, config=config)
+    do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@click.option(
+    "--accelerate/--no-accelerate",
+    default=True,
+    help="Use accelerate launch for multi-GPU training",
+)
+@add_options_from_dataclass(TrainerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+@filter_none_kwargs
+def train(config: str, accelerate: bool, **kwargs) -> None:
+    """
+    Train or fine-tune a model.
+
+    Args:
+        config: Path to `axolotl` config YAML file.
+        accelerate: Whether to use `accelerate` launcher.
+        kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
+            config options.
+    """
+    # Enable expandable segments for cuda allocation to improve VRAM usage
+    set_pytorch_cuda_alloc_conf()
+
+    if accelerate:
+        base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.train"]
+        if config:
+            base_cmd.append(config)
+        cmd = build_command(base_cmd, kwargs)
+        subprocess.run(cmd, check=True)  # nosec B603
     else:
-        from axolotl.cli.preprocess import do_cli
+        from axolotl.cli.train import do_cli
 
         do_cli(config=config, **kwargs)
 
 
-@cli.command()
-@click.argument("config", type=click.Path(exists=True, path_type=str))
-@click.option(
-    "--accelerate/--no-accelerate",
-    default=True,
-    help="Use accelerate launch for multi-GPU training",
-)
-@click.option("--cloud", default=None, type=click.Path(exists=True, path_type=str))
-@add_options_from_dataclass(TrainerCliArgs)
-@add_options_from_config(AxolotlInputConfig)
-def train(config: str, accelerate: bool, cloud: Optional[str], **kwargs):
-    """Train or fine-tune a model."""
-    kwargs = {k: v for k, v in kwargs.items() if v is not None}
-
-    # Enable expandable segments for cuda allocation to improve VRAM usage
-    set_pytorch_cuda_alloc_conf()
-    from axolotl.cli.cloud import do_cli_train
-
-    if accelerate:
-        if cloud:
-            do_cli_train(cloud_config=cloud, config=config, accelerate=True)
-        else:
-            base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.train"]
-            if config:
-                base_cmd.append(config)
-            cmd = build_command(base_cmd, kwargs)
-            subprocess.run(cmd, check=True)  # nosec B603
-    else:
-        if cloud:
-            do_cli_train(cloud_config=cloud, config=config, accelerate=False)
-        else:
-            from axolotl.cli.train import do_cli
-
-            do_cli(config=config, **kwargs)
-
-
 @cli.command()
 @click.argument("config", type=click.Path(exists=True, path_type=str))
 @click.option(
@@ -88,10 +88,17 @@ def train(config: str, accelerate: bool, cloud: Optional[str], **kwargs):
 )
 @add_options_from_dataclass(EvaluateCliArgs)
 @add_options_from_config(AxolotlInputConfig)
-def evaluate(config: str, accelerate: bool, **kwargs):
-    """Evaluate a model."""
-    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+@filter_none_kwargs
+def evaluate(config: str, accelerate: bool, **kwargs) -> None:
+    """
+    Evaluate a model.
 
+    Args:
+        config: Path to `axolotl` config YAML file.
+        accelerate: Whether to use `accelerate` launcher.
+        kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
+            config options.
+    """
     if accelerate:
         base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.evaluate"]
         if config:
@@ -111,81 +118,33 @@ def evaluate(config: str, accelerate: bool, **kwargs):
     default=False,
     help="Use accelerate launch for multi-GPU inference",
 )
-@click.option(
-    "--lora-model-dir",
-    type=click.Path(exists=True, path_type=str),
-    help="Directory containing LoRA model",
-)
-@click.option(
-    "--base-model",
-    type=click.Path(exists=True, path_type=str),
-    help="Path to base model for non-LoRA models",
-)
 @click.option("--gradio", is_flag=True, help="Launch Gradio interface")
-@click.option("--load-in-8bit", is_flag=True, help="Load model in 8-bit mode")
 @add_options_from_dataclass(TrainerCliArgs)
 @add_options_from_config(AxolotlInputConfig)
-def inference(
-    config: str,
-    accelerate: bool,
-    lora_model_dir: Optional[str] = None,
-    base_model: Optional[str] = None,
-    **kwargs,
-):
-    """Run inference with a trained model."""
-    kwargs = {k: v for k, v in kwargs.items() if v is not None}
-    del kwargs["inference"]  # interferes with inference.do_cli
-
-    if lora_model_dir:
-        kwargs["lora_model_dir"] = lora_model_dir
-    if base_model:
-        kwargs["base_model"] = base_model
+@filter_none_kwargs
+def inference(config: str, accelerate: bool, gradio: bool, **kwargs) -> None:
+    """
+    Run inference with a trained model.
 
+    Args:
+        config: Path to `axolotl` config YAML file.
+        accelerate: Whether to use `accelerate` launcher.
+        gradio: Whether to use Gradio browser interface or command line for inference.
+        kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
+            config options.
+    """
     if accelerate:
         base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.inference"]
         if config:
             base_cmd.append(config)
+        if gradio:
+            base_cmd.append("--gradio")
         cmd = build_command(base_cmd, kwargs)
         subprocess.run(cmd, check=True)  # nosec B603
     else:
         from axolotl.cli.inference import do_cli
 
-        do_cli(config=config, **kwargs)
-
-
-@cli.command()
-@click.argument("config", type=click.Path(exists=True, path_type=str))
-@click.option(
-    "--accelerate/--no-accelerate",
-    default=False,
-    help="Use accelerate launch for multi-GPU operations",
-)
-@click.option(
-    "--model-dir",
-    type=click.Path(exists=True, path_type=str),
-    help="Directory containing model weights to shard",
-)
-@click.option(
-    "--save-dir",
-    type=click.Path(path_type=str),
-    help="Directory to save sharded weights",
-)
-@add_options_from_dataclass(TrainerCliArgs)
-@add_options_from_config(AxolotlInputConfig)
-def shard(config: str, accelerate: bool, **kwargs):
-    """Shard model weights."""
-    kwargs = {k: v for k, v in kwargs.items() if v is not None}
-
-    if accelerate:
-        base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.shard"]
-        if config:
-            base_cmd.append(config)
-        cmd = build_command(base_cmd, kwargs)
-        subprocess.run(cmd, check=True)  # nosec B603
-    else:
-        from axolotl.cli.shard import do_cli
-
-        do_cli(config=config, **kwargs)
+        do_cli(config=config, gradio=gradio, **kwargs)
 
 
 @cli.command()
@@ -195,20 +154,19 @@ def shard(config: str, accelerate: bool, **kwargs):
     default=True,
     help="Use accelerate launch for weight merging",
 )
-@click.option(
-    "--model-dir",
-    type=click.Path(exists=True, path_type=str),
-    help="Directory containing sharded weights",
-)
-@click.option(
-    "--save-path", type=click.Path(path_type=str), help="Path to save merged weights"
-)
 @add_options_from_dataclass(TrainerCliArgs)
 @add_options_from_config(AxolotlInputConfig)
-def merge_sharded_fsdp_weights(config: str, accelerate: bool, **kwargs):
-    """Merge sharded FSDP model weights."""
-    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+@filter_none_kwargs
+def merge_sharded_fsdp_weights(config: str, accelerate: bool, **kwargs) -> None:
+    """
+    Merge sharded FSDP model weights.
 
+    Args:
+        config: Path to `axolotl` config YAML file.
+        accelerate: Whether to use `accelerate` launcher.
+        kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
+            config options.
+    """
     if accelerate:
         base_cmd = [
             "accelerate",
@@ -228,28 +186,19 @@ def merge_sharded_fsdp_weights(config: str, accelerate: bool, **kwargs):
 
 @cli.command()
 @click.argument("config", type=click.Path(exists=True, path_type=str))
-@click.option(
-    "--lora-model-dir",
-    type=click.Path(exists=True, path_type=str),
-    help="Directory containing the LoRA model to merge",
-)
-@click.option(
-    "--output-dir",
-    type=click.Path(path_type=str),
-    help="Directory to save the merged model",
-)
-def merge_lora(
-    config: str,
-    lora_model_dir: Optional[str] = None,
-    output_dir: Optional[str] = None,
-):
-    """Merge a trained LoRA into a base model"""
-    kwargs = {}
-    if lora_model_dir:
-        kwargs["lora_model_dir"] = lora_model_dir
-    if output_dir:
-        kwargs["output_dir"] = output_dir
+@add_options_from_dataclass(TrainerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+@filter_none_kwargs
+def merge_lora(config: str, **kwargs) -> None:
+    """
+    Merge trained LoRA adapters into a base model.
 
+    Args:
+        config: Path to `axolotl` config YAML file.
+        accelerate: Whether to use `accelerate` launcher.
+        kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
+            config options.
+    """
     from axolotl.cli.merge_lora import do_cli
 
     do_cli(config=config, **kwargs)
@@ -258,20 +207,21 @@ def merge_lora(
 @cli.command()
 @click.argument("directory", type=click.Choice(["examples", "deepspeed_configs"]))
 @click.option("--dest", help="Destination directory")
-def fetch(directory: str, dest: Optional[str]):
+def fetch(directory: str, dest: Optional[str]) -> None:
     """
     Fetch example configs or other resources.
 
     Available directories:
     - examples: Example configuration files
     - deepspeed_configs: DeepSpeed configuration files
+
+    Args:
+        directory: One of `examples`, `deepspeed_configs`.
+        dest: Optional destination directory.
     """
     fetch_from_github(f"{directory}/", dest)
 
 
-cli.add_command(lm_eval)
-
-
 def main():
     cli()
 

src/axolotl/cli/merge_lora.py

@@ -1,6 +1,6 @@
-"""
-CLI to run merge a trained LoRA into a base model
-"""
+"""CLI to merge a trained LoRA into a base model."""
+
+import logging
 from pathlib import Path
 from typing import Union
 
@@ -8,14 +8,58 @@ import fire
 import transformers
 from dotenv import load_dotenv
 
-from axolotl.cli import do_merge_lora, load_cfg, print_axolotl_text_art
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.cli.art import print_axolotl_text_art
+from axolotl.cli.config import load_cfg
+from axolotl.cli.utils import load_model_and_tokenizer
+from axolotl.utils.dict import DictDefault
+
+LOG = logging.getLogger(__name__)
 
 
-def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
-    # pylint: disable=duplicate-code
+def do_merge_lora(*, cfg: DictDefault) -> None:
+    """
+    Calls `transformers`' `merge_and_unload` on the model given in the `axolotl` config
+    along with the LoRA adapters to combine them into a single base model.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+    """
     print_axolotl_text_art()
-    parser = transformers.HfArgumentParser((TrainerCliArgs))
+
+    model, tokenizer = load_model_and_tokenizer(cfg=cfg)
+    safe_serialization = cfg.save_safetensors is True
+
+    LOG.info("Running merge of LoRA with base model...")
+    model = model.merge_and_unload(progressbar=True)
+    model.to(dtype=cfg.torch_dtype)
+    model.generation_config.do_sample = True
+
+    if cfg.local_rank == 0:
+        LOG.info(f"Saving merged model to: {str(Path(cfg.output_dir) / 'merged')}...")
+        model.save_pretrained(
+            str(Path(cfg.output_dir) / "merged"),
+            safe_serialization=safe_serialization,
+            progressbar=True,
+        )
+        tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
+
+
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
+    """
+    Parses `axolotl` config, CLI args, and calls `do_merge_lora`. Note that various
+    config values will be overwritten to allow the LoRA merge logic to work as expected
+    (`load_in_8bit=False`, `load_in4bit=False`, `flash_attention=False`, etc.).
+
+    Args:
+        config: Path to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments to override config file values.
+
+    Raises:
+        ValueError: If target directory for LoRA merged model does not exist.
+    """
+    # pylint: disable=duplicate-code
+    parser = transformers.HfArgumentParser(TrainerCliArgs)
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
@@ -46,7 +90,7 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     parsed_cfg.fsdp = None
     parsed_cfg.fsdp_config = None
 
-    do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    do_merge_lora(cfg=parsed_cfg)
 
 
 if __name__ == "__main__":

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -1,6 +1,5 @@
-"""
-This module provides a CLI to merge sharded FSDP model checkpoints into a single combined checkpoint
-"""
+"""CLI to merge sharded FSDP model checkpoints into a single combined checkpoint."""
+
 import json
 import logging
 import os
@@ -25,16 +24,15 @@ from huggingface_hub import split_torch_state_dict_into_shards
 from safetensors.torch import save_file as safe_save_file
 from torch.distributed.checkpoint.format_utils import _EmptyStateDictLoadPlanner
 
-from axolotl.cli import load_cfg, print_axolotl_text_art
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.cli.art import print_axolotl_text_art
+from axolotl.cli.config import load_cfg
 
-LOG = logging.getLogger("axolotl.cli.merge_sharded_fsdp_weights")
+LOG = logging.getLogger(__name__)
 
 
 class BFloat16CastPlanner(_EmptyStateDictLoadPlanner):
-    """
-    A custom planner to cast tensors to bfloat16 on the fly during loading.
-    """
+    """A custom planner to cast tensors to bfloat16 on the fly during loading."""
 
     def commit_tensor(self, read_item, tensor):  # pylint: disable=unused-argument
         tensor.copy_(tensor.to(torch.bfloat16))
@@ -45,11 +43,19 @@ def _distributed_checkpoint_to_merged_weights(
     save_path: str,
     safe_serialization: bool = False,
     max_shard_size: str = "5GB",
-):
+) -> Path:
     """
-    Passthrough to `torch.distributed.checkpoint.format_utils.dcp_to_torch_save`
+    Passthrough to `torch.distributed.checkpoint.format_utils.dcp_to_torch_save`. Will
+    save under `save_path` as either `model.safetensors` or `pytorch_model.bin`.
 
-    Will save under `save_path` as either `model.safetensors` or `pytorch_model.bin`.
+    Args:
+        checkpoint_dir: Directory where distributed checkpoint is saved.
+        save_path: Path to save model to.
+        safe_serialization: Whether to save in safetensors format.
+        max_shard_size: Max size of model shards to save.
+
+    Returns:
+        Path where model is saved.
     """
 
     state_dict: Dict = {}
@@ -79,6 +85,7 @@ def _distributed_checkpoint_to_merged_weights(
     state_dict_split = split_torch_state_dict_into_shards(
         state_dict, filename_pattern=filename_pattern, max_shard_size=max_shard_size
     )
+
     # Save index if sharded
     index = None
     if state_dict_split.is_sharded:
@@ -135,6 +142,9 @@ def merge_fsdp_weights(
             Whether to save the merged weights with safetensors (recommended).
         remove_checkpoint_dir (`bool`, *optional*, defaults to `False`):
             Whether to remove the checkpoint directory after merging.
+
+    Raises:
+        ValueError: If torch version < 2.3.0, or if `checkpoint_dir` does not exist.
     """
     checkpoint_dir_ = Path(checkpoint_dir)
     from accelerate.state import PartialState
@@ -178,18 +188,21 @@ def merge_fsdp_weights(
 
 
 def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
+    """
+    Parses `axolotl` config, CLI args, and calls `merge_fsdp_weights`.
+
+    Args:
+        config: Path to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments to override config file values.
+    """
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
-    parser = transformers.HfArgumentParser((TrainerCliArgs))
+    parser = transformers.HfArgumentParser(TrainerCliArgs)
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
     parsed_cli_args.merge_lora = True
-
-    parsed_cfg = load_cfg(
-        config,
-        **kwargs,
-    )
+    parsed_cfg = load_cfg(config, **kwargs)
 
     fsdp_dir = Path(parsed_cfg.output_dir) / "pytorch_model_fsdp_0"
     merge_fsdp_weights(

src/axolotl/cli/preprocess.py

@@ -1,6 +1,5 @@
-"""
-CLI to run training on a model
-"""
+"""CLI to run preprocessing of a dataset."""
+
 import logging
 import warnings
 from pathlib import Path
@@ -13,34 +12,31 @@ from colorama import Fore
 from dotenv import load_dotenv
 from transformers import AutoModelForCausalLM
 
-from axolotl.cli import (
-    check_accelerate_default_config,
-    check_user_token,
-    load_cfg,
-    load_datasets,
-    load_rl_datasets,
-    print_axolotl_text_art,
-)
-from axolotl.common.cli import PreprocessCliArgs
+from axolotl.cli.args import PreprocessCliArgs
+from axolotl.cli.art import print_axolotl_text_art
+from axolotl.cli.checks import check_accelerate_default_config, check_user_token
+from axolotl.cli.config import load_cfg
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
+from axolotl.common.datasets import load_datasets, load_preference_datasets
+from axolotl.utils.dict import DictDefault
 from axolotl.utils.trainer import disable_datasets_caching
 
-LOG = logging.getLogger("axolotl.cli.preprocess")
+LOG = logging.getLogger(__name__)
 
 
-def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
-    # pylint: disable=duplicate-code
+def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
+    """
+    Preprocesses dataset specified in axolotl config.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        cli_args: Preprocessing-specific CLI arguments.
+    """
     print_axolotl_text_art()
-    parsed_cfg = load_cfg(config, **kwargs)
-    parsed_cfg.is_preprocess = True
     check_accelerate_default_config()
     check_user_token()
-    parser = transformers.HfArgumentParser((PreprocessCliArgs))
-    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
-        return_remaining_strings=True
-    )
 
-    if not parsed_cfg.dataset_prepared_path:
+    if not cfg.dataset_prepared_path:
         msg = (
             Fore.RED
             + "preprocess CLI called without dataset_prepared_path set, "
@@ -48,16 +44,16 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
             + Fore.RESET
         )
         LOG.warning(msg)
-        parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
+        cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
     with disable_datasets_caching():
-        if parsed_cfg.rl:  # and parsed_cfg.rl != "orpo":
-            load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+        if cfg.rl:
+            load_preference_datasets(cfg=cfg, cli_args=cli_args)
         else:
-            load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+            load_datasets(cfg=cfg, cli_args=cli_args)
 
-    if parsed_cli_args.download:
-        model_name = parsed_cfg.base_model
+    if cli_args.download:
+        model_name = cfg.base_model
         with warnings.catch_warnings():
             # there are a bunch of useless UserWarnings about
             # "copying from a non-meta parameter in the checkpoint to a meta parameter in the current model"
@@ -74,11 +70,30 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
 
     LOG.info(
         Fore.GREEN
-        + f"Success! Preprocessed data path: `dataset_prepared_path: {parsed_cfg.dataset_prepared_path}`"
+        + f"Success! Preprocessed data path: `dataset_prepared_path: {cfg.dataset_prepared_path}`"
         + Fore.RESET
     )
 
 
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
+    """
+    Parses `axolotl` config, CLI args, and calls `do_preprocess`.
+
+    Args:
+        config: Path to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments to override config file values.
+    """
+    # pylint: disable=duplicate-code
+    parsed_cfg = load_cfg(config, **kwargs)
+    parsed_cfg.is_preprocess = True
+    parser = transformers.HfArgumentParser(PreprocessCliArgs)
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+
+    do_preprocess(parsed_cfg, parsed_cli_args)
+
+
 if __name__ == "__main__":
     load_dotenv()
     fire.Fire(do_cli)

src/axolotl/cli/shard.py

@@ -1,45 +0,0 @@
-"""
-CLI to shard a trained model into 10GiB chunks
-"""
-import logging
-from pathlib import Path
-from typing import Union
-
-import fire
-import transformers
-from dotenv import load_dotenv
-
-from axolotl.cli import load_cfg, print_axolotl_text_art
-from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
-from axolotl.utils.dict import DictDefault
-
-LOG = logging.getLogger("axolotl.scripts")
-
-
-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_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
-    # pylint: disable=duplicate-code
-    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
-    )
-    parsed_cli_args.shard = True
-
-    shard(cfg=parsed_cfg, cli_args=parsed_cli_args)
-
-
-if __name__ == "__main__":
-    load_dotenv()
-    fire.Fire(do_cli)

src/axolotl/cli/train.py

@@ -1,6 +1,5 @@
-"""
-CLI to run training on a model
-"""
+"""CLI to run training on a model."""
+
 import logging
 from pathlib import Path
 from typing import Union
@@ -9,42 +8,38 @@ import fire
 from dotenv import load_dotenv
 from transformers.hf_argparser import HfArgumentParser
 
-from axolotl.cli import (
-    check_accelerate_default_config,
-    check_user_token,
-    load_cfg,
-    load_datasets,
-    load_rl_datasets,
-    print_axolotl_text_art,
-)
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.cli.art import print_axolotl_text_art
+from axolotl.cli.checks import check_accelerate_default_config, check_user_token
+from axolotl.cli.config import load_cfg
+from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.integrations.base import PluginManager
 from axolotl.train import train
+from axolotl.utils.dict import DictDefault
 
-LOG = logging.getLogger("axolotl.cli.train")
+LOG = logging.getLogger(__name__)
 
 
-def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
-    # pylint: disable=duplicate-code
-    parsed_cfg = load_cfg(config, **kwargs)
-    parser = HfArgumentParser((TrainerCliArgs))
-    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
-        return_remaining_strings=True
-    )
-    return do_train(parsed_cfg, parsed_cli_args)
+def do_train(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
+    """
+    Trains a `transformers` model by first loading the dataset(s) specified in the
+    `axolotl` config, and then calling `axolotl.train.train`. Also runs the plugin
+    manager's `post_train_unload` once training completes.
 
-
-def do_train(cfg, cli_args) -> None:
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        cli_args: Training-specific CLI arguments.
+    """
     print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
 
-    if cfg.rl:  # and cfg.rl != "orpo":
-        dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
+    if cfg.rl:
+        dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-    model, tokenizer = train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+    model, tokenizer = train(cfg=cfg, dataset_meta=dataset_meta)
     plugin_manager = PluginManager.get_instance()
 
     del model
@@ -53,6 +48,24 @@ def do_train(cfg, cli_args) -> None:
     plugin_manager.post_train_unload(cfg)
 
 
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
+    """
+    Parses `axolotl` config, CLI args, and calls `do_train`.
+
+    Args:
+        config: Path to `axolotl` config YAML file.
+        kwargs: Additional keyword arguments to override config file values.
+    """
+    # pylint: disable=duplicate-code
+    parsed_cfg = load_cfg(config, **kwargs)
+    parser = HfArgumentParser(TrainerCliArgs)
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+
+    do_train(parsed_cfg, parsed_cli_args)
+
+
 if __name__ == "__main__":
     load_dotenv()
     fire.Fire(do_cli)

src/axolotl/cli/utils.py

@@ -1,32 +1,84 @@
-"""Utility methods for axoltl CLI."""
+"""Utility methods for axolotl CLI."""
+
 import concurrent.futures
 import dataclasses
 import hashlib
 import json
 import logging
+import typing
+from functools import wraps
 from pathlib import Path
 from types import NoneType
-from typing import Any, Dict, List, Optional, Tuple, Type, Union, get_args, get_origin
+from typing import Any, Callable, Type, Union, get_args, get_origin
 
 import click
 import requests
 from pydantic import BaseModel
+from transformers import PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast
 
-LOG = logging.getLogger("axolotl.cli.utils")
+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(__name__)
 
 
-def add_options_from_dataclass(config_class: Type[Any]):
-    """Create Click options from the fields of a dataclass."""
+def strip_optional_type(field_type: type | typing._SpecialForm | None):
+    """
+    Extracts the non-`None` type from an `Optional` / `Union` type.
 
-    def decorator(function):
+    Args:
+        field_type: Type of field for Axolotl CLI command.
+
+    Returns:
+        If the input type is `Union[T, None]` or `Optional[T]`, returns `T`. Otherwise
+            returns the input type unchanged.
+    """
+    if get_origin(field_type) is Union and type(None) in get_args(field_type):
+        field_type = next(
+            t for t in get_args(field_type) if not isinstance(t, NoneType)
+        )
+
+    return field_type
+
+
+def filter_none_kwargs(func: Callable) -> Callable:
+    """
+    Wraps function to remove `None`-valued `kwargs`.
+
+    Args:
+        func: Function to wrap.
+
+    Returns:
+        Wrapped function.
+    """
+
+    @wraps(func)
+    def wrapper(*args, **kwargs) -> Callable:
+        """Filters out `None`-valued `kwargs`."""
+        filtered_kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+        return func(*args, **filtered_kwargs)
+
+    return wrapper
+
+
+def add_options_from_dataclass(config_class: Type[Any]) -> Callable:
+    """
+    Create Click options from the fields of a dataclass.
+
+    Args:
+        config_class: Dataclass with fields to parse from the CLI.
+
+    Returns:
+        Function decorator for Axolotl CLI command.
+    """
+
+    def decorator(function: Callable) -> Callable:
         # Process dataclass fields in reverse order for correct option ordering
         for field in reversed(dataclasses.fields(config_class)):
-            field_type = field.type
-
-            if get_origin(field_type) is Union and type(None) in get_args(field_type):
-                field_type = next(
-                    t for t in get_args(field_type) if not isinstance(t, NoneType)
-                )
+            field_type = strip_optional_type(field.type)
 
             if field_type == bool:
                 field_name = field.name.replace("_", "-")
@@ -44,18 +96,29 @@ def add_options_from_dataclass(config_class: Type[Any]):
                     default=field.default,
                     help=field.metadata.get("description"),
                 )(function)
+
         return function
 
     return decorator
 
 
-def add_options_from_config(config_class: Type[BaseModel]):
-    """Create Click options from the fields of a Pydantic model."""
+def add_options_from_config(config_class: Type[BaseModel]) -> Callable:
+    """
+    Create Click options from the fields of a Pydantic model.
 
-    def decorator(function):
+    Args:
+        config_class: PyDantic model with fields to parse from the CLI
+
+    Returns:
+        Function decorator for Axolotl CLI command.
+    """
+
+    def decorator(function: Callable) -> Callable:
         # Process model fields in reverse order for correct option ordering
         for name, field in reversed(config_class.model_fields.items()):
-            if field.annotation == bool:
+            field_type = strip_optional_type(field.annotation)
+
+            if field_type == bool:
                 field_name = name.replace("_", "-")
                 option_name = f"--{field_name}/--no-{field_name}"
                 function = click.option(
@@ -66,13 +129,23 @@ def add_options_from_config(config_class: Type[BaseModel]):
                 function = click.option(
                     option_name, default=None, help=field.description
                 )(function)
+
         return function
 
     return decorator
 
 
-def build_command(base_cmd: List[str], options: Dict[str, Any]) -> List[str]:
-    """Build command list from base command and options."""
+def build_command(base_cmd: list[str], options: dict[str, Any]) -> list[str]:
+    """
+    Build command list from base command and options.
+
+    Args:
+        base_cmd: Command without options.
+        options: Options to parse and append to base command.
+
+    Returns:
+        List of strings giving shell command.
+    """
     cmd = base_cmd.copy()
 
     for key, value in options.items():
@@ -92,18 +165,18 @@ def build_command(base_cmd: List[str], options: Dict[str, Any]) -> List[str]:
 
 def download_file(
     file_info: tuple, raw_base_url: str, dest_path: Path, dir_prefix: str
-) -> Tuple[str, str]:
+) -> tuple[str, str]:
     """
     Download a single file and return its processing status.
 
     Args:
-        file_info: Tuple of (file_path, remote_sha)
-        raw_base_url: Base URL for raw GitHub content
-        dest_path: Local destination directory
-        dir_prefix: Directory prefix to filter files
+        file_info: Tuple of (file_path, remote_sha).
+        raw_base_url: Base URL for raw GitHub content.
+        dest_path: Local destination directory.
+        dir_prefix: Directory prefix to filter files.
 
     Returns:
-        Tuple of (file_path, status) where status is 'new', 'updated', or 'unchanged'
+        Tuple of (file_path, status) where status is 'new', 'updated', or 'unchanged'.
     """
     file_path, remote_sha = file_info
     raw_url = f"{raw_base_url}/{file_path}"
@@ -145,16 +218,17 @@ def download_file(
 
 
 def fetch_from_github(
-    dir_prefix: str, dest_dir: Optional[str] = None, max_workers: int = 5
+    dir_prefix: str, dest_dir: str | None = None, max_workers: int = 5
 ) -> None:
     """
     Sync files from a specific directory in the GitHub repository.
     Only downloads files that don't exist locally or have changed.
 
     Args:
-        dir_prefix: Directory prefix to filter files (e.g., 'examples/', 'deepspeed_configs/')
-        dest_dir: Local destination directory
-        max_workers: Maximum number of concurrent downloads
+        dir_prefix: Directory prefix to filter files (e.g., 'examples/',
+            'deepspeed_configs/').
+        dest_dir: Local destination directory.
+        max_workers: Maximum number of concurrent downloads.
     """
     api_url = "https://api.github.com/repos/axolotl-ai-cloud/axolotl/git/trees/main?recursive=1"
     raw_base_url = "https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main"
@@ -179,7 +253,7 @@ def fetch_from_github(
     dest_path = Path(dest_dir) if dest_dir else default_dest
 
     # Keep track of processed files for summary
-    files_processed: Dict[str, List[str]] = {
+    files_processed: dict[str, list[str]] = {
         "new": [],
         "updated": [],
         "unchanged": [],
@@ -216,3 +290,28 @@ def fetch_from_github(
     LOG.info(f"Unchanged files: {len(files_processed['unchanged'])}")
     if files_processed["error"]:
         LOG.info(f"Failed files: {len(files_processed['error'])}")
+
+
+def load_model_and_tokenizer(
+    *,
+    cfg: DictDefault,
+    inference: bool = False,
+) -> tuple[PreTrainedModel, PreTrainedTokenizer | PreTrainedTokenizerFast | Any]:
+    """
+    Helper function for loading a model and tokenizer specified in the given `axolotl`
+    config.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        inference: Boolean denoting inference mode.
+
+    Returns:
+        `transformers` model and tokenizer.
+    """
+    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
+    tokenizer = load_tokenizer(cfg)
+
+    LOG.info("loading model...")
+    model, _ = load_model(cfg, tokenizer, inference=inference)
+
+    return model, tokenizer

src/axolotl/common/cli.py

@@ -1,69 +0,0 @@
-"""
-shared module for cli specific things
-"""
-
-import logging
-from dataclasses import dataclass, field
-from typing import Optional
-
-import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
-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 PreprocessCliArgs:
-    """
-    dataclass representing arguments for preprocessing only
-    """
-
-    debug: bool = field(default=False)
-    debug_text_only: bool = field(default=False)
-    debug_num_examples: int = field(default=1)
-    prompter: Optional[str] = field(default=None)
-    download: Optional[bool] = field(default=True)
-
-
-@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=0)
-    inference: bool = field(default=False)
-    merge_lora: bool = field(default=False)
-    prompter: Optional[str] = field(default=None)
-    shard: bool = field(default=False)
-
-
-@dataclass
-class EvaluateCliArgs:
-    """
-    dataclass representing the various evaluation arguments
-    """
-
-    debug: bool = field(default=False)
-    debug_text_only: bool = field(default=False)
-    debug_num_examples: int = field(default=0)
-
-
-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...")
-    inference = getattr(cli_args, "inference", False)
-    model, _ = load_model(cfg, tokenizer, inference=inference)
-
-    return model, tokenizer

src/axolotl/common/datasets.py

@@ -0,0 +1,140 @@
+"""Dataset loading utilities."""
+
+import logging
+import math
+import random
+from dataclasses import dataclass
+from typing import Optional, Union
+
+from datasets import Dataset
+
+import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
+from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
+from axolotl.utils.data import prepare_dataset
+from axolotl.utils.data.rl import load_prepare_preference_datasets
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.models import load_processor, load_tokenizer
+from axolotl.utils.tokenization import check_dataset_labels
+
+LOG = logging.getLogger(__name__)
+
+
+@dataclass
+class TrainDatasetMeta:
+    """Dataclass with fields for training and validation datasets and metadata."""
+
+    train_dataset: Dataset
+    eval_dataset: Optional[Dataset] = None
+    total_num_steps: Optional[int] = None
+
+
+def sample_dataset(dataset: Dataset, num_samples: int) -> Dataset:
+    """
+    Randomly sample `num_samples` samples from `dataset`.
+
+    Args:
+        dataset: Dataset.
+        num_samples: Number of samples to return.
+
+    Returns:
+        Random sample (with replacement) of examples in `dataset`.
+    """
+    return dataset.select(
+        [random.randrange(0, len(dataset) - 1) for _ in range(num_samples)]  # nosec
+    )
+
+
+def load_datasets(
+    *,
+    cfg: DictDefault,
+    cli_args: Union[PreprocessCliArgs, TrainerCliArgs],
+) -> TrainDatasetMeta:
+    """
+    Loads one or more training or evaluation datasets, calling
+    `axolotl.utils.data.prepare_dataset`. Optionally, logs out debug information.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        cli_args: Command-specific CLI arguments.
+
+    Returns:
+        Dataclass with fields for training and evaluation datasets and the computed
+        `total_num_steps`.
+    """
+    tokenizer = load_tokenizer(cfg)
+    processor = load_processor(cfg, tokenizer=tokenizer) if cfg.processor_type else None
+
+    train_dataset, eval_dataset, total_num_steps, prompters = prepare_dataset(
+        cfg,
+        tokenizer,
+        processor=processor,
+    )
+
+    if (
+        cli_args.debug
+        or cfg.debug
+        or cli_args.debug_text_only
+        or int(cli_args.debug_num_examples) > 0
+    ):
+        LOG.info("check_dataset_labels...")
+
+        train_samples = sample_dataset(train_dataset, cli_args.debug_num_examples)
+        check_dataset_labels(
+            train_samples,
+            tokenizer,
+            num_examples=cli_args.debug_num_examples,
+            text_only=cli_args.debug_text_only,
+        )
+
+        LOG.info("printing prompters...")
+        for prompter in prompters:
+            LOG.info(prompter)
+
+    return TrainDatasetMeta(
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        total_num_steps=total_num_steps,
+    )
+
+
+def load_preference_datasets(
+    *,
+    cfg: DictDefault,
+    cli_args: Union[PreprocessCliArgs, TrainerCliArgs],
+) -> TrainDatasetMeta:
+    """
+    Loads one or more training or evaluation datasets for RL training using paired
+    preference data, calling `axolotl.utils.data.rl.load_prepare_preference_datasets`.
+    Optionally, logs out debug information.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        cli_args: Command-specific CLI arguments.
+
+    Returns:
+        Dataclass with fields for training and evaluation datasets and the computed
+        `total_num_steps`.
+    """
+    train_dataset, eval_dataset = load_prepare_preference_datasets(cfg)
+    total_num_steps = int(
+        math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
+    )
+
+    if cli_args.debug or cfg.debug:
+        LOG.info("check_dataset_labels...")
+
+        tokenizer = load_tokenizer(cfg)
+        train_samples = sample_dataset(train_dataset, cli_args.debug_num_examples)
+        check_dataset_labels(
+            train_samples,
+            tokenizer,
+            num_examples=cli_args.debug_num_examples,
+            text_only=cli_args.debug_text_only,
+            rl_mode=True,
+        )
+
+    return TrainDatasetMeta(
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        total_num_steps=total_num_steps,
+    )

src/axolotl/core/trainer_builder.py

@@ -22,7 +22,6 @@ from typing import Any, Dict, List, Literal, Optional, Type, Union
 import torch
 import transformers
 from datasets import Dataset
-from packaging import version
 from peft.optimizers import create_loraplus_optimizer
 from torch import nn
 from torch.optim.lr_scheduler import OneCycleLR
@@ -244,6 +243,10 @@ class AxolotlTrainingMixins:
         default=None,
         metadata={"help": "Scale the learning rate for the embedding layers."},
     )
+    lr_groups: Optional[list[dict]] = field(
+        default=None,
+        metadata={"help": "Specify learning rate groups for with different LRs."},
+    )
     embedding_lr: Optional[float] = field(
         default=None,
         metadata={"help": "absolute learning rate for the embedding layers."},
@@ -462,11 +465,95 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
             )
         return super()._wrap_model(model, training=training, dataloader=dataloader)
 
+    def create_optimizer_grouped_parameters(self, opt_model, optimizer_kwargs):
+        decay_parameters = self.get_decay_parameter_names(opt_model)
+        params = {
+            "to_weight_decay": {},  # LayerNorm and bias
+            "embeddings": {},  # lm_head, embed_tokens,
+            "no_weight_decay": {},
+        }
+        lr_groups_lookup = {}
+        lr_groups_learning_rates = {}
+        if self.args.lr_groups:
+            for lr_group in self.args.lr_groups:
+                group_name = lr_group["name"]
+                group_modules = lr_group["modules"]
+                for module in group_modules:
+                    lr_groups_lookup[module] = group_name
+                lr_groups_learning_rates[group_name] = lr_group["lr"]
+                params[f"to_weight_decay_{group_name}"] = {}
+
+        for name, param in opt_model.named_parameters():
+            if not param.requires_grad:
+                continue
+            if name.endswith("modules_to_save.default.weight") or any(
+                embed_name in name for embed_name in ["embed_tokens", "lm_head"]
+            ):
+                params["embeddings"][name] = param
+            elif name in decay_parameters:
+                lr_group_modules = [
+                    group_modules
+                    for group_modules in lr_groups_lookup
+                    if group_modules in name
+                ]
+                if lr_groups_lookup and any(lr_group_modules):
+                    lr_group_module = lr_group_modules[0]
+                    group_name = lr_groups_lookup[lr_group_module]
+                    params[f"to_weight_decay_{group_name}"][name] = param
+                else:
+                    params["to_weight_decay"][name] = param
+            else:
+                params["no_weight_decay"][name] = param
+        optimizer_grouped_parameters = []
+        if params["to_weight_decay"]:
+            optimizer_grouped_parameters.append(
+                {
+                    "params": list(params["to_weight_decay"].values()),
+                    "weight_decay": self.args.weight_decay,
+                    "lr": optimizer_kwargs["lr"],
+                }
+            )
+        if params["embeddings"]:
+            lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
+            if self.args.embedding_lr_scale:
+                lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
+            elif self.args.embedding_lr:
+                lr = self.args.embedding_lr  # pylint: disable=invalid-name
+            optimizer_grouped_parameters.append(
+                {
+                    "params": list(params["embeddings"].values()),
+                    "weight_decay": 0.0,
+                    "lr": lr,
+                }
+            )
+        if params["no_weight_decay"]:
+            optimizer_grouped_parameters.append(
+                {
+                    "params": list(params["no_weight_decay"].values()),
+                    "weight_decay": 0.0,
+                    "lr": optimizer_kwargs["lr"],
+                }
+            )
+        for group_name, group_lr in lr_groups_learning_rates.items():
+            if params[f"to_weight_decay_{group_name}"]:
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(
+                            params[f"to_weight_decay_{group_name}"].values()
+                        ),
+                        "weight_decay": self.args.weight_decay,
+                        "lr": group_lr,
+                    }
+                )
+
+        return optimizer_grouped_parameters
+
     def create_optimizer(self):
         if (
             self.args.loraplus_lr_ratio is None
             and self.args.embedding_lr_scale is None
             and self.args.embedding_lr is None
+            and self.args.lr_groups is None
             and self.args.alternate_optimizer
             not in [
                 "optimi_adamw",
@@ -480,59 +567,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
 
         opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
         if self.optimizer is None:  # pylint: disable=access-member-before-definition
-            decay_parameters = self.get_decay_parameter_names(opt_model)
-            params = {
-                "to_weight_decay": {},  # LayerNorm and bias
-                "embeddings": {},  # lm_head, embed_tokens,
-                "no_weight_decay": {},
-            }
-
             optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
                 self.args,
                 opt_model,
             )
-
-            for name, param in opt_model.named_parameters():
-                if not param.requires_grad:
-                    continue
-                if name.endswith("modules_to_save.default.weight") or any(
-                    embed_name in name for embed_name in ["embed_tokens", "lm_head"]
-                ):
-                    params["embeddings"][name] = param
-                elif name in decay_parameters:
-                    params["to_weight_decay"][name] = param
-                else:
-                    params["no_weight_decay"][name] = param
-            optimizer_grouped_parameters = []
-            if params["to_weight_decay"]:
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(params["to_weight_decay"].values()),
-                        "weight_decay": self.args.weight_decay,
-                        "lr": optimizer_kwargs["lr"],
-                    }
-                )
-            if params["embeddings"]:
-                lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
-                if self.args.embedding_lr_scale:
-                    lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
-                elif self.args.embedding_lr:
-                    lr = self.args.embedding_lr  # pylint: disable=invalid-name
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(params["embeddings"].values()),
-                        "weight_decay": 0.0,
-                        "lr": lr,
-                    }
-                )
-            if params["no_weight_decay"]:
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(params["no_weight_decay"].values()),
-                        "weight_decay": 0.0,
-                        "lr": optimizer_kwargs["lr"],
-                    }
-                )
+            optimizer_grouped_parameters = self.create_optimizer_grouped_parameters(
+                opt_model, optimizer_kwargs
+            )
 
             if self.args.loraplus_lr_ratio is not None:
                 loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
@@ -549,6 +590,7 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
             elif (
                 self.args.embedding_lr_scale is not None
                 or self.args.embedding_lr is not None
+                or self.args.lr_groups is not None
             ):
                 self.optimizer = (  # pylint: disable=attribute-defined-outside-init
                     optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
@@ -608,8 +650,14 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
                     self.state.train_batch_size or self.args.per_device_train_batch_size
                 )
                 batch_max_len = train_batch_size * self.args.max_seq_length
+
+            if self.args.curriculum_sampling:
+                sampler = SequentialSampler(self.train_dataset)
+            else:
+                sampler = RandomSampler(self.train_dataset)
+
             return MultipackBatchSampler(
-                RandomSampler(self.train_dataset),
+                sampler,
                 lengths=get_dataset_lengths(self.train_dataset),
                 packing_efficiency_estimate=self.args.sample_packing_efficiency,
                 batch_max_len=batch_max_len,
@@ -978,12 +1026,7 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
             logs[key] = torch.tensor(metrics).mean().item()
         del self._stored_metrics[train_eval]
 
-        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
-            try:
-                return super().log(logs, start_time)
-            except TypeError:
-                return super().log(logs)  # transformers<=4.46
-        return super().log(logs)  # transformers<=4.46
+        return super().log(logs, start_time)
 
     def store_metrics(
         self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train"
@@ -1079,6 +1122,7 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         super().__init__(*args, **kwargs)
         self.dataset_tags = dataset_tags
         self.optimizer = None
+        self.model_accepts_loss_kwargs = False
 
     def create_optimizer(self):
         if self.args.loraplus_lr_ratio is None:
@@ -1167,22 +1211,6 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         torch.cuda.empty_cache()
         return loss
 
-    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
-        # TODO remove once trl supports the updated to the Trainer.log method
-        # logs either has 'loss' or 'eval_loss'
-        train_eval = "train" if "loss" in logs else "eval"
-        # Add averaged stored metrics to logs
-        for key, metrics in self._stored_metrics[train_eval].items():
-            logs[key] = torch.tensor(metrics).mean().item()
-        del self._stored_metrics[train_eval]
-
-        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
-            return super(DPOTrainer, self).log(  # pylint: disable=bad-super-call
-                logs, start_time
-            )
-        # transformers<=4.46
-        return super(DPOTrainer, self).log(logs)  # pylint: disable=bad-super-call
-
 
 class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
     """
@@ -1191,22 +1219,6 @@ class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
 
     tag_names = ["axolotl", "orpo"]
 
-    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
-        # TODO remove once trl supports the updated to the Trainer.log method
-        # logs either has 'loss' or 'eval_loss'
-        train_eval = "train" if "loss" in logs else "eval"
-        # Add averaged stored metrics to logs
-        for key, metrics in self._stored_metrics[train_eval].items():
-            logs[key] = torch.tensor(metrics).mean().item()
-        del self._stored_metrics[train_eval]
-
-        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
-            return super(ORPOTrainer, self).log(  # pylint: disable=bad-super-call
-                logs, start_time
-            )
-        # transformers<=4.46
-        return super(ORPOTrainer, self).log(logs)  # pylint: disable=bad-super-call
-
 
 class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
     """
@@ -1215,49 +1227,6 @@ class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
 
     tag_names = ["axolotl", "kto"]
 
-    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
-        # TODO remove once trl supports the updated to the Trainer.log method
-        # logs either has 'loss' or 'eval_loss'
-        train_eval = "train" if "loss" in logs else "eval"
-        # train metrics should have no prefix, eval should have 'eval_'
-        prefix = "eval_" if train_eval == "eval" else ""
-        # accumulate average metrics from sums and lengths
-        for split in ["chosen", "rejected"]:
-            if f"count/{split}" in self._stored_metrics[train_eval]:
-                count_sum = (
-                    torch.Tensor(self._stored_metrics[train_eval][f"count/{split}"])
-                    .sum()
-                    .item()
-                )
-                for metric in ["rewards", "logps", "logits"]:
-                    logs[f"{prefix}{metric}/{split}"] = (
-                        torch.Tensor(
-                            self._stored_metrics[train_eval][f"{metric}/{split}_sum"]
-                        )
-                        .sum()
-                        .item()
-                        / count_sum
-                    )
-                    # delete obsolete metric
-                    del self._stored_metrics[train_eval][f"{metric}/{split}_sum"]
-                del self._stored_metrics[train_eval][f"count/{split}"]
-        # calculate reward margin
-        if f"{prefix}rewards/chosen" in logs and f"{prefix}rewards/rejected" in logs:
-            logs[f"{prefix}rewards/margins"] = (
-                logs[f"{prefix}rewards/chosen"] - logs[f"{prefix}rewards/rejected"]
-            )
-        # Add averaged stored metrics to logs
-        for key, metrics in self._stored_metrics[train_eval].items():
-            logs[f"{prefix}{key}"] = torch.Tensor(metrics).mean().item()
-        del self._stored_metrics[train_eval]
-
-        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
-            return super(KTOTrainer, self).log(  # pylint: disable=bad-super-call
-                logs, start_time
-            )
-        # transformers<=4.46
-        return super(KTOTrainer, self).log(logs)  # pylint: disable=bad-super-call
-
 
 class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
     """
@@ -1266,22 +1235,6 @@ class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
 
     tag_names = ["axolotl", "cpo"]
 
-    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
-        # TODO remove once trl supports the updated to the Trainer.log method
-        # logs either has 'loss' or 'eval_loss'
-        train_eval = "train" if "loss" in logs else "eval"
-        # Add averaged stored metrics to logs
-        for key, metrics in self._stored_metrics[train_eval].items():
-            logs[key] = torch.tensor(metrics).mean().item()
-        del self._stored_metrics[train_eval]
-
-        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
-            return super(CPOTrainer, self).log(  # pylint: disable=bad-super-call
-                logs, start_time
-            )
-        # transformers<=4.46
-        return super(CPOTrainer, self).log(logs)  # pylint: disable=bad-super-call
-
 
 class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
     """
@@ -1290,15 +1243,6 @@ class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
 
     tag_names = ["axolotl", "reward"]
 
-    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
-        # TODO remove once trl supports the updated to the Trainer.log method
-        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
-            return super(RewardTrainer, self).log(  # pylint: disable=bad-super-call
-                logs, start_time
-            )
-        # transformers<=4.46
-        return super(RewardTrainer, self).log(logs)  # pylint: disable=bad-super-call
-
 
 class TrainerBuilderBase(abc.ABC):
     """
@@ -1764,6 +1708,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         ] = self.cfg.loraplus_lr_embedding
         training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
         training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
+        training_arguments_kwargs["lr_groups"] = self.cfg.lr_groups
 
         if self.cfg.lr_scheduler in ["one_cycle", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"
@@ -1977,6 +1922,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
     ):
         if training_args.pretraining:
+            if self.cfg.pretraining_sample_concatenation is False:
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
+            if self.cfg.micro_batch_size > 1:
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
             return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/evaluate.py

@@ -9,7 +9,6 @@ from typing import Dict, Optional
 import torch
 from accelerate.logging import get_logger
 
-from axolotl.common.cli import TrainerCliArgs
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
 from axolotl.utils import set_pytorch_cuda_alloc_conf
@@ -62,16 +61,13 @@ def evaluate_dataset(
     return metrics
 
 
-def evaluate(
-    *, cfg: DictDefault, cli_args: TrainerCliArgs, dataset_meta: TrainDatasetMeta
-) -> Dict[str, float]:
+def evaluate(*, cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> Dict[str, float]:
     """
     Evaluate a model on training and validation datasets
 
     Args:
-        cfg: Configuration dictionary
-        cli_args: Command line arguments
-        dataset_meta: Dataset metadata containing training and evaluation datasets
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        dataset_meta: Dataset metadata containing training and evaluation datasets.
 
     Returns:
         Tuple containing:
@@ -102,9 +98,7 @@ def evaluate(
 
     # Load model
     LOG.debug("loading model for evaluation...")
-    model, _ = load_model(
-        cfg, tokenizer, processor=processor, inference=cli_args.inference
-    )
+    model, _ = load_model(cfg, tokenizer, processor=processor)
 
     # Set up trainer
     trainer = setup_trainer(

src/axolotl/integrations/base.py

@@ -48,9 +48,9 @@ class BasePlugin:
         Initializes the BasePlugin.
         """
 
-    def register(self, cfg):  # pylint: disable=unused-argument
+    def register(self):  # pylint: disable=unused-argument
         """
-        Registers the plugin with the given configuration.
+        Registers the plugin
 
         Parameters:
         cfg (dict): The configuration for the plugin.
@@ -274,6 +274,7 @@ class PluginManager:
         try:
             plugin = load_plugin(plugin_name)
             self.plugins[plugin_name] = plugin
+            plugin.register()
         except ImportError:
             logging.error(f"Failed to load plugin: {plugin_name}")
 

src/axolotl/integrations/liger/__init__.py

@@ -22,13 +22,6 @@ import inspect
 import logging
 import sys
 
-from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
-from liger_kernel.transformers.functional import liger_cross_entropy
-from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
-from liger_kernel.transformers.rms_norm import LigerRMSNorm
-from liger_kernel.transformers.rope import liger_rotary_pos_emb
-from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
-
 from axolotl.integrations.base import BasePlugin
 
 from ...utils.distributed import zero_only
@@ -46,6 +39,13 @@ class LigerPlugin(BasePlugin):
         return "axolotl.integrations.liger.LigerArgs"
 
     def pre_model_load(self, cfg):
+        from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
+        from liger_kernel.transformers.functional import liger_cross_entropy
+        from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
+        from liger_kernel.transformers.rms_norm import LigerRMSNorm
+        from liger_kernel.transformers.rope import liger_rotary_pos_emb
+        from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
+
         if cfg.model_config_type in MODEL_TYPE_TO_APPLY_LIGER_FN:
             apply_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[cfg.model_config_type]
             liger_fn_sig = inspect.signature(apply_liger_fn)

src/axolotl/integrations/lm_eval/__init__.py

@@ -2,9 +2,9 @@
 Module for the Plugin for LM Eval Harness
 """
 import subprocess  # nosec
+from datetime import datetime
 
 from axolotl.integrations.base import BasePlugin
-from axolotl.integrations.lm_eval.cli import build_lm_eval_command
 
 from .args import LMEvalArgs  # pylint: disable=unused-import. # noqa: F401
 
@@ -18,19 +18,25 @@ class LMEvalPlugin(BasePlugin):
         return "axolotl.integrations.lm_eval.LMEvalArgs"
 
     def post_train_unload(self, cfg):
-        if cfg.lm_eval_post_train:
-            # pylint: disable=duplicate-code
-            for lm_eval_args in build_lm_eval_command(
-                cfg.lm_eval_tasks,
-                bfloat16=cfg.bfloat16 or cfg.bf16,
-                flash_attention=cfg.flash_attention,
-                output_dir=cfg.output_dir,
-                batch_size=cfg.lm_eval_batch_size,
-                wandb_project=cfg.wandb_project,
-                wandb_entity=cfg.wandb_entity,
-                model=cfg.lm_eval_model or cfg.hub_model_id,
-            ):
-                subprocess.run(  # nosec
-                    lm_eval_args,
-                    check=True,
-                )
+        tasks = ",".join(cfg.lm_eval_tasks)
+        fa2 = ",attn_implementation=flash_attention_2" if cfg.flash_attention else ""
+        dtype = ",dtype=bfloat16" if cfg.bf16 else ",dtype=float16"
+        output_path = cfg.output_dir
+        output_path += "" if cfg.output_dir.endswith("/") else "/"
+        output_path += "lm_eval_results/" + datetime.now().strftime("%Y%m%d_%H%M%S")
+        subprocess.run(  # nosec
+            [
+                "lm_eval",
+                "--model",
+                "hf",
+                "--model_args",
+                f"pretrained={cfg.output_dir}{fa2}{dtype}",
+                "--tasks",
+                tasks,
+                "--batch_size",
+                str(cfg.lm_eval_batch_size),
+                "--output_path",
+                output_path,
+            ],
+            check=True,
+        )

src/axolotl/integrations/lm_eval/args.py

@@ -13,5 +13,3 @@ class LMEvalArgs(BaseModel):
 
     lm_eval_tasks: List[str] = []
     lm_eval_batch_size: Optional[int] = 8
-    lm_eval_post_train: Optional[bool] = True
-    lm_eval_model: Optional[str] = None

src/axolotl/integrations/lm_eval/cli.py

@@ -1,113 +0,0 @@
-"""
-axolotl CLI for running lm_eval tasks
-"""
-import subprocess  # nosec
-from collections import defaultdict
-from datetime import datetime
-from typing import Optional
-
-import click
-import yaml
-
-from axolotl.utils.dict import DictDefault
-
-
-def build_lm_eval_command(
-    tasks: list[str],
-    bfloat16=True,
-    flash_attention=False,
-    output_dir="./",
-    batch_size=8,
-    wandb_project=None,
-    wandb_entity=None,
-    model=None,
-    revision=None,
-    apply_chat_template=None,
-    fewshot_as_multiturn=None,
-):
-    tasks_by_num_fewshot: dict[str, list] = defaultdict(list)
-    for task in tasks:
-        num_fewshot = "-1"
-        task_parts = task.split(":")
-        task_name = task_parts[0]
-        if len(task_parts) == 2:
-            task_name, num_fewshot = task_parts
-        tasks_by_num_fewshot[str(num_fewshot)].append(task_name)
-
-    for num_fewshot, tasks_list in tasks_by_num_fewshot.items():
-        tasks_str = ",".join(tasks_list)
-        num_fewshot_val = num_fewshot if num_fewshot != "-1" else None
-        pretrained = "pretrained="
-        pretrained += model if model else output_dir
-        fa2 = ",attn_implementation=flash_attention_2" if flash_attention else ""
-        dtype = ",dtype=bfloat16" if bfloat16 else ",dtype=float16"
-        revision = f",revision={revision}" if revision else ""
-        output_path = output_dir
-        output_path += "" if output_dir.endswith("/") else "/"
-        output_path += "lm_eval_results/" + datetime.now().strftime("%Y%m%d_%H%M%S")
-        lm_eval_args = [
-            "lm_eval",
-            "--model",
-            "hf",
-            "--model_args",
-            f"{pretrained}{fa2}{dtype}{revision}",
-            "--tasks",
-            tasks_str,
-            "--batch_size",
-            str(batch_size),
-            "--output_path",
-            output_path,
-        ]
-        wandb_args = []
-        if wandb_project:
-            wandb_args.append(f"project={wandb_project}")
-        if wandb_entity:
-            wandb_args.append(f"entity={wandb_entity}")
-        if wandb_args:
-            lm_eval_args.append("--wandb_args")
-            lm_eval_args.append(",".join(wandb_args))
-        if apply_chat_template:
-            lm_eval_args.append("--apply_chat_template")
-        if num_fewshot_val:
-            lm_eval_args.append("--num_fewshot")
-            lm_eval_args.append(str(num_fewshot_val))
-            if apply_chat_template and fewshot_as_multiturn:
-                lm_eval_args.append("--fewshot_as_multiturn")
-
-        yield lm_eval_args
-
-
-@click.command()
-@click.argument("config", type=click.Path(exists=True, path_type=str))
-@click.option("--cloud", default=None, type=click.Path(exists=True, path_type=str))
-def lm_eval(config: str, cloud: Optional[str] = None):
-    """
-    use lm eval to evaluate a trained language model
-    """
-
-    if cloud:
-        from axolotl.cli.cloud import do_cli_lm_eval
-
-        do_cli_lm_eval(cloud_config=cloud, config=config)
-    else:
-        with open(config, encoding="utf-8") as file:
-            cfg: DictDefault = DictDefault(yaml.safe_load(file))
-
-        # pylint: disable=duplicate-code
-        for lm_eval_args in build_lm_eval_command(
-            cfg.lm_eval_tasks,
-            bfloat16=cfg.bfloat16 or cfg.bf16,
-            flash_attention=cfg.flash_attention,
-            output_dir=cfg.output_dir,
-            batch_size=cfg.lm_eval_batch_size,
-            wandb_project=cfg.wandb_project,
-            wandb_entity=cfg.wandb_entity,
-            model=cfg.lm_eval_model or cfg.hub_model_id,
-            revision=cfg.revision,
-            apply_chat_template=cfg.apply_chat_template,
-            fewshot_as_multiturn=cfg.fewshot_as_multiturn,
-        ):
-            subprocess.run(  # nosec
-                lm_eval_args,
-                check=True,
-            )

src/axolotl/integrations/rrt/__init__.py

@@ -0,0 +1,25 @@
+"""
+Axolotl Plugin for Relaxed Recursive Transformers
+"""
+
+import logging
+
+from axolotl.integrations.base import BasePlugin
+from axolotl.integrations.rrt.modeling import register_rrt_model
+
+LOG = logging.getLogger(__name__)
+
+
+class RelaxedRecursiveTransformerPlugin(BasePlugin):
+    """
+    Plugin for Relaxed Recursive Transformers integration with Axolotl
+    """
+
+    def get_input_args(self):
+        return "axolotl.integrations.rrt.args.RelaxedRecursiveTransformerArgs"
+
+    def register(self):
+        LOG.info(
+            "Registering Relaxed Recursive Transformers modeling with transformers"
+        )
+        register_rrt_model()

src/axolotl/integrations/rrt/args.py

@@ -0,0 +1,11 @@
+"""
+Axolotl config args for Relaxed Recursive Transformers plugin
+"""
+
+from pydantic import BaseModel
+
+
+class RelaxedRecursiveTransformerArgs(BaseModel):
+    """
+    Arguments pertaining to the Relaxed Recursive Transformer model.
+    """

src/axolotl/integrations/rrt/cli/convert.py

@@ -0,0 +1,370 @@
+"""
+cli script for converting a pretrained model to a relaxed recursive transformer model
+"""
+import json
+import logging
+import math
+import os
+import re
+from pathlib import Path
+from typing import Tuple
+
+import safetensors
+import torch
+from huggingface_hub import snapshot_download, split_torch_state_dict_into_shards
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers import AutoConfig, AutoTokenizer
+from transformers.utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME
+
+from axolotl.integrations.rrt.modeling.modeling_rrt_llama import (
+    RelaxedRecursiveLlamaConfig,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def extract_layer_number(key):
+    """Extract layer number from parameter key."""
+    match = re.search(r"layers\.(\d+)\.", key)
+    return int(match.group(1)) if match else None
+
+
+def iter_parameter_weights(model_path, device="mps"):
+    """
+    iterator over parameter weights in the model shards
+
+    :param model_path: Path to model shards
+    :param device: Computing device
+    :return: generator yielding (parameter key, parameter weight, layer index) tuples
+    """
+    shards = list(model_path.glob("model*.safetensors"))
+    if not shards:
+        raise ValueError(f"No model shards found in {model_path}")
+
+    for shard in tqdm(shards, desc="Processing shards"):
+        with safetensors.safe_open(shard, framework="pt", device=device) as f:
+            for key in f.keys():
+                layer_idx = extract_layer_number(key)
+                weight = f.get_tensor(key)
+                yield key, weight, layer_idx
+
+
+def iter_recursive_parameter_weights(
+    model_path, modules_to_recurse: list[str], device="mps", recurse_layers=12
+):
+    # setup placeholder state_dict for recursive weights, need to keep in float32 precision
+    # to avoid precision loss when averaging weights across layers
+    rrt_avg_model_state_dict: dict[str, list[torch.Tensor]] = {}
+
+    # iterate over all parameter weights in the model shards
+    for key, weight, layer_idx in iter_parameter_weights(model_path, device=device):
+        # get the matching module name in modules_to_recurse for the current parameter key
+        matched_module_name = next(
+            (module for module in modules_to_recurse if module in key), None
+        )
+        if matched_module_name is None:
+            continue
+
+        recurse_idx = layer_idx % recurse_layers
+        suffix = f"{recurse_idx}.{matched_module_name}"
+        if rrt_avg_model_state_dict.get(suffix) is None:
+            # setup as storage for suffix with torch.stack
+            rrt_avg_model_state_dict[suffix] = [weight.to(torch.float32).detach().cpu()]
+        else:
+            rrt_avg_model_state_dict[suffix].append(
+                weight.to(torch.float32).detach().cpu()
+            )
+
+    for module_name in modules_to_recurse:
+        for recurse_idx in range(recurse_layers):
+            suffix = f"{recurse_idx}.{module_name}"
+            prefix = f"model.layers.{suffix}"
+            avg_weight = torch.stack(rrt_avg_model_state_dict[suffix]).mean(dim=0)
+            yield f"{prefix}.weight_base", avg_weight
+
+    # compute the decomposed lora diff from the weight base to the actual weight for each module
+
+
+def low_rank_decomposition(
+    weight: torch.Tensor, max_rank: int
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Decompose a 2D matrix into low-rank matrices L and R using SVD.
+
+    :param weight: The matrix to decompose, of shape (H, W)
+    :param max_rank: The maximum rank of the decomposition
+    :return: A tuple of tensors (L, R)
+    """
+    # pylint: disable=invalid-name
+    assert (
+        weight.dim() == 2
+    ), f"Only support 2D matrix, but input has {weight.dim()} dimensions."
+    assert (
+        max_rank >= 1
+    ), f"Maximum rank must be a positive integer, but input max_rank={max_rank}."
+
+    dtype = weight.dtype
+
+    U, S, Vh = torch.linalg.svd(weight.float(), full_matrices=False)
+
+    # Distribute S to both to improve numerical precision
+    sqrt_S = torch.sqrt(torch.diag(S[:max_rank]))
+    A = sqrt_S @ Vh[:max_rank, :]  # shape: [r, cols]
+    B = U[:, :max_rank] @ sqrt_S  # shape: [rows, r]
+
+    return A.to(dtype), B.to(dtype)
+
+
+def get_weight_norm(weight, lora_weight, scaling) -> torch.Tensor:
+    # calculate L2 norm of weight matrix, column-wise
+    weight = weight + scaling * lora_weight
+    weight_norm = torch.linalg.norm(weight, dim=1).to(weight.dtype)
+    return weight_norm
+
+
+def decompose_delta_weight(layer_weight, avg_weight, alpha, rank, use_dora=True):
+    """
+    Decompose the difference in directions (ΔV) via SVD,
+    and return (magnitudes, L, R).
+    """
+    device = "cuda" if torch.cuda.is_available() else "mps"
+
+    # rslora
+    scaling = alpha / math.sqrt(rank)
+
+    base_weight = avg_weight.to(device)
+    final_weight = layer_weight.to(device)
+
+    delta_for_svd = final_weight - base_weight
+
+    # Low-rank factorization of the delta direction
+    lora_A, lora_B = low_rank_decomposition(  # pylint: disable=invalid-name
+        delta_for_svd, rank
+    )
+
+    if use_dora:
+        lora_weight = lora_B @ lora_A
+        weight_norm = get_weight_norm(
+            base_weight.to(lora_A.device), lora_weight, scaling
+        )
+        return lora_A.cpu(), lora_B.cpu(), weight_norm.cpu()
+
+    # let's rescale the lora weight to have the same magnitude as the base weight
+
+    return lora_A.cpu(), lora_B.cpu(), None
+
+
+def iter_dora_parameter_weights(
+    model_path,
+    avg_recursive_weights,
+    modules_to_recurse: list[str],
+    alpha,
+    rank,
+    device="mps",
+    recurse_layers=12,
+    use_dora=True,
+):
+    # iterate over all parameter weights in the model shards
+    for key, weight, layer_idx in iter_parameter_weights(model_path, device=device):
+        # get the matching module name in modules_to_recurse for the current parameter key
+        matched_module_name = next(
+            (module for module in modules_to_recurse if module in key), None
+        )
+        if matched_module_name is None:
+            if "input_layernorm" in key:
+                # map to input_layernorm_list in the recursive layers and account for the layer_idx and loop_idx
+                loop_idx = layer_idx // recurse_layers
+                layer_idx = layer_idx % recurse_layers
+                layernorm_key = (
+                    f"model.layers.{layer_idx}.input_layernorm_list.{loop_idx}.weight"
+                )
+                yield layernorm_key, weight
+            elif "post_attention_layernorm" in key:
+                # map to input_layernorm_list in the recursive layers and account for the layer_idx and loop_idx
+                loop_idx = layer_idx // recurse_layers
+                layer_idx = layer_idx % recurse_layers
+                layernorm_key = f"model.layers.{layer_idx}.post_attention_layernorm_list.{loop_idx}.weight"
+                yield layernorm_key, weight
+            else:
+                yield key, weight
+            continue
+
+        # figure out the base weight layer for this key
+        loop_idx = layer_idx // recurse_layers
+        layer_idx = layer_idx % recurse_layers
+        suffix = f"{layer_idx}.{matched_module_name}"
+        prefix = f"model.layers.{suffix}.weight_base"
+        avg_weight = avg_recursive_weights[prefix]
+        lora_a_key = f"model.layers.{suffix}.lora_A_list.{loop_idx}"
+        lora_b_key = f"model.layers.{suffix}.lora_B_list.{loop_idx}"
+        lora_magnitude_key = (
+            f"model.layers.{suffix}.lora_magnitude_vector_list.{loop_idx}"
+        )
+        lora_a, lora_b, lora_magnitude = decompose_delta_weight(
+            weight,
+            avg_weight,
+            alpha,
+            rank,
+            use_dora=use_dora,
+        )
+        yield lora_a_key, lora_a
+        yield lora_b_key, lora_b
+        if use_dora:
+            yield lora_magnitude_key, lora_magnitude
+
+
+def save_state_dict_to_safetensors(state_dict, save_directory):
+    os.makedirs(save_directory, exist_ok=True)
+    weights_name = SAFE_WEIGHTS_NAME
+
+    filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(
+        ".safetensors", "{suffix}.safetensors"
+    )
+    state_dict_split = split_torch_state_dict_into_shards(
+        state_dict, filename_pattern=filename_pattern, max_shard_size="1GB"
+    )
+    # pylint: disable=duplicate-code
+    # Save index if sharded
+    index = None
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+
+    # Clean the folder from a previous save
+    for filename in os.listdir(save_directory):
+        full_filename = os.path.join(save_directory, filename)
+        # If we have a shard file that is not going to be replaced, we delete it, but only from the main process
+        # in distributed settings to avoid race conditions.
+        weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
+
+        # make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
+        filename_no_suffix = filename.replace(".bin", "").replace(".safetensors", "")
+        reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
+
+        if (
+            filename.startswith(weights_no_suffix)
+            and os.path.isfile(full_filename)
+            and filename not in state_dict_split.filename_to_tensors.keys()
+            and reg.fullmatch(filename_no_suffix) is not None
+        ):
+            os.remove(full_filename)
+
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in filename_to_tensors:
+        shard = {}
+        for tensor in tensors:
+            shard[tensor] = state_dict[tensor].contiguous()
+            del state_dict[tensor]
+
+        save_file(
+            shard, os.path.join(save_directory, shard_file), metadata={"format": "pt"}
+        )
+
+    del state_dict
+
+    if index is None:
+        path_to_weights = os.path.join(save_directory, weights_name)
+        logger.info(f"Model weights saved in {path_to_weights}")
+    else:
+        save_index_file = SAFE_WEIGHTS_INDEX_NAME
+        save_index_file = os.path.join(save_directory, save_index_file)
+        # Save the index as well
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def convert_llama_to_rrt(
+    model_name,
+    output_dir,
+    recurse_layers: int = 12,
+    rank=32,
+    alpha=32,
+    device=None,
+    use_dora=True,
+):
+    if not device:
+        if torch.backends.mps.is_available():
+            device = "mps"
+        elif torch.cuda.is_available():
+            device = "cuda"
+        else:
+            device = "cpu"
+
+    modules_to_recurse = [
+        "self_attn.q_proj",
+        "self_attn.k_proj",
+        "self_attn.v_proj",
+        "self_attn.o_proj",
+        "mlp.down_proj",
+        "mlp.gate_proj",
+        "mlp.up_proj",
+    ]
+
+    config = AutoConfig.from_pretrained(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    num_hidden_layers = config.num_hidden_layers
+    if num_hidden_layers % recurse_layers != 0:
+        raise ValueError(
+            f"The number of hidden layers ({num_hidden_layers}) in the model must be "
+            f"divisible by the recurse layers ({recurse_layers})"
+        )
+
+    config = RelaxedRecursiveLlamaConfig.from_dict(
+        {
+            **config.to_dict(),
+            "recurse_layers": recurse_layers,
+            "rank": rank,
+            "alpha": alpha,
+            "use_dora": use_dora,
+        }
+    )
+    config.save_pretrained(output_dir)
+    tokenizer.save_pretrained(output_dir)
+    model_path = Path(snapshot_download(model_name, ignore_patterns="*.pth"))
+
+    # create a new state_dict to store the RRT model weights
+    rrt_model_state_dict = {}
+
+    logger.info("Calculating average recursive weights...")
+    for key, weight in iter_recursive_parameter_weights(
+        model_path, modules_to_recurse, device=device, recurse_layers=recurse_layers
+    ):
+        rrt_model_state_dict[key] = weight.to(torch.bfloat16).detach().cpu()
+
+    logger.info("Calculating decomposed lora diff...")
+    # now that we have the average weights, we need to loop over the shards again to calculate the decomposed lora diff
+    rrt_lora_state_dict = {}
+    for key, weight in iter_dora_parameter_weights(
+        model_path,
+        rrt_model_state_dict,
+        modules_to_recurse,
+        alpha=32,
+        rank=rank,
+        device=device,
+        recurse_layers=recurse_layers,
+        use_dora=use_dora,
+    ):
+        rrt_lora_state_dict[key] = weight.to(torch.bfloat16).detach().cpu()
+
+    # combine state dicts into a single state_dict
+    rrt_model_state_dict.update(rrt_lora_state_dict)
+
+    # save state dict as sharded safetensors to disk using split_torch_state_dict_into_shards
+    save_state_dict_to_safetensors(rrt_model_state_dict, output_dir)
+
+
+if __name__ == "__main__":
+    # meta-llama/Llama-3.2-1B has 16 hidden layers
+    # meta-llama/Llama-3.2-3B has 28 hidden layers
+    convert_llama_to_rrt(
+        "meta-llama/Llama-3.2-3B",
+        "/tmp/rrt_model",  # nosec
+        recurse_layers=4,
+        rank=256,
+        alpha=512,
+        use_dora=False,
+    )

src/axolotl/integrations/rrt/modeling/__init__.py

@@ -0,0 +1,25 @@
+"""
+module for modeling relaxed recursive transformers model
+"""
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from .configuration_rrt_llama import RelaxedRecursiveLlamaConfig
+from .modeling_rrt_llama import (
+    RelaxedRecursiveLlamaForCausalLM,
+    RelaxedRecursiveLlamaModel,
+)
+
+
+def register_rrt_model():
+    """
+    Register Relaxed Recursive Transformers model with transformers
+    """
+
+    # Register configs
+    AutoConfig.register("llama-rrt", RelaxedRecursiveLlamaConfig)
+
+    # Register models
+    AutoModel.register(RelaxedRecursiveLlamaConfig, RelaxedRecursiveLlamaModel)
+    AutoModelForCausalLM.register(
+        RelaxedRecursiveLlamaConfig, RelaxedRecursiveLlamaForCausalLM
+    )

src/axolotl/integrations/rrt/modeling/configuration_rrt_llama.py

@@ -0,0 +1,16 @@
+"""
+module for custom configuration for relaxed recursive transformers model
+"""
+from transformers import LlamaConfig
+
+
+class RelaxedRecursiveLlamaConfig(LlamaConfig):
+    """
+    Configuration for Relaxed Recursive Llama.
+    """
+
+    model_type: str = "llama-rrt"
+    recurse_layers: int = 4
+    rank: int
+    alpha: int
+    use_dora: bool = True

src/axolotl/integrations/rrt/modeling/linear.py

@@ -0,0 +1,116 @@
+"""
+module for the shared linear layer for the relaxed recursive transformers model
+"""
+import math
+
+import torch
+import torch.nn.functional as F
+from peft.utils import transpose
+from torch import nn
+
+
+class RelaxedRecursiveDoraLinear(nn.Module):
+    """
+    A single linear layer that is "shared" across multiple loop iterations,
+    but each iteration has its own DoRA offsets (A_i, B_i, magnitude_i).
+
+    The constructor expects you to specify:
+      - in_features, out_features
+      - B: number of loop iterations (i.e., how many times we "unroll")
+      - fan_in_fan_out: pass True if your underlying base weight is transposed, etc.
+
+    The forward(...) expects an additional argument "loop_idx" in [0..B-1],
+    which picks out the iteration-specific DoRA offsets.
+    """
+
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        B: int,  # pylint: disable=invalid-name
+        rank: int,
+        alpha: int,
+        fan_in_fan_out: bool = False,
+        bias: bool = True,
+        use_dora: bool = True,
+    ):
+        super().__init__()
+        self.B = B  # pylint: disable=invalid-name
+        self.fan_in_fan_out = fan_in_fan_out
+
+        self.weight_base = nn.Parameter(torch.empty(out_features, in_features))
+
+        self.use_bias = bias
+        if self.use_bias:
+            self.bias = nn.Parameter(torch.zeros(out_features))
+        else:
+            self.register_parameter("bias", None)
+
+        self.lora_A_list = nn.ParameterList(  # pylint: disable=invalid-name
+            [nn.Parameter(torch.zeros(rank, in_features)) for _ in range(B)]
+        )
+        self.lora_B_list = nn.ParameterList(  # pylint: disable=invalid-name
+            [nn.Parameter(torch.zeros(out_features, rank)) for _ in range(B)]
+        )
+        # rslora
+        self.scaling = alpha / math.sqrt(rank)
+        self.use_dora = use_dora
+        if use_dora:
+            self.lora_magnitude_vector_list = nn.ParameterList(
+                [nn.Parameter(torch.ones(out_features)) for _ in range(B)]
+            )
+
+    def get_weight_norm(self, weight, lora_weight, scaling) -> torch.Tensor:
+        # calculate L2 norm of weight matrix, column-wise
+        weight = transpose(weight, self.fan_in_fan_out)
+        weight = weight + scaling * lora_weight
+        weight_norm = torch.linalg.norm(weight, dim=1).to(weight.dtype)
+        return weight_norm
+
+    def forward(self, x, loop_idx: int):
+        """
+
+        :param x: hidden state of shape (batch_size, seq_len, in_features)
+        :param loop_idx:
+        :return:
+        """
+        eps = 1e-6
+        w_base = self.weight_base
+        w_base = w_base.to(x.dtype)
+
+        lora_A: torch.Tensor = self.lora_A_list[  # pylint: disable=invalid-name
+            loop_idx
+        ]
+        lora_B: torch.Tensor = self.lora_B_list[  # pylint: disable=invalid-name
+            loop_idx
+        ]
+
+        base_out: torch.Tensor = F.linear(x, w_base, self.bias)
+        lora_out: torch.Tensor = F.linear(F.linear(x, lora_A), lora_B) * self.scaling
+
+        if self.use_dora:
+            x_eye: torch.Tensor = torch.eye(
+                lora_A.shape[1], device=lora_A.device, dtype=x.dtype
+            )
+            tmp = F.linear(x_eye, lora_A)  # [hidden_size, rank]
+            w_dora_full: torch.Tensor = F.linear(tmp, lora_B)
+            w_dora_full = w_dora_full.t()
+
+            magnitude_vector: torch.Tensor = self.lora_magnitude_vector_list[loop_idx]
+            w_dora_norm: torch.Tensor = self.get_weight_norm(
+                w_base, w_dora_full.detach(), self.scaling
+            )
+            w_dora_norm = w_dora_norm.detach()
+            scale_factor = (magnitude_vector / w_dora_norm).unsqueeze(
+                0
+            )  # shape [1, out_features]
+
+            result_dora = (scale_factor - 1) * base_out + scale_factor * lora_out
+            return result_dora
+
+        # scale the lora norm to prevent gradient explosion
+        orig_norm = torch.linalg.norm(w_base)
+        update_norm = torch.linalg.norm(lora_out)
+        scale = orig_norm / (update_norm + eps)
+
+        return base_out + lora_out * scale

src/axolotl/integrations/rrt/modeling/modeling_rrt_llama.py

@@ -0,0 +1,471 @@
+import logging
+from typing import Callable, Optional, Tuple, Union, Unpack
+
+import torch
+from torch import nn
+from transformers import Cache, DynamicCache, LlamaConfig
+from transformers.activations import ACT2FN
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+from transformers.models.llama.modeling_llama import (
+    LlamaForCausalLM,
+    LlamaModel,
+    LlamaRMSNorm,
+    LlamaRotaryEmbedding,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+)
+
+from axolotl.integrations.rrt.modeling.linear import RelaxedRecursiveDoraLinear
+
+from .configuration_rrt_llama import RelaxedRecursiveLlamaConfig
+
+logger = logging.getLogger(__name__)
+
+
+# pylint: skip-file
+# mypy: ignore-errors
+
+
+class RelaxedRecursiveLlamaMLP(nn.Module):
+    def __init__(self, config: RelaxedRecursiveLlamaConfig):
+        super().__init__()
+        recurse_loops = config.num_hidden_layers // config.recurse_layers
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = RelaxedRecursiveDoraLinear(
+            self.hidden_size,
+            self.intermediate_size,
+            recurse_loops,
+            config.rank,
+            config.alpha,
+            bias=config.mlp_bias,
+            use_dora=config.use_dora,
+        )
+        self.up_proj = RelaxedRecursiveDoraLinear(
+            self.hidden_size,
+            self.intermediate_size,
+            recurse_loops,
+            config.rank,
+            config.alpha,
+            bias=config.mlp_bias,
+            use_dora=config.use_dora,
+        )
+        self.down_proj = RelaxedRecursiveDoraLinear(
+            self.intermediate_size,
+            self.hidden_size,
+            recurse_loops,
+            config.rank,
+            config.alpha,
+            bias=config.mlp_bias,
+            use_dora=config.use_dora,
+        )
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x, loop_idx: int):
+        down_proj = self.down_proj(
+            self.act_fn(self.gate_proj(x, loop_idx)) * self.up_proj(x, loop_idx),
+            loop_idx,
+        )
+        return down_proj
+
+
+class RelaxedRecursiveLlamaAttention(nn.Module):
+    """
+    A single attention layer of the Relaxed Recursive Llama.
+    """
+
+    def __init__(self, config: RelaxedRecursiveLlamaConfig, layer_idx: int):
+        super().__init__()
+        recurse_loops = config.num_hidden_layers // config.recurse_layers
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(
+            config, "head_dim", config.hidden_size // config.num_attention_heads
+        )
+        self.num_key_value_groups = (
+            config.num_attention_heads // config.num_key_value_heads
+        )
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = RelaxedRecursiveDoraLinear(
+            config.hidden_size,
+            config.num_attention_heads * self.head_dim,
+            recurse_loops,
+            config.rank,
+            config.alpha,
+            bias=config.attention_bias,
+            use_dora=config.use_dora,
+        )
+        self.k_proj = RelaxedRecursiveDoraLinear(
+            config.hidden_size,
+            config.num_key_value_heads * self.head_dim,
+            recurse_loops,
+            config.rank,
+            config.alpha,
+            bias=config.attention_bias,
+            use_dora=config.use_dora,
+        )
+        self.v_proj = RelaxedRecursiveDoraLinear(
+            config.hidden_size,
+            config.num_key_value_heads * self.head_dim,
+            recurse_loops,
+            config.rank,
+            config.alpha,
+            bias=config.attention_bias,
+            use_dora=config.use_dora,
+        )
+        self.o_proj = RelaxedRecursiveDoraLinear(
+            config.num_attention_heads * self.head_dim,
+            config.hidden_size,
+            recurse_loops,
+            config.rank,
+            config.alpha,
+            bias=config.attention_bias,
+            use_dora=config.use_dora,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        loop_idx: int,
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],  # pylint: disable=misc
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = (
+            self.q_proj(hidden_states, loop_idx).view(hidden_shape).transpose(1, 2)
+        )
+        key_states = (
+            self.k_proj(hidden_states, loop_idx).view(hidden_shape).transpose(1, 2)
+        )
+        value_states = (
+            self.v_proj(hidden_states, loop_idx).view(hidden_shape).transpose(1, 2)
+        )
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin
+        )
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get(
+                "output_attentions", False
+            ):
+                logger.warning(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[
+                    self.config._attn_implementation
+                ]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output, loop_idx)
+        return attn_output, attn_weights  # pylint: disable=return-value
+
+
+class RelaxedRecursiveLlamaDecoderLayer(nn.Module):
+    """
+    A single layer of the Relaxed Recursive Llama decoder.
+    """
+
+    def __init__(self, config: LlamaConfig, layer_idx: int):
+        super().__init__()
+        recurse_loops = config.num_hidden_layers // config.recurse_layers
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = RelaxedRecursiveLlamaAttention(
+            config=config, layer_idx=layer_idx
+        )
+
+        self.mlp = RelaxedRecursiveLlamaMLP(config)
+
+        self.input_layernorm_list = nn.ModuleList(
+            [
+                LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+                for _ in range(recurse_loops)
+            ]
+        )
+        self.post_attention_layernorm_list = nn.ModuleList(
+            [
+                LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+                for _ in range(recurse_loops)
+            ]
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        loop_idx: int,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[
+            Tuple[torch.Tensor, torch.Tensor]
+        ] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],  # pylint: disable=misc
+    ) -> Tuple[
+        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm_list[loop_idx](hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            loop_idx=loop_idx,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm_list[loop_idx](hidden_states)
+        hidden_states = self.mlp(hidden_states, loop_idx)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class RelaxedRecursiveLlamaModel(LlamaModel):
+    config_class = RelaxedRecursiveLlamaConfig
+
+    def __init__(self, config):
+        super(LlamaModel, self).__init__(config)
+        self.recurse_loops = config.num_hidden_layers // config.recurse_layers
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx
+        )
+        self.layers = nn.ModuleList(
+            [
+                RelaxedRecursiveLlamaDecoderLayer(config, layer_idx)
+                for layer_idx in range(config.recurse_layers)
+            ]
+        )
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = LlamaRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = (
+                past_key_values.get_seq_length() if past_key_values is not None else 0
+            )
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask,
+            inputs_embeds,
+            cache_position,
+            past_key_values,
+            output_attentions,
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        for loop_idx in range(self.recurse_loops):
+            for decoder_layer in self.layers[: self.config.recurse_layers]:
+                if output_hidden_states:
+                    all_hidden_states += (hidden_states,)
+
+                if self.gradient_checkpointing and self.training:
+                    layer_outputs = self._gradient_checkpointing_func(
+                        decoder_layer.__call__,
+                        hidden_states,
+                        loop_idx,
+                        causal_mask,
+                        position_ids,
+                        past_key_values,
+                        output_attentions,
+                        use_cache,
+                        cache_position,
+                        position_embeddings,
+                    )
+                else:
+                    layer_outputs = decoder_layer(
+                        hidden_states,
+                        loop_idx,
+                        attention_mask=causal_mask,
+                        position_ids=position_ids,
+                        past_key_value=past_key_values,
+                        output_attentions=output_attentions,
+                        use_cache=use_cache,
+                        cache_position=cache_position,
+                        position_embeddings=position_embeddings,
+                        **flash_attn_kwargs,
+                    )
+
+                hidden_states = layer_outputs[0]
+
+                if output_attentions:
+                    all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        output = BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+        return output if return_dict else output.to_tuple()
+
+
+class RelaxedRecursiveLlamaForCausalLM(LlamaForCausalLM):
+    config_class = RelaxedRecursiveLlamaConfig
+
+    def __init__(self, config):
+        super(LlamaForCausalLM, self).__init__(config)
+        self.model = RelaxedRecursiveLlamaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_nb_trainable_parameters(self) -> tuple[int, int, int]:
+        r"""
+        Returns the number of trainable parameters and the number of all parameters in the model.
+        """
+        trainable_params = 0
+        all_param = 0
+        lora_params = 0
+        for name, param in self.named_parameters():
+            num_params = param.numel()
+            # if using DS Zero 3 and the weights are initialized empty
+            if num_params == 0 and hasattr(param, "ds_numel"):
+                num_params = param.ds_numel
+
+            # Due to the design of 4bit linear layers from bitsandbytes
+            # one needs to multiply the number of parameters by 2 to get
+            # the correct number of parameters
+            if param.__class__.__name__ == "Params4bit":
+                if hasattr(param, "element_size"):
+                    num_bytes = param.element_size()
+                elif not hasattr(param, "quant_storage"):
+                    num_bytes = 1
+                else:
+                    num_bytes = param.quant_storage.itemsize
+                num_params = num_params * 2 * num_bytes
+
+            all_param += num_params
+            if param.requires_grad:
+                trainable_params += num_params
+            if "lora_" in name:
+                lora_params += num_params
+
+        return trainable_params, all_param, lora_params

src/axolotl/monkeypatch/trainer_fsdp_optim.py

@@ -6,7 +6,7 @@ import logging
 
 from transformers import Trainer
 
-from axolotl.monkeypatch.unsloth_ import detab_code
+from axolotl.monkeypatch.utils import detab_code
 
 LOG = logging.getLogger("axolotl.monkeypatch.trainer_fsdp_save")
 

src/axolotl/monkeypatch/trainer_grad_accum.py

@@ -1,308 +0,0 @@
-"""
-fix for FSDP gradient accumulation
-see https://github.com/huggingface/transformers/pull/35128
-"""
-import inspect
-import logging
-
-from transformers import LlamaForCausalLM, Trainer
-from transformers.modeling_flash_attention_utils import _flash_attention_forward
-
-from axolotl.monkeypatch.unsloth_ import detab_code
-
-LOG = logging.getLogger("axolotl.monkeypatch.trainer_grad_accum")
-
-ORIGINAL_CONTEXT_CODE = """
-    with self.compute_loss_context_manager():
-        loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
-"""
-
-PATCHED_CONTEXT_CODE = """
-    with self.compute_loss_context_manager():
-        if self.model_accepts_loss_kwargs:
-            loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
-        else:
-            loss = self.compute_loss(model, inputs)
-"""
-
-ORIGINAL_LLAMA_FCLM_CODE = """
-    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-    output_hidden_states = (
-        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-    )
-    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-    outputs = self.model(
-        input_ids=input_ids,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        use_cache=use_cache,
-        output_attentions=output_attentions,
-        output_hidden_states=output_hidden_states,
-        return_dict=return_dict,
-        cache_position=cache_position,
-        **kwargs,
-    )
-
-    hidden_states = outputs[0]
-    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
-    logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-
-    loss = None
-    if labels is not None:
-        loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
-"""
-
-PATCHED_LLAMA_FCLM_CODE = """
-    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-    output_hidden_states = (
-        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-    )
-    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-    # remove num_items_in_batch otherwise self.model attempts to pass it to flash_attention
-    num_items_in_batch = kwargs.pop("num_items_in_batch", None)
-
-    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-    outputs = self.model(
-        input_ids=input_ids,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        use_cache=use_cache,
-        output_attentions=output_attentions,
-        output_hidden_states=output_hidden_states,
-        return_dict=return_dict,
-        cache_position=cache_position,
-        **kwargs,
-    )
-    hidden_states = outputs[0]
-    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
-    logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-
-    loss = None
-    if labels is not None:
-        loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, num_items_in_batch=num_items_in_batch, **kwargs)
-"""
-
-
-def get_training_step_code() -> str:
-    training_step = inspect.getsource(
-        Trainer.training_step  # pylint: disable=protected-access
-    )
-    return training_step
-
-
-def check_training_step_is_patchable() -> bool:
-    training_step = get_training_step_code()
-    training_step, _ = detab_code(training_step)
-    return ORIGINAL_CONTEXT_CODE in training_step
-
-
-def patch_training_step_for_ga():
-    """
-    monkeypatch for fixing the training loop for gradient accumulation
-    """
-
-    try:
-        training_step = get_training_step_code()
-    except OSError:
-        return
-    Trainer._original_training_step = training_step  # pylint: disable=protected-access
-    training_step, _ = detab_code(training_step)
-    if ORIGINAL_CONTEXT_CODE not in training_step:
-        return
-    # assert (
-    #     ORIGINAL_CONTEXT_CODE in training_step
-    # ), "Original training_step code not found"
-
-    training_step = training_step.replace(ORIGINAL_CONTEXT_CODE, PATCHED_CONTEXT_CODE)
-    training_step = training_step.replace(
-        "def training_step(",
-        "def _fixed_training_step(",
-        1,
-    )
-
-    # load imports necessary
-    import transformers.trainer
-
-    items_to_import = []
-    for item in dir(transformers.trainer):
-        if item in training_step:
-            items_to_import.append(item)
-
-    exec(  # pylint: disable=exec-used  # nosec B102
-        "from transformers.trainer import ("
-        + ", ".join(x for x in items_to_import)
-        + ")",
-        globals(),
-    )
-    exec(training_step, globals())  # pylint: disable=exec-used  # nosec B102
-    LOG.info("patching training_step")
-    Trainer.training_step = (  # pylint: disable=protected-access
-        _fixed_training_step  # pylint: disable=undefined-variable  # noqa: F821
-    )
-
-
-def get_model_forward_code() -> str:
-    forward = inspect.getsource(
-        LlamaForCausalLM.forward  # pylint: disable=protected-access
-    )
-    return forward
-
-
-def check_forward_is_patchable() -> bool:
-    forward = get_model_forward_code()
-    forward, _ = detab_code(forward)
-    return ORIGINAL_LLAMA_FCLM_CODE in forward
-
-
-def patch_forward_for_ga():
-    """
-    monkeypatch for fixing the training loop for gradient accumulation
-    """
-
-    try:
-        forward = get_model_forward_code()
-    except OSError:
-        return
-    LlamaForCausalLM._original_forward = forward  # pylint: disable=protected-access
-    forward, _ = detab_code(forward)
-    if ORIGINAL_LLAMA_FCLM_CODE not in forward:
-        return
-    # assert ORIGINAL_LLAMA_FCLM_CODE in forward, "Original forward code not found"
-
-    forward = forward.replace(ORIGINAL_LLAMA_FCLM_CODE, PATCHED_LLAMA_FCLM_CODE)
-    forward = forward.replace(
-        "def forward(",
-        "def _fixed_forward(",
-        1,
-    )
-
-    # load imports necessary
-    import transformers.models.llama.modeling_llama
-
-    items_to_import = []
-    for item in dir(transformers.models.llama.modeling_llama):
-        if item in forward:
-            items_to_import.append(item)
-
-    exec(  # pylint: disable=exec-used  # nosec B102
-        "from transformers.models.llama.modeling_llama import ("
-        + ", ".join(x for x in items_to_import)
-        + ")",
-        globals(),
-    )
-    exec(forward, globals())  # pylint: disable=exec-used  # nosec B102
-    LOG.info("patching forward")
-    LlamaForCausalLM.forward = (  # pylint: disable=protected-access
-        _fixed_forward  # pylint: disable=undefined-variable  # noqa: F821
-    )
-
-
-ORIGINAL_TRAINER_CODE = """
-                context = (
-                    functools.partial(self.accelerator.no_sync, model=model)
-                    if i != len(batch_samples) - 1
-                    else contextlib.nullcontext
-                )
-                with context():
-                    tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
-"""
-
-PATCHED_TRAINER_CODE = """
-                disable_deepspeed_no_sync = (
-                        self.accelerator.distributed_type == DistributedType.DEEPSPEED
-                        # and self.accelerator.deepspeed_engine_wrapped.engine.zero_optimization_partition_gradients()
-                )
-                context = (
-                    functools.partial(self.accelerator.no_sync, model=model)
-                    if i != len(batch_samples) - 1 and not disable_deepspeed_no_sync
-                    else contextlib.nullcontext
-                )
-                with context():
-                    tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
-"""
-
-
-def get_training_loop_code() -> str:
-    training_loop = inspect.getsource(
-        Trainer._inner_training_loop  # pylint: disable=protected-access
-    )
-    return training_loop
-
-
-def check_training_loop_is_patchable() -> bool:
-    training_loop = get_training_loop_code()
-    training_loop, _ = detab_code(training_loop)
-    return ORIGINAL_TRAINER_CODE in training_loop
-
-
-def patch_training_loop_for_deepspeed_0_16_x():
-    """
-    monkeypatch for fixing the training loop for deepspeed GA
-
-    see https://github.com/huggingface/transformers/pull/35157
-    """
-
-    try:
-        training_loop = get_training_loop_code()
-    except OSError:
-        return
-    Trainer._original_inner_training_loop = (  # pylint: disable=protected-access
-        training_loop
-    )
-    training_loop, _ = detab_code(training_loop)
-    if ORIGINAL_TRAINER_CODE not in training_loop:
-        return
-
-    training_loop = training_loop.replace(ORIGINAL_TRAINER_CODE, PATCHED_TRAINER_CODE)
-    training_loop = training_loop.replace(
-        "def _inner_training_loop(",
-        "def _fixed_inner_training_loop(",
-        1,
-    )
-
-    # load imports necessary
-    import transformers.trainer
-
-    items_to_import = []
-    for item in dir(transformers.trainer):
-        if item in training_loop:
-            items_to_import.append(item)
-
-    exec(  # pylint: disable=exec-used  # nosec B102
-        "from transformers.trainer import ("
-        + ", ".join(x for x in items_to_import)
-        + ")",
-        globals(),
-    )
-    exec(training_loop, globals())  # pylint: disable=exec-used  # nosec B102
-    LOG.info("patching _inner_training_loop for fsdp optimizer save")
-    Trainer._inner_training_loop = (  # pylint: disable=protected-access
-        _fixed_inner_training_loop  # pylint: disable=undefined-variable  # noqa: F821
-    )
-
-
-def patch_flash_attention_forward():
-    """
-    monkeypatch for fixing the forward pass for flash attention to ignore num_items_in_batch
-    """
-
-    import transformers.modeling_flash_attention_utils
-
-    def proxy_flash_attention_forward(*args, **kwargs):
-        kwargs.pop("num_items_in_batch", None)
-
-        return _flash_attention_forward(*args, **kwargs)
-
-    transformers.modeling_flash_attention_utils._flash_attention_forward = (  # pylint: disable=protected-access
-        proxy_flash_attention_forward
-    )
-    transformers.models.llama.modeling_llama._flash_attention_forward = (  # pylint: disable=protected-access
-        proxy_flash_attention_forward
-    )

src/axolotl/monkeypatch/transformers_fa_utils.py

@@ -0,0 +1,67 @@
+"""
+see https://github.com/huggingface/transformers/pull/35834
+"""
+
+import logging
+from functools import partial
+from typing import Optional
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+
+def fixed_fa_peft_integration_check(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    target_dtype: Optional[torch.dtype] = None,
+    preferred_dtype: Optional[torch.dtype] = None,
+):
+    """
+    PEFT usually casts the layer norms in float32 for training stability reasons
+    therefore the input hidden states gets silently casted in float32. Hence, we need
+    cast them back in float16 / bfloat16 just to be sure everything works as expected.
+    This might slowdown training & inference so it is recommended to not cast the LayerNorms!
+
+    Args:
+        query (`torch.Tensor`):
+            Input query states to be passed to Flash Attention API
+        key (`torch.Tensor`):
+            Input key states to be passed to Flash Attention API
+        value (`torch.Tensor`):
+            Input value states to be passed to Flash Attention API
+        target_dtype (`torch.dtype`, *optional*):
+            The dtype to convert the attention tensors to. Conversion can be ignored by
+            not providing the target dtype.
+        preferred_dtype (`torch.dtype`, *optional*):
+            The preferred dtype to convert the attention tensors to regardless of the
+            target dtype.
+    """
+    if target_dtype is None and preferred_dtype is None:
+        return query, key, value
+
+    if preferred_dtype and target_dtype != preferred_dtype:
+        target_dtype = preferred_dtype
+
+    # check if any of query, key, or value are in float32. If so, cast them back to target dtype.
+    if any(module.dtype == torch.float32 for module in [query, key, value]):
+        logger.warning_once(
+            f"The input hidden states seems to be silently casted in float32, this might be related to"
+            f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+            f" {target_dtype}."
+        )
+
+        query = query.to(target_dtype)
+        key = key.to(target_dtype)
+        value = value.to(target_dtype)
+
+    return query, key, value
+
+
+def patch_fa_peft_integration():
+    import transformers.modeling_flash_attention_utils
+
+    transformers.modeling_flash_attention_utils.fa_peft_integration_check = partial(
+        fixed_fa_peft_integration_check, preferred_dtype=None
+    )

src/axolotl/monkeypatch/unsloth_.py

@@ -1,9 +1,7 @@
 """module for patching with unsloth optimizations"""
 
 import inspect
-import re
 import types
-from typing import Tuple
 
 import torch
 from accelerate.logging import get_logger
@@ -11,6 +9,8 @@ from peft import PeftModelForCausalLM
 from torch import nn
 from transformers.models.llama.modeling_llama import LlamaFlashAttention2
 
+from axolotl.monkeypatch.utils import detab_code
+
 LOG = get_logger("axolotl.monkeypatch.unsloth")
 
 ORIGINAL_QKV_CODE = """
@@ -93,15 +93,6 @@ def integrate_cross_entropy_loss_patch(model_type: str = "llama") -> None:
         raise ValueError("Unsupported model type")
 
 
-def detab_code(code: str) -> Tuple[str, str]:
-    try:
-        spaces = re.match(r"([\s\t]{1,})", code).group(0)
-        code = re.sub(r"^" + spaces, "", code, flags=re.MULTILINE)
-    except AttributeError:
-        return code, ""
-    return code, spaces
-
-
 self_attn_lora_patched = False  # pylint: disable=invalid-name
 
 

src/axolotl/monkeypatch/utils.py

@@ -1,7 +1,8 @@
 """
 Shared utils for the monkeypatches
 """
-from typing import Optional
+import re
+from typing import Optional, Tuple
 
 import torch
 import torch.nn.functional as F
@@ -223,3 +224,12 @@ def patched_prepare_4d_causal_attention_mask_for_sdpa(
         mask_2d_to_4d(attention_mask, dtype=dtype),
         *args,
     )
+
+
+def detab_code(code: str) -> Tuple[str, str]:
+    try:
+        spaces = re.match(r"([\s\t]{1,})", code).group(0)
+        code = re.sub(r"^" + spaces, "", code, flags=re.MULTILINE)
+    except AttributeError:
+        return code, ""
+    return code, spaces

src/axolotl/train.py

@@ -5,21 +5,19 @@ import os
 import signal
 import sys
 import weakref
-from dataclasses import dataclass
 from pathlib import Path
-from typing import Optional, Tuple, Union
+from typing import Tuple, Union
 
 import torch
 import transformers.modelcard
 from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
-from datasets import Dataset
 from peft import PeftModel
 from pkg_resources import get_distribution  # type: ignore
 from transformers import PreTrainedModel, PreTrainedTokenizer
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.common.datasets import TrainDatasetMeta
 from axolotl.contribs.lgpl.unsloth import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
@@ -39,22 +37,11 @@ src_dir = os.path.join(project_root, "src")
 sys.path.insert(0, src_dir)
 
 configure_logging()
-LOG = get_logger("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
+LOG = get_logger(__name__)
 
 
 def train(
-    *, cfg: DictDefault, cli_args: TrainerCliArgs, dataset_meta: TrainDatasetMeta
+    *, cfg: DictDefault, dataset_meta: TrainDatasetMeta
 ) -> Tuple[Union[PeftModel, PreTrainedModel], PreTrainedTokenizer]:
     # Load tokenizer
     LOG.debug(
@@ -93,9 +80,7 @@ def train(
     if cfg.adapter:
         msg += " and peft_config..."
     LOG.debug(msg)
-    model, peft_config = load_model(
-        cfg, tokenizer, processor=processor, inference=cli_args.inference
-    )
+    model, peft_config = load_model(cfg, tokenizer, processor=processor)
     if model.generation_config is not None:
         model.generation_config.do_sample = True
 
@@ -107,9 +92,7 @@ def train(
             model_ref = None  # explicit setting to None
         else:
             # load the model again for model_ref/baseline
-            model_ref, _ = load_model(
-                cfg, tokenizer, inference=cli_args.inference, reference_model=True
-            )
+            model_ref, _ = load_model(cfg, tokenizer, reference_model=True)
 
     safe_serialization = cfg.save_safetensors is True
 

src/axolotl/utils/callbacks/lisa.py

@@ -43,7 +43,7 @@ def lisa_callback_factory(trainer: "AxolotlTrainer"):
                 getattr, self.layers_attribute.split("."), self.trainer.model
             )
             LOG.info(
-                f"LISA will activate {self.n_layers}/{len(layers)} layers ({self.n_layers*100/len(layers)}%) every {self.step_interval} steps"
+                f"LISA will activate {self.n_layers}/{len(layers)} layers ({self.n_layers * 100 / len(layers)}%) every {self.step_interval} steps"
             )
 
         def freeze_all_layers(self):

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -128,6 +128,8 @@ class PretrainingDataset(BaseModel):
     text_column: Optional[str] = "text"
     type: Optional[str] = "pretrain"
     trust_remote_code: Optional[bool] = False
+    data_files: Optional[str] = None
+    skip: Optional[int] = None
 
 
 class UserDefinedPrompterType(BaseModel):
@@ -145,6 +147,14 @@ class UserDefinedPrompterType(BaseModel):
     field: Optional[str] = None
 
 
+class LrGroup(BaseModel):
+    """Custom learning rate group configuration"""
+
+    name: str
+    modules: List[str]
+    lr: float
+
+
 class SFTDataset(BaseModel):
     """SFT configuration subset"""
 
@@ -366,6 +376,13 @@ class LoraConfig(BaseModel):
             loraplus_lr_embedding = float(loraplus_lr_embedding)
         return loraplus_lr_embedding
 
+    @model_validator(mode="before")
+    @classmethod
+    def validate_lora_dropout(cls, data):
+        if data.get("adapter") is not None and data.get("lora_dropout") is None:
+            data["lora_dropout"] = 0.0
+        return data
+
 
 class ReLoRAConfig(BaseModel):
     """ReLoRA configuration subset"""
@@ -466,6 +483,7 @@ class HyperparametersConfig(BaseModel):
     cosine_min_lr_ratio: Optional[float] = None
     cosine_constant_lr_ratio: Optional[float] = None
     lr_div_factor: Optional[float] = None
+    lr_groups: Optional[List[LrGroup]] = None
 
     adam_epsilon: Optional[float] = None
     adam_beta1: Optional[float] = None
@@ -697,6 +715,12 @@ class AxolotlInputConfig(
     pad_to_sequence_len: Optional[bool] = None
     curriculum_sampling: Optional[bool] = None
     multipack_real_batches: Optional[bool] = None
+    pretraining_sample_concatenation: Optional[bool] = Field(
+        default=None,
+        json_schema_extra={
+            "description": "whether to soft pack/concatenate samples during pretraining",
+        },
+    )
 
     batch_flattening: Optional[Union[Literal["auto"], bool]] = None
 

src/axolotl/utils/data/__init__.py

@@ -5,7 +5,7 @@ from axolotl.utils.data.pretraining import (  # noqa: F401
     encode_pretraining,
     wrap_pretraining_dataset,
 )
-from axolotl.utils.data.rl import load_prepare_dpo_datasets  # noqa: F401
+from axolotl.utils.data.rl import load_prepare_preference_datasets  # noqa: F401
 from axolotl.utils.data.sft import (  # noqa: F401
     get_dataset_wrapper,
     load_prepare_datasets,

src/axolotl/utils/data/pretraining.py

@@ -18,10 +18,14 @@ LOG = logging.getLogger("axolotl")
 
 
 def encode_pretraining(
-    tokenizer: PreTrainedTokenizerBase, max_tokens: int, examples: Dict[str, List]
+    tokenizer: PreTrainedTokenizerBase,
+    max_tokens: int,
+    examples: Dict[str, List],
+    text_column: str = "text",
+    concatenate: bool = True,
 ) -> Dict[str, List]:
     res = tokenizer(
-        examples["text"],
+        examples[text_column],
         truncation=True,
         max_length=max_tokens - 2,
         add_special_tokens=True,
@@ -30,6 +34,13 @@ def encode_pretraining(
     input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
     targets = [torch.tensor(seq) for seq in res["input_ids"]]
     attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
+    if not concatenate:
+        return {
+            "input_ids": [seq.tolist() for seq in input_ids],
+            "labels": [seq.tolist() for seq in targets],
+            "attention_mask": [seq.tolist() for seq in attention_mask],
+        }
+
     new_input_ids = []
     new_labels = []
     new_attention_mask = []
@@ -180,7 +191,7 @@ def wrap_pretraining_dataset(
             tokenizer,
             return_tensors="pt",
             padding=True,
-            pad_to_multiple_of=max_tokens * batch_size,
+            pad_to_multiple_of=max_tokens,
             multipack_attn=cfg.pretrain_multipack_attn,
         )
         encode = functools.partial(
@@ -190,13 +201,17 @@ def wrap_pretraining_dataset(
             max_seq_length=max_tokens,
             batch_size=batch_size,
             multipack_attn=cfg.pretrain_multipack_attn,
-            group_size=cfg.sample_packing_group_size,
-            bin_size=cfg.sample_packing_bin_size,
         )
         # set this to 1 so downstream data_loader doesn't try to increase the batch again
         cfg.micro_batch_size = 1
     else:
-        encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
+        encode = functools.partial(
+            encode_pretraining,
+            tokenizer,
+            max_tokens,
+            text_column=cfg.pretraining_dataset[0].text_column or "text",
+            concatenate=cfg.pretraining_sample_concatenation is True,
+        )
 
     if cfg.shuffle_merged_datasets:
         dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
@@ -230,9 +245,7 @@ def encode_packed_pretraining(
     examples: Dict[str, List],
     max_seq_length: int = 2048,
     batch_size: int = 4,
-    multipack_attn: Optional[bool] = False,
-    group_size: int = 100000,
-    bin_size: int = 200,
+    multipack_attn: Optional[bool] = True,
 ) -> Dict[str, List]:
     # pylint: disable=duplicate-code
     # tokenize all the examples
@@ -243,6 +256,9 @@ def encode_packed_pretraining(
         train_dataset,
         max_seq_length,
         skip_position_ids=not multipack_attn,
+        # FIXME using attention mask unpad/pad with trainer and packed pretraining is broken atm
+        # workaround by using the position id logic for now in trainer
+        drop_attention_mask=multipack_attn,
     )
 
     sampler = MultipackBatchSampler(
@@ -250,8 +266,6 @@ def encode_packed_pretraining(
         lengths=get_dataset_lengths(train_dataset),
         batch_size=1,
         batch_max_len=batch_size * max_seq_length,
-        group_size=group_size,
-        bin_size=bin_size,
         drop_last=True,
     )
 

src/axolotl/utils/data/rl.py

@@ -115,7 +115,7 @@ def drop_long_rl_seq(
     raise ValueError("Unknown RL type")
 
 
-def load_prepare_dpo_datasets(cfg):
+def load_prepare_preference_datasets(cfg):
     def load_split(dataset_cfgs, _cfg):
         split_datasets: List[Any] = []
         for i, ds_cfg in enumerate(dataset_cfgs):

src/axolotl/utils/data/sft.py

@@ -88,14 +88,19 @@ def prepare_dataset(cfg, tokenizer, processor=None):
         path = cfg.pretraining_dataset
         split = "train"
         name = None
+        data_files = None
+        skip = 0
         if isinstance(cfg.pretraining_dataset, list) and isinstance(
             cfg.pretraining_dataset[0], dict
         ):
             path = cfg.pretraining_dataset[0]["path"]
             name = cfg.pretraining_dataset[0]["name"]
+            skip = cfg.pretraining_dataset[0]["skip"]
             if "split" in cfg.pretraining_dataset[0]:
                 split = cfg.pretraining_dataset[0]["split"]
 
+            data_files = cfg.pretraining_dataset[0].get("data_files")
+
         ds_wrapper_partial = functools.partial(
             get_dataset_wrapper,
             cfg.pretraining_dataset[0],
@@ -104,8 +109,14 @@ def prepare_dataset(cfg, tokenizer, processor=None):
             cfg.pretraining_dataset[0]["type"] or "pretrain",
         )
 
+        iter_ds = load_dataset(
+            path, streaming=True, split=split, name=name, data_files=data_files
+        )
+        if skip:
+            LOG.info(f"Skipping {skip} samples from the dataset")
+            iter_ds = iter_ds.skip(skip)
         train_dataset = wrap_pretraining_dataset(
-            load_dataset(path, streaming=True, split=split, name=name),
+            iter_ds,
             tokenizer,
             cfg,
             ds_wrapper_partial,

src/axolotl/utils/data/shared.py

@@ -107,6 +107,13 @@ def load_dataset_w_config(config_dataset, auth_token):
     except (FileNotFoundError, ConnectionError):
         pass
 
+    # gather extra args from the config
+    load_ds_kwargs = {}
+    if config_dataset.split:
+        load_ds_kwargs["split"] = config_dataset.split
+    else:
+        load_ds_kwargs["split"] = None
+
     # prefer local dataset, even if hub exists
     local_path = Path(config_dataset.path)
     if local_path.exists():
@@ -118,7 +125,7 @@ def load_dataset_w_config(config_dataset, auth_token):
                     name=config_dataset.name,
                     data_files=config_dataset.data_files,
                     streaming=False,
-                    split=None,
+                    **load_ds_kwargs,
                 )
             else:
                 try:
@@ -130,7 +137,7 @@ def load_dataset_w_config(config_dataset, auth_token):
                         config_dataset.path,
                         name=config_dataset.name,
                         streaming=False,
-                        split=None,
+                        **load_ds_kwargs,
                     )
         elif local_path.is_file():
             ds_type = get_ds_type(config_dataset)
@@ -140,16 +147,13 @@ def load_dataset_w_config(config_dataset, auth_token):
                 name=config_dataset.name,
                 data_files=config_dataset.path,
                 streaming=False,
-                split=None,
+                **load_ds_kwargs,
             )
         else:
             raise ValueError(
                 "unhandled dataset load: local path exists, but is neither a directory or a file"
             )
     elif ds_from_hub:
-        load_ds_kwargs = {}
-        if config_dataset.split:
-            load_ds_kwargs["split"] = config_dataset.split
         ds = load_dataset(
             config_dataset.path,
             name=config_dataset.name,
@@ -173,9 +177,9 @@ def load_dataset_w_config(config_dataset, auth_token):
                 name=config_dataset.name,
                 data_files=config_dataset.path,
                 streaming=False,
-                split=None,
                 storage_options=storage_options,
                 trust_remote_code=config_dataset.trust_remote_code,
+                **load_ds_kwargs,
             )
     elif config_dataset.path.startswith("https://"):
         ds_type = get_ds_type(config_dataset)
@@ -184,9 +188,9 @@ def load_dataset_w_config(config_dataset, auth_token):
             name=config_dataset.name,
             data_files=config_dataset.path,
             streaming=False,
-            split=None,
             storage_options=storage_options,
             trust_remote_code=config_dataset.trust_remote_code,
+            **load_ds_kwargs,
         )
     else:
         if isinstance(config_dataset.data_files, str):
@@ -214,7 +218,7 @@ def load_dataset_w_config(config_dataset, auth_token):
             name=config_dataset.name,
             data_files=fp,
             streaming=False,
-            split=None,
+            **load_ds_kwargs,
         )
     if not ds:
         raise ValueError("unhandled dataset load")

src/axolotl/utils/model_shard_quant.py

@@ -270,7 +270,7 @@ def load_sharded_model_quant(
     model.hf_quantizer = AutoHfQuantizer.from_config(quantization_config)
 
     if cfg.local_rank == 0 and verbose:
-        print(f"Loaded model weights in {time.time()-start:.3f} seconds")
+        print(f"Loaded model weights in {time.time() - start:.3f} seconds")
     # cleanup any extra memory usage from parallel loading
     torch.cuda.empty_cache()
 

src/axolotl/utils/models.py

@@ -380,23 +380,19 @@ class ModelLoader:
         plugin_manager = PluginManager.get_instance()
         plugin_manager.pre_model_load(self.cfg)
 
+        if self.cfg.adapter:
+            from axolotl.monkeypatch.transformers_fa_utils import (
+                patch_fa_peft_integration,
+            )
+
+            patch_fa_peft_integration()
+
         if self.cfg.gradient_checkpointing == "unsloth":
             transformers.modeling_utils.checkpoint = hf_grad_checkpoint_unsloth_wrapper
 
         if self.cfg.flash_attention:
             self.patch_attention()
 
-        if self.cfg.model_config_type == "llama":
-            from axolotl.monkeypatch.trainer_grad_accum import (
-                patch_flash_attention_forward,
-                patch_forward_for_ga,
-                patch_training_step_for_ga,
-            )
-
-            patch_flash_attention_forward()
-            patch_forward_for_ga()
-            patch_training_step_for_ga()
-
         if self.cfg.sample_packing and self.cfg.s2_attention:
             raise ValueError(
                 "Received `sample_packing=true` and `s2_attention=true`; however, \
@@ -1057,7 +1053,7 @@ class ModelLoader:
         )
         if (
             hasattr(self.model, "get_input_embeddings")
-            and self.model.get_input_embeddings().num_embeddings < embeddings_len
+            and self.model.get_input_embeddings().num_embeddings != embeddings_len
         ):
             resize_kwargs = {}
             if self.cfg.mean_resizing_embeddings is not None:

src/axolotl/utils/trainer.py

@@ -196,7 +196,7 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         if eval_dataset:
             eval_dataset = eval_dataset.remove_columns("attention_mask")
 
-    if cfg.model_config_type == "falcon":
+    if cfg.model_config_type in ["falcon", "mistral"]:
         LOG.info("dropping token_type_ids column if it exists")
         if "token_type_ids" in train_dataset.column_names:
             train_dataset = train_dataset.remove_columns("token_type_ids")
@@ -310,19 +310,22 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
 
 
 def process_pretraining_datasets_for_packing(
-    train_dataset, sequence_len, skip_position_ids=True
+    train_dataset, sequence_len, skip_position_ids=True, drop_attention_mask=False
 ):
     drop_long = partial(drop_long_seq, sequence_len=sequence_len)
 
     train_dataset = train_dataset.filter(
         drop_long,
         desc="Dropping Long Sequences",
+        load_from_cache_file=False,
     )
-    if skip_position_ids:
+    if not skip_position_ids:
         train_dataset = train_dataset.map(
             add_position_ids,
             desc="Add position_id column (Pretraining Sample Packing)",
         )
+    if drop_attention_mask:
+        train_dataset = train_dataset.remove_columns("attention_mask")
 
     return train_dataset
 

tests/cli/conftest.py

@@ -1,4 +1,5 @@
 """Shared pytest fixtures for cli module."""
+
 import pytest
 from click.testing import CliRunner
 

tests/cli/test_cli_fetch.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI fetch command."""
+
 from unittest.mock import patch
 
 from axolotl.cli.main import fetch

tests/cli/test_cli_inference.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI inference command."""
+
 from unittest.mock import patch
 
 from axolotl.cli.main import cli

tests/cli/test_cli_interface.py

@@ -1,4 +1,5 @@
 """General pytest tests for axolotl.cli.main interface."""
+
 from axolotl.cli.main import build_command, cli
 
 

tests/cli/test_cli_merge_lora.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI merge_lora command."""
+
 from unittest.mock import patch
 
 from axolotl.cli.main import cli

tests/cli/test_cli_merge_sharded_fsdp_weights.py

@@ -1,5 +1,6 @@
 """pytest tests for axolotl CLI merge_sharded_fsdp_weights command."""
 # pylint: disable=duplicate-code
+
 from unittest.mock import patch
 
 from axolotl.cli.main import cli
@@ -15,46 +16,3 @@ def test_merge_sharded_fsdp_weights_no_accelerate(cli_runner, config_path):
         assert mock.called
         assert mock.call_args.kwargs["config"] == str(config_path)
         assert result.exit_code == 0
-
-
-def test_merge_sharded_fsdp_weights_with_model_dir(cli_runner, config_path, tmp_path):
-    """Test merge_sharded_fsdp_weights command with model_dir option"""
-    model_dir = tmp_path / "model"
-    model_dir.mkdir()
-
-    with patch("axolotl.cli.merge_sharded_fsdp_weights.do_cli") as mock:
-        result = cli_runner.invoke(
-            cli,
-            [
-                "merge-sharded-fsdp-weights",
-                str(config_path),
-                "--no-accelerate",
-                "--model-dir",
-                str(model_dir),
-            ],
-        )
-
-        assert mock.called
-        assert mock.call_args.kwargs["config"] == str(config_path)
-        assert mock.call_args.kwargs["model_dir"] == str(model_dir)
-        assert result.exit_code == 0
-
-
-def test_merge_sharded_fsdp_weights_with_save_path(cli_runner, config_path):
-    """Test merge_sharded_fsdp_weights command with save_path option"""
-    with patch("axolotl.cli.merge_sharded_fsdp_weights.do_cli") as mock:
-        result = cli_runner.invoke(
-            cli,
-            [
-                "merge-sharded-fsdp-weights",
-                str(config_path),
-                "--no-accelerate",
-                "--save-path",
-                "/path/to/save",
-            ],
-        )
-
-        assert mock.called
-        assert mock.call_args.kwargs["config"] == str(config_path)
-        assert mock.call_args.kwargs["save_path"] == "/path/to/save"
-        assert result.exit_code == 0

tests/cli/test_cli_preprocess.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI preprocess command."""
+
 import shutil
 from pathlib import Path
 from unittest.mock import patch

tests/cli/test_cli_shard.py

@@ -1,76 +0,0 @@
-"""pytest tests for axolotl CLI shard command."""
-# pylint: disable=duplicate-code
-from unittest.mock import patch
-
-from axolotl.cli.main import cli
-
-
-def test_shard_with_accelerate(cli_runner, config_path):
-    """Test shard command with accelerate"""
-    with patch("subprocess.run") as mock:
-        result = cli_runner.invoke(cli, ["shard", str(config_path), "--accelerate"])
-
-        assert mock.called
-        assert mock.call_args.args[0] == [
-            "accelerate",
-            "launch",
-            "-m",
-            "axolotl.cli.shard",
-            str(config_path),
-            "--debug-num-examples",
-            "0",
-        ]
-        assert mock.call_args.kwargs == {"check": True}
-        assert result.exit_code == 0
-
-
-def test_shard_no_accelerate(cli_runner, config_path):
-    """Test shard command without accelerate"""
-    with patch("axolotl.cli.shard.do_cli") as mock:
-        result = cli_runner.invoke(cli, ["shard", str(config_path), "--no-accelerate"])
-
-        assert mock.called
-        assert result.exit_code == 0
-
-
-def test_shard_with_model_dir(cli_runner, config_path, tmp_path):
-    """Test shard command with model_dir option"""
-    model_dir = tmp_path / "model"
-    model_dir.mkdir()
-
-    with patch("axolotl.cli.shard.do_cli") as mock:
-        result = cli_runner.invoke(
-            cli,
-            [
-                "shard",
-                str(config_path),
-                "--no-accelerate",
-                "--model-dir",
-                str(model_dir),
-            ],
-            catch_exceptions=False,
-        )
-
-        assert mock.called
-        assert mock.call_args.kwargs["config"] == str(config_path)
-        assert mock.call_args.kwargs["model_dir"] == str(model_dir)
-        assert result.exit_code == 0
-
-
-def test_shard_with_save_dir(cli_runner, config_path):
-    with patch("axolotl.cli.shard.do_cli") as mock:
-        result = cli_runner.invoke(
-            cli,
-            [
-                "shard",
-                str(config_path),
-                "--no-accelerate",
-                "--save-dir",
-                "/path/to/save",
-            ],
-        )
-
-        assert mock.called
-        assert mock.call_args.kwargs["config"] == str(config_path)
-        assert mock.call_args.kwargs["save_dir"] == "/path/to/save"
-        assert result.exit_code == 0

tests/cli/test_cli_version.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI --version"""
+
 from axolotl.cli.main import cli
 
 

tests/cli/test_utils.py

@@ -1,5 +1,6 @@
 """pytest tests for axolotl CLI utils."""
 # pylint: disable=redefined-outer-name
+
 import json
 from unittest.mock import Mock, patch
 

tests/conftest.py

@@ -120,13 +120,12 @@ def temp_dir():
 @pytest.fixture(scope="function", autouse=True)
 def cleanup_monkeypatches():
     from transformers import Trainer
-    from transformers.models.llama.modeling_llama import (
+    from transformers.models.llama.modeling_llama import (  # LlamaFlashAttention2,
         LlamaAttention,
-        LlamaFlashAttention2,
         LlamaForCausalLM,
     )
 
-    original_fa2_forward = LlamaFlashAttention2.forward
+    # original_fa2_forward = LlamaFlashAttention2.forward
     original_llama_attn_forward = LlamaAttention.forward
     original_llama_forward = LlamaForCausalLM.forward
     original_trainer_inner_training_loop = (
@@ -136,7 +135,7 @@ def cleanup_monkeypatches():
     # monkey patches can happen inside the tests
     yield
     # Reset LlamaFlashAttention2 forward
-    LlamaFlashAttention2.forward = original_fa2_forward
+    # LlamaFlashAttention2.forward = original_fa2_forward
     LlamaAttention.forward = original_llama_attn_forward
     LlamaForCausalLM.forward = original_llama_forward
     Trainer._inner_training_loop = (  # pylint: disable=protected-access
@@ -149,7 +148,10 @@ def cleanup_monkeypatches():
         ("transformers.models.llama",),
         (
             "transformers.models.llama.modeling_llama",
-            ["LlamaFlashAttention2", "LlamaAttention"],
+            [
+                # "LlamaFlashAttention2",
+                "LlamaAttention",
+            ],
         ),
         ("transformers.trainer",),
         ("transformers", ["Trainer"]),

tests/e2e/integrations/test_cut_cross_entropy.py

@@ -2,17 +2,17 @@
 Simple end-to-end test for Cut Cross Entropy integration
 """
 
-from pathlib import Path
-
 import pytest
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils import get_pytorch_version
 from axolotl.utils.config import normalize_config, prepare_plugins
 from axolotl.utils.dict import DictDefault
 
+from ..utils import check_model_output_exists
+
 # pylint: disable=duplicate-code
 
 
@@ -64,10 +64,10 @@ class TestCutCrossEntropyIntegration:
         major, minor, _ = get_pytorch_version()
         if (major, minor) < (2, 4):
             with pytest.raises(ImportError):
-                train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+                train(cfg=cfg, dataset_meta=dataset_meta)
         else:
-            train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-            assert (Path(temp_dir) / "model.safetensors").exists()
+            train(cfg=cfg, dataset_meta=dataset_meta)
+            check_model_output_exists(temp_dir, cfg)
 
     @pytest.mark.parametrize(
         "attention_type",
@@ -92,7 +92,7 @@ class TestCutCrossEntropyIntegration:
         major, minor, _ = get_pytorch_version()
         if (major, minor) < (2, 4):
             with pytest.raises(ImportError):
-                train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+                train(cfg=cfg, dataset_meta=dataset_meta)
         else:
-            train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-            assert (Path(temp_dir) / "model.safetensors").exists()
+            train(cfg=cfg, dataset_meta=dataset_meta)
+            check_model_output_exists(temp_dir, cfg)

tests/e2e/integrations/test_liger.py

@@ -1,43 +1,41 @@
 """
 Simple end-to-end test for Liger integration
 """
-import unittest
-from pathlib import Path
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from e2e.utils import require_torch_2_4_1
+
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config, prepare_plugins
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists
 
 
-class LigerIntegrationTestCase(unittest.TestCase):
+class LigerIntegrationTestCase:
     """
     e2e tests for liger integration with Axolotl
     """
 
-    @with_temp_dir
+    @require_torch_2_4_1
     def test_llama_wo_flce(self, temp_dir):
+        # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "plugins": [
                     "axolotl.integrations.liger.LigerPlugin",
                 ],
                 "liger_rope": True,
                 "liger_rms_norm": True,
-                "liger_swiglu": True,
+                "liger_glu_activation": True,
                 "liger_cross_entropy": True,
                 "liger_fused_linear_cross_entropy": False,
                 "sequence_len": 1024,
-                "val_set_size": 0.1,
+                "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -46,15 +44,15 @@ class LigerIntegrationTestCase(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "micro_batch_size": 8,
-                "gradient_accumulation_steps": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",
                 "save_safetensors": True,
                 "bf16": "auto",
-                "max_steps": 10,
+                "max_steps": 5,
             }
         )
         prepare_plugins(cfg)
@@ -62,29 +60,27 @@ class LigerIntegrationTestCase(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "model.safetensors").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
-    @with_temp_dir
+    @require_torch_2_4_1
     def test_llama_w_flce(self, temp_dir):
+        # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "plugins": [
                     "axolotl.integrations.liger.LigerPlugin",
                 ],
                 "liger_rope": True,
                 "liger_rms_norm": True,
-                "liger_swiglu": True,
+                "liger_glu_activation": True,
                 "liger_cross_entropy": False,
                 "liger_fused_linear_cross_entropy": True,
                 "sequence_len": 1024,
-                "val_set_size": 0.1,
+                "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -93,15 +89,15 @@ class LigerIntegrationTestCase(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "micro_batch_size": 8,
-                "gradient_accumulation_steps": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",
                 "save_safetensors": True,
                 "bf16": "auto",
-                "max_steps": 10,
+                "max_steps": 5,
             }
         )
         prepare_plugins(cfg)
@@ -109,5 +105,5 @@ class LigerIntegrationTestCase(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "model.safetensors").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/multigpu/test_llama.py

@@ -63,6 +63,7 @@ class TestMultiGPULlama:
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
                 "use_tensorboard": True,
+                "bf16": True,
             }
         )
 
@@ -127,6 +128,7 @@ class TestMultiGPULlama:
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
                 "use_tensorboard": True,
+                "bf16": True,
             }
         )
 
@@ -201,6 +203,7 @@ class TestMultiGPULlama:
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
                 "use_tensorboard": True,
+                "bf16": True,
             }
         )
 
@@ -223,8 +226,12 @@ class TestMultiGPULlama:
             ]
         )
 
+        loss_threshold = 2.3
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs",
+            "train/train_loss",
+            loss_threshold,
+            "Train Loss is too high",
         )
 
     def test_dpo_qlora_ddp(self, temp_dir):
@@ -275,6 +282,7 @@ class TestMultiGPULlama:
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
                 "use_tensorboard": True,
+                "bf16": True,
             }
         )
 
@@ -297,8 +305,12 @@ class TestMultiGPULlama:
             ]
         )
 
+        loss_threshold = 2.3
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs",
+            "train/train_loss",
+            loss_threshold,
+            "Train Loss is too high",
         )
 
     @pytest.mark.parametrize(

tests/e2e/patched/test_4d_multipack_llama.py

@@ -5,15 +5,14 @@ E2E tests for multipack fft llama using 4d attention masks
 import logging
 import os
 import unittest
-from pathlib import Path
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import require_torch_2_3_1, with_temp_dir
+from ..utils import check_model_output_exists, require_torch_2_3_1, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -66,8 +65,8 @@ class Test4dMultipackLlama(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
     @with_temp_dir
     def test_torch_lora_packing(self, temp_dir):
@@ -110,5 +109,5 @@ class Test4dMultipackLlama(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_cli_integrations.py

@@ -5,7 +5,7 @@ from pathlib import Path
 
 import yaml
 
-from axolotl.cli import load_cfg
+from axolotl.cli.config import load_cfg
 from axolotl.utils.dict import DictDefault
 
 

tests/e2e/patched/test_fa_xentropy.py

@@ -4,18 +4,17 @@ E2E tests for lora llama
 
 import logging
 import os
-from pathlib import Path
 
 import pytest
 from transformers.utils import is_torch_bf16_gpu_available
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_tensorboard
+from ..utils import check_model_output_exists, check_tensorboard
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -81,8 +80,8 @@ class TestFAXentropyLlama:
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
         check_tensorboard(
             temp_dir + "/runs", "train/train_loss", 1.5, "Train Loss is too high"

tests/e2e/patched/test_falcon_samplepack.py

@@ -5,15 +5,14 @@ E2E tests for falcon
 import logging
 import os
 import unittest
-from pathlib import Path
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -68,8 +67,8 @@ class TestFalconPatched(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
     @with_temp_dir
     def test_ft(self, temp_dir):
@@ -108,5 +107,5 @@ class TestFalconPatched(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "pytorch_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_fused_llama.py

@@ -5,18 +5,17 @@ E2E tests for lora llama
 import logging
 import os
 import unittest
-from pathlib import Path
 
 import pytest
 from transformers.utils import is_torch_bf16_gpu_available
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -72,5 +71,5 @@ class TestFusedLlama(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "pytorch_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_llama_s2_attention.py

@@ -5,17 +5,16 @@ E2E tests for llama w/ S2 attn
 import logging
 import os
 import unittest
-from pathlib import Path
 
 import pytest
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -70,8 +69,8 @@ class TestLlamaShiftedSparseAttention(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
     @with_temp_dir
     def test_fft_s2_attn(self, temp_dir):
@@ -110,5 +109,5 @@ class TestLlamaShiftedSparseAttention(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "pytorch_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_lora_llama_multipack.py

@@ -5,18 +5,17 @@ E2E tests for lora llama
 import logging
 import os
 import unittest
-from pathlib import Path
 
 import pytest
 from transformers.utils import is_auto_gptq_available, is_torch_bf16_gpu_available
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -75,8 +74,8 @@ class TestLoraLlama(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
     @pytest.mark.skipif(not is_auto_gptq_available(), reason="auto-gptq not available")
     @with_temp_dir
@@ -125,5 +124,5 @@ class TestLoraLlama(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_mistral_samplepack.py

@@ -5,15 +5,14 @@ E2E tests for lora llama
 import logging
 import os
 import unittest
-from pathlib import Path
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -68,8 +67,8 @@ class TestMistral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
     @with_temp_dir
     def test_ft_packing(self, temp_dir):
@@ -109,5 +108,5 @@ class TestMistral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "pytorch_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_mixtral_samplepack.py

@@ -5,15 +5,14 @@ E2E tests for mixtral
 import logging
 import os
 import unittest
-from pathlib import Path
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -65,8 +64,8 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
     @with_temp_dir
     def test_ft(self, temp_dir):
@@ -103,9 +102,5 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        model, _ = train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (
-            "MixtralFlashAttention2"
-            in model.model.layers[0].self_attn.__class__.__name__
-        )
-        assert (Path(temp_dir) / "pytorch_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_model_patches.py

@@ -6,7 +6,6 @@ import unittest
 
 import transformers
 
-from axolotl.common.cli import TrainerCliArgs
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
@@ -49,14 +48,8 @@ class TestModelPatches(unittest.TestCase):
             }
         )
         normalize_config(cfg)
-        cli_args = TrainerCliArgs()
         tokenizer = load_tokenizer(cfg)
-        model, _ = load_model(cfg, tokenizer, inference=cli_args.inference)
-
-        assert (
-            "MixtralFlashAttention2"
-            in model.model.layers[0].self_attn.__class__.__name__
-        )
+        load_model(cfg, tokenizer, inference=False)
 
     @with_temp_dir
     def test_mistral_multipack(self, temp_dir):
@@ -87,9 +80,8 @@ class TestModelPatches(unittest.TestCase):
             }
         )
         normalize_config(cfg)
-        cli_args = TrainerCliArgs()
         tokenizer = load_tokenizer(cfg)
-        load_model(cfg, tokenizer, inference=cli_args.inference)
+        load_model(cfg, tokenizer, inference=False)
 
         assert (
             "torch.jit"

tests/e2e/patched/test_phi_multipack.py

@@ -5,15 +5,14 @@ E2E tests for lora llama
 import logging
 import os
 import unittest
-from pathlib import Path
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_model_output_exists, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -68,8 +67,8 @@ class TestPhiMultipack(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "pytorch_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
     @with_temp_dir
     def test_qlora_packed(self, temp_dir):
@@ -119,5 +118,5 @@ class TestPhiMultipack(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_resume.py

@@ -6,17 +6,16 @@ import logging
 import os
 import re
 import subprocess
-from pathlib import Path
 
 from transformers.utils import is_torch_bf16_gpu_available
 
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import most_recent_subdir
+from ..utils import check_model_output_exists, most_recent_subdir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -72,7 +71,7 @@ class TestResumeLlama:
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        train(cfg=cfg, dataset_meta=dataset_meta)
 
         resume_cfg = cfg | DictDefault(
             {
@@ -82,8 +81,8 @@ class TestResumeLlama:
         normalize_config(resume_cfg)
         cli_args = TrainerCliArgs()
 
-        train(cfg=resume_cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()
+        train(cfg=resume_cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
 
         tb_log_path_1 = most_recent_subdir(temp_dir + "/runs")
         cmd = f"tensorboard --inspect  --logdir {tb_log_path_1}"

tests/e2e/patched/test_unsloth_integration.py

@@ -1,13 +1,18 @@
 """Test module for checking whether the integration of Unsloth with Hugging Face Transformers is working as expected."""
 import unittest
 
-from axolotl.monkeypatch.unsloth_ import check_self_attn_is_patchable
+import pytest
 
 
+@pytest.mark.skip(
+    reason="Unsloth integration will be broken going into latest transformers"
+)
 class TestUnslothIntegration(unittest.TestCase):
     """Unsloth monkeypatch integration tests."""
 
     def test_is_self_attn_patchable(self):
+        from axolotl.monkeypatch.unsloth_ import check_self_attn_is_patchable
+
         # ensures the current version of transformers has loss code that matches our patching code
         self.assertTrue(
             check_self_attn_is_patchable(),