Liger Kernel integration (#1861)

* add initial plugin support w Liger kernel patches

* integrate the input args classes

* fix liger plugin and dynamic configuration class

* drop untrainable samples and refactor config plugins integration

* fix incorrect inputs and circular imports

* fix bool comparison

* fix for dropping untraibable tokens

* fix licensing so liger integration is Apache 2.0

* add jamba support

* pylint ignore

.mypy.ini

@@ -11,6 +11,9 @@ ignore_errors = True
 [mypy-axolotl.models.mixtral.*]
 ignore_errors = True
 
+[mypy-axolotl.integrations.liger.models.*]
+ignore_errors = True
+
 [mypy-axolotl.models.phi.*]
 ignore_errors = True
 

requirements.txt

@@ -33,6 +33,8 @@ gradio==3.50.2
 tensorboard
 python-dotenv==1.0.1
 autoawq>=0.2.5
+triton>=2.3.0
+liger-kernel
 
 mamba-ssm==1.2.0.post1
 

src/axolotl/cli/__init__.py

@@ -27,6 +27,7 @@
 from transformers.utils.import_utils import _is_package_available
 
 from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
+from axolotl.integrations.base import PluginManager
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
 from axolotl.utils.config import (
@@ -365,6 +366,11 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs):
 
     cfg.axolotl_config_path = config
 
+    if cfg.get("plugins"):
+        plugin_manager = PluginManager.get_instance()
+        for plugin_name in cfg["plugins"]:
+            plugin_manager.register(plugin_name)
+
     try:
         device_props = torch.cuda.get_device_properties("cuda")
         gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)