fix the torch dtype check

* deprecated wandb.save

* also use wandb.save for axolotl yaml

* chore: lint

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>

examples/qwen/qwen2-moe-qlora.yaml

@@ -16,7 +16,7 @@ sequence_len: 1024  # supports up to 32k
 sample_packing: false
 pad_to_sequence_len: false
 
-adapter: lora
+adapter: qlora
 lora_model_dir:
 lora_r: 32
 lora_alpha: 16

src/axolotl/cli/__init__.py

@@ -24,6 +24,7 @@ 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
@@ -62,6 +63,20 @@ def print_axolotl_text_art(suffix=None):
     if is_main_process():
         print(ascii_art)
 
+    print_dep_versions()
+
+
+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:
+            version = _is_package_available(pkg, return_version=True)
+            print(f"{pkg: >{max_len}}: {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

src/axolotl/cli/merge_lora.py

@@ -8,6 +8,7 @@ import transformers
 
 from axolotl.cli import do_merge_lora, load_cfg, print_axolotl_text_art
 from axolotl.common.cli import TrainerCliArgs
+from axolotl.utils.dict import DictDefault
 
 
 def do_cli(config: Path = Path("examples/"), **kwargs):
@@ -27,21 +28,26 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         flash_attention=False,
         **kwargs,
     )
+    cfg = modify_cfg_for_merge(parsed_cfg)
 
-    if not parsed_cfg.lora_model_dir and parsed_cfg.output_dir:
-        parsed_cfg.lora_model_dir = parsed_cfg.output_dir
-    if not Path(parsed_cfg.lora_model_dir).exists():
+    do_merge_lora(cfg=cfg, cli_args=parsed_cli_args)
+
+
+def modify_cfg_for_merge(cfg: DictDefault) -> DictDefault:
+    if not cfg.lora_model_dir and cfg.output_dir:
+        cfg.lora_model_dir = cfg.output_dir
+    if not Path(cfg.lora_model_dir).exists():
         raise ValueError(
-            f"Target directory for merge: `{parsed_cfg.lora_model_dir}` does not exist."
+            f"Target directory for merge: `{cfg.lora_model_dir}` does not exist."
         )
 
-    parsed_cfg.load_in_4bit = False
-    parsed_cfg.load_in_8bit = False
-    parsed_cfg.flash_attention = False
-    parsed_cfg.deepspeed = None
-    parsed_cfg.fsdp = None
+    cfg.load_in_4bit = False
+    cfg.load_in_8bit = False
+    cfg.flash_attention = False
+    cfg.deepspeed = None
+    cfg.fsdp = None
 
-    do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    return cfg
 
 
 if __name__ == "__main__":

src/axolotl/core/trainer_builder.py

@@ -36,6 +36,7 @@ from trl.trainer.utils import pad_to_length
 from axolotl.loraplus import create_loraplus_optimizer
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
+from axolotl.utils import is_mlflow_available
 from axolotl.utils.callbacks import (
     EvalFirstStepCallback,
     GPUStatsCallback,
@@ -71,10 +72,6 @@ except ImportError:
 LOG = logging.getLogger("axolotl.core.trainer_builder")
 
 
-def is_mlflow_available():
-    return importlib.util.find_spec("mlflow") is not None
-
-
 def _sanitize_kwargs_for_tagging(tag_names, kwargs=None):
     if isinstance(tag_names, str):
         tag_names = [tag_names]
@@ -943,7 +940,16 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         callbacks = []
         if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
             LogPredictionCallback = log_prediction_callback_factory(
-                trainer, self.tokenizer
+                trainer, self.tokenizer, "wandb"
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+        if (
+            self.cfg.use_mlflow
+            and is_mlflow_available()
+            and self.cfg.eval_table_size > 0
+        ):
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer, "mlflow"
             )
             callbacks.append(LogPredictionCallback(self.cfg))
 

src/axolotl/utils/__init__.py

@@ -0,0 +1,8 @@
+"""
+Basic utils for Axolotl
+"""
+import importlib
+
+
+def is_mlflow_available():
+    return importlib.util.find_spec("mlflow") is not None

src/axolotl/utils/callbacks/__init__.py

@@ -6,7 +6,7 @@ import logging
 import os
 from shutil import copyfile
 from tempfile import NamedTemporaryFile
-from typing import TYPE_CHECKING, Dict, List
+from typing import TYPE_CHECKING, Any, Dict, List
 
 import evaluate
 import numpy as np
@@ -27,7 +27,9 @@ from transformers import (
 )
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, IntervalStrategy
 
+from axolotl.utils import is_mlflow_available
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.config.models.input.v0_4_1 import AxolotlInputConfig
 from axolotl.utils.distributed import (
     barrier,
     broadcast_dict,
@@ -540,7 +542,7 @@ def causal_lm_bench_eval_callback_factory(trainer: Trainer, tokenizer):
     return CausalLMBenchEvalCallback
 
 
-def log_prediction_callback_factory(trainer: Trainer, tokenizer):
+def log_prediction_callback_factory(trainer: Trainer, tokenizer, logger: str):
     class LogPredictionCallback(TrainerCallback):
         """Callback to log prediction values during each evaluation"""
 
@@ -597,15 +599,13 @@ def log_prediction_callback_factory(trainer: Trainer, tokenizer):
                 return ranges
 
             def log_table_from_dataloader(name: str, table_dataloader):
-                table = wandb.Table(  # type: ignore[attr-defined]
-                    columns=[
-                        "id",
-                        "Prompt",
-                        "Correct Completion",
-                        "Predicted Completion (model.generate)",
-                        "Predicted Completion (trainer.prediction_step)",
-                    ]
-                )
+                table_data: Dict[str, List[Any]] = {
+                    "id": [],
+                    "Prompt": [],
+                    "Correct Completion": [],
+                    "Predicted Completion (model.generate)": [],
+                    "Predicted Completion (trainer.prediction_step)": [],
+                }
                 row_index = 0
 
                 for batch in tqdm(table_dataloader):
@@ -709,16 +709,29 @@ def log_prediction_callback_factory(trainer: Trainer, tokenizer):
                     ) in zip(
                         prompt_texts, completion_texts, predicted_texts, pred_step_texts
                     ):
-                        table.add_data(
-                            row_index,
-                            prompt_text,
-                            completion_text,
-                            prediction_text,
-                            pred_step_text,
+                        table_data["id"].append(row_index)
+                        table_data["Prompt"].append(prompt_text)
+                        table_data["Correct Completion"].append(completion_text)
+                        table_data["Predicted Completion (model.generate)"].append(
+                            prediction_text
                         )
+                        table_data[
+                            "Predicted Completion (trainer.prediction_step)"
+                        ].append(pred_step_text)
                         row_index += 1
+                if logger == "wandb":
+                    wandb.run.log({f"{name} - Predictions vs Ground Truth": pd.DataFrame(table_data)})  # type: ignore[attr-defined]
+                elif logger == "mlflow" and is_mlflow_available():
+                    import mlflow
 
-                wandb.run.log({f"{name} - Predictions vs Ground Truth": table})  # type: ignore[attr-defined]
+                    tracking_uri = AxolotlInputConfig(
+                        **self.cfg.to_dict()
+                    ).mlflow_tracking_uri
+                    mlflow.log_table(
+                        data=table_data,
+                        artifact_file="PredictionsVsGroundTruth.json",
+                        tracking_uri=tracking_uri,
+                    )
 
             if is_main_process():
                 log_table_from_dataloader("Eval", eval_dataloader)
@@ -748,6 +761,11 @@ class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
                     mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
                 ) as temp_file:
                     copyfile(self.axolotl_config_path, temp_file.name)
+                    artifact = wandb.Artifact(
+                        f"config-{wandb.run.id}", type="axolotl-config"
+                    )
+                    artifact.add_file(temp_file.name)
+                    wandb.log_artifact(artifact)
                     wandb.save(temp_file.name)
                 LOG.info(
                     "The Axolotl config has been saved to the WandB run under files."

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

@@ -98,6 +98,7 @@ class SFTDataset(BaseModel):
     ds_type: Optional[str] = None
     train_on_split: Optional[str] = None
 
+    field: Optional[str] = None
     field_human: Optional[str] = None
     field_model: Optional[str] = None
 

src/axolotl/utils/data/sft.py

@@ -379,14 +379,15 @@ def load_tokenized_prepared_datasets(
                 d_base_type = d_type_split[0]
                 d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
 
-            if config_dataset.split and config_dataset.split in ds:
-                ds = ds[config_dataset.split]
-            elif split in ds:
-                ds = ds[split]
-            elif isinstance(ds, DatasetDict):
-                raise ValueError(
-                    f"no {split} split found for dataset {config_dataset.path}, you may specify a split with 'split: `"
-                )
+            if isinstance(ds, DatasetDict):
+                if config_dataset.split and config_dataset.split in ds:
+                    ds = ds[config_dataset.split]
+                elif split in ds:
+                    ds = ds[split]
+                else:
+                    raise ValueError(
+                        f"no {split} split found for dataset {config_dataset.path}, you may specify a split with 'split: `"
+                    )
 
             # support for using a subset of the data
             if config_dataset.shards:

src/axolotl/utils/trainer.py

@@ -198,7 +198,7 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
             .apply(lambda x: len(x))  # pylint: disable=unnecessary-lambda
             .values
         )
-        LOG.debug(f"total_num_tokens: {total_num_tokens}", main_process_only=True)
+        LOG.debug(f"total_num_tokens: {total_num_tokens:_}", main_process_only=True)
         if update:
             cfg.total_num_tokens = total_num_tokens
 
@@ -212,7 +212,7 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
             .sum()
         )
         LOG.debug(
-            f"`total_supervised_tokens: {total_supervised_tokens}`",
+            f"`total_supervised_tokens: {total_supervised_tokens:_}`",
             main_process_only=True,
         )
         if update:
@@ -239,7 +239,7 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
                 * cfg.num_epochs
             )
             LOG.debug(
-                f"total_num_tokens: {cfg.total_num_tokens}, total_num_steps: {total_num_steps}",
+                f"total_num_tokens: {cfg.total_num_tokens:_}, total_num_steps: {total_num_steps:_}",
                 main_process_only=True,
             )
         else:

tests/e2e/patched/test_llama_s2_attention.py

@@ -7,8 +7,6 @@ import os
 import unittest
 from pathlib import Path
 
-import pytest
-
 from axolotl.cli import load_datasets
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.train import train
@@ -21,7 +19,6 @@ LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-@pytest.mark.skip("Skipping test due to timeout.")
 class TestLlamaShiftedSparseAttention(unittest.TestCase):
     """
     Test case for Llama models using S2 Attn

tests/e2e/test_lora_llama.py

@@ -1,13 +1,16 @@
 """
 E2E tests for lora llama
 """
-
+import json
 import logging
 import os
 import unittest
 from pathlib import Path
 
-from axolotl.cli import load_datasets
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.cli import do_merge_lora, load_datasets
+from axolotl.cli.merge_lora import modify_cfg_for_merge
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.train import train
 from axolotl.utils.config import normalize_config
@@ -39,11 +42,6 @@ class TestLoraLlama(unittest.TestCase):
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "val_set_size": 0.1,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
                 "datasets": [
                     {
                         "path": "mhenrichsen/alpaca_2k_test",
@@ -57,6 +55,7 @@ class TestLoraLlama(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",
+                "max_steps": 10,
             }
         )
         normalize_config(cfg)
@@ -65,3 +64,67 @@ class TestLoraLlama(unittest.TestCase):
 
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+    @with_temp_dir
+    def test_lora_merge(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 32,
+                "lora_alpha": 64,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "max_steps": 10,
+                "bf16": "auto",
+            }
+        )
+        normalize_config(cfg)
+        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()
+
+        cfg.lora_model_dir = cfg.output_dir
+        cfg.load_in_4bit = False
+        cfg.load_in_8bit = False
+        cfg.flash_attention = False
+        cfg.deepspeed = None
+        cfg.fsdp = None
+
+        cfg = modify_cfg_for_merge(cfg)
+        cfg.merge_lora = True
+
+        cli_args = TrainerCliArgs(merge_lora=True)
+
+        do_merge_lora(cfg=cfg, cli_args=cli_args)
+        assert (Path(temp_dir) / "merged/pytorch_model.bin").exists()
+
+        with open(
+            Path(temp_dir) / "merged/config.json", "r", encoding="utf-8"
+        ) as f_handle:
+            config = f_handle.read()
+        config = json.loads(config)
+        if is_torch_bf16_gpu_available():
+            assert config["torch_dtype"] == "bfloat16"
+        else:
+            assert config["torch_dtype"] == "float16"