fix(vlm): handle legacy conversation data format and check image in data (#2018) [skip ci]

* fix: handle legacy conversation data format and check image in data

* feat: add test for llama vision

* feat: add max_steps to test

* fix: incorrect indent and return preprocess

* feat: use smaller model and dataset

* chore: add extra config for sharegpt dataset

tests/e2e/test_llama_vision.py

@@ -0,0 +1,116 @@
+"""
+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.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestLlamaVision(unittest.TestCase):
+    """
+    Test case for Llama Vision models
+    """
+
+    @with_temp_dir
+    def test_lora_llama_vision_text_only_dataset(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "axolotl-ai-co/Llama-3.2-39M-Vision",
+                "processor_type": "AutoProcessor",
+                "skip_prepare_dataset": True,
+                "remove_unused_columns": False,
+                "sample_packing": False,
+                "sequence_len": 1024,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_modules": r"language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj",
+                "val_set_size": 0,
+                "chat_template": "llama3_2_vision",
+                "datasets": [
+                    {
+                        "path": "LDJnr/Puffin",
+                        "type": "chat_template",
+                        "field_messages": "conversations",
+                        "message_field_role": "from",
+                        "message_field_content": "value",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "save_safetensors": True,
+                "bf16": True,
+            }
+        )
+        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.safetensors").exists()
+
+    @with_temp_dir
+    def test_lora_llama_vision_multimodal_dataset(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "axolotl-ai-co/Llama-3.2-39M-Vision",
+                "processor_type": "AutoProcessor",
+                "skip_prepare_dataset": True,
+                "remove_unused_columns": False,
+                "sample_packing": False,
+                "sequence_len": 1024,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_modules": r"language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj",
+                "val_set_size": 0,
+                "chat_template": "llama3_2_vision",
+                "datasets": [
+                    {
+                        "path": "axolotl-ai-co/llava-instruct-mix-vsft-small",
+                        "type": "chat_template",
+                        "split": "train",
+                        "field_messages": "messages",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "save_safetensors": True,
+                "bf16": True,
+            }
+        )
+        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.safetensors").exists()

tests/e2e/test_lora_llama.py

@@ -57,6 +57,7 @@ class TestLoraLlama(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",
+                "max_steps": 20,
             }
         )
         normalize_config(cfg)

tests/e2e/test_optimizers.py

@@ -56,6 +56,7 @@ class TestCustomOptimizers(unittest.TestCase):
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "optimi_adamw",
+                "max_steps": 5,
                 "lr_scheduler": "cosine",
             }
         )

tests/e2e/test_relora_llama.py

@@ -58,6 +58,7 @@ class TestReLoraLlama(unittest.TestCase):
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
+                "max_steps": 5,
                 "lr_scheduler": "cosine",
             }
         )