Multimodal Vision Llama - rudimentary support (#1940)

---------

Co-authored-by: Sunny <[email protected]>
Co-authored-by: sunny <[email protected]>

docs/input_output.qmd

@@ -205,7 +205,7 @@ ds = load_from_disk(f'last_run_prepared/{directory[0]}/')
     hi there!.  goodbye  farewell</s>
 ```
 
-We can check that the right tokens are ingored by comparing the labels
+We can check that the right tokens are ignored by comparing the labels
 to each token:
 
 ```python

docs/multimodal.qmd

@@ -0,0 +1,28 @@
+# MultiModal / Vision Language Models (BETA)
+
+### Supported Models
+
+- Mllama, i.e. llama with vision models
+
+### Usage
+
+Currently multimodal support is limited and doesn't have full feature parity. To finetune a multimodal Llama w/ LoRA,
+you'll need to use the following in YAML in combination with the rest of the required hyperparams.
+
+```yaml
+base_model: alpindale/Llama-3.2-11B-Vision-Instruct
+processor_type: AutoProcessor
+skip_prepare_dataset: true
+
+chat_template: llama3_2_vision
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+remove_unused_columns: false
+sample_packing: false
+
+# only finetune the Language model, leave the vision model and vision tower frozen
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+```

examples/llama-3-vision/lora-11b.yaml

@@ -0,0 +1,63 @@
+base_model: alpindale/Llama-3.2-11B-Vision-Instruct
+processor_type: AutoProcessor
+strict: false
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+chat_template: llama3_2_vision
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 8192
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+local_rank:
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:

src/axolotl/cli/__init__.py

@@ -40,7 +40,7 @@
 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_tokenizer
+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
@@ -430,9 +430,12 @@ def load_datasets(
     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
+        cfg,
+        tokenizer,
+        processor=processor,
     )
 
     if cli_args.debug or cfg.debug:

src/axolotl/core/trainer_builder.py

@@ -61,12 +61,14 @@
     log_prediction_callback_factory,
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
     MambaDataCollator,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
+from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
 from axolotl.utils.models import ensure_dtype
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 from axolotl.utils.schedulers import (
@@ -250,6 +252,10 @@ class AxolotlTrainingMixins:
             "help": "workaround to pass an alternate lr scheduler to the HF trainer"
         },
     )
+    chat_template: Optional[str] = field(
+        default=None,
+        metadata={"help": "Chat template converting chat messages to text"},
+    )
 
 
 @dataclass
@@ -1043,10 +1049,11 @@ class TrainerBuilderBase(abc.ABC):
     _model_ref = None
     _peft_config = None
 
-    def __init__(self, cfg, model, tokenizer):
+    def __init__(self, cfg, model, tokenizer, processor=None):
         self.cfg = cfg
         self.model = model
         self.tokenizer = tokenizer
+        self.processor = processor
 
         # in case the model supports tagging, add the axolotl tag.
         # This makes sure the tag is correctly pushed even if a user calls
@@ -1515,6 +1522,10 @@ def build(self, total_num_steps):
         )
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
+        if self.cfg.chat_template:
+            training_arguments_kwargs["chat_template"] = chat_templates(
+                self.cfg.chat_template
+            )
 
         if self.cfg.rl == "orpo":
             training_arguments_kwargs["orpo_alpha"] = self.cfg.orpo_alpha
@@ -1661,7 +1672,12 @@ def build_collator(
             else:
                 collator = BatchSamplerDataCollatorForSeq2Seq
         else:
-            collator = DataCollatorForSeq2Seq
+            if self.cfg.processor_type and self.processor:
+                collator = MultiModalChatDataCollator
+                kwargs["processor"] = self.processor
+                kwargs["chat_template"] = training_args.chat_template
+            else:
+                collator = DataCollatorForSeq2Seq
 
         return collator(
             self.tokenizer,