WIP Multimodal support

docs/config.qmd

@@ -31,7 +31,17 @@ tokenizer_legacy:
 # This is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
 
+# Optional processor configuration path in case you want to use a different processor
+# than the one defined in the base model
+processor_config:
+# Corresponding processor for the model AutoProcessor is a good choice
+processor_type: AutoProcessor
+
+
 # (Internal use only)
+# To identify whether the model is multimodal
+is_multimodal:
+
 # Used to identify which the model is based on
 is_falcon_derived_model:
 is_llama_derived_model:

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

examples/mllama/ft_test.yaml

@@ -0,0 +1,64 @@
+base_model: meta-llama/Llama-3.2-11B-Vision
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: llava
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: false
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_modules_to_save:
+  - embed_tokens
+  - lm_head
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: false
+s2_attention: false
+sdp_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+   pad_token: <|end_of_text|>

src/axolotl/cli/__init__.py

@@ -39,7 +39,7 @@
 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
@@ -407,9 +407,13 @@ def load_datasets(
     cli_args: TrainerCliArgs,
 ) -> TrainDatasetMeta:
     tokenizer = load_tokenizer(cfg)
+    tokenizer_processor = tokenizer
+    if cfg.is_multimodal:
+        processor = load_processor(cfg, tokenizer)
+        tokenizer_processor = processor
 
     train_dataset, eval_dataset, total_num_steps, prompters = prepare_dataset(
-        cfg, tokenizer
+        cfg, tokenizer_processor
     )
 
     if cli_args.debug or cfg.debug: