Add LLaMa-2 Chat Backbone

prismatic/conf/models.py

@@ -259,6 +259,19 @@ class Exp_13B_Llama2(Exp_13B_One_Stage):
     llm_backbone_id: str = "llama2-13b-pure"
 
 
+# ~ Additional LLM Backbones ~
+@dataclass
+class Ext_Exp_7B_Llama2_Chat(Exp_7B_One_Stage):
+    model_id: str = "llama2-chat+7b"
+    llm_backbone_id: str = "llama2-7b-chat"
+
+
+@dataclass
+class Ext_Exp_13B_Llama2_Chat(Exp_13B_One_Stage):
+    model_id: str = "llama2-chat+13b"
+    llm_backbone_id: str = "llama2-13b-chat"
+
+
 # Section 4.3B :: ✌️ --> Co-training on Language-only Data
 #   =>> Note :: Run with `--dataset.type "llava-multimodal" (multimodal data only / no co-training)
 @dataclass
@@ -501,6 +514,11 @@ class ModelRegistry(Enum):
     EXP_LLAMA2_7B = Exp_7B_Llama2
     EXP_LLAMA2_13B = Exp_13B_Llama2
 
+    # ~ Additional LLM Backbone Experiments :: LLaMa-2 Chat ~
+    EXT_EXP_LLAMA2_CHAT_7B = Ext_Exp_7B_Llama2_Chat
+    EXT_EXP_LLAMA2_CHAT_13B = Ext_Exp_13B_Llama2_Chat
+
+    # Cotraining w/ Unimodal Data
     EXP_VICUNA_NO_COTRAINING_7B = Exp_7B_Vicuna_No_Cotraining
     EXP_LLAMA2_NO_COTRAINING_7B = Exp_7B_Llama2_No_Cotraining
 

scripts/hf-hub/upload.py

@@ -78,7 +78,7 @@ def upload(cfg: UploadConfig) -> None:
     subprocess.run(
         f"HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload {cfg.hub_repo} {convert_path!s} {cfg.model_id}/",
         shell=True,
-        check=True
+        check=True,
     )
 
     # Done

scripts/pretrain.py

@@ -51,7 +51,7 @@ class PretrainConfig:
 
     # ModelConfig (`prismatic/conf/models.py`); override with --model.type `ModelRegistry.<MODEL>.model_id`
     model: ModelConfig = field(
-        default_factory=ModelConfig.get_choice_class(ModelRegistry.PRISM_DINOSIGLIP_CONTROLLED_7B.model_id)
+        default_factory=ModelConfig.get_choice_class(ModelRegistry.EXT_EXP_LLAMA2_CHAT_13B.model_id)
     )
 
     # DatasetConfig (`prismatic/conf/datasets.py`); override with --dataset.type `DatasetRegistry.<DATASET>.dataset_id`