better llamaboard

* easily resume from checkpoint
* support full and freeze checkpoints
* faster ui


Former-commit-id: 84cfb2452cc86b037ccddee6e833f8eb7c129fa4

src/llamafactory/extras/constants.py

@@ -2,6 +2,19 @@ from collections import OrderedDict, defaultdict
 from enum import Enum
 from typing import Dict, Optional
 
+from peft.utils import SAFETENSORS_WEIGHTS_NAME as SAFE_ADAPTER_WEIGHTS_NAME
+from peft.utils import WEIGHTS_NAME as ADAPTER_WEIGHTS_NAME
+from transformers.utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME, WEIGHTS_INDEX_NAME, WEIGHTS_NAME
+
+
+CHECKPOINT_NAMES = {
+    SAFE_ADAPTER_WEIGHTS_NAME,
+    ADAPTER_WEIGHTS_NAME,
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_INDEX_NAME,
+    WEIGHTS_NAME,
+}
 
 CHOICES = ["A", "B", "C", "D"]
 
@@ -26,9 +39,9 @@ LAYERNORM_NAMES = {"norm", "ln"}
 
 METHODS = ["full", "freeze", "lora"]
 
-MOD_SUPPORTED_MODELS = ["bloom", "falcon", "gemma", "llama", "mistral", "mixtral", "phi", "starcoder2"]
+MOD_SUPPORTED_MODELS = {"bloom", "falcon", "gemma", "llama", "mistral", "mixtral", "phi", "starcoder2"}
 
-PEFT_METHODS = ["lora"]
+PEFT_METHODS = {"lora"}
 
 RUNNING_LOG = "running_log.txt"
 
@@ -49,9 +62,9 @@ TRAINING_STAGES = {
     "Pre-Training": "pt",
 }
 
-STAGES_USE_PAIR_DATA = ["rm", "dpo", "orpo"]
+STAGES_USE_PAIR_DATA = {"rm", "dpo"}
 
-SUPPORTED_CLASS_FOR_S2ATTN = ["llama"]
+SUPPORTED_CLASS_FOR_S2ATTN = {"llama"}
 
 V_HEAD_WEIGHTS_NAME = "value_head.bin"