[New feature] 支持多教师在线蒸馏（MOPD）

src/twinkle/data_format/sampling.py

@@ -10,6 +10,7 @@
 @dataclass
 class SamplingParams:
     max_tokens: Optional[int] = None
+    min_tokens: Optional[int] = None
     seed: Optional[int] = None
     stop: Union[str, Sequence[str], Sequence[int], None] = None
     temperature: float = 1.0
@@ -59,6 +60,16 @@ def __post_init__(self):
             if self.max_tokens < 0:
                 raise ValueError(f'max_tokens must be >= 1, got {self.max_tokens}')
 
+        if self.min_tokens is not None:
+            if not isinstance(self.min_tokens, int):
+                raise ValueError(f'min_tokens must be an int or None, got {type(self.min_tokens)}')
+            if self.min_tokens < 0:
+                raise ValueError(f'min_tokens must be >= 0, got {self.min_tokens}')
+            if self.max_tokens is not None and self.min_tokens > self.max_tokens:
+                raise ValueError(
+                    f'min_tokens ({self.min_tokens}) must be <= max_tokens ({self.max_tokens})'
+                )
+
         if not isinstance(self.repetition_penalty, (int, float)):
             raise ValueError(f'repetition_penalty must be a number, got {type(self.repetition_penalty)}')
         if self.repetition_penalty <= 0:
@@ -79,6 +90,9 @@ def to_vllm(self, **kwargs):
         if self.max_tokens is not None:
             kwargs['max_tokens'] = self.max_tokens
 
+        if self.min_tokens is not None:
+            kwargs['min_tokens'] = self.min_tokens
+
         if self.seed is not None:
             kwargs['seed'] = self.seed
 

src/twinkle/loss/__init__.py

@@ -8,6 +8,8 @@
 from .infonce import InfonceLoss
 from .liger_fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from .mse import MSELoss
+from .gold import GOLDLoss
+from .cross_token import CrossTokenLoss
 from .value import PPOValueLoss
 
 torch_loss_mapping = {
@@ -33,4 +35,6 @@
     'orpo': ORPOLoss,
     # Embedding / contrastive losses
     'infonce': InfonceLoss,
+    'gold': GOLDLoss,
+    'cross_token': CrossTokenLoss
 }