已合并
docs: add loss_ctx parameter docstring for DeepseekV4 and MiniMaxM2 CausalLM #4930
docs: add loss_ctx parameter docstring for DeepseekV4 and MiniMaxM2 CausalLM #4930
已合并
Rostellaria创建于 20 天前
2 个文件变更+8-0
@@ -1701,6 +1701,10 @@ class DeepseekV4ForCausalLM(DeepseekV4PreTrainedModel, GenerationMixin):
1701 Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,1701 Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
1702 config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored1702 config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
1703 (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.1703 (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
1704+ loss_ctx (`callable`, *optional*):
1705+ A callable with signature `(hidden_states, weight, bias) -> loss` that lets the model compute the
1706+ loss directly without materialising the full logits. When provided, `labels` is ignored for the main loss
1707+ path and the returned `logits` are `None`. Note that `labels` is still used for the MTP loss path.
1704 1708 
1705 Example:1709 Example:
1706 1710 
@@ -820,6 +820,10 @@ class MiniMaxM2ForCausalLM(MiniMaxM2PreTrainedModel, GenerationMixin):
820 Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,820 Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
821 config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored821 config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
822 (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.822 (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
823+ loss_ctx (`callable`, *optional*):
824+ A callable with signature `(hidden_states, weight, bias) -> loss` that lets the model compute the
825+ loss directly without materialising the full logits. When provided, `labels` is ignored for the main loss
826+ path and the returned `logits` are `None`.
823 827 
824 Example:828 Example:
825 829