make-lora-stateless

block.py

@@ -233,7 +233,7 @@ class Block(nn.Module):
                     is_rms_norm=isinstance(self.norm1, RMSNorm),
                 )
             if not isinstance(self.mlp, nn.Identity):
-                mlp_out = self.mlp(hidden_states)
+                mlp_out = self.mlp(hidden_states, task_type=mixer_kwargs.get('task_type'))
                 if self.return_residual:  # mlp out is actually a pair here
                     mlp_out, hidden_states = mlp_out
                 if not self.fused_dropout_add_ln:

configuration_xlm_roberta.py

@@ -23,6 +23,7 @@ class XLMRobertaFlashConfig(PretrainedConfig):
             use_cache=True,
             classifier_dropout=None,
             lora_adaptations=None,
+            lora_prompts=None,
             lora_rank=4,
             lora_dropout_p=0.0,
             lora_alpha=1,
@@ -55,6 +56,7 @@ class XLMRobertaFlashConfig(PretrainedConfig):
         self.classifier_dropout = classifier_dropout
         self.load_trained_adapters = load_trained_adapters
         self.lora_adaptations = lora_adaptations
+        self.lora_prompts = lora_prompts
         self.lora_rank = lora_rank
         self.lora_dropout_p = lora_dropout_p
         self.lora_alpha = lora_alpha

embedding.py

@@ -40,14 +40,15 @@ class XLMRobertaEmbeddings(nn.Module):
         if self.type_vocab_size > 0:
             self.token_type_embeddings = nn.Embedding(type_vocab_size, embed_dim, **factory_kwargs)
 
-    def forward(self, input_ids, position_ids=None, token_type_ids=None):
+    def forward(self, input_ids, position_ids=None, token_type_ids=None, task_type=None):
         """
         input_ids: (batch, seqlen)
         position_ids: (batch, seqlen)
         token_type_ids: (batch, seqlen)
         """
         batch_size, seqlen = input_ids.shape
-        embeddings = self.word_embeddings(input_ids)
+        lora_kwargs = {'task_type': task_type} if task_type is not None else {}
+        embeddings = self.word_embeddings(input_ids, **lora_kwargs)
         if self.max_position_embeddings > 0:
             if position_ids is None:
                 position_ids = create_position_ids_from_input_ids(input_ids, padding_idx=self.word_embeddings.padding_idx).to(input_ids.device)
@@ -57,6 +58,6 @@ class XLMRobertaEmbeddings(nn.Module):
         if self.type_vocab_size > 0:
             if token_type_ids is None:
                 token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
-            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            token_type_embeddings = self.token_type_embeddings(token_type_ids, **lora_kwargs)
             embeddings = embeddings + token_type_embeddings
         return embeddings

mha.py

@@ -450,6 +450,7 @@ class MHA(nn.Module):
 
         if fused_bias_fc and FusedDense is None:
             raise ImportError("fused_dense is not installed")
+
         linear_cls = nn.Linear if not fused_bias_fc else FusedDense
         linear_resid_cls = (
             LinearResidual if not fused_bias_fc else partial(FusedDense, return_residual=True)
@@ -589,6 +590,7 @@ class MHA(nn.Module):
         max_seqlen=None,
         mixer_subset=None,
         inference_params=None,
+        task_type=None,
         **kwargs,
     ):
         """
@@ -643,10 +645,14 @@ class MHA(nn.Module):
         batch, seqlen = x.shape[:2]
         if not self.cross_attn and self.num_heads_kv == self.num_heads:
             assert x_kv is None and mixer_subset is None
+            lora_kwargs = {'task_type': task_type} if task_type is not None else {}
             if not self.return_residual:
-                qkv = self.Wqkv(x)
+                qkv = self.Wqkv(x, **lora_kwargs)
             else:
-                qkv, x = self.Wqkv(x)
+                if lora_kwargs:
+                    lora_kwargs['residual'] = True
+                qkv, x = self.Wqkv(x, **lora_kwargs)
+
             if self.dwconv:
                 qkv = rearrange(
                     self.dwconv_qkv(rearrange(qkv, "b s d -> b d s"))[..., :-2], "b d s -> b s d"
@@ -731,5 +737,7 @@ class MHA(nn.Module):
                     context = self._update_kvcache_attention(q, kv, inference_params)
             else:
                 context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
-        out = self.out_proj(rearrange(context, "... h d -> ... (h d)"))
+
+        lora_kwargs.pop('residual', None)
+        out = self.out_proj(rearrange(context, "... h d -> ... (h d)"), **lora_kwargs)
         return out if not self.return_residual else (out, x)

mlp.py

@@ -47,10 +47,11 @@ class Mlp(nn.Module):
         self.activation = activation
         self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)
 
-    def forward(self, x):
-        y = self.fc1(x)
+    def forward(self, x, task_type=None):
+        lora_kwargs = {'task_type': task_type} if task_type is not None else {}
+        y = self.fc1(x, **lora_kwargs)
         y = self.activation(y)
-        y = self.fc2(y)
+        y = self.fc2(y, **lora_kwargs)
         return y if not self.return_residual else (y, x)
 
 

modeling_lora.py

@@ -9,14 +9,12 @@ import torch
 import torch.nn.utils.parametrize as parametrize
 from torch import nn
 from torch.nn import Parameter
+from torch.nn import functional as F
 from transformers import PretrainedConfig
 
 from .modeling_xlm_roberta import XLMRobertaFlashConfig, XLMRobertaModel, XLMRobertaPreTrainedModel
 
 
-LORA_NO_UPDATE = '__lora_no_update__'
-
-
 def initialized_weights(
     shape: Tuple[int], num_adaptations: int, init: str = "kaiming"
 ) -> torch.Tensor:
@@ -91,22 +89,19 @@ class LoRAParametrization(nn.Module):
             torch.ones(self.swap((1, fan_in)), dtype=self.lora_A.dtype),
             persistent=False,
         )
-        self.forward_fn = lambda x: x
-        self.current_task = None
 
     def _dropout(self, A):
         # to mimic the original implementation: A @ dropout(x), we do (A * dropout(ones)) @ x
         return A * self.lora_dropout(self.lora_dropout_mask)
 
-    def lora_forward(self, X):
-        assert self.current_task is not None
+    def lora_forward(self, X, current_task):
         return (
             X
             + torch.matmul(
                 *self.swap(
                     (
-                        self.lora_B[self.current_task],
-                        self.dropout_fn(self.lora_A[self.current_task]),
+                        self.lora_B[current_task],
+                        self.dropout_fn(self.lora_A[current_task]),
                     )
                 )
             ).view(X.shape)
@@ -114,19 +109,7 @@ class LoRAParametrization(nn.Module):
         )
 
     def forward(self, X):
-        return self.forward_fn(X)
-
-    @property
-    def current_task(self):
-        return self._current_task
-
-    @current_task.setter
-    def current_task(self, task: Union[None, int]):
-        self._current_task = task
-        if task is None:
-            self.forward_fn = lambda x: x
-        else:
-            self.forward_fn = self.lora_forward
+        return X
 
     @classmethod
     def from_linear(
@@ -178,6 +161,7 @@ class LoRAParametrization(nn.Module):
         rank: int,
         dropout_p: float,
         alpha: float,
+        adaptation_map: dict,
     ):
         if isinstance(layer, nn.Linear):
             parametrize.register_parametrization(
@@ -191,6 +175,22 @@ class LoRAParametrization(nn.Module):
                     alpha=alpha,
                 ),
             )
+
+            def new_forward(self, input, task_type, residual=False):
+                task_idx = adaptation_map[task_type] if task_type else None
+                if task_idx is not None:
+                    weights = self.parametrizations.weight[0].lora_forward(self.weight, current_task=task_idx)
+                else:
+                    weights = self.weight
+
+                out = F.linear(input, weights, self.bias)
+
+                if residual:
+                    return out, input
+                return out
+
+            layer.forward = new_forward.__get__(layer, layer.__class__)
+
         elif isinstance(layer, nn.Embedding):
             parametrize.register_parametrization(
                 layer,
@@ -204,10 +204,20 @@ class LoRAParametrization(nn.Module):
                 ),
             )
 
-    @staticmethod
-    def select_task_for_layer(layer: nn.Module, task_idx: Optional[int] = None):
-        if isinstance(layer, LoRAParametrization):
-            layer.current_task = task_idx
+            def new_forward(self, input, task_type):
+                task_idx = adaptation_map[task_type] if task_type else None
+                if task_idx is not None:
+                    weights = self.parametrizations.weight[0].lora_forward(self.weight, current_task=task_idx)
+                else:
+                    weights = self.weight
+
+                out = F.embedding(
+                    input, weights, self.padding_idx, self.max_norm,
+                    self.norm_type, self.scale_grad_by_freq, self.sparse)
+
+                return out
+
+            layer.forward = new_forward.__get__(layer, layer.__class__)
 
 
 class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
@@ -231,6 +241,16 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
             raise ValueError(
                 f'`lora_adaptations` must be a list and contain at least one element'
             )
+        self._lora_prompts = config.lora_prompts
+        if (
+            not isinstance(self._lora_prompts, dict)
+            or len(self._lora_prompts) != len(self._lora_adaptations)
+            or not all([v in self._lora_adaptations for v in self._lora_prompts.keys()])
+        ):
+            raise ValueError(
+                f'`lora_prompts` must be a dict and contain the same number of elements '
+                f'as `lora_adaptations` with all keys in `lora_prompts` present in `lora_adaptations`.'
+        )
         self._adaptation_map = {
             name: idx for idx, name in enumerate(self._lora_adaptations)
         }
@@ -244,9 +264,7 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
             alpha=self._alpha,
         )
         self.main_params_trainable = config.lora_main_params_trainable
-        self._task_idx = None
-        # By default, disable LoRA until it's specified which adapter/task to use
-        self.current_task = None
+
 
     @property
     def main_params_trainable(self):
@@ -300,42 +318,11 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
                 rank=rank,
                 dropout_p=dropout_p,
                 alpha=alpha,
+                adaptation_map=self._adaptation_map,
             )
         )
 
-    @property
-    def current_task(self):
-        """Which LoRA is currently selected
-        :return: Integer or None (when LoRA is disabled)
-        """
-        return self._task_idx
-
-    @current_task.setter
-    def current_task(self, task_name: Union[None, str]):
-        """Set the LoRA that is to be used.
-        The LoRA is specified by `task_idx`, which may be an integer >= 0,
-        indexing the available LoRAs. If it is None, no LoRA is used.
-        :param task_name: Which LoRA to use
-        :return:
-        """
-        if task_name and task_name not in self._lora_adaptations:
-            raise ValueError(
-                f"Unsupported task '{task_name}'. "
-                f"Supported tasks are: {', '.join(self.config.lora_adaptations)}."
-                f"Alternatively, set `task` to `None` if you want to disable LoRA."
-            )
-        task_idx = self._adaptation_map[task_name] if task_name else None
-        if self._task_idx != task_idx:
-            # In this case, we need to update the LoRAs everywhere
-            self._task_idx = task_idx
-            self.apply(
-                partial(LoRAParametrization.select_task_for_layer, task_idx=task_idx)
-            )
-
-    def forward(self, *args, task: Union[str, None] = LORA_NO_UPDATE, **kwargs):
-        if task != LORA_NO_UPDATE:
-            self.current_task = task
-
+    def forward(self, *args, **kwargs):
         return self.roberta(*args, **kwargs)
 
     def parameters(self, recurse: bool = True) -> Iterator[Parameter]:
@@ -355,27 +342,22 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
     def encode(
         self,
         *args,
-        task: Union[str, None] = LORA_NO_UPDATE,
+        task_type: Optional[str] = None,
         **kwargs,
     ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
         """
         Computes sentence embeddings
 
-        task(`str`, *optional*, defaults to `LORA_NO_UPDATE`):
-            Specifies the task for which the encoding is intended. This parameter controls the
-            use of specialized LoRA adapters that are tuned for specific tasks. If `task` is set
-            to `LORA_NO_UPDATE`, there will be no update to the current task, retaining the
-            existing adapter configuration. If `task` is explicitly set to `None`, all LoRA
-            adapters are disabled, and the model reverts to its original, general-purpose weights.
-            If `task` is set to a specific LoRA adaptation, that adaptation is activated.
+        task_type(`str`, *optional*, defaults to `None`):
+            Specifies the task for which the encoding is intended. If `task_type` is not provide,
+            all LoRA adapters are disabled, and the model reverts to its original,
+            general-purpose weights.
         """
-        if task != LORA_NO_UPDATE:
-            if not task:
-                warnings.warn(
-                    f"Task-specific embeddings are disabled. To enable, specify the `task` "
-                    f"argument with one of the supported tasks: {', '.join(self.config.lora_adaptations)}",
-                    category=UserWarning,
-                )
-            self.current_task = task
+        if task_type and task_type not in self._lora_adaptations:
+            raise ValueError(
+                f"Unsupported task '{task_type}'. "
+                f"Supported tasks are: {', '.join(self.config.lora_adaptations)}."
+                f"Alternatively, don't pass the `task_type` argument to disable LoRA."
+            )
 
-        return self.roberta.encode(*args, **kwargs)
+        return self.roberta.encode(*args, task_type=task_type, **kwargs)

modeling_xlm_roberta.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 from einops import rearrange
-from transformers import PretrainedConfig
+from transformers import PretrainedConfig, AutoTokenizer
 from transformers.modeling_utils import PreTrainedModel
 from transformers.modeling_outputs import MaskedLMOutput,SequenceClassifierOutput
 from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead
@@ -204,7 +204,7 @@ class XLMRobertaEncoder(nn.Module):
     def gradient_checkpointing(self, value):
         self._grad_checkpointing = value
 
-    def forward(self, hidden_states, key_padding_mask=None, subset_mask=None):
+    def forward(self, hidden_states, key_padding_mask=None, subset_mask=None, task_type=None):
         """If subset_mask is not None, we only want output for the subset of the sequence.
         This means that we only compute the last layer output for these tokens.
         subset_mask: (batch, seqlen), dtype=torch.bool
@@ -215,6 +215,7 @@ class XLMRobertaEncoder(nn.Module):
                 if key_padding_mask is not None
                 else None
             )
+            mixer_kwargs['task_type'] = task_type
             for layer in self.layers:
                 if self._grad_checkpointing:
                     hidden_states = torch.utils.checkpoint.checkpoint(
@@ -232,7 +233,7 @@ class XLMRobertaEncoder(nn.Module):
             hidden_states, indices, cu_seqlens, max_seqlen_in_batch = unpad_input(
                 hidden_states, key_padding_mask
             )
-            mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch}
+            mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch, "task_type": task_type}
             if subset_mask is None:
                 for layer in self.layers:
                     if self._grad_checkpointing:
@@ -309,11 +310,13 @@ class XLMRobertaPooler(nn.Module):
         self.dense = linear_cls(config.hidden_size, config.hidden_size)
         self.activation = nn.Tanh()
 
-    def forward(self, hidden_states, pool=True):
+    def forward(self, hidden_states, pool=True, task_type=None):
         # We "pool" the model by simply taking the hidden state corresponding
         # to the first token.
+        lora_kwargs = {'task_type': task_type} if task_type is not None else {}
+
         first_token_tensor = hidden_states[:, 0] if pool else hidden_states
-        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.dense(first_token_tensor, **lora_kwargs)
         pooled_output = self.activation(pooled_output)
         return pooled_output
 
@@ -440,7 +443,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
         self.pooler = XLMRobertaPooler(config) if add_pooling_layer else None
 
         self.apply(partial(_init_weights, initializer_range=config.initializer_range))
-
+        self.tokenizer = AutoTokenizer.from_pretrained(self.name_or_path, trust_remote_code=True)
 
     @torch.inference_mode()
     def encode(
@@ -454,6 +457,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
         device: Optional[torch.device] = None,
         normalize_embeddings: bool = False,
         truncate_dim: Optional[int] = None,
+        task_type: Optional[str] = None,
         **tokenizer_kwargs,
     ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
         """
@@ -492,12 +496,6 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
             If convert_to_tensor, a stacked tensor is returned.
             If convert_to_numpy, a numpy matrix is returned.
         """
-        from transformers import AutoTokenizer
-
-        self.tokenizer = AutoTokenizer.from_pretrained(
-            self.name_or_path, trust_remote_code=True
-        )
-
         is_training = self.training
         self.eval()
 
@@ -544,14 +542,14 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
             )
         else:
             range_iter = range(0, len(sentences), batch_size)
-
+        lora_kwargs = {'task_type': task_type} if task_type is not None else {}
         for i in range_iter:
             encoded_input = self.tokenizer(
                 sentences[i : i + batch_size],
                 return_tensors='pt',
                 **tokenizer_kwargs,
             ).to(self.device)
-            token_embs = self.forward(**encoded_input)[0]
+            token_embs = self.forward(**encoded_input, **lora_kwargs)[0]
 
             # Accumulate in fp32 to avoid overflow
             token_embs = token_embs.float()
@@ -639,7 +637,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
         layer output for these tokens.
         masked_tokens_mask: (batch, seqlen), dtype=torch.bool
         """
-
+        task_type = kwargs.pop('task_type', None)
         if kwargs:
             for key, value in kwargs.items():
                 if value is not None:
@@ -653,7 +651,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
         )
 
         hidden_states = self.embeddings(
-            input_ids, position_ids=position_ids, token_type_ids=token_type_ids
+            input_ids, position_ids=position_ids, token_type_ids=token_type_ids, task_type=task_type
         )
         # TD [2022-12:18]: Don't need to force residual in fp32
         # BERT puts embedding LayerNorm before embedding dropout.
@@ -677,12 +675,12 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
             subset_mask = None
 
         sequence_output = self.encoder(
-            hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask
+            hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask, task_type=task_type
         )
 
         if masked_tokens_mask is None:
             pooled_output = (
-                self.pooler(sequence_output) if self.pooler is not None else None
+                self.pooler(sequence_output, task_type=task_type) if self.pooler is not None else None
             )
         else:
             # TD [2022-03-01]: the indexing here is very tricky.
@@ -696,7 +694,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
                 pool_input = sequence_output[first_col_mask[subset_mask]]
                 sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
             pooled_output = (
-                self.pooler(pool_input, pool=False) if self.pooler is not None else None
+                self.pooler(pool_input, pool=False, task_type=task_type) if self.pooler is not None else None
             )
 
         if not return_dict:
@@ -1278,4 +1276,4 @@ class XLMRobertaForSequenceClassification(XLMRobertaPreTrainedModel):
             logits=logits,
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
-        )
+        )

rotary.py

@@ -6,11 +6,13 @@ from typing import Optional, Tuple, Union
 
 import torch
 from einops import rearrange, repeat
-try:
-    from flash_attn.ops.triton.rotary import apply_rotary
-except ImportError:
-    def apply_rotary(*args, **kwargs):
-        raise RuntimeError('RoPE requires flash-attention to be installed')
+
+if torch.cuda.is_available():
+    try:
+        from flash_attn.ops.triton.rotary import apply_rotary
+    except ImportError:
+        def apply_rotary(*args, **kwargs):
+            raise RuntimeError('RoPE requires flash-attention to be installed')
 
 
 def rotate_half(x, interleaved=False):
@@ -29,6 +31,10 @@ def apply_rotary_emb_torch(x, cos, sin, interleaved=False):
     """
     ro_dim = cos.shape[-1] * 2
     assert ro_dim <= x.shape[-1]
+    cos, sin = (
+        cos[:x.shape[1]],
+        sin[:x.shape[1]],
+    )
     cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
     sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
     return torch.cat(
@@ -60,6 +66,7 @@ class ApplyRotaryEmb(torch.autograd.Function):
             interleaved=interleaved,
             inplace=inplace,
         )
+
         if isinstance(seqlen_offsets, int):
             ctx.save_for_backward(cos, sin, cu_seqlens)  # Can't save int with save_for_backward
             ctx.seqlen_offsets = seqlen_offsets
@@ -82,6 +89,7 @@ class ApplyRotaryEmb(torch.autograd.Function):
         # "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
         if not ctx.interleaved and not ctx.inplace:
             do = do.clone()
+
         dx = apply_rotary(
             do,
             cos,
@@ -150,21 +158,37 @@ class ApplyRotaryEmbQKV_(torch.autograd.Function):
         # batch, seqlen, three, nheads, headdim = qkv.shape
         assert qkv.shape[-3] == 3
         if cos_k is None and sin_k is None and qkv.is_contiguous():
-            # Call 1 kernel instead of 2 kernels
-            # We need qkv to be contiguous so that when we reshape to combine (3, nheads)
-            # dimensions, we get the same tensor
-            qk = rearrange(qkv[..., :2, :, :], "... t h d -> ... (t h) d")
-            # qk = qkv[:, :, :2].reshape(batch, seqlen, -1, headdim)
-            apply_rotary(
-                qk,
-                cos,
-                sin,
-                seqlen_offsets=seqlen_offsets,
-                interleaved=interleaved,
-                inplace=True,
-                cu_seqlens=cu_seqlens,
-                max_seqlen=max_seqlen,
-            )
+
+            if torch.cuda.is_available():
+                # Call 1 kernel instead of 2 kernels
+                # We need qkv to be contiguous so that when we reshape to combine (3, nheads)
+                # dimensions, we get the same tensor
+                qk = rearrange(qkv[..., :2, :, :], "... t h d -> ... (t h) d")
+                # qk = qkv[:, :, :2].reshape(batch, seqlen, -1, headdim)
+                apply_rotary(
+                    qk,
+                    cos,
+                    sin,
+                    seqlen_offsets=seqlen_offsets,
+                    interleaved=interleaved,
+                    inplace=True,
+                    cu_seqlens=cu_seqlens,
+                    max_seqlen=max_seqlen,
+                )
+            else:
+                q_rot = apply_rotary_emb_torch(
+                    qkv[:, :, 0],
+                    cos,
+                    sin,
+                    interleaved=interleaved,
+                )
+                k_rot = apply_rotary_emb_torch(
+                    qkv[:, :, 1],
+                    cos,
+                    sin,
+                    interleaved=interleaved,
+                )
+                qkv = torch.stack((q_rot, k_rot, qkv[:, :, 2]), dim=2)
         else:
             cos_k = cos if cos_k is None else cos_k
             sin_k = sin if sin_k is None else sin_k
@@ -228,7 +252,6 @@ class ApplyRotaryEmbQKV_(torch.autograd.Function):
             sin_k = sin if sin_k is None else sin_k
             dq, dk = dqkv[..., 0, :, :], dqkv[..., 1, :, :]
             apply_rotary(
-
                 dq,
                 cos,
                 sin,