refactor: finalize impl

Signed-off-by: jupyterjazz <saba.sturua@jina.ai>

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=mixer_kwargs.get('task'))
                 if self.return_residual:  # mlp out is actually a pair here
                     mlp_out, hidden_states = mlp_out
                 if not self.fused_dropout_add_ln:

embedding.py

@@ -40,14 +40,17 @@ 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=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, task='sts')
+        lora_kwargs = {}
+        if task is not None:
+            lora_kwargs['task'] = task
+        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 +60,6 @@ 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, task='sts')
+            token_type_embeddings = self.token_type_embeddings(token_type_ids, **lora_kwargs)
             embeddings = embeddings + token_type_embeddings
         return embeddings

mha.py

@@ -340,9 +340,8 @@ class CrossAttention(nn.Module):
 class LinearResidual(nn.Linear):
     """Wrap nn.Linear to return the residual as well. For compatibility with FusedDense."""
 
-    def forward(self, input: torch.Tensor, task=None) -> torch.Tensor:
-        print('aq vafshe ar modis?')
-        return super().forward(input, task=task), input
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        return super().forward(input), input
 
 
 def _update_kv_cache(kv, inference_params, layer_idx):
@@ -591,6 +590,7 @@ class MHA(nn.Module):
         max_seqlen=None,
         mixer_subset=None,
         inference_params=None,
+        task=None,
         **kwargs,
     ):
         """
@@ -645,10 +645,15 @@ 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 = {}
+            if task:
+                lora_kwargs['task'] = task
+                lora_kwargs['residual'] = self.return_residual
+
             if not self.return_residual:
-                qkv = self.Wqkv(x)
+                qkv = self.Wqkv(x, **lora_kwargs)
             else:
-                qkv, x = self.Wqkv(x, task='query', residual=True)
+                qkv, x = self.Wqkv(x, **lora_kwargs)
 
             if self.dwconv:
                 qkv = rearrange(
@@ -734,5 +739,6 @@ 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)"), task='passage')
+        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 +47,13 @@ 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, task='clustering')
+    def forward(self, x, task):
+        lora_kwargs = {}
+        if task:
+            lora_kwargs['task'] = task
+        y = self.fc1(x, **lora_kwargs)
         y = self.activation(y)
-        y = self.fc2(y, task='sts')
+        y = self.fc2(y, **lora_kwargs)
         return y if not self.return_residual else (y, x)
 
 

modeling_lora.py

@@ -92,8 +92,6 @@ 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
@@ -116,18 +114,6 @@ class LoRAParametrization(nn.Module):
     def forward(self, X):
         return 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
-
     @classmethod
     def from_linear(
         cls,
@@ -239,12 +225,6 @@ class LoRAParametrization(nn.Module):
             layer.forward = new_forward.__get__(layer, layer.__class__)
 
 
-    @staticmethod
-    def select_task_for_layer(layer: nn.Module, task_idx: Optional[int] = None):
-        if isinstance(layer, LoRAParametrization):
-            layer.current_task = task_idx
-
-
 class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
     def __init__(
         self,
@@ -279,9 +259,6 @@ 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
@@ -340,39 +317,7 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
             )
         )
 
-    @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]:

modeling_xlm_roberta.py

@@ -215,6 +215,7 @@ class XLMRobertaEncoder(nn.Module):
                 if key_padding_mask is not None
                 else None
             )
+            mixer_kwargs['task'] = task
             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": task}
             if subset_mask is None:
                 for layer in self.layers:
                     if self._grad_checkpointing:
@@ -309,11 +310,15 @@ 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=None):
         # We "pool" the model by simply taking the hidden state corresponding
         # to the first token.
+        lora_kwargs = {}
+        if task:
+            lora_kwargs['task'] = task
+
         first_token_tensor = hidden_states[:, 0] if pool else hidden_states
-        pooled_output = self.dense(first_token_tensor, task='passage')
+        pooled_output = self.dense(first_token_tensor, **lora_kwargs)
         pooled_output = self.activation(pooled_output)
         return pooled_output
 
@@ -639,7 +644,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
         layer output for these tokens.
         masked_tokens_mask: (batch, seqlen), dtype=torch.bool
         """
-
+        task = kwargs.pop('task', None)
         if kwargs:
             for key, value in kwargs.items():
                 if value is not None:
@@ -653,7 +658,7 @@ 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=task
         )
         # TD [2022-12:18]: Don't need to force residual in fp32
         # BERT puts embedding LayerNorm before embedding dropout.
@@ -677,12 +682,12 @@ 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=task
         )
 
         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=task) if self.pooler is not None else None
             )
         else:
             # TD [2022-03-01]: the indexing here is very tricky.
@@ -696,7 +701,7 @@ 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=task) if self.pooler is not None else None
             )
 
         if not return_dict: