fix version to 4.46

src/transformers/models/bloom/modeling_bloom.py

@@ -900,17 +900,15 @@ def prepare_inputs_for_generation(
             # the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
             model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format), "inputs_embeds": None}
 
+        # This part differs from other models because BLOOM needs a 2D mask to construct alibi tensor
+        # The only difference is the usage of 2D instead of 4D mask, but the shape will be static
         if isinstance(past_key_values, StaticCache) and attention_mask is not None:
             target_length = past_key_values.get_max_length()
             batch_size, seq_length = attention_mask.shape
             diff = target_length - seq_length
-            attention_mask = torch.cat(
-                [
-                    attention_mask,
-                    torch.zeros(batch_size, diff, device=attention_mask.device, dtype=attention_mask.dtype),
-                ],
-                dim=-1,
-            )
+
+            new_attn_mask = torch.zeros(batch_size, diff, device=attention_mask.device, dtype=attention_mask.dtype)
+            attention_mask = torch.cat([attention_mask, new_attn_mask,], dim=-1,)
 
         model_inputs.update(
             {

src/transformers/models/falcon/modeling_falcon.py

@@ -256,7 +256,7 @@ class FalconLinearScalingRotaryEmbedding(FalconRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`FalconLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`FalconLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`FalconRotaryEmbedding`, which now also does linear scaling (simply pass the model config to __init__)."
         )
         kwargs["rope_type"] = "linear"
@@ -269,7 +269,7 @@ class FalconDynamicNTKScalingRotaryEmbedding(FalconRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`FalconDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`FalconDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`FalconRotaryEmbedding`, which now also does dynamic ntk scaling (simply pass the model config to "
             "__init__)."
         )

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -578,7 +578,7 @@ def __init__(
         if config is None:
             logger.warning_once(
                 "`GPTNeoXRotaryEmbedding` can now be fully parameterized by passing the model config through the "
-                "`config` argument. All other arguments will be removed in v4.45"
+                "`config` argument. All other arguments will be removed in v4.46"
             )
             self.rope_kwargs = {
                 "rope_type": rope_type,
@@ -654,7 +654,7 @@ class GPTNeoXLinearScalingRotaryEmbedding(GPTNeoXRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`GPTNeoXLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`GPTNeoXLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`GPTNeoXRotaryEmbedding`, which now also does linear scaling (simply pass the model config to __init__)."
         )
         kwargs["rope_type"] = "linear"
@@ -667,7 +667,7 @@ class GPTNeoXDynamicNTKScalingRotaryEmbedding(GPTNeoXRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`GPTNeoXDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`GPTNeoXDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`GPTNeoXRotaryEmbedding`, which now also does dynamic ntk scaling (simply pass the model config to "
             "__init__)."
         )

src/transformers/models/llama/modeling_llama.py

@@ -149,7 +149,7 @@ def __init__(
         if config is None:
             logger.warning_once(
                 "`LlamaRotaryEmbedding` can now be fully parameterized by passing the model config through the "
-                "`config` argument. All other arguments will be removed in v4.45"
+                "`config` argument. All other arguments will be removed in v4.46"
             )
             self.rope_kwargs = {
                 "rope_type": rope_type,
@@ -224,7 +224,7 @@ class LlamaLinearScalingRotaryEmbedding(LlamaRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`LlamaLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`LlamaLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`LlamaRotaryEmbedding`, which now also does linear scaling (simply pass the model config to __init__)."
         )
         kwargs["rope_type"] = "linear"
@@ -236,7 +236,7 @@ class LlamaDynamicNTKScalingRotaryEmbedding(LlamaRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`LlamaDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`LlamaDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`LlamaRotaryEmbedding`, which now also does dynamic ntk scaling (simply pass the model config to "
             "__init__)."
         )
@@ -353,7 +353,7 @@ def __init__(self, config: LlamaConfig, layer_idx: Optional[int] = None):
         self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
         self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
 
-        # TODO (joao): remove in v4.45 (RoPE is computed in the model, not in the decoder layers)
+        # TODO (joao): remove in v4.46 (RoPE is computed in the model, not in the decoder layers)
         self.rotary_emb = LlamaRotaryEmbedding(config=self.config)
 
     def forward(
@@ -365,7 +365,7 @@ def forward(
         output_attentions: bool = False,
         use_cache: bool = False,
         cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
         bsz, q_len, _ = hidden_states.size()
@@ -400,7 +400,7 @@ def forward(
             logger.warning_once(
                 "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
                 "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
-                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
                 "removed and `position_embeddings` will be mandatory."
             )
             cos, sin = self.rotary_emb(value_states, position_ids)
@@ -473,7 +473,7 @@ def forward(
         output_attentions: bool = False,
         use_cache: bool = False,
         cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
         if isinstance(past_key_value, StaticCache):
             raise ValueError(
@@ -500,7 +500,7 @@ def forward(
             logger.warning_once(
                 "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
                 "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
-                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
                 "removed and `position_embeddings` will be mandatory."
             )
             cos, sin = self.rotary_emb(value_states, position_ids)
@@ -586,7 +586,7 @@ def forward(
         output_attentions: bool = False,
         use_cache: bool = False,
         cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
         if output_attentions:
@@ -620,7 +620,7 @@ def forward(
             logger.warning_once(
                 "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
                 "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
-                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
                 "removed and `position_embeddings` will be mandatory."
             )
             cos, sin = self.rotary_emb(value_states, position_ids)
@@ -695,7 +695,7 @@ def forward(
         output_attentions: Optional[bool] = False,
         use_cache: Optional[bool] = False,
         cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
         **kwargs,
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """

src/transformers/models/nemotron/modeling_nemotron.py

@@ -589,7 +589,7 @@ def forward(
         output_attentions: Optional[bool] = False,
         use_cache: Optional[bool] = False,
         cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
         **kwargs,
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """

src/transformers/models/persimmon/modeling_persimmon.py

@@ -119,7 +119,7 @@ def __init__(
         if config is None:
             logger.warning_once(
                 "`PersimmonRotaryEmbedding` can now be fully parameterized by passing the model config through the "
-                "`config` argument. All other arguments will be removed in v4.45"
+                "`config` argument. All other arguments will be removed in v4.46"
             )
             self.rope_kwargs = {
                 "rope_type": rope_type,
@@ -195,7 +195,7 @@ class PersimmonLinearScalingRotaryEmbedding(PersimmonRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`PersimmonLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`PersimmonLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`PersimmonRotaryEmbedding`, which now also does linear scaling (simply pass the model config to __init__)."
         )
         kwargs["rope_type"] = "linear"
@@ -208,7 +208,7 @@ class PersimmonDynamicNTKScalingRotaryEmbedding(PersimmonRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`PersimmonDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`PersimmonDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`PersimmonRotaryEmbedding`, which now also does dynamic ntk scaling (simply pass the model config to "
             "__init__)."
         )

src/transformers/models/phi/modeling_phi.py

@@ -131,7 +131,7 @@ def __init__(
         if config is None:
             logger.warning_once(
                 "`PhiRotaryEmbedding` can now be fully parameterized by passing the model config through the "
-                "`config` argument. All other arguments will be removed in v4.45"
+                "`config` argument. All other arguments will be removed in v4.46"
             )
             self.rope_kwargs = {
                 "rope_type": rope_type,
@@ -207,7 +207,7 @@ class PhiLinearScalingRotaryEmbedding(PhiRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`PhiLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`PhiLinearScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`PhiRotaryEmbedding`, which now also does linear scaling (simply pass the model config to __init__)."
         )
         kwargs["rope_type"] = "linear"
@@ -220,7 +220,7 @@ class PhiDynamicNTKScalingRotaryEmbedding(PhiRotaryEmbedding):
 
     def __init__(self, *args, **kwargs):
         logger.warning_once(
-            "`PhiDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.45. Please use "
+            "`PhiDynamicNTKScalingRotaryEmbedding` is deprecated an will be removed in v4.46. Please use "
             "`PhiRotaryEmbedding`, which now also does dynamic ntk scaling (simply pass the model config to "
             "__init__)."
         )

src/transformers/models/qwen2/modeling_qwen2.py

@@ -154,7 +154,7 @@ def __init__(
         if config is None:
             logger.warning_once(
                 "`Qwen2RotaryEmbedding` can now be fully parameterized by passing the model config through the "
-                "`config` argument. All other arguments will be removed in v4.45"
+                "`config` argument. All other arguments will be removed in v4.46"
             )
             self.rope_kwargs = {
                 "rope_type": rope_type,

src/transformers/models/qwen2_moe/modeling_qwen2_moe.py

@@ -230,7 +230,7 @@ def __init__(
         if config is None:
             logger.warning_once(
                 "`Qwen2MoeRotaryEmbedding` can now be fully parameterized by passing the model config through the "
-                "`config` argument. All other arguments will be removed in v4.45"
+                "`config` argument. All other arguments will be removed in v4.46"
             )
             self.rope_kwargs = {
                 "rope_type": rope_type,