Make default RoPE path explicit in RotaryEmbedding classes

src/transformers/models/apertus/modeling_apertus.py

@@ -89,7 +89,10 @@ def __init__(self, config: ApertusConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/arcee/modeling_arcee.py

@@ -96,7 +96,10 @@ def __init__(self, config: ArceeConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/aria/modeling_aria.py

@@ -629,7 +629,10 @@ def __init__(self, config: AriaTextConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/bamba/modeling_bamba.py

@@ -207,7 +207,10 @@ def __init__(self, config: BambaConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/bitnet/modeling_bitnet.py

@@ -284,7 +284,10 @@ def __init__(self, config: BitNetConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/blt/modeling_blt.py

@@ -99,7 +99,10 @@ def __init__(self, config: BltConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/cohere/modeling_cohere.py

@@ -79,7 +79,10 @@ def __init__(self, config: CohereConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/cohere2/modeling_cohere2.py

@@ -54,7 +54,10 @@ def __init__(self, config: Cohere2Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/csm/modeling_csm.py

@@ -130,7 +130,10 @@ def __init__(self, config: CsmConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/dbrx/modeling_dbrx.py

@@ -53,7 +53,10 @@ def __init__(self, config: DbrxConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/deepseek_v3/modeling_deepseek_v3.py

@@ -67,7 +67,10 @@ def __init__(self, config: DeepseekV3Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/dia/modeling_dia.py

@@ -154,7 +154,10 @@ def __init__(self, config: DiaConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/diffllama/modeling_diffllama.py

@@ -568,7 +568,10 @@ def __init__(self, config: DiffLlamaConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/doge/modeling_doge.py

@@ -84,7 +84,10 @@ def __init__(self, config: DogeConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/dots1/modeling_dots1.py

@@ -76,7 +76,10 @@ def __init__(self, config: Dots1Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/emu3/modeling_emu3.py

@@ -1121,7 +1121,10 @@ def __init__(self, config: Emu3Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/ernie4_5/modeling_ernie4_5.py

@@ -53,7 +53,10 @@ def __init__(self, config: Ernie4_5Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/ernie4_5_moe/modeling_ernie4_5_moe.py

@@ -92,7 +92,10 @@ def __init__(self, config: Ernie4_5_MoeConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/evolla/modeling_evolla.py

@@ -1093,7 +1093,10 @@ def __init__(self, config: EvollaConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/exaone4/modeling_exaone4.py

@@ -82,7 +82,10 @@ def __init__(self, config: Exaone4Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/falcon/modeling_falcon.py

@@ -98,7 +98,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return q_embed, k_embed
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Falcon
+# Adapted from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Falcon
 class FalconRotaryEmbedding(nn.Module):
     inv_freq: torch.Tensor  # fix linting for `register_buffer`
 

src/transformers/models/falcon_h1/modeling_falcon_h1.py

@@ -236,7 +236,10 @@ def __init__(self, config: FalconH1Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/flex_olmo/modeling_flex_olmo.py

@@ -75,7 +75,10 @@ def __init__(self, config: FlexOlmoConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/gemma/modeling_gemma.py

@@ -93,7 +93,10 @@ def __init__(self, config: GemmaConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/gemma2/modeling_gemma2.py

@@ -97,7 +97,10 @@ def __init__(self, config: Gemma2Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/gemma3/modeling_gemma3.py

@@ -157,7 +157,10 @@ def __init__(self, config: Gemma3TextConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/gemma3n/modeling_gemma3n.py

@@ -1158,7 +1158,10 @@ def __init__(self, config: Gemma3nTextConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/glm/modeling_glm.py

@@ -253,7 +253,10 @@ def __init__(self, config: GlmConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/glm4/modeling_glm4.py

@@ -303,7 +303,10 @@ def __init__(self, config: Glm4Config, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/glm4_moe/modeling_glm4_moe.py

@@ -439,7 +439,10 @@ def __init__(self, config: Glm4MoeConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/glm4v/modeling_glm4v.py

@@ -399,7 +399,10 @@ def __init__(self, config: Glm4vTextConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/glm4v_moe/modeling_glm4v_moe.py

@@ -748,7 +748,10 @@ def __init__(self, config: Glm4vMoeTextConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -259,7 +259,10 @@ def __init__(self, config: GPTNeoXConfig, device=None):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type == "default":
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
+        else:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)