FIX: TF32 warning (#43012)

src/transformers/models/afmoe/modeling_afmoe.py

@@ -99,7 +99,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/apertus/modeling_apertus.py

@@ -134,7 +134,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/arcee/modeling_arcee.py

@@ -139,7 +139,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/aria/modeling_aria.py

@@ -676,7 +676,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/bamba/modeling_bamba.py

@@ -240,7 +240,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/bitnet/modeling_bitnet.py

@@ -327,7 +327,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/chameleon/modeling_chameleon.py

@@ -124,7 +124,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/csm/modeling_csm.py

@@ -176,7 +176,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/cwm/modeling_cwm.py

@@ -98,7 +98,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/dbrx/modeling_dbrx.py

@@ -98,7 +98,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/deepseek_v3/modeling_deepseek_v3.py

@@ -111,7 +111,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/dia/modeling_dia.py

@@ -193,7 +193,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/diffllama/modeling_diffllama.py

@@ -126,7 +126,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/doge/modeling_doge.py

@@ -128,7 +128,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/dots1/modeling_dots1.py

@@ -120,7 +120,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/emu3/modeling_emu3.py

@@ -1172,7 +1172,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/evolla/modeling_evolla.py

@@ -1028,7 +1028,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/exaone4/modeling_exaone4.py

@@ -126,7 +126,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/falcon/modeling_falcon.py

@@ -162,7 +162,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/falcon_h1/modeling_falcon_h1.py

@@ -281,7 +281,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/gemma/modeling_gemma.py

@@ -138,7 +138,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/gemma2/modeling_gemma2.py

@@ -139,7 +139,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/gemma2/modular_gemma2.py

@@ -254,7 +254,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/glm/modeling_glm.py

@@ -121,7 +121,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/glm4/modeling_glm4.py

@@ -326,7 +326,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/glm4_moe/modeling_glm4_moe.py

@@ -102,7 +102,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -108,7 +108,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py

@@ -118,7 +118,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/granite/modeling_granite.py

@@ -377,7 +377,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/granitemoe/modeling_granitemoe.py

@@ -120,7 +120,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py

@@ -959,7 +959,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/granitemoeshared/modeling_granitemoeshared.py

@@ -534,7 +534,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling

src/transformers/models/helium/modeling_helium.py

@@ -119,7 +119,7 @@ def forward(self, x, position_ids):
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            freqs = (inv_freq_expanded.float() * position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling