Add XPU type for work-around -inf mask causing sdpa NaN issue in mode…

…ling files (#35647)

* add xpu for unmask

* change modular for generated matching

* add lastest modeling for helium

examples/modular-transformers/modeling_dummy.py

@@ -639,7 +639,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

examples/modular-transformers/modeling_multimodal1.py

@@ -639,7 +639,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

examples/modular-transformers/modeling_my_new_model2.py

@@ -644,7 +644,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

examples/modular-transformers/modeling_super.py

@@ -561,7 +561,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/aria/modeling_aria.py

@@ -1056,7 +1056,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/bamba/modeling_bamba.py

@@ -1360,7 +1360,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/bamba/modular_bamba.py

@@ -1101,7 +1101,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/bloom/modeling_bloom.py

@@ -785,7 +785,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/chameleon/modeling_chameleon.py

@@ -1431,7 +1431,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/codegen/modeling_codegen.py

@@ -628,7 +628,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/cohere/modeling_cohere.py

@@ -706,7 +706,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/dbrx/modeling_dbrx.py

@@ -1159,7 +1159,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/diffllama/modeling_diffllama.py

@@ -945,7 +945,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/emu3/modeling_emu3.py

@@ -1524,7 +1524,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/falcon/modeling_falcon.py

@@ -1088,7 +1088,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/gemma/modeling_gemma.py

@@ -677,7 +677,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/glm/modeling_glm.py

@@ -687,7 +687,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py

@@ -927,7 +927,7 @@ def forward(
                     # [batch_size, target_length, 1, source_length], not compatible with SDPA, hence this transpose.
                     self_attention_mask = self_attention_mask.transpose(1, 2)
 
-                if query_length > 1 and attention_mask is not None and attention_mask.device.type == "cuda":
+                if query_length > 1 and attention_mask is not None and attention_mask.device.type in ["cuda", "xpu"]:
                     # From PyTorch 2.1 onwards, F.scaled_dot_product_attention with the memory-efficient attention backend
                     # produces nans if sequences are completely unattended in the attention mask. Details: https://github.com/pytorch/pytorch/issues/110213
                     self_attention_mask = AttentionMaskConverter._unmask_unattended(

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -837,7 +837,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -681,7 +681,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py

@@ -708,7 +708,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/gptj/modeling_gptj.py

@@ -936,7 +936,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/granite/modeling_granite.py

@@ -690,7 +690,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/granitemoe/modeling_granitemoe.py

@@ -1177,7 +1177,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/helium/modeling_helium.py

@@ -674,7 +674,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/idefics/modeling_idefics.py

@@ -1408,7 +1408,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/jamba/modeling_jamba.py

@@ -1381,7 +1381,7 @@ def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.

src/transformers/models/jetmoe/modeling_jetmoe.py

@@ -1170,7 +1170,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/llama/modeling_llama.py

@@ -676,7 +676,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/longt5/modeling_longt5.py

@@ -1645,7 +1645,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/mimi/modeling_mimi.py

@@ -1133,7 +1133,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/mistral/modeling_mistral.py

@@ -667,7 +667,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/mistral/modular_mistral.py

@@ -185,7 +185,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/mixtral/modeling_mixtral.py

@@ -801,7 +801,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/mllama/modeling_mllama.py

@@ -1123,7 +1123,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/moonshine/modeling_moonshine.py

@@ -1040,7 +1040,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/moshi/modeling_moshi.py

@@ -1363,7 +1363,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
@@ -1675,7 +1675,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/mt5/modeling_mt5.py

@@ -1237,7 +1237,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/nemotron/modeling_nemotron.py

@@ -924,7 +924,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/olmo/modeling_olmo.py

@@ -652,7 +652,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/olmo2/modeling_olmo2.py

@@ -653,7 +653,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/olmoe/modeling_olmoe.py

@@ -1084,7 +1084,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when

src/transformers/models/persimmon/modeling_persimmon.py

@@ -725,7 +725,7 @@ def _update_causal_mask(
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
-            and attention_mask.device.type == "cuda"
+            and attention_mask.device.type in ["cuda", "xpu"]
             and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when