Flax dtype-dependent numerical masking (#21197)

src/transformers/models/albert/modeling_flax_albert.py

@@ -245,7 +245,7 @@ class FlaxAlbertSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/bart/modeling_flax_bart.py

@@ -371,7 +371,7 @@ class FlaxBartAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/bert/modeling_flax_bert.py

@@ -358,7 +358,7 @@ class FlaxBertSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/big_bird/modeling_flax_big_bird.py

@@ -380,7 +380,7 @@ class FlaxBigBirdSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/blenderbot/modeling_flax_blenderbot.py

@@ -359,7 +359,7 @@ class FlaxBlenderbotAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/blenderbot_small/modeling_flax_blenderbot_small.py

@@ -371,7 +371,7 @@ class FlaxBlenderbotSmallAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/clip/modeling_flax_clip.py

@@ -315,7 +315,7 @@ class FlaxCLIPAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e4).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/electra/modeling_flax_electra.py

@@ -326,7 +326,7 @@ class FlaxElectraSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/gpt2/modeling_flax_gpt2.py

@@ -255,7 +255,7 @@ class FlaxGPT2Attention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e4).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/gpt_neo/modeling_flax_gpt_neo.py

@@ -223,7 +223,7 @@ class FlaxGPTNeoSelfAttention(nn.Module):
         attention_bias = lax.select(
             attention_mask > 0,
             jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-            jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+            jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
         )
 
         # usual dot product attention

src/transformers/models/gptj/modeling_flax_gptj.py

@@ -270,7 +270,7 @@ class FlaxGPTJAttention(nn.Module):
         attention_bias = lax.select(
             attention_mask > 0,
             jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-            jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+            jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
         )
 
         # usual dot product attention

src/transformers/models/marian/modeling_flax_marian.py

@@ -381,7 +381,7 @@ class FlaxMarianAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/mbart/modeling_flax_mbart.py

@@ -383,7 +383,7 @@ class FlaxMBartAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/opt/modeling_flax_opt.py

@@ -245,7 +245,7 @@ class FlaxOPTAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/pegasus/modeling_flax_pegasus.py

@@ -375,7 +375,7 @@ class FlaxPegasusAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e9).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/roberta/modeling_flax_roberta.py

@@ -319,7 +319,7 @@ class FlaxRobertaSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/roberta_prelayernorm/modeling_flax_roberta_prelayernorm.py

@@ -321,7 +321,7 @@ class FlaxRobertaPreLayerNormSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/roformer/modeling_flax_roformer.py

@@ -240,7 +240,7 @@ class FlaxRoFormerSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/wav2vec2/modeling_flax_wav2vec2.py

@@ -497,7 +497,7 @@ class FlaxWav2Vec2Attention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, float("-inf")).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None

src/transformers/models/xlm_roberta/modeling_flax_xlm_roberta.py

@@ -329,7 +329,7 @@ class FlaxXLMRobertaSelfAttention(nn.Module):
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
-                jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
             )
         else:
             attention_bias = None