Unverified Commit cbaaa2f6 authored by Joao Gante's avatar Joao Gante Committed by GitHub
Browse files

Flax dtype-dependent numerical masking (#21197)

parent 0b86e330
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
......@@ -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
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment