I am investigating some gradients and noticed that
julia> gradients == gradients
false
and when I look at them I get different values from time to time
julia> gradients.head_ibot.chain.dense_1.bias
2048-element ConcreteIFRTArray{Float32,1} with "mhlo.sharding = {replicated}":
59177.938
0.0
-5.183805f-34
0.0
⋮
0.0
30460.812
0.0
julia> gradients.head_ibot.chain.dense_1.bias
2048-element ConcreteIFRTArray{Float32,1} with "mhlo.sharding = {replicated}":
59177.938
0.0
-45811.688
0.0
⋮
0.0
0.0
0.0
julia> gradients.head_ibot.chain.dense_1.bias
2048-element ConcreteIFRTArray{Float32,1} with "mhlo.sharding = {replicated}":
-2.6467664f-30
0.0
-1.1724075f16
0.0
⋮
0.0
0.0
0.0
julia> gradients.head_ibot.chain.dense_1.bias
2048-element ConcreteIFRTArray{Float32,1} with "mhlo.sharding = {replicated}":
1.0809131f31
0.0
-1.1724075f16
0.0
⋮
NaN
1.54f-43
0.0
I am trying to work out an MWE
I am investigating some gradients and noticed that
julia> gradients == gradients falseand when I look at them I get different values from time to time
I am trying to work out an MWE