`optim_adam()` results differ across platforms even with `torch_manual_seed()`

[yiqi-liu]

Hi there,

I'd like to use optim_adam() for a use case that requires reproducible results across platforms. I set both set.seed() and torch_manual_seed(), but still got different results on macOS (ARM) and Linux (x86_64). Here’s a simple example taken from the optimizer tutorial:

library(torch)
set.seed(0)
torch_manual_seed(0)

### generate training data -----------------------------------------------------

# input dimensionality (number of input features)
d_in <- 3
# output dimensionality (number of predicted features)
d_out <- 1
# number of observations in training set
n <- 100


# create random data
x <- torch_randn(n, d_in)
y <- x[, 1, NULL] * 0.2 - x[, 2, NULL] * 1.3 - x[, 3, NULL] * 0.5 + torch_randn(n, 1)



### define the network ---------------------------------------------------------

# dimensionality of hidden layer
d_hidden <- 32

model <- nn_sequential(
  nn_linear(d_in, d_hidden),
  nn_relu(),
  nn_linear(d_hidden, d_out)
)

### network parameters ---------------------------------------------------------

# for adam, need to choose a much higher learning rate in this problem
learning_rate <- 0.08

optimizer <- optim_adam(model$parameters, lr = learning_rate)

### training loop --------------------------------------------------------------

for (t in 1:200) {
  
  ### -------- Forward pass -------- 
  
  y_pred <- model(x)
  
  ### -------- compute loss -------- 
  loss <- nnf_mse_loss(y_pred, y, reduction = "sum")
  if (t %% 10 == 0)
    cat("Epoch: ", t, "   Loss: ", loss$item(), "\n")
  
  ### -------- Backpropagation -------- 
  
  # Still need to zero out the gradients before the backward pass, only this time,
  # on the optimizer object
  optimizer$zero_grad()
  
  # gradients are still computed on the loss tensor (no change here)
  loss$backward()
  
  ### -------- Update weights -------- 
  
  # use the optimizer to update model parameters
  optimizer$step()
}

# CHECK REPRODUCIBILITY: Print final bias of last layer
sprintf("%.20f", model$parameters$`2.bias`)

Results:

• macOS (Apple Silicon)

> sessionInfo()
R version 4.4.1 (2024-06-14)
Platform: aarch64-apple-darwin20
Running under: macOS Sonoma 14.6.1

> sprintf("%.20f", model$parameters$`2.bias`)
[1] "0.12490314245223999023"

• Linux (x86_64)

> sessionInfo()
R version 4.4.2 (2024-10-31)
Platform: x86_64-pc-linux-gnu
Running under: Ubuntu 24.04.1 LTS

> sprintf("%.20f", model$parameters$`2.bias`)
[1] "0.13555328547954559326"

Even though the seed is fixed, results still differ across platforms. Is full cross-platform reproducibility currently feasible with torch in R? Or is there something additional I should configure (e.g., thread settings, environment variables)?

Thanks so much for making this amazing package available!