compare to official llama2

tests/testthat/test-llama.R

@@ -1,13 +1,14 @@
 test_that("Can create a llama model", {
   skip_on_ci() # this is too big for the github runners.
-  model <- llama_from_pretrained("huggyllama/llama-7b")
+  model <- llama_from_pretrained("meta-llama/Llama-2-7b-chat-hf")
   model$to(dtype=torch_float32())
   model$eval()
   with_no_grad({
     pred <- model(torch_ones(1, 3, dtype="long") + 1L)
   })
 
   out <- pred[1, 1, 1:5]
-  reference <- c(-12.7782, -28.6373,   0.9082,  -6.1501,  -4.3769)
-  expect_equal(as.numeric(out), reference, tolerance = 1e-5)
+
+  reference <- c(0.2226,  0.0299,  0.2729, -0.7919,  1.6164)
+  expect_equal(as.numeric(out), reference, tolerance = 1e-4)
 })