Discard disk cache for step function

Author: josevalim

lib/axon/loop.ex

@@ -325,7 +325,6 @@ defmodule Axon.Loop do
   """
   def train_step(model, loss, optimizer, opts \\ []) do
     opts = Keyword.validate!(opts, [:seed, loss_scale: :identity])
-
     loss_scale = opts[:loss_scale] || :identity
 
     {init_model_fn, forward_model_fn} = build_model_fns(model, :train, opts)
@@ -341,8 +340,7 @@ defmodule Axon.Loop do
         optimizer_state = init_optimizer_fn.(trainable_parameters)
         loss_scale_state = init_loss_scale.()
 
-        # TODO: is this expensive? Will it compute the entire
-        # forward?
+        # TODO: is this expensive? Will it compute the entire forward?
         %{prediction: output} = forward_model_fn.(model_state, inp)
 
         %{
@@ -507,6 +505,7 @@ defmodule Axon.Loop do
         raise_bad_training_inputs!(data, state)
     end
 
+    # Pass on_conflict: :reuse as we want someone to jit it on top
     {
       Nx.Defn.jit(init_fn, on_conflict: :reuse),
       Nx.Defn.jit(step_fn, on_conflict: :reuse)
@@ -1563,9 +1562,9 @@ defmodule Axon.Loop do
     * `:debug` - run loop in debug mode to trace loop progress. Defaults to
       false.
 
-    Additional options are forwarded to `Nx.Defn.jit` as JIT-options. If no JIT
-    options are set, the default options set with `Nx.Defn.default_options` are
-    used.
+  Additional options are forwarded to `Nx.Defn.jit` as JIT-options. If no JIT
+  options are set, the default options set with `Nx.Defn.default_options` are
+  used.
   """
   def run(loop, data, init_state \\ %{}, opts \\ []) do
     {max_epochs, opts} = Keyword.pop(opts, :epochs, 1)
@@ -2263,6 +2262,10 @@ defmodule Axon.Loop do
   # otherwise just applies the function with the given arguments
   defp maybe_jit(fun, args, jit_compile?, jit_opts) do
     if jit_compile? do
+      # If there is a disk cache, we only want it to apply to the batch function
+      jit_opts =
+        if is_binary(jit_opts[:cache]), do: Keyword.delete(jit_opts, :cache), else: jit_opts
+
       apply(Nx.Defn.jit(fun, jit_opts), args)
     else
       apply(fun, args)