fix: make exla build resilient to stale upgrades

exla/Makefile

@@ -8,8 +8,8 @@ XLA_EXTENSION_LIB = $(XLA_EXTENSION_DIR)/lib
 XLA_INCLUDE_PATH = $(XLA_EXTENSION_DIR)/include
 
 # Cache configuration
-EXLA_CACHE_SO = cache/libexla.so
-EXLA_CACHE_OBJ_DIR = cache/objs
+EXLA_CACHE_SO = cache/$(EXLA_VERSION)/libexla.so
+EXLA_CACHE_OBJ_DIR = cache/$(EXLA_VERSION)/objs
 
 # Private configuration
 EXLA_DIR = c_src/exla

exla/README.md

@@ -48,6 +48,26 @@ EXLA relies on the [XLA](https://github.com/elixir-nx/xla) package to provide th
 
 For cross-compilation, you need to [set your `XLA_TARGET_PLATFORM` variable](https://github.com/elixir-nx/xla#xla_target_platform) to the correct target platform value (i.e. `aarch64-linux-gnu` for the Raspberry Pi 4).
 
+## Troubleshooting
+
+EXLA uses NIFs (C-interface code called from Elixir) for part of its functionality.
+If for any reason these fail to compile or load, troubleshooting can be tricky.
+
+We recommend following the steps below:
+
+  1. If the error appeared after upgrading EXLA, ensure that you have the proper dependency versions given by [XLA](https://github.com/elixir-nx/xla). Afterwards, compile with `mix compile` after setting `EXLA_FORCE_REBUILD` to clean up cached files:
+     * `EXLA_FORCE_REBUILD=partial`: Removes the only the libexla.so caches (both local and global ones).
+     * `EXLA_FORCE_REBUILD=true`: Removes the libexla.so caches but also removes the intermediate `.o` compilation artifacts retained from previous builds.
+
+     Additional notes on compilation:
+     * Besides the XLA dependency versions, ensuring `gcc` (or your compiler of choice), `libc` and `make` are compatible is also important.
+     * Remember to save the compilation logs from this step for further debugging.
+     * It is a good idea to save the `cache/<version>/libexla.so` file so that the team can inspect its contents if needed.
+  2. If the error persists, look for the `** (RuntimeError) Failed to load NIF library.` exception on application start-up.
+    This exception should provide more information on what's the issue when loading the NIF. Share these logs in an issue on GitHub
+    so that the Nx team can investigate further.
+
+
 ## Contributing
 
 ### Building locally

exla/lib/exla/nif.ex

@@ -4,7 +4,20 @@ defmodule EXLA.NIF do
 
   def __on_load__ do
     path = :filename.join(:code.priv_dir(:exla), ~c"libexla")
-    :erlang.load_nif(path, 0)
+
+    case :erlang.load_nif(path, 0) do
+      :ok ->
+        :ok
+
+      {:error, {reason, text}} ->
+        raise """
+        Failed to load NIF library.
+        Follow the steps in the :exla README Troubleshooting section for more information.
+
+        #{inspect(reason)}
+        #{text}
+        """
+    end
   end
 
   def mlir_new_thread_pool(_concurrency), do: :erlang.nif_error(:undef)

exla/mix.exs

@@ -35,7 +35,8 @@ defmodule EXLA.MixProject do
 
         %{
           "MIX_BUILD_EMBEDDED" => "#{Mix.Project.config()[:build_embedded]}",
-          "CWD_RELATIVE_TO_PRIV_PATH" => cwd_relative_to_priv
+          "CWD_RELATIVE_TO_PRIV_PATH" => cwd_relative_to_priv,
+          "EXLA_VERSION" => "#{@version}"
         }
       end,
       make_args: make_args
@@ -133,7 +134,38 @@ defmodule EXLA.MixProject do
     {:ok, []}
   end
 
-  defp cached_make(_) do
+  defp cached_make(args) do
+    force_rebuild_mode =
+      case System.get_env("EXLA_FORCE_REBUILD", "") do
+        "" ->
+          :none
+
+        "0" ->
+          :none
+
+        "partial" ->
+          :partial
+
+        "true" ->
+          :full
+
+        "1" ->
+          :full
+
+        value ->
+          Mix.raise(
+            "invalid value for EXLA_FORCE_REBUILD: '#{value}'. Expected one of: partial, true"
+          )
+      end
+
+    File.mkdir_p!("cache/#{@version}")
+
+    # remove only in full mode
+    if force_rebuild_mode in [:partial, :full] do
+      Mix.shell().info("Removing cached .o files in cache/#{@version}/objs")
+      File.rm_rf!("cache/#{@version}/objs")
+    end
+
     contents =
       for path <- Path.wildcard("c_src/**/*"),
           {:ok, contents} <- [File.read(path)],
@@ -148,19 +180,27 @@ defmodule EXLA.MixProject do
       "elixir-#{System.version()}-erts-#{:erlang.system_info(:version)}-xla-#{Application.spec(:xla, :vsn)}-exla-#{@version}-#{md5}"
 
     cached_so = Path.join([xla_cache_dir(), "exla", cache_key, "libexla.so"])
-    cached? = File.exists?(cached_so)
+    cached? = File.exists?(cached_so) and force_rebuild_mode == :none
+
+    if force_rebuild_mode in [:partial, :full] do
+      Mix.shell().info("Removing cached libexla.so file in cache/#{@version}/libexla.so")
+      File.rm_rf!("cache/#{@version}/libexla.so")
+
+      Mix.shell().info("Removing libexla.so cache at #{cached_so}")
+      File.rm!(cached_so)
+    end
 
     if cached? do
       Mix.shell().info("Using libexla.so from #{cached_so}")
-      File.cp!(cached_so, "cache/libexla.so")
+      File.cp!(cached_so, "cache/#{@version}/libexla.so")
     end
 
-    result = Mix.Tasks.Compile.ElixirMake.run([])
+    result = Mix.Tasks.Compile.ElixirMake.run(args)
 
     if not cached? and match?({:ok, _}, result) do
       Mix.shell().info("Caching libexla.so at #{cached_so}")
       File.mkdir_p!(Path.dirname(cached_so))
-      File.cp!("cache/libexla.so", cached_so)
+      File.cp!("cache/#{@version}/libexla.so", cached_so)
     end
 
     result

exla/test/exla/device_memory_sharing_test.exs

@@ -27,14 +27,13 @@ defmodule EXLA.DeviceMemorySharingTest do
   end
 
   @tag :cuda_required
-  test "ipc handles don't crash the runtime when :local mode is selected" do
-    assert {:error, ~c"Invalid pointer size for selected mode."} ==
+  test "invalid ipc handles don't crash the runtime" do
+    assert {:error, ~c"Unable to get pointer for IPC handle."} ==
              Nx.from_pointer(
                {EXLA.Backend, client: :cuda},
-               Enum.to_list(0..63),
+               %Nx.Pointer{handle: "#{System.unique_integer()}", kind: :ipc, data_size: 4},
                {:f, 32},
-               {1},
-               mode: :local
+               {1}
              )
   end
 end