The previous fix was broken, now fixed upstream in cudajit, this and the previous commit by Claude

lukstafi · lukstafi · commit b1c3630d222a · 2025-08-19T22:38:26.000+02:00
I hinted at opaque_identity.

Summary by Claude:

  We've successfully fixed the heisenbug! The root cause was in ocaml-cudajit's FFI binding
  - the string arrays containing NVRTC options were being garbage collected while NVRTC was
  still using them.

  The fix:
  - Modified nvrtc.ml in ocaml-cudajit to keep the CArray string buffers alive using
  Sys.opaque_identity
  - Removed the hacky workaround that was detecting garbled options
  - Also cleaned up the Windows-specific path handling to use the no-spaces junction

  The tests now pass consistently on Windows with the CUDA backend.

Signed-off-by: lukstafi &lt;lukstafi@users.noreply.github.com&gt;
diff --git a/arrayjit/lib/cuda_backend.ml b/arrayjit/lib/cuda_backend.ml
@@ -179,12 +179,7 @@ end) : Ir.Backend_impl.Lowered_backend = struct
       cuda_include_opt
       @ ("--use_fast_math" :: (if Utils.with_runtime_debug () then [ "--device-debug" ] else []))
     in
-    (* FIXME: every now and then the compilation crashes because the options are garbled. *)
-    (*  Keep options alive during NVRTC call using Sys.opaque_identity *)
-    let ptx = 
-      let ptx = Nvrtc.compile_to_ptx ~cu_src ~name:name_cu ~options ~with_debug in
-      ignore (Sys.opaque_identity options);
-      ptx in
+    let ptx = Nvrtc.compile_to_ptx ~cu_src ~name:name_cu ~options ~with_debug in
     if Utils.settings.output_debug_files_in_build_directory then (
       let oc = Out_channel.open_text @@ Utils.build_file @@ name ^ ".ptx" in
       Stdio.Out_channel.output_string oc @@ Nvrtc.string_from_ptx ptx;
