set cuda env vars for tensorflow

R/install.R

@@ -149,7 +149,7 @@ function(method = c("auto", "virtualenv", "conda"),
   ))
 
   if(is_linux() && version %in% c("default", "release", "2.13"))
-    packages <- c("nvidia-cudnn-cu11==8.6.*")
+    packages <- c(packages, "nvidia-cudnn-cu11==8.6.*")
 
   if (isTRUE(new_env)) {
 

R/package.R

@@ -31,6 +31,46 @@ tf_v2 <- function() {
 .globals <- new.env(parent = emptyenv())
 .globals$tensorboard <- NULL
 
+set_cuda_env_vars <- function(python_too = FALSE) {
+
+  # For TF 2.13, this assumes that someone already has cudn 11-8 installed,
+  # e.g., on ubuntu:
+  # sudo apt install cuda-toolkit-11-8
+
+  python <- py_discover_config()$python
+  cudnn_module_path <- suppressWarnings(system2(python, c("-c",
+    shQuote("import nvidia.cudnn;print(nvidia.cudnn.__file__)")),
+    stdout = TRUE, stderr = TRUE))
+  if (!is.null(attr(cudnn_module_path, "status")) ||
+      !is_string(cudnn_module_path) ||
+      !file.exists(cudnn_module_path))
+    return()
+
+  cudnn_path <- dirname(cudnn_module_path)
+  ld_library_path <- paste0(c(paste0(cudnn_path, "/lib"),
+                              Sys.getenv("LD_LIBRARY_PATH")),
+                            collapse = ":")
+
+  vars <- list(CUDNN_PATH = cudnn_path,
+               LD_LIBRARY_PATH = ld_library_path)
+
+  out <- Sys.getenv(names(vars), unset = NA, names = TRUE)
+
+  do.call(Sys.setenv, vars)
+
+  if(py_available()) {
+    # none of these 3 approaches seem to do anything for tensorflow
+    import("os", convert = FALSE)$environ$update(vars)
+    # environ <- (os <- import("os", convert = FALSE))$environ
+    # for(i in seq_along(vars)) {
+    #   py_set_item(environ, names(vars)[[i]], vars[[i]])
+    #   os$putenv(names(vars)[[i]], vars[[i]])
+    # }
+  }
+
+  invisible(out)
+}
+
 .onLoad <- function(libname, pkgname) {
 
   # if TENSORFLOW_PYTHON is defined then forward it to RETICULATE_PYTHON
@@ -44,13 +84,30 @@ tf_v2 <- function() {
   if (!is.null(cpp_log_opt))
     Sys.setenv(TF_CPP_MIN_LOG_LEVEL = max(min(cpp_log_opt, 1), 0))
 
+  if (is_linux()) {
+    if (!py_available())
+      options("reticulate.python.beforeInitialized" = set_cuda_env_vars)
+    else {
+      # This is likely too little too late: it would seem that LD_LIBRARY_PATH
+      # *must* be set before python symbols are loaded (that is, before reticulate
+      # initialized python). calling os.putenv()/os.environ.set() after python
+      # has initialized but before importing the tensorflow module (e.g., from a
+      # before_load() module import hook) don't seem to get the job done.
+      set_cuda_env_vars()
+    }
+  }
+
   # delay load tensorflow
   tf <<- import("tensorflow", delay_load = list(
 
-    priority = 5,
+    priority = 5, # keras sets priority = 10
 
     environment = "r-tensorflow",
 
+    # before_load = function() {
+    #
+    # },
+
     on_load = function() {
 
       # register warning suppression handler
@@ -116,6 +173,9 @@ tf_v2 <- function() {
 }
 
 
+is_string <- function(x) {
+  is.character(x) && length(x) == 1L && !is.na(x)
+}
 
 #' TensorFlow configuration information
 #'