Revert CUDA PJRT plugin version to 0.4.38 to address performance regression on XLA master.

runtimes/cuda/cuda.bzl

@@ -208,8 +208,9 @@ def _cuda_impl(mctx):
     http_archive(
         name = "libpjrt_cuda",
         build_file = "libpjrt_cuda.BUILD.bazel",
-        url = "https://github.com/zml/pjrt-artifacts/releases/download/v5.0.0/pjrt-cuda_linux-amd64.tar.gz",
+        url = "https://files.pythonhosted.org/packages/90/43/ac2c369e202e3e3e7e5aa7929b197801ba02eaf11868437adaa5341704e4/jax_cuda12_pjrt-0.4.38-py3-none-manylinux2014_x86_64.whl",
-        sha256 = "1c3ca76d887d112762d03ebb28f17a08beebf6338453c3044a36225e1678a113",
+        type = "zip",
+        sha256 = "83be4c59fbcf30077a60085d98e7d59dc738b1c91e0d628e4ac1779fde15ac2b",
     )
 
     return mctx.extension_metadata(

runtimes/cuda/cuda.zig

@@ -1,10 +1,16 @@
 const builtin = @import("builtin");
 const std = @import("std");
-const asynk = @import("async");
-const pjrt = @import("pjrt");
-const c = @import("c");
 
-const nvidiaLibsPath = "/usr/local/cuda/lib64";
+const asynk = @import("async");
+const bazel_builtin = @import("bazel_builtin");
+const c = @import("c");
+const pjrt = @import("pjrt");
+const runfiles = @import("runfiles");
+const stdx = @import("stdx");
+
+const nvidiaLibsPath = "/cuda/";
+
+const log = std.log.scoped(.@"zml/runtime/cuda");
 
 pub fn isEnabled() bool {
     return @hasDecl(c, "ZML_RUNTIME_CUDA");
@@ -22,7 +28,24 @@ fn hasCudaPathInLDPath() bool {
         return false;
     }
 
-    return std.mem.indexOf(u8, std.mem.span(ldLibraryPath), nvidiaLibsPath) != null;
+    return std.ascii.indexOfIgnoreCase(std.mem.span(ldLibraryPath), nvidiaLibsPath) != null;
+}
+
+fn setupXlaGpuCudaDirFlag() !void {
+    var arena = std.heap.ArenaAllocator.init(std.heap.c_allocator);
+    defer arena.deinit();
+
+    var r_ = try runfiles.Runfiles.create(.{ .allocator = arena.allocator() }) orelse {
+        stdx.debug.panic("Unable to find CUDA directory", .{});
+    };
+
+    const source_repo = bazel_builtin.current_repository;
+    const r = r_.withSourceRepo(source_repo);
+    const cuda_data_dir = (try r.rlocationAlloc(arena.allocator(), "libpjrt_cuda/sandbox")).?;
+    const xla_flags = std.process.getEnvVarOwned(arena.allocator(), "XLA_FLAGS") catch "";
+    const new_xla_flagsZ = try std.fmt.allocPrintZ(arena.allocator(), "{s} --xla_gpu_cuda_data_dir={s}", .{ xla_flags, cuda_data_dir });
+
+    _ = c.setenv("XLA_FLAGS", new_xla_flagsZ, 1);
 }
 
 pub fn load() !*const pjrt.Api {
@@ -36,8 +59,12 @@ pub fn load() !*const pjrt.Api {
         return error.Unavailable;
     }
     if (hasCudaPathInLDPath()) {
-        std.log.warn("Detected {s} in LD_LIBRARY_PATH. This can lead to undefined behaviors and crashes", .{nvidiaLibsPath});
+        log.warn("Detected {s} in LD_LIBRARY_PATH. This can lead to undefined behaviors and crashes", .{nvidiaLibsPath});
     }
 
+    // CUDA path has to be set _before_ loading the PJRT plugin.
+    // See https://github.com/openxla/xla/issues/21428
+    try setupXlaGpuCudaDirFlag();
+
     return try asynk.callBlocking(pjrt.Api.loadFrom, .{"libpjrt_cuda.so"});
 }

runtimes/cuda/libpjrt_cuda.BUILD.bazel

@@ -25,7 +25,8 @@ copy_to_directory(
 cc_import(
     name = "libpjrt_cuda",
     data = [":sandbox"],
-    shared_library = "libpjrt_cuda.so",
+    shared_library = "jax_plugins/xla_cuda12/xla_cuda_plugin.so",
+    soname = "libpjrt_cuda.so",
     add_needed = ["libzmlxcuda.so.0"],
     rename_dynamic_symbols = {
         "dlopen": "zmlxcuda_dlopen",

zml/module.zig

@@ -900,7 +900,7 @@ fn compileModuleToPjrtExecutable(arena: std.mem.Allocator, platform: Platform, m
         }
     }
     switch (platform.target) {
-        .cuda => cuda_dir: {
+        .cuda => {
             // NVIDIA recommends these settings
             // https://github.com/NVIDIA/JAX-Toolbox?tab=readme-ov-file#environment-variables
             setFlag(&options, "xla_gpu_enable_triton_gemm", false);
@@ -914,17 +914,6 @@ fn compileModuleToPjrtExecutable(arena: std.mem.Allocator, platform: Platform, m
             //  setFlag(&options, "xla_gpu_enable_dynamic_slice_fusion", true);
             //  setFlag(&options, "xla_gpu_enable_while_loop_double_buffering", true);
             //  setFlag(&options, "xla_gpu_use_runtime_fusion", true);
-
-            var r_ = try runfiles.Runfiles.create(.{ .allocator = arena }) orelse {
-                log.warn("Bazel runfile not found !", .{});
-                break :cuda_dir;
-            };
-            defer r_.deinit(arena);
-            const source_repo = @import("bazel_builtin").current_repository;
-            const r = r_.withSourceRepo(source_repo);
-            const cuda_data_dir = (try r.rlocationAlloc(arena, "libpjrt_cuda/sandbox")).?;
-            log.debug("xla_gpu_cuda_data_dir: {s}", .{cuda_data_dir});
-            setFlag(&options, "xla_gpu_cuda_data_dir", cuda_data_dir);
         },
         .rocm => {
             // Disable Triton GEMM on ROCM. For some reason it's much, much slower when