diff --git a/runtimes/common/packages.bzl b/runtimes/common/packages.bzl
index b5e20cf..bf30258 100644
--- a/runtimes/common/packages.bzl
+++ b/runtimes/common/packages.bzl
@@ -10,6 +10,22 @@ def _kwargs(**kwargs):
 def _cc_import(**kwargs):
     return """cc_import({})""".format(_kwargs(**kwargs))
 
+def _cc_import_glob_hdrs(name, hdrs_glob, shared_library, deps = []):
+    return """\
+filegroup(
+    name = "{name}_files",
+    srcs = glob(["{hdrs_glob}"]),
+    visibility = ["//visibility:public"],
+)
+cc_import(
+    name = "{name}",
+    shared_library = {shared_library},
+    hdrs = [":{name}_files"],
+    deps = {deps},
+    visibility = ["//visibility:public"],
+)
+""".format(name = name, hdrs_glob = hdrs_glob, shared_library = repr(shared_library), deps = repr(deps))
+
 def _filegroup(**kwargs):
     return """filegroup({})""".format(_kwargs(**kwargs))
 
@@ -59,6 +75,7 @@ common_apt_packages = module_extension(
 packages = struct(
     read = _read,
     cc_import = _cc_import,
+    cc_import_glob_hdrs = _cc_import_glob_hdrs,
     filegroup = _filegroup,
     load_ = _load,
 )
diff --git a/runtimes/cuda/cuda.bzl b/runtimes/cuda/cuda.bzl
index 58fec19..82d1e44 100644
--- a/runtimes/cuda/cuda.bzl
+++ b/runtimes/cuda/cuda.bzl
@@ -25,6 +25,11 @@ CUDA_PACKAGES = {
         name = "cupti",
         shared_library = "lib/libcupti.so.12",
     ),
+    "cuda_nvtx": packages.cc_import_glob_hdrs(
+        name = "nvtx",
+        hdrs_glob = "include/nvtx3/**/*.h",
+        shared_library = "lib/libnvToolsExt.so.1",
+    ),
     "libcufft": packages.cc_import(
         name = "cufft",
         shared_library = "lib/libcufft.so.11",
diff --git a/runtimes/cuda/libpjrt_cuda.BUILD.bazel b/runtimes/cuda/libpjrt_cuda.BUILD.bazel
index 4412cd5..8881a01 100644
--- a/runtimes/cuda/libpjrt_cuda.BUILD.bazel
+++ b/runtimes/cuda/libpjrt_cuda.BUILD.bazel
@@ -35,6 +35,7 @@ cc_import(
         ":zmlxcuda",
         "@cuda_cudart//:cudart",
         "@cuda_cupti//:cupti",
+        "@cuda_nvtx//:nvtx",
         "@cuda_nvcc//:nvptxcompiler",
         "@cuda_nvcc//:nvvm",
         "@cuda_nvrtc//:nvrtc",
diff --git a/zml/tools/tracer.zig b/zml/tools/tracer.zig
index ed7d7fe..0576956 100644
--- a/zml/tools/tracer.zig
+++ b/zml/tools/tracer.zig
@@ -1,13 +1,50 @@
 const builtin = @import("builtin");
+const c = @import("c");
 
 pub const Tracer = switch (builtin.os.tag) {
     .macos => MacOsTracer,
+    .linux => if (@hasDecl(c, "ZML_RUNTIME_CUDA")) CudaTracer else FakeTracer,
     else => FakeTracer,
 };
 
-const MacOsTracer = struct {
-    const c = @import("c");
+const CudaTracer = struct {
+    extern fn cudaProfilerStart() c_int;
+    extern fn cudaProfilerStop() c_int;
 
+    extern fn nvtxMarkA(message: [*:0]const u8) void;
+    extern fn nvtxRangeStartA(message: [*:0]const u8) c_int;
+    extern fn nvtxRangeEnd(id: c_int) void;
+
+    pub fn init(name: [:0]const u8) CudaTracer {
+        _ = name;
+        _ = cudaProfilerStart();
+        return .{};
+    }
+
+    pub fn deinit(self: *const CudaTracer) void {
+        _ = self;
+        _ = cudaProfilerStop();
+    }
+
+    pub fn event(self: *const CudaTracer, message: [:0]const u8) void {
+        _ = self;
+        nvtxMarkA(message.ptr);
+    }
+
+    pub fn frameStart(self: *const CudaTracer, message: [:0]const u8) u64 {
+        _ = self;
+        return @intCast(nvtxRangeStartA(message.ptr));
+    }
+
+    pub fn frameEnd(self: *const CudaTracer, interval_id: u64, message: [:0]const u8) void {
+        _ = self;
+        _ = message;
+        nvtxRangeEnd(@intCast(interval_id));
+        return;
+    }
+};
+
+const MacOsTracer = struct {
     logger: c.os_log_t,
 
     pub fn init(name: [:0]const u8) MacOsTracer {
@@ -40,7 +77,7 @@ const FakeTracer = struct {
         return .{};
     }
 
-    pub fn event(self: *const MacOsTracer, message: [:0]const u8) void {
+    pub fn event(self: *const FakeTracer, message: [:0]const u8) void {
         _ = self;
         _ = message;
         return;