Add transformer engine package (#43982)

* Add py-flash-attn@2.4.2
* Add py-transfomer-engine package

---------

Co-authored-by: Tamara Dahlgren <35777542+tldahlgren@users.noreply.github.com>

var/spack/repos/builtin/packages/py-transformer-engine/package.py

@@ -0,0 +1,48 @@
+# Copyright 2013-2024 Lawrence Livermore National Security, LLC and other
+# Spack Project Developers. See the top-level COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+from spack.package import *
+
+
+class PyTransformerEngine(PythonPackage):
+    """
+    A library for accelerating Transformer models on NVIDIA GPUs, including fp8 precision on Hopper
+    GPUs.
+    """
+
+    homepage = "https://github.com/NVIDIA/TransformerEngine"
+    url = "https://github.com/NVIDIA/TransformerEngine/archive/refs/tags/v0.0.tar.gz"
+    git = "https://github.com/NVIDIA/TransformerEngine.git"
+    maintainers("aurianer")
+
+    license("Apache-2.0")
+
+    version("1.4", tag="v1.4", submodules=True)
+    version("main", branch="main", submodules=True)
+
+    variant("userbuffers", default=True, description="Enable userbuffers, this option needs MPI.")
+
+    depends_on("py-setuptools", type="build")
+    depends_on("cmake@3.18:")
+    depends_on("py-pydantic")
+    depends_on("py-importlib-metadata")
+
+    with default_args(type=("build", "run")):
+        depends_on("py-accelerate")
+        depends_on("py-datasets")
+        depends_on("py-flash-attn@2.2:2.4.2")
+        depends_on("py-packaging")
+        depends_on("py-torchvision")
+        depends_on("py-transformers")
+        depends_on("mpi", when="+userbuffers")
+
+    with default_args(type=("build", "link", "run")):
+        depends_on("py-torch+cuda+cudnn")
+
+    def setup_build_environment(self, env):
+        env.set("NVTE_FRAMEWORK", "pytorch")
+        if self.spec.satisfies("+userbuffers"):
+            env.set("NVTE_WITH_USERBUFFERS", "1")
+            env.set("MPI_HOME", self.spec["mpi"].prefix)