google-ml-infra · ybaturina · Jan 14, 2026 · Jan 9, 2026 · Jan 12, 2026 · Jan 12, 2026
diff --git a/MODULE.bazel b/MODULE.bazel
@@ -18,13 +18,13 @@ use_repo(pybind11_internal_configure, "pybind11")
 ##############################################################
 # CUDA
 
-cuda_configure_ext = use_extension("//third_party/extensions:cuda_configure.bzl", "cuda_configure_ext")
+cuda_configure_ext = use_extension("//extensions:cuda_configure.bzl", "cuda_configure_ext")
 use_repo(cuda_configure_ext, "local_config_cuda")
 
-cuda_json_ext = use_extension("//third_party/extensions:cuda_json.bzl", "cuda_json_ext")
+cuda_json_ext = use_extension("//extensions:cuda_json.bzl", "cuda_json_ext")
 use_repo(cuda_json_ext, "cuda_redist_json")
 
-cuda_redist_init_ext = use_extension("//third_party/extensions:cuda_redist_init.bzl", "cuda_redist_init_ext")
+cuda_redist_init_ext = use_extension("//extensions:cuda_redist_init.bzl", "cuda_redist_init_ext")
 use_repo(
     cuda_redist_init_ext,
     "cuda_cccl",
@@ -51,22 +51,22 @@ use_repo(
 ##############################################################
 # NCCL configuration
 
-nccl_redist = use_extension("//third_party/extensions:nccl_redist.bzl", "nccl_redist_ext")
+nccl_redist = use_extension("//extensions:nccl_redist.bzl", "nccl_redist_ext")
 use_repo(nccl_redist, "cuda_nccl")
 
-nccl_configure = use_extension("//third_party/extensions:nccl_configure.bzl", "nccl_configure_ext")
+nccl_configure = use_extension("//extensions:nccl_configure.bzl", "nccl_configure_ext")
 use_repo(nccl_configure, "local_config_nccl")
 
 ##############################################################
 # SYCL configuration
 
-sycl_configure = use_extension("//third_party/extensions:sycl_configure.bzl", "sycl_configure_ext")
+sycl_configure = use_extension("//extensions:sycl_configure.bzl", "sycl_configure_ext")
 use_repo(sycl_configure, "local_config_sycl")
 
 ##############################################################
 # Hermetic toolchain configuration
 
-toolchain = use_extension("//third_party/extensions:toolchain.bzl", "toolchain_ext")
+toolchain = use_extension("//extensions:toolchain.bzl", "toolchain_ext")
 use_repo(
     toolchain,
     "llvm18_linux_aarch64",

diff --git a/cc/deps/cc_toolchain_deps.bzl b/cc/deps/cc_toolchain_deps.bzl
@@ -17,7 +17,7 @@ load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
 load("@bazel_tools//tools/build_defs/repo:local.bzl", "new_local_repository")
 load("//common:mirrored_http_archive.bzl", "mirrored_http_archive")
 load("//common:tar_extraction_utils.bzl", "tool_archive")
-load("//third_party:repo.bzl", "tf_mirror_urls")
+load("//common:repo.bzl", "tf_mirror_urls")
 load("//cc/llvms:llvm.bzl", "llvm")
 load("//cc/sysroots:sysroot.bzl", "sysroot")
 

diff --git a/cc/impls/linux_aarch64_linux_aarch64_cuda/wrappers/nvcc_wrapper b/cc/impls/linux_aarch64_linux_aarch64_cuda/wrappers/nvcc_wrapper
@@ -30,8 +30,8 @@ DESCRIPTION:
 
 NOTES:
   Changes to the contents of this file must be propagated from
-  //third_party/gpus/crosstool/crosstool_wrapper_is_not_gcc to
-  //third_party/gpus/crosstool/v*/*/clang/bin/crosstool_wrapper_is_not_gcc
+  //gpu/cuda/legacy/crosstool/crosstool_wrapper_is_not_gcc to
+  //gpu/cuda/legacy/crosstool/v*/*/clang/bin/crosstool_wrapper_is_not_gcc
 """
 
 __author__ = 'keveman@google.com (Manjunath Kudlur)'

diff --git a/cc/impls/linux_x86_64_linux_x86_64_cuda/wrappers/nvcc_wrapper b/cc/impls/linux_x86_64_linux_x86_64_cuda/wrappers/nvcc_wrapper
@@ -30,8 +30,8 @@ DESCRIPTION:
 
 NOTES:
   Changes to the contents of this file must be propagated from
-  //third_party/gpus/crosstool/crosstool_wrapper_is_not_gcc to
-  //third_party/gpus/crosstool/v*/*/clang/bin/crosstool_wrapper_is_not_gcc
+  //gpu/cuda/legacy/crosstool/crosstool_wrapper_is_not_gcc to
+  //gpu/cuda/legacy/crosstool/v*/*/clang/bin/crosstool_wrapper_is_not_gcc
 """
 
 __author__ = 'keveman@google.com (Manjunath Kudlur)'

diff --git a/cc/llvms/llvm.bzl b/cc/llvms/llvm.bzl
@@ -14,7 +14,7 @@
 # ==============================================================================
 
 load(
-    "//third_party/remote_config:common.bzl",
+    "//common:common.bzl",
     "get_host_environ",
 )
 

diff --git a/cc/sysroots/sysroot.bzl b/cc/sysroots/sysroot.bzl
@@ -14,7 +14,7 @@
 # ==============================================================================
 
 load(
-    "//third_party/remote_config:common.bzl",
+    "//common:common.bzl",
     "get_host_environ",
 )
 

diff --git a/common/BUILD b/common/BUILD
@@ -75,6 +75,7 @@ config_setting(
         "@platforms//os:linux",
         "@platforms//cpu:x86_64",
     ],
+    visibility = ["//visibility:public"],
 )
 
 config_setting(
@@ -83,6 +84,27 @@ config_setting(
         "@platforms//os:linux",
         "@platforms//cpu:aarch64",
     ],
+    visibility = ["//visibility:public"],
+)
+
+config_setting(
+    name = "is_linux_armhf",
+    constraint_values =
+        [
+            "@platforms//cpu:armv7e-mf",
+            "@platforms//os:linux",
+        ],
+    visibility = ["//visibility:public"],
+)
+
+config_setting(
+    name = "is_linux_ppc64le",
+    constraint_values =
+        [
+            "@platforms//cpu:ppc64le",
+            "@platforms//os:linux",
+        ],
+    visibility = ["//visibility:public"],
 )
 
 #######################################################
@@ -137,6 +159,19 @@ config_setting(
     flag_values = {":enable_cuda": "False"},
 )
 
+# Config setting for CUDA stub rpaths.
+bool_flag(
+    name = "cuda_rpath",
+    build_setting_default = True,
+)
+
+config_setting(
+    name = "enable_cuda_rpath",
+    flag_values = {
+        ":cuda_rpath": "True",
+        "@local_config_cuda//:enable_cuda": "True"
+    },
+)
 #######################################################
 # Enable SYCL support flags
 

diff --git a/third_party/gpus/crosstool/LICENSE → common/LICENSE b/third_party/gpus/crosstool/LICENSE → common/LICENSE
diff --git a/third_party/remote_config/common.bzl → common/common.bzl b/third_party/remote_config/common.bzl → common/common.bzl
diff --git a/third_party/repo.bzl → common/repo.bzl b/third_party/repo.bzl → common/repo.bzl
@@ -139,22 +139,3 @@ def tf_http_archive(name, sha256, urls, **kwargs):
         urls = urls,
         **kwargs
     )
-
-def _tf_vendored_impl(repository_ctx):
-    parent_path = repository_ctx.path(repository_ctx.attr.parent).dirname
-
-    # get_child doesn't allow slashes. Yes this is silly. bazel_skylib paths
-    # doesn't work with path objects.
-    relpath_parts = repository_ctx.attr.relpath.split("/")
-    vendored_path = parent_path
-    for part in relpath_parts:
-        vendored_path = vendored_path.get_child(part)
-    repository_ctx.symlink(vendored_path, ".")
-
-tf_vendored = repository_rule(
-    implementation = _tf_vendored_impl,
-    attrs = {
-        "parent": attr.label(default = "//:WORKSPACE"),
-        "relpath": attr.string(),
-    },
-)
diff --git a/common/tar_extraction_utils.bzl b/common/tar_extraction_utils.bzl
@@ -14,7 +14,7 @@
 # ==============================================================================
 
 load(
-    "//third_party/remote_config:common.bzl",
+    "//common:common.bzl",
     "execute",
     "get_bash_bin",
 )

diff --git a/third_party/extensions/BUILD → extensions/BUILD b/third_party/extensions/BUILD → extensions/BUILD
diff --git a/third_party/extensions/cuda_configure.bzl → extensions/cuda_configure.bzl b/third_party/extensions/cuda_configure.bzl → extensions/cuda_configure.bzl
@@ -15,7 +15,7 @@
 """CUDA module extension."""
 
 load(
-    "//third_party/gpus/cuda/hermetic:cuda_configure.bzl",
+    "//gpu/cuda:cuda_configure.bzl",
     "cuda_configure",
 )
 

diff --git a/third_party/extensions/cuda_json.bzl → extensions/cuda_json.bzl b/third_party/extensions/cuda_json.bzl → extensions/cuda_json.bzl
diff --git a/third_party/extensions/cuda_redist_init.bzl → extensions/cuda_redist_init.bzl b/third_party/extensions/cuda_redist_init.bzl → extensions/cuda_redist_init.bzl
diff --git a/third_party/extensions/nccl_configure.bzl → extensions/nccl_configure.bzl b/third_party/extensions/nccl_configure.bzl → extensions/nccl_configure.bzl
diff --git a/third_party/extensions/nccl_redist.bzl → extensions/nccl_redist.bzl b/third_party/extensions/nccl_redist.bzl → extensions/nccl_redist.bzl
diff --git a/third_party/extensions/sycl_configure.bzl → extensions/sycl_configure.bzl b/third_party/extensions/sycl_configure.bzl → extensions/sycl_configure.bzl
diff --git a/third_party/extensions/toolchain.bzl → extensions/toolchain.bzl b/third_party/extensions/toolchain.bzl → extensions/toolchain.bzl
diff --git a/third_party/gpus/compiler_common_tools.bzl → gpu/compiler_common_tools.bzl b/third_party/gpus/compiler_common_tools.bzl → gpu/compiler_common_tools.bzl
@@ -1,7 +1,7 @@
 """Common compiler functions. """
 
 load(
-    "//third_party/remote_config:common.bzl",
+    "//common:common.bzl",
     "err_out",
     "raw_exec",
     "realpath",

diff --git a/third_party/gpus/crosstool/BUILD.bazel → gpu/cuda/BUILD.bazel b/third_party/gpus/crosstool/BUILD.bazel → gpu/cuda/BUILD.bazel
diff --git a/third_party/gpus/cuda/hermetic/BUILD.tpl → gpu/cuda/BUILD.tpl b/third_party/gpus/cuda/hermetic/BUILD.tpl → gpu/cuda/BUILD.tpl
diff --git a/third_party/gpus/cuda/LICENSE → gpu/cuda/LICENSE b/third_party/gpus/cuda/LICENSE → gpu/cuda/LICENSE
diff --git a/third_party/gpus/cuda/build_defs.bzl.tpl → gpu/cuda/build_defs.bzl.tpl b/third_party/gpus/cuda/build_defs.bzl.tpl → gpu/cuda/build_defs.bzl.tpl
diff --git a/third_party/gpus/cuda/BUILD.bazel → gpu/cuda/build_templates/BUILD.bazel b/third_party/gpus/cuda/BUILD.bazel → gpu/cuda/build_templates/BUILD.bazel
diff --git a/...ty/gpus/cuda/hermetic/cuda_cccl.BUILD.tpl → gpu/cuda/build_templates/cuda_cccl.BUILD.tpl b/...ty/gpus/cuda/hermetic/cuda_cccl.BUILD.tpl → gpu/cuda/build_templates/cuda_cccl.BUILD.tpl
@@ -1,6 +1,6 @@
 licenses(["restricted"])  # NVIDIA proprietary license
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_lib_header_prefix"
 )
 load("@cuda_cudart//:version.bzl", _cudart_version = "VERSION")

diff --git a/.../cuda/hermetic/cuda_cccl_github.BUILD.tpl → ...uild_templates/cuda_cccl_github.BUILD.tpl b/.../cuda/hermetic/cuda_cccl_github.BUILD.tpl → ...uild_templates/cuda_cccl_github.BUILD.tpl
diff --git a/...rty/gpus/cuda/hermetic/cuda_crt.BUILD.tpl → gpu/cuda/build_templates/cuda_crt.BUILD.tpl b/...rty/gpus/cuda/hermetic/cuda_crt.BUILD.tpl → gpu/cuda/build_templates/cuda_crt.BUILD.tpl
diff --git a/.../gpus/cuda/hermetic/cuda_cublas.BUILD.tpl → ...uda/build_templates/cuda_cublas.BUILD.tpl b/.../gpus/cuda/hermetic/cuda_cublas.BUILD.tpl → ...uda/build_templates/cuda_cublas.BUILD.tpl
@@ -4,7 +4,7 @@ load(
      "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/.../gpus/cuda/hermetic/cuda_cudart.BUILD.tpl → ...uda/build_templates/cuda_cudart.BUILD.tpl b/.../gpus/cuda/hermetic/cuda_cudart.BUILD.tpl → ...uda/build_templates/cuda_cudart.BUILD.tpl
@@ -4,7 +4,7 @@ load(
      "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/...y/gpus/cuda/hermetic/cuda_cudnn.BUILD.tpl → ...cuda/build_templates/cuda_cudnn.BUILD.tpl b/...y/gpus/cuda/hermetic/cuda_cudnn.BUILD.tpl → ...cuda/build_templates/cuda_cudnn.BUILD.tpl
@@ -1,6 +1,6 @@
 licenses(["restricted"])  # NVIDIA proprietary license
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/.../gpus/cuda/hermetic/cuda_cudnn8.BUILD.tpl → ...uda/build_templates/cuda_cudnn8.BUILD.tpl b/.../gpus/cuda/hermetic/cuda_cudnn8.BUILD.tpl → ...uda/build_templates/cuda_cudnn8.BUILD.tpl
@@ -1,6 +1,6 @@
 licenses(["restricted"])  # NVIDIA proprietary license
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/...y/gpus/cuda/hermetic/cuda_cufft.BUILD.tpl → ...cuda/build_templates/cuda_cufft.BUILD.tpl b/...y/gpus/cuda/hermetic/cuda_cufft.BUILD.tpl → ...cuda/build_templates/cuda_cufft.BUILD.tpl
@@ -4,7 +4,7 @@ load(
      "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/...y/gpus/cuda/hermetic/cuda_cupti.BUILD.tpl → ...cuda/build_templates/cuda_cupti.BUILD.tpl b/...y/gpus/cuda/hermetic/cuda_cupti.BUILD.tpl → ...cuda/build_templates/cuda_cupti.BUILD.tpl
@@ -5,7 +5,7 @@ load(
     "if_version_equal_or_greater_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/.../gpus/cuda/hermetic/cuda_curand.BUILD.tpl → ...uda/build_templates/cuda_curand.BUILD.tpl b/.../gpus/cuda/hermetic/cuda_curand.BUILD.tpl → ...uda/build_templates/cuda_curand.BUILD.tpl
@@ -4,7 +4,7 @@ load(
     "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/...pus/cuda/hermetic/cuda_cusolver.BUILD.tpl → ...a/build_templates/cuda_cusolver.BUILD.tpl b/...pus/cuda/hermetic/cuda_cusolver.BUILD.tpl → ...a/build_templates/cuda_cusolver.BUILD.tpl
@@ -4,7 +4,7 @@ load(
     "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/...pus/cuda/hermetic/cuda_cusparse.BUILD.tpl → ...a/build_templates/cuda_cusparse.BUILD.tpl b/...pus/cuda/hermetic/cuda_cusparse.BUILD.tpl → ...a/build_templates/cuda_cusparse.BUILD.tpl
@@ -4,7 +4,7 @@ load(
     "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/.../gpus/cuda/hermetic/cuda_driver.BUILD.tpl → ...uda/build_templates/cuda_driver.BUILD.tpl b/.../gpus/cuda/hermetic/cuda_driver.BUILD.tpl → ...uda/build_templates/cuda_driver.BUILD.tpl
diff --git a/...ty/gpus/cuda/hermetic/cuda_nvcc.BUILD.tpl → gpu/cuda/build_templates/cuda_nvcc.BUILD.tpl b/...ty/gpus/cuda/hermetic/cuda_nvcc.BUILD.tpl → gpu/cuda/build_templates/cuda_nvcc.BUILD.tpl
diff --git a/...ty/gpus/cuda/hermetic/cuda_nvdisasm.BUILD → gpu/cuda/build_templates/cuda_nvdisasm.BUILD b/...ty/gpus/cuda/hermetic/cuda_nvdisasm.BUILD → gpu/cuda/build_templates/cuda_nvdisasm.BUILD
diff --git a/...us/cuda/hermetic/cuda_nvjitlink.BUILD.tpl → .../build_templates/cuda_nvjitlink.BUILD.tpl b/...us/cuda/hermetic/cuda_nvjitlink.BUILD.tpl → .../build_templates/cuda_nvjitlink.BUILD.tpl
@@ -4,7 +4,7 @@ load(
     "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/...ty/gpus/cuda/hermetic/cuda_nvml.BUILD.tpl → gpu/cuda/build_templates/cuda_nvml.BUILD.tpl b/...ty/gpus/cuda/hermetic/cuda_nvml.BUILD.tpl → gpu/cuda/build_templates/cuda_nvml.BUILD.tpl
diff --git a/...rty/gpus/cuda/hermetic/cuda_nvprune.BUILD → gpu/cuda/build_templates/cuda_nvprune.BUILD b/...rty/gpus/cuda/hermetic/cuda_nvprune.BUILD → gpu/cuda/build_templates/cuda_nvprune.BUILD
diff --git a/...y/gpus/cuda/hermetic/cuda_nvrtc.BUILD.tpl → ...cuda/build_templates/cuda_nvrtc.BUILD.tpl b/...y/gpus/cuda/hermetic/cuda_nvrtc.BUILD.tpl → ...cuda/build_templates/cuda_nvrtc.BUILD.tpl
@@ -4,7 +4,7 @@ load(
     "if_cuda_newer_than",
 )
 load(
-    "@rules_ml_toolchain//third_party/gpus:nvidia_common_rules.bzl",
+    "@rules_ml_toolchain//gpu:nvidia_common_rules.bzl",
     "cuda_rpath_flags",
 )
 

diff --git a/...ty/gpus/cuda/hermetic/cuda_nvtx.BUILD.tpl → gpu/cuda/build_templates/cuda_nvtx.BUILD.tpl b/...ty/gpus/cuda/hermetic/cuda_nvtx.BUILD.tpl → gpu/cuda/build_templates/cuda_nvtx.BUILD.tpl
diff --git a/..._party/gpus/cuda/hermetic/cuda_nvvm.BUILD → gpu/cuda/build_templates/cuda_nvvm.BUILD b/..._party/gpus/cuda/hermetic/cuda_nvvm.BUILD → gpu/cuda/build_templates/cuda_nvvm.BUILD
diff --git a/...pus/cuda/hermetic/cuda_profiler.BUILD.tpl → ...a/build_templates/cuda_profiler.BUILD.tpl b/...pus/cuda/hermetic/cuda_profiler.BUILD.tpl → ...a/build_templates/cuda_profiler.BUILD.tpl
diff --git a/third_party/gpus/cuda/cuda_config.h.tpl → gpu/cuda/cuda_config.h.tpl b/third_party/gpus/cuda/cuda_config.h.tpl → gpu/cuda/cuda_config.h.tpl
diff --git a/third_party/gpus/cuda/cuda_config.py.tpl → gpu/cuda/cuda_config.py.tpl b/third_party/gpus/cuda/cuda_config.py.tpl → gpu/cuda/cuda_config.py.tpl