feat: Enable TensorRT-RTX build for linux-aarch64 (SBSA / DGX Spark)

[narendasan]

TensorRT-RTX 1.5 ships a Linux aarch64 (SBSA) distribution, but the repo only wired up the x86_64 and Windows RTX archives. This adds the aarch64 RTX archive and threads a rtx_sbsa config_setting (aarch64 + linux + compute_libs=rtx) through the TensorRT selects, so USE_TRT_RTX=1 builds resolve @tensorrt_rtx_sbsa on aarch64.

Build wiring:

• MODULE.bazel: add tensorrt_rtx_sbsa http_archive (aarch64 tarball).
• third_party/tensorrt_rtx/{archive,local}/BUILD: rtx_sbsa config_setting
  • lib/header/nvinfer selects (lib/libtensorrt_rtx.so, @cuda//:cudart).
• ~20 repo BUILD files: rtx_sbsa config_setting + a :rtx_sbsa select entry mirroring :rtx_x86_64 (sourced from @tensorrt_rtx_sbsa).

Basic CI:

• .github/workflows/build-test-linux-aarch64_rtx.yml: mirrors the standard SBSA build (build-test-linux-aarch64.yml) with use-rtx: true. BUILD-ONLY (no aarch64 GPU test runners) — it validates the RTX aarch64 wheel builds. The full manifest-based wiring will layer on in the CI redesign (test-dx).

Description

Please include a summary of the change and which issue is fixed. Please also include relevant motivation and context. List any dependencies that are required for this change.

Fixes # (issue)

Type of change

Please delete options that are not relevant and/or add your own.

• New feature (non-breaking change which adds functionality)

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:26:58.191844+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:27:16.313461+00:00
@@ -2,11 +2,10 @@

import torch
from torch.fx import GraphModule, Node

from .pass_utils import clean_up_graph_after_modifications
-

_INT64_MAX = 2**63 - 1
_SYM_MIN = getattr(torch, "sym_min", None)

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:26:59.786592+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:27:20.904802+00:00
@@ -2,11 +2,10 @@

import torch
from torch.fx import GraphModule, Node

from .pass_utils import clean_up_graph_after_modifications
-

_INT64_MAX = 2**63 - 1
_SYM_MIN = getattr(torch, "sym_min", None)

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:26:58.610133+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:27:22.936809+00:00
@@ -2,11 +2,10 @@

import torch
from torch.fx import GraphModule, Node

from .pass_utils import clean_up_graph_after_modifications
-

_INT64_MAX = 2**63 - 1
_SYM_MIN = getattr(torch, "sym_min", None)

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:26:58.542315+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-22 20:27:22.244056+00:00
@@ -2,11 +2,10 @@

import torch
from torch.fx import GraphModule, Node

from .pass_utils import clean_up_graph_after_modifications
-

_INT64_MAX = 2**63 - 1
_SYM_MIN = getattr(torch, "sym_min", None)

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-23 01:33:23.576086+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-23 01:33:48.284954+00:00
@@ -2,11 +2,10 @@

import torch
from torch.fx import GraphModule, Node

from .pass_utils import clean_up_graph_after_modifications
-

_INT64_MAX = 2**63 - 1
_SYM_MIN = getattr(torch, "sym_min", None)

[lanluo-nvidia]

LGTM

[tp5uiuc]

IIUC this calls into setup.py. From my understanding of the code, we may need to modify that one as well?

• The if IS_SBSA: branch is tested before the USE_TRT_RTX branch (which is in x86_64) during compilation. This means an aarch64+RTX build resolves the pybind include dir to @tensorrt_sbsa (which is enterprise) while the Bazel core is modified to link @tensorrt_rtx_sbsa. We should have a tensorrt_rtx_sbsa_external_dir lambda and use it when IS_SBSA and USE_TRT_RTX.
• Similarly the SBSA wheels have the wrong dependencies. get_sbsa_requirements today unconditionally appends tensorrt>=11.0.0,<11.1.0 and does not have a USE_TRT_RTX branch (unlike get_x86_64_requirements which emits tensorrt_rtx>=1.5.0.114,<1.6.0.0)
• These wheels also can get pushed to pypi as it is right now : should we gate it till we see everything passes?

[narendasan]

These wheels also can get pushed to pypi as it is right now : should we gate it till we see everything passes?

I think its ok since we dont run tests on this channel in CI, there are no gpu runners

[tp5uiuc]

I see this in CI

# Smoke test is intentionally disabled.
# The issue was smoke test installs the built torch_tensorrt wheel file and checks `import torch_tensorrt; print(torch_tensorrt.__version__)`
# Since tensorrt cannot be pip installable in CI, the smoke test will fail.
# One way we tried to handle it is manually install tensorrt wheel while by extracting from the tarball.
# However, the TensorRT-10.8.0.43/lib path doesn't seem to show up in LD_LIBRARY_PATH even if we explicitly set it.
# TODO: Implement a custom smoke_test script to verify torch_tensorrt installation.

is this still true? I checked it because of the comment below (where malformed wheels can get pushed to pypi potentially). Why is tensorrt not pip installable in CI (because of GPU requirements or something)? An import torch_tensorrt; assert ENABLED_FEATURES.tensorrt_rtx check would suffice here I think