enable back jetpack build (#3720)

Author: lanluo-nvidia

.github/workflows/build-test-linux-aarch64-jetpack.yml

@@ -1,16 +1,16 @@
 name: Build and test Linux aarch64 wheels for Jetpack
 
 on:
-  # pull_request:
-  # push:
-  #   branches:
-  #     - main
-  #     - nightly
-  #     - release/*
-  #   tags:
-  #     # NOTE: Binary build pipelines should only get triggered on release candidate builds
-  #     # Release candidate tags look like: v1.11.0-rc1
-  #     - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
+  pull_request:
+  push:
+    branches:
+      - main
+      - nightly
+      - release/*
+    tags:
+      # NOTE: Binary build pipelines should only get triggered on release candidate builds
+      # Release candidate tags look like: v1.11.0-rc1
+      - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
   workflow_dispatch:
 
 jobs:

.github/workflows/build_wheels_linux_aarch64.yml

@@ -264,7 +264,7 @@ jobs:
           if [[ ${{ inputs.is-jetpack }} == false ]]; then
             ${CONDA_RUN} python setup.py bdist_wheel
           else
-            ${CONDA_RUN} python setup.py bdist_wheel --jetpack --plat-name=linux_tegra_aarch64
+            ${CONDA_RUN} python setup.py bdist_wheel --jetpack
           fi
       - name: Repair Manylinux_2_28 Wheel
         shell: bash -l {0}
@@ -336,8 +336,8 @@ jobs:
   upload:
     needs: build
     uses: pytorch/test-infra/.github/workflows/_binary_upload.yml@main
-    # for jetpack builds, only upload to pytorch index for nightly builds
-    if: ${{ inputs.is-jetpack == false || (github.event_name == 'push' && startsWith(github.event.ref, 'refs/heads/nightly')) }}
+    # for jetpack builds, do not upload to pytorch nightly index, only upload to https://pypi.jetson-ai-lab.io/ manually for each release
+    if: ${{ inputs.is-jetpack == false }}
     with:
       repository: ${{ inputs.repository }}
       ref: ${{ inputs.ref }}

MODULE.bazel

@@ -90,10 +90,9 @@ http_archive(
 http_archive(
     name = "torch_l4t",
     build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "6eff643c0a7acda92734cc798338f733ff35c7df1a4434576f5ff7c66fc97319",
     strip_prefix = "torch",
     type = "zip",
-    urls = ["https://pypi.jetson-ai-lab.dev/jp6/cu126/+f/6ef/f643c0a7acda9/torch-2.7.0-cp310-cp310-linux_aarch64.whl"],
+    urls = ["https://pypi.jetson-ai-lab.io/jp6/cu126/+f/62a/1beee9f2f1470/torch-2.8.0-cp310-cp310-linux_aarch64.whl"],
 )
 
 # Download these tarballs manually from the NVIDIA website

docsrc/getting_started/jetpack.rst

@@ -90,22 +90,22 @@ Build Environment Setup
    .. code-block:: sh
 
       # Can only install the torch and torchvision wheel from the JPL repo which is built specifically for JetPack 6.2
-      python -m pip install torch==2.7.0 torchvision==0.22.0  --index-url=https://pypi.jetson-ai-lab.dev/jp6/cu126/
+      python -m pip install torch==2.8.0 torchvision==0.23.0  --index-url=https://pypi.jetson-ai-lab.io/jp6/cu126
 
 
 Building the Wheel
 ==================
 
 .. code-block:: sh
-   python setup.py bdist_wheel
+   python setup.py bdist_wheel --jetpack
 
 Installation
 ============
 
 .. code-block:: sh
-   # you will be able to find the wheel in the dist directory, has platform name linux_tegra_aarch64
+   # you will be able to find the wheel in the dist directory
    cd dist
-   python -m pip install torch_tensorrt-2.8.0.dev0+d8318d8fc-cp310-cp310-linux_tegra_aarch64.whl
+   python -m pip install torch_tensorrt-2.8.0.dev0+d8318d8fc-cp310-cp310-linux_aarch64.whl
 
 Post-Installation Verification
 ==============================

packaging/pre_build_script.sh

@@ -43,7 +43,7 @@ pip uninstall -y torch torchvision
 
 if [[ ${IS_JETPACK} == true ]]; then
     # install torch 2.7 for jp6.2
-    pip install torch==2.7.0 --index-url=https://pypi.jetson-ai-lab.dev/jp6/cu126/
+    pip install torch==2.8.0 --index-url=https://pypi.jetson-ai-lab.io/jp6/cu126/
 else
     TORCH=$(grep "^torch>" py/requirements.txt)
     INDEX_URL=https://download.pytorch.org/whl/${CHANNEL}/${CU_VERSION}

py/torch_tensorrt/_enums.py

@@ -8,6 +8,7 @@
 import tensorrt as trt
 import torch
 from torch_tensorrt._features import ENABLED_FEATURES, needs_torch_tensorrt_runtime
+from torch_tensorrt._utils import is_tensorrt_version_supported
 
 
 class dtype(Enum):
@@ -199,8 +200,6 @@ def _from(
                 return dtype.i8
             elif t == trt.DataType.FP8:
                 return dtype.f8
-            elif t == trt.DataType.FP4:
-                return dtype.fp4
             elif t == trt.DataType.INT32:
                 return dtype.i32
             elif t == trt.DataType.INT64:
@@ -214,6 +213,9 @@ def _from(
             elif t == trt.DataType.BF16:
                 return dtype.bf16
             else:
+                if is_tensorrt_version_supported("10.8.0"):
+                    if t == trt.DataType.FP4:
+                        return dtype.fp4
                 raise TypeError(
                     f"Provided an unsupported data type as a data type for translation (support: bool, int, half, float, bfloat16), got: {t}"
                 )
@@ -409,11 +411,12 @@ def to(
                 return trt.DataType.BOOL
             elif self == dtype.bf16:
                 return trt.DataType.BF16
-            elif self == dtype.f4:
-                return trt.DataType.FP4
             elif use_default:
                 return trt.DataType.FLOAT
             else:
+                if is_tensorrt_version_supported("10.8.0"):
+                    if self == dtype.f4:
+                        return trt.DataType.FP4
                 raise TypeError("Unsupported tensorrt dtype")
 
         elif t == np.dtype:

py/torch_tensorrt/_utils.py

@@ -24,3 +24,29 @@ def check_cross_compile_trt_win_lib() -> bool:
         target_lib = ".*libnvinfer_builder_resource_win.so.*"
         return any(re.match(target_lib, lib) for lib in loaded_libs)
     return False
+
+
+def is_tensorrt_version_supported(min_version: str = "10.8.0") -> bool:
+    """
+    Check if the installed TensorRT version supports the specified minimum version.
+
+    Args:
+        min_version (str): Minimum required TensorRT version (default: "10.8.0" for FP4 support)
+
+    Returns:
+        bool: True if TensorRT version is >= min_version, False otherwise
+
+    Example:
+        >>> if is_tensorrt_version_supported("10.8.0"):
+        ...     # Use FP4 features
+        ...     pass
+    """
+    try:
+        from importlib import metadata
+
+        from packaging.version import Version
+
+        return bool(Version(metadata.version("tensorrt")) >= Version(min_version))
+    except (ImportError, ValueError):
+        # If tensorrt is not installed or version cannot be determined
+        return False

py/torch_tensorrt/dynamo/backend/backends.py

@@ -81,7 +81,7 @@ def aot_torch_tensorrt_aten_backend(
             decompositions=settings_aot_autograd["decompositions"],
         )(gm, sample_inputs)
 
-    if is_tegra_platform():
+    if not is_tegra_platform():
         from torch.distributed.tensor import DTensor
 
         if any(isinstance(tensor, DTensor) for tensor in sample_inputs):

py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py

@@ -7,6 +7,7 @@
 import numpy as np
 import torch
 from torch.fx.node import Argument, Node, Target
+from torch_tensorrt._utils import is_tensorrt_version_supported
 from torch_tensorrt.dynamo._settings import CompilationSettings
 from torch_tensorrt.dynamo._SourceIR import SourceIR
 from torch_tensorrt.dynamo.conversion import impl
@@ -619,40 +620,41 @@ def aten_ops_quantize_op(
         )
 
 
-try:
-    import modelopt.torch.quantization as mtq  # noqa: F401
+if is_tensorrt_version_supported("10.8.0"):
+    try:
+        import modelopt.torch.quantization as mtq  # noqa: F401
 
-    assert torch.ops.tensorrt.dynamic_block_quantize_op.default
-except Exception as e:
-    _LOGGER.warning(
-        "Unable to import quantize op. Please install modelopt library (https://github.com/NVIDIA/TensorRT-Model-Optimizer?tab=readme-ov-file#installation) to add support for compiling quantized models"
-    )
-else:
+        assert torch.ops.tensorrt.dynamic_block_quantize_op.default
+    except Exception as e:
+        _LOGGER.warning(
+            "Unable to import quantize op. Please install modelopt library (https://github.com/NVIDIA/TensorRT-Model-Optimizer?tab=readme-ov-file#installation) to add support for compiling quantized models"
+        )
+    else:
 
-    @dynamo_tensorrt_converter(
-        torch.ops.tensorrt.dynamic_block_quantize_op.default,
-        supports_dynamic_shapes=True,
-    )
-    def aten_ops_dynamic_block_quantize_op(
-        ctx: ConversionContext,
-        target: Target,
-        args: Tuple[Argument, ...],
-        kwargs: Dict[str, Argument],
-        name: str,
-    ) -> Union[TRTTensor, Sequence[TRTTensor]]:
-        return impl.dynamic_block_quantize.quantize(
-            ctx,
-            target,
-            SourceIR.ATEN,
-            name,
-            args[0],
-            args[1],
-            args[2],
-            args[3],
-            args[4],
-            args[5],
-            args[6],
+        @dynamo_tensorrt_converter(
+            torch.ops.tensorrt.dynamic_block_quantize_op.default,
+            supports_dynamic_shapes=True,
         )
+        def aten_ops_dynamic_block_quantize_op(
+            ctx: ConversionContext,
+            target: Target,
+            args: Tuple[Argument, ...],
+            kwargs: Dict[str, Argument],
+            name: str,
+        ) -> Union[TRTTensor, Sequence[TRTTensor]]:
+            return impl.dynamic_block_quantize.quantize(
+                ctx,
+                target,
+                SourceIR.ATEN,
+                name,
+                args[0],
+                args[1],
+                args[2],
+                args[3],
+                args[4],
+                args[5],
+                args[6],
+            )
 
 
 @dynamo_tensorrt_converter(torch.ops.aten.squeeze.dim, supports_dynamic_shapes=True)

py/torch_tensorrt/dynamo/conversion/converter_utils.py

@@ -24,6 +24,7 @@
 from torch.fx.node import Argument, Target
 from torch.fx.passes.shape_prop import TensorMetadata
 from torch_tensorrt import _enums
+from torch_tensorrt._utils import is_tensorrt_version_supported
 from torch_tensorrt.dynamo._settings import CompilationSettings
 from torch_tensorrt.dynamo._SourceIR import SourceIR
 from torch_tensorrt.dynamo.conversion._ConversionContext import ConversionContext
@@ -447,30 +448,35 @@ def create_constant(
         if torch_value is not None:
 
             if torch_value.dtype == torch.uint8:
-                if (
-                    target_quantized_type is None
-                    or target_quantized_type != trt.DataType.FP4
-                ):
-                    # Iconstant layer does not support Uint8, it only support that FP4 data packed in uint8
+                if is_tensorrt_version_supported("10.8.0"):
+                    if (
+                        target_quantized_type is None
+                        or target_quantized_type != trt.DataType.FP4
+                    ):
+                        # Iconstant layer does not support Uint8, it only support that FP4 data packed in uint8
+                        raise ValueError(
+                            "Currently supported target_quantized_type for uint8 is FP4, got {target_quantized_type=}"
+                        )
+                    shape[-1] = shape[-1] * 2
+                    weights = to_trt_weights(
+                        ctx,
+                        torch_value,
+                        name,
+                        "CONSTANT",
+                        "CONSTANT",
+                        dtype=trt.DataType.FP4,
+                        count=torch_value.numel() * 2,
+                    )
+                    constant = ctx.net.add_constant(
+                        shape,
+                        weights,
+                    )
+                    constant.name = name
+                    return constant.get_output(0)
+                else:
                     raise ValueError(
-                        "Currently supported target_quantized_type for uint8 is FP4, got {target_quantized_type=}"
+                        "Currently FP4 is only supported in TensorRT 10.8.0 and above"
                     )
-                shape[-1] = shape[-1] * 2
-                weights = to_trt_weights(
-                    ctx,
-                    torch_value,
-                    name,
-                    "CONSTANT",
-                    "CONSTANT",
-                    dtype=trt.DataType.FP4,
-                    count=torch_value.numel() * 2,
-                )
-                constant = ctx.net.add_constant(
-                    shape,
-                    weights,
-                )
-                constant.name = name
-                return constant.get_output(0)
 
             # Record the weight in ctx for refit and cpu memory reference