[Dependency] Add torch-c-dlpack-ext to project requirements (#1403)

* [Dependency] Add torch-c-dlpack-ext to project requirements

* Added torch-c-dlpack-ext to both pyproject.toml and requirements.txt to provide prebuilt torch extensions, which may prevent JIT compilation on first import of TVM FFI.

* [Build] Update manylinux images in project configuration

* Changed the manylinux image for x86_64 from "manylinux2014" to "manylinux_2_28" in both pyproject.toml and the Dockerfile to align with updated standards for compatibility and performance.

* [Build] Update CUDA repository configuration in pyproject.toml

* Changed the package manager command from `yum-config-manager` to `dnf config-manager` for adding the CUDA repository, ensuring compatibility with newer systems.

* fix

* [Build] Update CUDA repository to RHEL 8

* Changed the CUDA repository configuration in both pyproject.toml and the manylinux Dockerfile from RHEL 7 to RHEL 8, ensuring compatibility with newer systems.

* test: run out of space

* use cu130 to reduce size

* upd

* upd comment

* upd

---------
Co-authored-by: Your Name <wenji.yyc@alibaba-inc.com>

.github/workflows/dist.yml

@@ -106,7 +106,7 @@ jobs:
     strategy:
       matrix:
         target:
-          - { runner: ubuntu-latest, toolkit: "CUDA-12.1" }
+          - { runner: ubuntu-latest, toolkit: "CUDA-12.8" }
           - { runner: ubuntu-24.04-arm, toolkit: "CUDA-12.8" }
           - { runner: macos-latest, toolkit: "Metal" }
         python-version:

maint/scripts/pypi.manylinux.Dockerfile

-FROM quay.io/pypa/manylinux2014_x86_64 AS builder_amd64
+FROM quay.io/pypa/manylinux_2_28_x86_64 AS builder_amd64
 
-RUN yum-config-manager --add-repo https://developer.download.nvidia.cn/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
+RUN dnf config-manager --add-repo https://developer.download.nvidia.cn/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
 
-ARG CUDA_VERSION=12.1
+ARG CUDA_VERSION=12.8
 ENV CUDA_VERSION=${CUDA_VERSION}
 
 FROM quay.io/pypa/manylinux_2_28_aarch64 AS builder_arm64

pyproject.toml

@@ -32,6 +32,9 @@ dependencies = [
     # should be removed after our tvm's update.
     # See discussion in tilelang#1373 and apache/tvm-ffi#307
     "apache-tvm-ffi>=0.1.2",
+    # torch-c-dlpack-ext provides prebuilt torch extensions.
+    # Without it, TVM FFI may require JIT compilation on first import.
+    "torch-c-dlpack-ext",
     "cloudpickle",
     "ml-dtypes",
     "numpy>=1.23.5",
@@ -218,12 +221,10 @@ environment.PYTHONDEVMODE = "1"
 environment.PYTHONUNBUFFERED = "1"
 environment.PATH = "/usr/local/cuda/bin:$PATH"
 environment.LD_LIBRARY_PATH = "/usr/local/cuda/lib64:/usr/local/cuda/lib64/stubs:$LD_LIBRARY_PATH"
-# Pin to glibc 2.17 for x86 and 2.28 for aarch64 for now
-# TODO: upgrade to manylinux_2_28 at some time
-manylinux-x86_64-image = "manylinux2014"   # CentOS 7
-manylinux-aarch64-image = "manylinux_2_28" # AlmaLinux 8
+manylinux-x86_64-image  = "manylinux_2_28"  # AlmaLinux 8
+manylinux-aarch64-image = "manylinux_2_28"  # AlmaLinux 8
 # Install CUDA runtime and stub driver library
-# manylinux_2_28 uses gcc 14, which needs CUDA 12.8
+# manylinux_2_28 uses gcc 14, which needs CUDA >=12.8
 before-all = """
 set -eux
 
@@ -232,8 +233,8 @@ uname -a
 
 case "$(uname -m)" in
     "x86_64")
-        DEFAULT_CUDA_VERSION="12.1"
-        yum-config-manager --add-repo https://developer.download.nvidia.cn/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
+        DEFAULT_CUDA_VERSION="12.8"
+        dnf config-manager --add-repo https://developer.download.nvidia.cn/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
         ;;
     "aarch64")
         DEFAULT_CUDA_VERSION="12.8"
@@ -247,6 +248,7 @@ esac
 cudaver="$(echo "${CUDA_VERSION:-$DEFAULT_CUDA_VERSION}" | cut -d '.' -f-2)"
 v="${cudaver//./-}"
 yum install -y "cuda-minimal-build-${v}" "cuda-driver-devel-${v}" "cuda-nvrtc-devel-${v}" nvidia-driver-cuda-libs
+yum clean all
 """
 repair-wheel-command = [
     "auditwheel -v repair --exclude libtvm_ffi.so --exclude libcuda.so.1 --exclude '/usr/local/cuda*' -w {dest_dir} {wheel}",
@@ -261,7 +263,8 @@ repair-wheel-command = [
 
 [[tool.cibuildwheel.overrides]]
 select = "*linux*x86_64*"
-# CentOS 7 is too old to run import test. Do wheel installation test only.
-test-command = [
-    "echo 'Wheel is installed successfully'",
+# x86_64 runners in GitHub Actions have limited storage,
+# pre-install torch without caching to reduce disk usage during install tilelang.
+before-test = [
+    "pip install torch --no-cache-dir",
 ]

requirements.txt

 # Runtime requirements
 apache-tvm-ffi>=0.1.2
+torch-c-dlpack-ext
 cloudpickle
 ml-dtypes
 numpy>=1.23.5