[Deps] Retire the CUDA 12 lane (#38404)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>

python/sglang/kernels/aot/Dockerfile

@@ -98,8 +98,7 @@ RUN --mount=type=cache,id=sgl-kernel-pip,target=/root/.cache/pip \
     set -eux; \
     case "${CUDA_VERSION}" in \
       13.0) TORCH_VER=2.13.0; CU_TAG=cu130 ;; \
-      12.9) TORCH_VER=2.13.0; CU_TAG=cu129 ;; \
-      *)    TORCH_VER=2.13.0; CU_TAG=cu126 ;; \
+      *) echo "Unsupported CUDA version: ${CUDA_VERSION}" && exit 1 ;; \
     esac; \
     ${PYTHON_ROOT_PATH}/bin/pip install torch==${TORCH_VER} --index-url ${PYTORCH_INDEX_BASE}/${CU_TAG}; \
     ${PYTHON_ROOT_PATH}/bin/pip install ninja setuptools==75.0.0 wheel==0.41.0 numpy uv scikit-build-core --index-url ${PIP_DEFAULT_INDEX}

python/sglang/kernels/aot/python/sgl_kernel/load_utils.py

@@ -167,12 +167,7 @@ def _load_architecture_specific_ops():
 
     # All attempts failed
     cuda_version = torch.version.cuda
-    if cuda_version and cuda_version.startswith("12"):
-        install_hint = (
-            "pip install sglang-kernel --index-url https://docs.sglang.ai/whl/cu129/"
-        )
-    else:
-        install_hint = "pip install --upgrade sglang-kernel"
+    install_hint = "pip install --upgrade sglang-kernel"
 
     error_msg = f"""
 [sgl_kernel] CRITICAL: Could not load any common_ops library!

python/sglang/kernels/aot/rename_wheels.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# Align CUDA wheel filenames (+cu124/+cu129/+cu130) with internal METADATA Version and
+# Align CUDA wheel filenames (+cu130) with internal METADATA Version and
 # WHEEL tags after build (fixes pip "inconsistent version" when only the .whl name changed).
 # Unpack → patch WHEEL/METADATA → wheel pack (RECORD regenerated; no hand-editing).
 set -ex
@@ -7,11 +7,7 @@ set -ex
 WHEEL_DIR="dist"
 
 detect_cuda_suffix() {
-    if ls /usr/local/ 2>/dev/null | grep -q "12.4"; then
-        echo "+cu124"
-    elif ls /usr/local/ 2>/dev/null | grep -q "12.9"; then
-        echo "+cu129"
-    elif ls /usr/local/ 2>/dev/null | grep -q "13.0"; then
+    if ls /usr/local/ 2>/dev/null | grep -q "13.0"; then
         echo "+cu130"
     else
         echo ""

python/sglang/kernels/jit/utils/arch.py

@@ -42,7 +42,7 @@ def _jit_cuda_version() -> tuple[int, ...]:
     """CUDA version of the nvcc that JIT builds actually run.
 
     The target has to match the compiler, not the toolkit PyTorch was built
-    against: a cu129 wheel on a CUDA 12.8 toolkit would otherwise select
+    against: a cu130 wheel on a CUDA 12.8 toolkit would otherwise select
     `sm_120f`, which nvcc 12.8 rejects. Resolve nvcc the way tvm-ffi does
     (`CUDA_HOME` / `CUDA_PATH`, then `$PATH`, then `/usr/local/cuda`) and fall
     back to `torch.version.cuda` when it cannot be probed.