Downgrade to CUDA 12.8.1 for vLLM compatibility

cuMemcpyBatchAsync API changed in CUDA 13 - removed fail_idx parameter.
vLLM code targets CUDA 12.8 API. Downgrade to CUDA 12.8.1.

vllm/Dockerfile

@@ -1,4 +1,4 @@
-ARG CUDA_VERSION=13.0.1
+ARG CUDA_VERSION=12.8.1
 ARG IMAGE_DISTRO=ubuntu24.04
 ARG PYTHON_VERSION=3.12
 
@@ -9,7 +9,7 @@ FROM nvcr.io/nvidia/cuda:${CUDA_VERSION}-devel-${IMAGE_DISTRO} AS base
 # 'a' suffix is not forward compatible but enables all optimizations
 ARG TORCH_CUDA_ARCH_LIST="9.0a"
 ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}
-ENV UV_TORCH_BACKEND=cu130
+ENV UV_TORCH_BACKEND=cu128
 ARG VLLM_FA_CMAKE_GPU_ARCHES="90a-real"
 ENV VLLM_FA_CMAKE_GPU_ARCHES=${VLLM_FA_CMAKE_GPU_ARCHES}
 
@@ -55,7 +55,7 @@ RUN apt-get update && apt install -y wget
 
 RUN uv pip install numpy==2.0.0
 # Install pytorch nightly
-RUN uv pip install torch==2.9.0+cu130 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
+RUN uv pip install torch==2.7.0+cu128 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128
 
 FROM base AS build-base
 RUN mkdir /wheels
@@ -81,7 +81,7 @@ RUN mkdir -p /wheels && \
 FROM build-base AS build-flashinfer
 ARG FLASHINFER_ENABLE_AOT=1
 ARG FLASHINFER_REF=v0.6.6
-ARG FLASHINFER_BUILD_SUFFIX=cu130
+ARG FLASHINFER_BUILD_SUFFIX=cu128
 ENV FLASHINFER_LOCAL_VERSION=${FLASHINFER_BUILD_SUFFIX:-}
 RUN git clone https://github.com/flashinfer-ai/flashinfer.git
 RUN cd flashinfer && \