The RDNA 4 (RX 9070 XT) PyTorch & vLLM Build Guide

RDNA4_ROCm7.2_Build_Guide.md

🚀 The RDNA 4 (RX 9070 XT) PyTorch & vLLM Build Guide

⚠️ ALPHA / EXPERIMENTAL RELEASE This guide outlines a "bleeding-edge" bare-metal compilation process for the AMD Radeon RX 9070 XT (GFX1201) using ROCm 7.2. These patches bypass undocumented compiler strictness changes and API mismatches between PyTorch, vLLM, and HuggingFace. It is provided "as-is" for the community. Use at your own risk.

If you own an AMD Radeon RX 9070 XT and want to run native local AI, you cannot use standard PyTorch binaries or Docker containers. You must compile from source against ROCm 7.2 using the gfx1201 architecture flag.

This guide contains the exact surgical patches required to bypass the bleeding-edge compiler errors.

Phase 1: PyTorch 2.4.0 Compilation

Note: You must use PyTorch 2.4.0 to maintain compatibility with stable vLLM releases.

1. Clone PyTorch

   git clone --recursive -b v2.4.0 https://github.com/pytorch/pytorch
   cd pytorch

2. The "Wavefront 32" Hardware Patch ROCm 7.2 dynamically evaluates warpSize, breaking PyTorch's constexpr assertions. You must hardcode it for RDNA 4:

   sed -i 's/constexpr int kCUDABlockReduceMaxThreads = C10_WARP_SIZE \* C10_WARP_SIZE;/constexpr int kCUDABlockReduceMaxThreads = 1024;/' aten/src/ATen/native/hip/block_reduce.cuh
   sed -i 's/shared\[C10_WARP_SIZE\]/shared[64]/g' aten/src/ATen/native/hip/Normalization.cuh
   sed -i 's/#define C10_WARP_SIZE warpSize/#define C10_WARP_SIZE 32/g' c10/macros/Macros.h

3. The Triton Deprecation Patch Triton uses -Werror which fails on modern ROCm 7.2 deprecation warnings:

   find build/aotriton/src/third_party/triton -name "CMakeLists.txt" -exec sed -i "s/-Werror/-Wno-deprecated-declarations/g" {} +
   sed -i "1i #include <stdbool.h>" build/aotriton/src/third_party/triton/python/triton/runtime/backends/hip.c

4. The C/C++ Linkage Patch ROCm 7.2 compiler strictly separates C and C++ standards.

   # In CMakeLists.txt and cmake/Dependencies.cmake, inject:
   # set(CMAKE_PREFIX_PATH "/opt/rocm;/opt/rocm/lib/cmake/hipblas-common;/opt/rocm/lib/cmake/hipblaslt")

5. Build

   export PYTORCH_ROCM_ARCH="gfx1201"
   export USE_ROCM=1
   export ROCM_PATH="/opt/rocm"
   export CXX="/opt/rocm/llvm/bin/amdclang++"
   export CC="/opt/rocm/llvm/bin/amdclang"
   python3 setup.py bdist_wheel

Phase 2: vLLM 0.6.2 Compilation

1. Clone and Patch

   git clone https://github.com/vllm-project/vllm.git
   cd vllm
   git checkout v0.6.2

2. Add GFX1201 to the Whitelist

   sed -i 's/gfx1100/gfx1100;gfx1201/' CMakeLists.txt

3. The vLLM Wavefront Patch

   sed -i 's/#define WARP_SIZE warpSize/#define WARP_SIZE 32/g' csrc/cuda_compat.h
   sed -i 's/#define WARP_SIZE warpSize/#define WARP_SIZE 32/g' csrc/attention/attention_kernels.cu

4. Build (Ensure you have removed PyTorch dependencies from requirements.txt so it doesn't overwrite your custom build)

   export PYTORCH_ROCM_ARCH="gfx1201"
   export VLLM_TARGET_DEVICE="rocm"
   python3 setup.py bdist_wheel

Phase 3: The HuggingFace RoPE Bug

Newer HuggingFace Transformers libraries break vLLM 0.6.2 when parsing Qwen 2.5 metadata. You must manually patch vllm/model_executor/layers/rotary_embedding.py inside your python site-packages to safely fallback to a standard RotaryEmbedding if the mrope config fails.

---

Authored by Project Apollo