# ROCm + RX 9070 XT (gfx1201/RDNA4) working via librocdxg on WSL2 (Ubuntu 26.04)

[miguelpassinho]

Summary

Confirming that ROCm HIP compute (not just Vulkan) works on an AMD Radeon RX 9070 XT (gfx1201, RDNA4) inside WSL2, via librocdxg. Tested and benchmarked end-to-end, since public documentation for this GPU generation + WSL combination appears to be limited — most existing guides predate RDNA4 consumer card support.

Posting the recipe, a gotcha that cost significant debugging time, and real benchmark numbers in case it saves someone else the trouble.

Hardware / environment

• CPU: AMD Ryzen 7 9700X
• GPU: AMD Radeon RX 9070 XT (gfx1201, RDNA4), 16GB VRAM
• Host OS: Windows, WSL2
• Fan curve / power limit: fixed via BIOS (not left to dynamic Windows/driver management) — noting this since it may affect thermal behavior reproducibility

Recipe — Ubuntu 26.04, ROCm 7.13 pre-release

sudo amdgpu-install --usecase=rocm --no-dkms
sudo apt install amdrocm-profiler7.13 amdrocm-opencl amdrocm-gfx120x

Notes:

• No amdgpu-dkms needed — it was purged in testing and never reinstalled; this makes sense since WSL has no native kernel GPU driver anyway.
• HSA_ENABLE_DXG_DETECTION was not required — ROCm 7.13 removes this requirement.
• librocdxg installed via the official .deb (rocdxg-roct_1.2.0_amd64.deb), not built from source.
• Survived a full reboot with no persistence tricks needed — rocminfo worked cleanly on a fresh session (Runtime Version 1.21).

Gotcha: missing HIP dev/CMake package

Compiling llama.cpp with -DGGML_HIP=ON initially failed:

CMake Error: The ROCm root directory: /opt/rocm/core-7.13
  does not contain the HIP runtime CMake package, expected at:
  /opt/rocm/core-7.13/lib/cmake/hip-lang/hip-lang-config.cmake

Root cause: the amdrocm-profiler7.13 / amdrocm-opencl / amdrocm-gfx120x set only installs the runtime, not HIP development files. The missing piece was:

sudo apt install amdrocm-runtime-dev7.13

This installs hip-lang-config.cmake and the rest of the CMake package files needed for enable_language(HIP). Not obvious from package naming — worth documenting since the runtime packages alone give a false impression that HIP dev support is complete.

One caveat observed: CMake configure step produced repeated warnings about libhipblas.so.3 / librocblas.so.5 existing both under /opt/rocm/lib (new AMD 7.13 packages) and /usr/lib/x86_64-linux-gnu (older Ubuntu-repo 7.1 packages, pulled in as automatic dependencies earlier). Did not cause a build failure, but flagging as a potential source of runtime library confusion for others.

Benchmark results (llama.cpp, HIP backend)

Built with:

HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1201 -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j$(nproc)

Confirmed backend load:

ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16200 MiB):
  Device 0: AMD Radeon RX 9070 XT, gfx1201 (0x1201), VMM: no, Wave Size: 32, VRAM: 16200 MiB
load_backend: loaded ROCm backend from /usr/lib/x86_64-linux-gnu/ggml/backends0/libggml-hip.so

TinyLlama 1B Q4_K:

test	t/s
pp512	19837.99 ± 4575.55
tg128	280.70 ± 4.38

10 consecutive runs (-r 3 -n 512 -p 512, no pause between iterations) showed stable throughput, 271–281 t/s tg, no degradation or errors.

Mistral-Nemo 12B Q4_K_M (heavier sustained-load test, no pauses between iterations, -n 1024 -p 2048):

• Windows Task Manager (host side, external confirmation): GPU Compute engine at 94% utilization, 9.4GB dedicated VRAM in use, temperature stable at 52-53°C throughout, no throttling observed
• Throughput held consistent across all 15 iterations with minimal variance
• No crashes, no HIP errors, no "zombie state" behavior observed under this load (relevant since this failure mode has been reported by other RX 9070 XT + ROCm + WSL users elsewhere)

Conclusion

Ubuntu 26.04 with ROCm 7.13 pre-release works reliably for gfx1201 under WSL2 via librocdxg, with no HSA_ENABLE_DXG_DETECTION needed and no persistence issues across reboots.

Happy to provide more detail, full package lists, or test additional configurations if useful.

I'm Miguel, I live in Salvador BA, Brazil
Thanks!

[miguelpassinho]

Stress Test Validation: Arch-based Linux (Garuda + CachyOS tooling) + RDNA4 (gfx1201) + ROCm 7.2.4

Test Summary

• Duration: 50 minutes 21 seconds
• Models Tested: TinyLlama, Mistral-Nemo, gemma4 (9.6GB), laguna-xs-2.1 (20GB)
• Parallel Instances: Up to 7 concurrent processes
• Result: ZERO crashes, ZERO "zombie state", ZERO OOM errors
• Environment: Native Linux (NOT WSL2)

Hardware Configuration

• GPU: AMD Radeon RX 9070 XT OC (gfx1201, RDNA4)
• CPU: Ryzen 7 9700X
• RAM: 32GB
• OS: Garuda Linux (Arch-based) with CachyOS tooling integration
• ROCm: 7.2.4
• Kernel: Linux native (no WSL2 compatibility layer)

Design Decisions & Personal Responsibility

I'm aware of the repository policies regarding CachyOS and Garuda tooling separation. However, I deliberately chose to integrate librocdxg from CachyOS tooling into my Garuda/Arch setup.

I take full responsibility for this decision. My knowledge is limited in scope and depth, and there may have been other approaches I wasn't aware of. However, I made a conscious choice to prioritize functionality over distro purity, and I own that decision completely.

The result is documented below: RDNA4 (gfx1201) running stable LLM inference on native Linux with ROCm 7.2.4. Whether this approach is "correct" by project standards is for the maintainers to decide, but the technical outcome speaks for itself.

Key Findings

1. Stability: 50+ minutes continuous operation without degradation
2. Memory Management: VRAM allocation stable (66-98% depending on model)
3. Power Management: GPU correctly enters/exits low-power state between iterations
4. Performance:
   • TinyLlama: 280 t/s
   • Mistral-Nemo: 64 t/s
   • Prompt processing: 1240-1600 tok/s
5. Thermal: Junction temp peaked at 46°C under sustained load, stable at 38-46°C typical

Critical Setup Notes

• librocdxg via CachyOS tooling required for RDNA4 support
• HSA_ENABLE_DXG_DETECTION=1 mandatory for ROCm < 7.13
• Native Linux environment - no WSL2 compatibility layer involved
• No official documentation exists for RDNA4 (gfx1201) on Arch-based systems

Conclusion

This validates that consumer-grade RDNA4 hardware can run sustained LLM inference workloads on native Linux (Garuda + CachyOS tooling hybrid) with ROCm 7.2.4 at production-level stability. The cross-distro tooling integration was necessary, deliberate, and successful.

Full log available upon request
stress_20260711_194635.log
stress_20260711_193650.log
stress_20260711_193929.log
stress_20260711_194621.log
stress_20260711_194635.log
stress.log

Miguel Passinho
Bahia, Brasil

[miguelpassinho]

Verified benchmarks — RX 9070 XT (gfx1201/RDNA4)
WSL2 (Ubuntu 26.04, ROCm 7.13 pre-release, via librocdxg):

TinyLlama 1B Q4_K: pp512 = 19,837.99 t/s, tg128 = 280.70 t/s (stable across 10 runs)
Mistral-Nemo 12B Q4_K_M: 94% GPU compute utilization, 52-53°C sustained (Windows Task Manager, external confirmation), stable throughput across 15 consecutive runs, no crashes

Native Arch Linux (ROCm 7.2.4, Ollama):

Qwen2.5-Coder 7B (my actual daily-driver model for dev work, not a synthetic pick), 2050-token prompt: prefill = 4253.56 t/s (481.94ms), decode = 96.99–100.16 t/s
GPU temps during stress test (rocm-smi log): edge 36–41°C, junction 38–46°C, memory 56–62°C