#646·ZLUDA

Hard crash (Access Violation 0xc0000005) during initialization on RX 9070 (RDNA 4 / gfx1201) with ROCm 7.1

Author: Filters83Created Jun 4, 2026Updated Aug 11, 2026

zluda_trace logs (tarball/zip file)

Environment

  • OS: Windows 11 64-bit
  • GPU: AMD Radeon RX 9070 (RDNA 4 / gfx1201)
  • Drivers: AMD Adrenalin + HIP SDK / ROCm 7.1
  • ZLUDA Version: v6-preview.76 (and earlier)
  • Application: Koharu (using Hugging Face candle framework / cuBLASLt)

Bug Description

When attempting to run CUDA-based inference (Candle framework) via ZLUDA on the new RX 9070, the application hard crashes instantly before any computation begins.

Windows Event Viewer logs an Application Error with Exception Code 0xc0000005 (Access Violation) in an unknown module.

Troubleshooting & Symptoms

  1. Empty PTX Dump: I set $env:ZLUDA_DUMP_DIR to trace the failing kernel, but the directory remains completely empty. This indicates the crash happens during the initial API handshake, context creation (hipInit), or library setup (MIOpen/hipBLAS), well before any PTX to ISA compilation occurs.
  2. Library Loading: cuda_check.exe via ZLUDA launcher shows OK for nvml and cufft, meaning the basic DLL injection works.
  3. ABI Break Suspected: Since the crash happens instantly with an Access Violation without any PTX dump, it strongly points to an ABI mismatch/struct size difference between what ZLUDA expects and what the ROCm 7.1 driver expects for gfx1201.

Let me know if you need me to run any specific debug build to provide further traces!

Description

Using the manga translator Koharu from Mayocream on an AMD RX 9070

Steps to reproduce

Steps to Reproduce

  1. Install HIP SDK 7.1 on a system with an RX 9000 series GPU.
  2. Inject ZLUDA v6-preview.76 into an app using the candle ML framework (e.g., Koharu).
  3. Try to initialize the Device::Cuda backend.
  4. Instant Segfault.

ZLUDA version

Version 6-preview.76

Operating System

Windows 11 64 bit pro

GPU

RX 9700