#5789 [Engine] Windows HIP (gfx1201, RX 9070): `strata-device.exe --selftest` crashes 0xC0000005 in amdhip64_7.dll — engine 0.1.58 and 0.1.59, Adrenalin 26.10.2
open · @hetaohetaoht-source · 1 commentaires · Sur GitHub
Setup & installServer & APIAMD / HIPModels & quantsWindows
Description
### Area
[Engine] (speed, kernels, models, the server and API)
### Desktop app, Terminal or source?
Terminal (the `strata` command); reproduces with the engine package alone.
### What happened
Engine 0.1.59 (`windows-x86_64-hip-rx9070`) detects the card, but every path that creates GPU work dies with an access violation inside the bundled HIP runtime. No model needed:
```
PS> .\engine\0.1.59-windows-x86_64-hip-rx9070\strata-device.exe --selftest
strata: hipMemGetInfo failed (invalid argument); retrying once after 250 ms
strata: hipMemGetInfo retry: ok
strata: Windows budgets 15450 of this card's 16304 MiB for this process; ...
exit -1073741819 (0xC0000005)
```
Event Log (Application, ID 1000), every time:
```
Faulting module: ...\engine\0.1.59-windows-x86_64-hip-rx9070\amdhip64_7.dll, version 10.0.3686.0
Exception code: 0xc0000005
Fault offset: 0x0000000000453df0
```
`strata-device.exe` **without** `--selftest` works (device, VRAM, memory plan: FITS).
### Smallest steps to reproduce
1. Install engine 0.1.59 (or 0.1.58) `windows-x86_64-hip-rx9070`.
2. `.\strata-device.exe --selftest` from the engine folder.
3. Expected `selftest OK`; actual exit `-1073741819`, Event ID 1000 as above.
### Evidence the HIP runtime cannot create/launch kernels
`AMD_LOG_LEVEL=2`:
```
devprogram.cpp :1357: Error: Creating the executable from ISA assembly text failed.
devprogram.cpp :1357: Error: Failed to compile source (from CL or HIP source to LLVM IR).
paldevice.cpp :2918: Couldn't create blit kernels!
palvirtual.cpp :1056: Could not create BlitManager!
paldevice.cpp :1163: Couldn't create the device transfer manager!
```
`AMD_LOG_LEVEL=3` shows `hipMalloc(64 MiB)` returning `hipSuccess`, then `hipLaunchKernel` (`hip_fatbin.cpp: Forcing SPIRV: false`) and the crash. `amd_comgr.dll` 3.5 loads fine.
### What was tried (all identical failure)
- Engine **0.1.58** from release v0.2.12 (SHA-256 checked): same module, same offset.
- **Adrenalin 26.10.2** (AMD Windows Driver 32.0.32015.12012), rebooted: same. (Started on 32.0.22042.14002, 2026-04-30.)
- `strata-device.exe` copied into `rocm\bin` (binds that copy of the runtime): same.
- `--max-context 2048` / `8192`, `--no-graph`, `--model` vs `--native`, `rocm\bin` on `PATH`, `STRATA_WDDM_BUDGET=0`, model load via `strata serve`: same crash.
- Started from a Task Scheduler task (outside the calling process tree): same crash.
### Cross-check: the same GGUF runs on this PC with llama.cpp's Vulkan backend
To rule out the machine, the same model (`Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf`, 21.28 GiB) was served with llama.cpp b12110 (Windows Vulkan x64, official release):
```
Available devices:
Vulkan0: AMD Radeon RX 9070 (16304 MiB, 15439 MiB free)
llama-cli -m Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -ngl 99 -cmoe -n 24 -c 1024
[ Prompt: 9.9 t/s | Generation: 18.7 t/s ] (GPU attention, MoE experts on CPU)
same flags with -ngl 0 (CPU only): [ Prompt: 19.5 t/s | Generation: 9.4 t/s ]
```
So the card, driver and model are fine; the failure is specific to the engine's ROCm 10.0.3686 build on this machine.
### Reproduction status
Reproduced locally (2 engine versions, 2 drivers, in and out of the parent process tree).
### Strata version
Terminal/Control 0.2.13 (control 0.1.59); engines **0.1.59** and **0.1.58**, package `windows-x86_64-hip-rx9070`; feed sequence 124.
### Graphics card, memory and system
- AMD Radeon RX 9070 16 GB, `gfx1201` wave32
- Driver: **32.0.32015.12012** (Adrenalin 26.10.2, 2026-10-28); was 32.0.22042.14002 (2026-04-30) when first seen
- HIP runtime in package: 10.0.3686.0; `System32\amdhip64_7.dll`: 10.0.3661.0 → 10.0.3683.0 after the driver update
- CPU: Ryzen 7 5800X3D (8C/16T, AVX2); RAM 32 GB; Windows 11 Pro for Workstations 10.0.26200; ASUS PRIME X570-PRO
### Model, size and settings
Qwen3.6-35B-A3B UD-Q4_K_XL (Unsloth), 22.85 GB + 0.90 GB vision encoder; download, SHA-256 and tokenizer preparation all succeed. The crash happens before settings matter.
### Suggestion
- Report the failure instead of crashing: if the first kernels cannot be created, print that and the reason (and maybe the `AMD_LOG_LEVEL` hint) rather than dying with `0xC0000005`.
- `engine-component.json` declares `min_driver: {}` for this backend; a HIP minimum (or a startup kernel smoke test) would catch this class of problem at install time.Sur le site
Liens install, modèles, releases.