#5608 [Engine] RX 9070 XT gfx1201: repeatable GPU page faults during IQ3_S inference on HIP 0.1.57/0.1.58
open · @sirbottoms98 · 11 コメント · GitHub で見る
AMD / HIPModels & quantsSecurityWindowsLinux
本文
**Summary** Strata's HIP inference engine repeatedly crashes during Qwen3.8-Flash-Next IQ3_S inference on my RX 9070 XT under CachyOS Linux, using Strata Desktop 0.2.11. The failure reproduces through Strata's built-in Web UI without using the Desktop agent or tools. The engine terminates with exit code -13, accompanied by AMDGPU [gfxhub] page faults involving process strata. **System** GPU: AMD Radeon RX 9070 XT, 16 GB, gfx1201 CPU: AMD Ryzen 7 9800X3D RAM: 64 GB DDR5 Motherboard: Gigabyte B650 GAMING X AX V2 BIOS: F42d OS: CachyOS Linux, KDE Plasma Kernels used: 7.3.1-1-cachyos and 6.18.55-2-cachyos-lts. The attached Desktop diagnostics contain a crash recorded under the LTS kernel. Strata Desktop: 0.2.11 AppImage Engine versions tested: 0.1.57 and 0.1.58, Linux HIP RX9070 builds Model: Qwen3.8-Flash-Next GSQ-RCO IQ3_S **Relevant engine configuration** --expert-cache auto --prefill auto --spec 4 --max-context 32768 --kv int8 --vision --pool-workers 7 --suffix-draft 0 --vram-reserve-mib 1536 Also reproduced after disabling several optimizations: STRATA_HIP_WMMA=0 STRATA_PF_FUSED=0 STRATA_GFX12_CHUNK=0 Additional diagnostic trials included: AMD_SERIALIZE_KERNEL=3 HIP_LAUNCH_BLOCKING=1 These environment variables were verified in the running engine process. Strata reports approximately 9.55 GiB of GPU expert cache and 38–41 GiB of pinned system-RAM expert cache, depending on the run. **Reproduction** Launch Strata Desktop and load the IQ3_S or IQ3_XXS model. Wait for the inference server to become ready. Open its built-in Web UI at http://127.0.0.1:18500/. Authenticate using the locally generated API key. Submit: "Explain how a five-stage CPU pipeline handles data hazards, forwarding, and pipeline stalls. Include worked instruction examples and provide approximately 1,800 words of explanation. Do not use tools." Allow inference to continue. **Expected** The engine completes the response without crashing or causing GPU memory-access faults. **Actual** The inference engine repeatedly terminates during generation or reasoning. The Web UI reports: the engine stopped unexpectedly (exit code -13); the next request restarts it In one direct Web UI run, the engine stopped after approximately 28 seconds. In a later run, it stopped after approximately 9 seconds. The later failure coincided with these kernel messages: Nov 08 13:36:21 amdgpu 0000:03:00.0: [gfxhub] page fault (src_id:0 ring:24 vmid:8 pasid:215) Process strata pid 5766 in page starting at address 0x0000000000055000 GCVM_L2_PROTECTION_FAULT_STATUS:0x00801031 Faulty UTCL2 client ID: TCP (0x8) MORE_FAULTS: 0x1 WALKER_ERROR: 0x0 PERMISSION_FAULTS: 0x3 MAPPING_ERROR: 0x0 RW: 0x0 Additional faults were reported at addresses 0x58000, 0x5a000, and 0x5d000. **Troubleshooting performed** Reproduced using engines 0.1.57 and 0.1.58. Reproduced through the direct Web UI, independently of the Desktop agent. Disabled HIP WMMA, fused prefill, and gfx12 chunk optimizations. Tested serialized/blocking HIP execution. Disabled suffix drafting. Verified the GPU's physical PCIe link at Gen4 ×16, matching the motherboard's advertised specification. Confirmed the failure persists at Gen4. Previous searches of kernel logs did not identify a corresponding OOM-killer event. The RX 9070 XT is otherwise stable in Windows and other applications. **Questions** Could this indicate an invalid pointer or GPU virtual-memory mapping in the HIP inference path, potentially involving pinned host memory or expert-cache operations? Are there existing diagnostics or engine patches that would help isolate the failing GPU operation? I can provide sanitized engine logs, kernel logs, and additional controlled reproductions if needed. Possibly related: #4912 #1709, #4862. Unlike #5559, this reproduces on both 0.1.57 and 0.1.58 Additional diagnostics are attached. They also record a separate Strata Desktop window crash involving Mesa's libgallium and an amdgpu command-submission rejection. This is distinct from the reproducible HIP inference page faults described above. [strata-diagnostics-sanitized.txt](https://github.com/user-attachments/files/34553812/strata-diagnostics-sanitized.txt)
関連リンク
インストール・モデル・リリースへのサイト内リンク。