qwen3-vl:8b-instruct 0xc0000005 on Vulkan AMD RX 6750 XT after multi-model load (Windows 0.34.1)

qwen3-vl:8b-instruct crashes llama-server (0xc0000005) on Vulkan AMD RX 6750 XT after multi-model load — Windows

Summary

On Ollama 0.34.1 / Windows 11, qwen3-vl:8b-instruct loads and runs on Vulkan + AMD Radeon RX 6750 XT (12 GB) until a second VL model is loaded alongside the first. After that trigger, every subsequent GPU load of any Qwen3-VL variant crashes during handle_qwen3vl_clip: translating with exit status 0xc0000005. The failure persists across reboot, PnP GPU reset, Ollama reinstall (winget), and fresh ollama pull of the same model.

CPU inference (OLLAMA_VULKAN=false) still works.

Environment

Item	Value
Ollama	0.34.1 (winget, latest)
OS	Windows 10.0.26200
CPU	AMD Ryzen 7 5700X
RAM	32 GB
GPU	AMD Radeon RX 6750 XT, 12 GB VRAM
Driver	Adrenalin 32.0.21045.5002
Backend	Vulkan (OLLAMA_VULKAN=true) — ROCm reports "AMD driver is too old" (HIP 6 vs 7), not used
Model	qwen3-vl:8b-instruct (official library)

Server env (repro)

OLLAMA_VULKAN=true
OLLAMA_MAX_LOADED_MODELS=1
OLLAMA_FLASH_ATTENTION=0
OLLAMA_KEEP_ALIVE=30m

(No LLAMA_ARG_MMPROJ_OFFLOAD — removed after testing.)

Timeline (2026-09-16)

1. 09:06–14:01 — Fresh Ollama 0.34.1 install. qwen3-vl:8b-instruct loads on Vulkan, handle_qwen3vl_clip: translating → handle_qwen3vl: applying compatibility fixes → model loaded. Vision inference ~25 s/image.
2. 14:05:31 — With 8B already in VRAM, load of qwen3-vl:2b-instruct attempted (fits alongside existing models, 5.8 GiB free). Crash at clip translate → 0xc0000005.
3. 14:05+ — All further GPU loads of Qwen3-VL (8B, 2B, 4B) fail at the same point, even alone after unload.
4. 19:03 — Full reboot + GPU PnP disable/enable — no fix.
5. 19:xx — Reinstall Ollama 0.34.1 via winget — no fix.
6. 19:44 — ollama rm + ollama pull qwen3-vl:8b-instruct (fresh blobs) — still crashes on GPU. CPU OK.

Reproduction steps

1. Windows + RX 6750 XT + Ollama 0.34.1 + OLLAMA_VULKAN=true.
2. ollama pull qwen3-vl:8b-instruct
3. Run a vision request — works on clean session (confirmed earlier today).
4. With 8B still loaded, request qwen3-vl:2b-instruct (or any second VL model) with OLLAMA_MAX_LOADED_MODELS=0 or >1.
5. Observe crash during mmproj translation.
6. Stop all models / restart Ollama / reboot.
7. Request only qwen3-vl:8b-instruct again — crashes at step below.

Minimal repro after step 6:

ollama run qwen3-vl:8b-instruct
# or POST /v1/chat/completions with any prompt

Expected

GPU load succeeds; ollama ps shows 100% GPU.

Actual

handle_qwen3vl_clip: detected Ollama-format qwen3vl GGUF used as mmproj; translating
Load failed ... exit status 0xc0000005

Fresh pull (post-ollama rm) — same blob path, same crash:

model=C:\Users\gusta\.ollama\models\blobs\sha256-1329cd5ab37e5ef6faef37cd8ff8660ffc96ed5b47b34fddf8eb0912e6c82466
handle_qwen3vl_clip: detected Ollama-format qwen3vl GGUF used as mmproj; translating
Load failed ... 0xc0000005

Log excerpts

Success (morning, before trigger)

handle_qwen3vl_clip: detected Ollama-format qwen3vl GGUF used as mmproj; translating
handle_qwen3vl: detected Ollama-format qwen3vl GGUF; applying compatibility fixes
srv  llama_server: model loaded
inference compute library=Vulkan name=Vulkan0 description="AMD Radeon RX 6750 XT" total="12.0 GiB" available="11.2 GiB"

Trigger (14:05:31)

updated VRAM based on existing loaded models gpu=0 library=Vulkan total="12.0 GiB" available="5.8 GiB"
llama-server model fits alongside existing models predicted="2.2 GiB" ... available="5.8 GiB"
handle_qwen3vl_clip: detected Ollama-format qwen3vl GGUF used as mmproj; translating
Load failed ... 0xc0000005

After fresh pull (19:44, only 8B, MAX_LOADED_MODELS=1)

OLLAMA_VULKAN:true OLLAMA_MAX_LOADED_MODELS:1
handle_qwen3vl_clip: detected Ollama-format qwen3vl GGUF used as mmproj; translating
Load failed ... 0xc0000005

Full logs: %LOCALAPPDATA%\Ollama\server-4.log (success + trigger), server.log (post-reinstall).

Related issues

• #13029 — Vulkan 0xc0000005 with qwen2.5vl on same GPU (RX 6750 XT)
• #13585 — qwen3-vl:8b run error (multimodal / Vulkan)
• #13791 — Vulkan access violation on AMD Windows

Workaround

• OLLAMA_VULKAN=false — CPU inference works (~19 s cold / ~1 s warm for text-only ping).
• Avoid loading multiple VL models on GPU; set OLLAMA_MAX_LOADED_MODELS=1.
• Application-side: skip vision captchas until GPU restored.

Notes

• LLAMA_ARG_MMPROJ_OFFLOAD=0 did not fix; only moved crash later to common_params_fit_impl.
• "AMD driver is too old" warning appears but Vulkan path is active; morning GPU success had the same warning.
• Not a corrupted download — fresh pull reproduces on identical blob hash.