Vulkan on AMD Strix Halo (gfx1151) does not load

Row: BACKEND-VULKAN

Tried FP8:

$ ./vllm.cpp/build/examples/vllm-cli --model ./models/Qwen3.6-27B-FP8 --prompt "Draw an SVG image of a skateboarding cat" 
2>&1 | tee cat.svg
vllm-cli: loading model from ./models/Qwen3.6-27B-FP8
vllm-cli: model load failed (status 2): vllm_engine_load: vt: dense loader: expected BF16 for model.language_model.layers.
0.linear_attn.in_proj_qkv.weight at /home/brent/vllm.cpp/include/vllm/model_executor/models/dense_weight_loaders.h:109    

Then BF16:

$ ./vllm.cpp/build/examples/vllm-cli --model ./models/Qwen3.6-27B --prompt "Draw an SVG image of a skateboarding cat" 2>&1
 | tee cat.svg
vllm-cli: loading model from ./models/Qwen3.6-27B
vllm.cpp: Asynchronous scheduling is enabled (max_concurrent_batches=2)
engine-fatal: EngineCore busy loop threw: vt: vulkan: embedding: table points outside every Vulkan allocation — Vulkan ker
nels can only bind memory obtained from vt::GetBackend(DeviceType::kVULKAN).Alloc() at /home/brent/vllm.cpp/src/vt/vulkan/
vulkan_backend.cpp:144
async-llm: output handler saw engine death: EngineCore encountered an issue. See stack trace (above) for the root cause. [
vt: vulkan: embedding: table points outside every Vulkan allocation — Vulkan kernels can only bind memory obtained from vt
::GetBackend(DeviceType::kVULKAN).Alloc() at /home/brent/vllm.cpp/src/vt/vulkan/vulkan_backend.cpp:144]
vllm-cli: completion failed (status 3): vllm_complete: EngineCore encountered an issue. See stack trace (above) for the ro
ot cause. [vt: vulkan: embedding: table points outside every Vulkan allocation — Vulkan kernels can only bind memory obtai
ned from vt::GetBackend(DeviceType::kVULKAN).Alloc() at /home/brent/vllm.cpp/src/vt/vulkan/vulkan_backend.cpp:144]

I am on Slackware-current. I have the Vulkan libraries (vulkan-sdk-1.4.341.1-x86_64-2 ) and am able to use llama.cpp's Vulkan backend.

$ ls /usr/lib64/libvulkan*
/usr/lib64/libvulkan.so          /usr/lib64/libvulkan_intel.so        /usr/lib64/libvulkan_nouveau.so
/usr/lib64/libvulkan.so.1        /usr/lib64/libvulkan_intel_hasvk.so  /usr/lib64/libvulkan_radeon.so
/usr/lib64/libvulkan.so.1.4.341  /usr/lib64/libvulkan_lvp.so          /usr/lib64/libvulkan_virtio.so

I usually run llama.cpp like this:

GGML_VK_PREFER_HOST_MEMORY=1 AMD_VULKAN_ICD=RADV ~/llama.cpp/build/bin/llama-server \
  --host 0.0.0.0 \
  --no-ui \
  --metrics \
  --no-models-autoload \
  --models-max 1 \
  --models-preset $HOME/resources/presets.ini 2>&1 | tee $HOME/local/var/log/llama.cpp.log

And load Qwen3.6 with these presets, for example:

[11]
hf = unsloth/Qwen3.6-35B-A3B-MTP-GGUF:UD-Q4_K_M
host = 0.0.0.0
c = 262144
spec-type = draft-mtp
spec-draft-n-max  = 2
chat-template-file = /home/brent/resources/froggeric_tpl/chat_template.jinja
chat-template-kwargs = {"preserve_thinking":true}
np = 1
ctk = q4_0
ctv = q4_0
ctkd = q4_0
ctvd = q4_0
temp = 1.0
top-k = 20
top-p = 0.95
min-p = 0

$ uname -a
Linux beelink.home 7.1.2 #1 SMP PREEMPT_DYNAMIC Sat Jun 27 17:46:51 CDT 2026 x86_64 AMD RYZEN AI MAX+ 395 w/ Radeon 8060S AuthenticAMD GNU/Linux

Cheers

Reproduce the failure with the Qwen3.6 FP8 and BF16 commands on the reported AMD Strix Halo setup. Start at include/vllm/model_executor/models/dense_weight_loaders.h:109 and src/vt/vulkan/vulkan_backend.cpp:144, then trace the two reported load and allocation errors. Done means the Vulkan backend loads and runs the reported models without either fatal error.