Support request: Bielik-11B-v3.0-Instruct-heretic-MPOA on Ryzen AI NPU

[Mikikudi]

Hello FastFlowLM team,

I would like to request NPU support for the Polish Llama-based model Bielik-11B-v3.0-Instruct-heretic-MPOA.

Model sources:

• Base model: https://huggingface.co/speakleash/Bielik-11B-v3-Base-20250730
• Q6_K GGUF: https://huggingface.co/mradermacher/Bielik-PL-11B-v3.0-Instruct-heretic-GGUF

I converted the Q6_K GGUF to Q4NX_FLL successfully. The resulting model has 50 layers, hidden size 4096, intermediate size 14336, 32 attention heads, and 8 key/value heads.

System:

• Linux
• FastFlowLM 1.0.4
• AMD RyzenAI-npu6 / XDNA2
• XRT detects the NPU and flm validate --json reports ready: true.

I tested the existing Llama-3.1-8B-NPU2 xclbins as a diagnostic only. The model loads and acquires the NPU, but generation fails with:

runlist failed execution (ERT_CMD_STATE_TIMEOUT)
Kernel Instance: MLIR_AIE

Could you please advise:

1. Is there an existing compatible xclbin set for this Bielik/Llama architecture?
2. If not, could support be added, including the required attn.xclbin, dequant.xclbin, layer.xclbin, and mm.xclbin?
3. Is there a supported public workflow for compiling FLM-compatible xclbins for a custom Q4NX_FLL model?

I can provide the Q4NX tensor metadata, model configuration, and reproducible logs if helpful.

Thank you.