Qwen3.5-0.8B GGUF crashes on IQ-9075 NPU with QGPUInstructionValidator assertion

[Eric851127]

Describe the bug

Running unsloth/Qwen3.5-0.8B-GGUF with the GenieX NPU backend on a Dragonwing IQ-9075 EVK consistently crashes.

The same model runs successfully with the CPU backend.

When --compute npu is used, the process aborts inside the Qualcomm Adreno shader compiler with a QGPUInstructionValidator assertion.

The issue is reproducible on Qualcomm Linux Reference Distro 2.0.

---

To Reproduce

1. Install GenieX on a Dragonwing IQ-9075 EVK.

2. Verify the GenieX version:

geniex version

Output:

GenieX CLI Version:     v0.3.14
QAIRT Runtime Version:  v2.45.0.260326
LlamaCPP Runtime Hash:  be4a6a6

3. Download the model:

geniex pull unsloth/Qwen3.5-0.8B-GGUF

The selected chipset is:

Dragonwing IQ-9075 EVK

The model downloads successfully and is automatically detected as:

Detected model type: vlm

4. Verify that CPU inference succeeds:

geniex infer unsloth/Qwen3.5-0.8B-GGUF \
  --compute cpu \
  --ngl 0 \
  --think=false \
  --max-tokens 128 \
  --prompt "Please list five common inventory anomalies."

CPU inference completes successfully.

Observed performance:

Approximately 29 tokens/s
Approximately 0.1 seconds to first token

A Traditional Chinese prompt also works correctly:

geniex infer unsloth/Qwen3.5-0.8B-GGUF \
  --compute cpu \
  --ngl 0 \
  --think=false \
  --max-tokens 128 \
  --prompt "請使用繁體中文列出五項常見庫存異常。"

5. Run the same model using the NPU backend:

geniex infer unsloth/Qwen3.5-0.8B-GGUF \
  --compute npu \
  --ngl 999 \
  --think=false \
  --max-tokens 128 \
  --prompt "Please list five common inventory anomalies."

6. The process crashes with the following error:

NumConstRegsError: # of constant registers must be < 2.
    In inst# 1: %R6_c<def> = IMINrr %C22.x, %UR0_a, 1

Assertion failed: false && "back-end instruction validation failed",
file /usr/src/debug/qcom-adreno/1.855.4/shadercompiler/llvm/lib/Target/Oxili/QGPUInstructionValidator.cpp,
line 440

FATAL ERROR: false && "back-end instruction validation failed"

---

Expected behavior

The model should run successfully using the IQ-9075 NPU backend and generate a text response.

If this model, VLM configuration, or one of its operators is not supported by the NPU backend, GenieX should return a clear unsupported-model or unsupported-operator error instead of terminating with a shader compiler assertion.

---

Screenshots

Not applicable.

The full terminal output is included above.

---

Environment

• Device: Dragonwing IQ-9075 EVK
• SoC: Qualcomm QCS9075
• Architecture: aarch64
• OS: Qualcomm Linux Reference Distro 2.0
• Kernel: 6.18.21-g7368c42b9cca-dirty
• GenieX CLI: v0.3.14
• QAIRT Runtime: v2.45.0.260326
• LlamaCPP Runtime Hash: be4a6a6
• Adreno shader compiler version: 1.855.4
• Model: unsloth/Qwen3.5-0.8B-GGUF
• Precision: Q4_0
• Model type detected by GenieX: VLM

OS information

ID=qcom-distro
NAME="Qualcomm Linux Reference Distro"
VERSION="2.0"
VERSION_ID=2.0
PRETTY_NAME="Qualcomm Linux Reference Distro 2.0"
CPE_NAME="cpe:/o:openembedded:qcom-distro:2.0"

Kernel information

Linux iq-9075-evk 6.18.21-g7368c42b9cca-dirty #1 SMP PREEMPT Tue May 5 09:29:10 UTC 2026 aarch64 GNU/Linux

---

Additional context

• The model runs correctly using --compute cpu.
• The failure occurs only when hardware acceleration is enabled.
• The crash occurs inside the Adreno shader compiler even though --compute npu is explicitly selected.
• Qwen3.5-0.8B is automatically detected as a VLM.
• The test is text-only and does not include an image input.
• The same issue was previously reproduced on Qualcomm Linux 1.6 with Adreno version 1.808.11.
• Reflashing the board to Qualcomm Linux Reference Distro 2.0 upgraded the Adreno stack to 1.855.4, but the same failure remains.

Could you please confirm:

1. Is unsloth/Qwen3.5-0.8B-GGUF Q4_0 officially supported by the GenieX NPU backend on IQ-9075 / QCS9075?
2. Is text-only NPU inference supported when this model is detected as a VLM?
3. Why does --compute npu fail inside the Adreno shader compiler?
4. Is this NumConstRegsError a known issue?
5. Is there a recommended GenieX version, model export, GGUF file, or BSP configuration for this model?