ggml-hexagon: QCS6490 (Hexagon V68) crashes during prompt prefill with dspqueue_read failed: 0x0000002e

[rahulbtc]

Summary

Running GGUF models through the GenieX Hexagon backend on a Radxa Dragon Q6A (Qualcomm QCS6490, Hexagon V68) consistently crashes during prompt prefill.

The same models run correctly on the CPU backend.

The same hardware also successfully runs Llama 3.2 1B on the NPU using Qualcomm's QAIRT runtime, suggesting that the DSP, firmware and FastRPC stack are functional.

The failure appears specific to the ggml-hexagon execution path.

---

Environment

Board:
Radxa Dragon Q6A

SoC:
Qualcomm QCS6490

NPU:
Hexagon V68

OS:
Armbian 26.2.4 (Ubuntu 24.04)

Kernel:
6.18.2-current-qcs6490

GenieX:
v0.3.14

QAIRT:
v2.45.0.260326

---

Models Tested

Model	CPU	Hybrid/NPU
unsloth/Llama-3.2-1B-Instruct-GGUF	✅	❌
google/gemma-4-E2B-it-qat-q4_0-gguf	✅	❌

---

Minimal Reproduction

geniex infer \
  unsloth/Llama-3.2-1B-Instruct-GGUF \
  --compute hybrid \
  --prompt "What is 2+2?"

Expected:

4

Actual:

encoding...
ggml-hexagon.cpp:2233:
ggml-hex: dspqueue_read failed: 0x0000002e
SIGABRT

---

Additional observations

CPU execution succeeds.

Example:

geniex infer \
  unsloth/Llama-3.2-1B-Instruct-GGUF \
  --compute cpu

Output is correct (~7.9 tok/s).

The same board also successfully runs Llama 3.2 1B on the Hexagon NPU through a QAIRT-compiled model, indicating that the hardware and firmware are operational.

---

Questions

1. Is QCS6490 / Hexagon V68 officially supported by the current ggml-hexagon backend?

2. Is dspqueue_read failed: 0x0000002e a known issue?

3. Has anyone successfully run GGUF inference through GenieX on a Hexagon V68 device?

4. Is additional firmware or driver configuration required for Linux ARM64 systems?