Hi,
I’m running a Raspberry Pi 5 + AI HAT+ 2 (Hailo-10H) with HailoRT 5.2.0 and trying to load the Qwen2-1.5B-Instruct-Function-Calling-v1 model.
The Problem
Loading the model via the Python GenAI API fails with HAILO_TIMEOUT(4):
[HailoRT] [error] CHECK_SUCCESS failed with status=HAILO_TIMEOUT(4) - Failed to create LLM
[HailoRT] [error] Failed to create LLM with status HAILO_TIMEOUT(4). Check device logs for more details.
hailo_platform.pyhailort._pyhailort.HailoRTStatusException: 4
Code:
from hailo_platform.genai import VDevice, LLM
vd = VDevice()
llm = LLM(vd, “/opt/axon-agent/models/Qwen2-1.5B-Instruct-Function-Calling-v1.hef”)
What works
The base Qwen2-1.5B-Instruct.hef (downloaded from the v5.2.0 link) loads and generates perfectly on the same setup:
llm = LLM(vd, “/opt/axon-agent/models/Qwen2-1.5B-Instruct.hef”)
r = llm.generate_all([{“role”: “user”, “content”: “Hello”}], temperature=0.1, max_generated_tokens=10)
# Output: “Hello! How can I assist you today?”
Root cause (I think)
The [MODELS.rst]( hailo_model_zoo_genai/docs/MODELS.rst at main · hailo-ai/hailo_model_zoo_genai · GitHub ) lists Qwen2-1.5B-Instruct-Function-Calling-v1 as requiring HailoRT > 5.2.0, but the pre-compiled HEF is only available on the v5.3.0 link path:
- https://dev-public.hailo.ai/v5.3.0/blob/Qwen2-1.5B-Instruct-Function-Calling-v1.hef — **200 OK** (this is what I downloaded)
- https://dev-public.hailo.ai/v5.2.0/blob/Qwen2-1.5B-Instruct-Function-Calling-v1.hef — **404 Not Found**
So the HEF appears to be compiled with the v5.3.0 toolchain and isn’t backwards-compatible with the 5.2.0 runtime.
My environment
- Hardware: Raspberry Pi 5 + AI HAT+ 2 (Hailo-10H)
- OS: DietPi v10.2.3 (Debian Trixie, aarch64)
- Kernel: 6.12.75+rpt-rpi-2712
- HailoRT: 5.2.0
- Driver: hailort-pcie-driver 5.2.0 (hailo1x_pci)
- Firmware: 5.2.0
- Python: 3.10.20 (pyenv) with hailort-5.2.0-cp310 wheel
- Device confirms working: `hailortcli fw-control identify` reports Hailo-10H, firmware 5.2.0
Questions
- Is there a v5.2.0-compatible HEF for Qwen2-1.5B-Instruct-Function-Calling-v1?
- If not, when will HailoRT 5.3.0 packages be publicly available? (Currently only 5.2.0 is on the developer zone)
- As a workaround, is there a way to compile the function-calling model’s HEF using DFC 5.2.0 and the Qwen2-1.5B HAR?
Thanks for any help!