n8dgr8/ov_intent_analysis_sft-RKLLM-RK3588

ovintentanalysis_sft — RKLLM (RK3588) conversion

Pre-converted W8A8 RKLLM runtime file of `guoxuter/ov_intent_analysis_sft` (the OpenViking retrieval intent-analysis / query-planner model, fine-tuned from Qwen3.5-0.8B).

Run the same tuned query planner on your RK3588 NPU — no x86 conversion rig required.

Why this exists

OpenViking's docs recommend this model for the query_planner slot: it decides whether a search needs context retrieval, skips chitchat (no queries → no token spend), and emits structured skill / resource / memory queries. The stock model ships as HF safetensors; to run it on the RK3588 NPU you need a .rkllm file, which can only be produced by rkllm-toolkit (x86_64-only). This repo is that conversion, done once, so RK3588 owners can skip the whole rig.

File

Runtime requirements: librkllmrt.so 1.3.0 (rkllm-toolkit 1.3.0 generation). Verified on kernel 6.1 vendor with rknpu driver 0.9.8.

How to serve it

Any RKLLM-capable server works. Two options:

Option A — full rkllama server (Ollama API; recommended for OpenViking)

pip install rkllama   # python 3.9–3.12
rkllama_server --models /path/to/models

Place the .rkllm in your models dir. This gives an Ollama-compatible API, which is what OpenViking's query_planner speaks.

Option B — minimal OpenAI+Ollama server (no transformers/torch)

The same ctypes wrapper, no heavy deps (see the conversion recipe below for the gist). Serves /v1/* and /api/*.

OpenViking wiring

{
  "query_planner": {
    "provider": "litellm",
    "model": "ollama/guoxuter/ov_intent_analysis_sft:v7_q8",
    "api_base": "http://127.0.0.1:8091",
    "temperature": 0.0,
    "timeout": 60,
    "extra_request_body": { "think": false }
  }
}

Keep the model string exactly as-is — OpenViking auto-matches the bundled v7 prompt by string (retrieval.ov_intent_analysis_sft_v7 in intent_analyzer.py). Only api_base changes: point it at your RKLLM server instead of Ollama.

Benchmark (RK3588, same prompt)

Query planning is prefill-dominated, which is exactly where the NPU wins. Decode is memory-bandwidth-bound, so don't expect magic on long generations — this model's outputs are short JSON.

Conversion recipe (for reproducing / other models)

The converter is x86-only, so run it on an x86 box (any Linux, or a serverless cloud like Modal):

# rkllm-toolkit 1.3.0 (wheel from airockchip/rknn-llm release-v1.3.0,
# rkllm-toolkit/packages/rkllm_toolkit-1.3.0-cp311-cp311-linux_x86_64.whl)
# deps pinned from that release's requirements.txt (torch 2.6.0, transformers 5.8.0, ...)

from rkllm.api import RKLLM

llm = RKLLM()
llm.load_huggingface(model="guoxuter/ov_intent_analysis_sft", device="cpu")  # or cuda
llm.build(
    do_quantization=True,
    optimization_level=1,
    quantized_dtype="W8A8",
    quantized_algorithm="normal",
    target_platform="RK3588",
    num_npu_core=3,
    dataset="data_quant.json",   # calibration: input/target pairs
    hybrid_rate=0,               # REQUIRED arg in 1.3.0
    max_context=4096,            # NOTE: `max_context`, NOT `max_context_len`
)
llm.export_rkllm("ov_intent_analysis_sft_v7_w8a8_rk3588.rkllm")

Gotchas hit along the way (so you don't):

• —build() takes max_context, not max_context_len (raises TypeError: unexpected keyword argument).
• —hybrid_rate=0 is required in the 1.3.0 signature.
• —The toolkit only ships x86_64 wheels — there is no aarch64 path; don't fight it on an ARM SBC.
• —Keep the toolkit major version aligned with your runtime's librkllmrt.so (1.3.0 ↔ 1.3.0).

Attribution & license

• —Base model: `guoxuter/ov_intent_analysis_sft` — Apache-2.0, which its card states covers the fine-tuned checkpoint (same license as Qwen3.5-0.8B).
• —Conversion performed with Rockchip's rkllm-toolkit 1.3.0 (airockchip/rknn-llm).
• —This conversion is published under Apache-2.0.

Big thanks to guoxuter for the tuned model and the OpenViking team for the recommended workflow.